Saatja: Kaupo Sempelson <
[email protected]>
Saadetud: 02.07.2020 16:09
Adressaat: Kaupo Sempelson <
[email protected]>
Koopia: Triin Tõnisson <
[email protected]>; Kristiina Padu
<
[email protected]>; Priit Tinits <
[email protected]>
Teema: Maailmapanga analüüs varase hoiatuse võimalustest Eestis - infoks
ja arvamuse avaldamiseks
Head partnerid ja huvigruppide esindajad!
26. juunil 2019. a avaldati Euroopa Liidu Teatajas Euroopa Parlamendi ja
nõukogu direktiiv (EL) 2019/1023, 20. juuni 2019, mis käsitleb ennetava
saneerimise raamistikke, võlgadest vabastamist ja äritegevuse keeldu ning
saneerimis-, maksejõuetus- ja võlgadest vabastamise menetluste tõhususe
suurendamise meetmeid, ning millega muudetakse direktiivi (EL) 2017/1132
(saneerimise ja maksejõuetuse direktiiv), mis nõuab liikmesriikidelt, et
nad tagaksid võlgnikele juurdepääsu ühele või mitmele selgele ja
läbipaistvale varajase hoiatamise vahendile, mis aitavad tuvastada
asjaolusid, mis võivad põhjustada maksejõuetuse tõenäolisust, ja anda
võlgnikule märku vajadusest viivitamatult tegutseda.
Maailmapank on koostöös Euroopa Komisjoniga koostanud analüüsi, milles
keskendutakse varajase hoiatamise meetmete süsteemsele lahendusele
Eestis. Lisaks on Maailmapank planeerinud korraldada vastavasisulise
ümarlaua Eesti huvigruppidega (kuupäev on hetkel veel selgumisel),
arutamaks muuhulgas nende poolt koostatud analüüsi ja pakutud lahendusi.
Eelnevast tulenevalt saadame Teile kõnesoleva analüüsi nii tutvumiseks,
kuid eelkõige võimalusel ka tagasiside andmiseks, tagamaks võimalikult
laiaulatusliku huvigruppide seisukohtadega arvestamise juba praeguses
faasis. Sooviksime Teie tagasisidet, mille palume saata e-posti
aadressile , 24. augustiks 2020.
Lisa: Maailmapanga analüüs (86 lehel).
Parimat soovides,
Kaupo Sempelson
Majandusarengu osakonna ekspert
+372 625 6350 | Skype: kaupo.sempelson
| Suur-Ameerika 1, Tallinn 10122
Majandus- ja Kommunikatsiooniministeerium
P Säästa loodust ja ära prindi seda e-kirja!
1
Estonia / Early Warning Tools
Technical Note / May 2020
ESTONIA
EARLY WARNING
TOOLS
TECHNICAL NOTE
MAY 2020
CONTENTS
GLOSSARY............................................................................................................................... 4
ACKNOWLEDGMENT............................................................................................................. 6
1. EXECUTIVE SUMMARY...................................................................................................... 7
2. INITIAL CONSIDERATIONS............................................................................................. 10
2.1 Definition................................................................................................................. 11
2.2 Purpose................................................................................................................... 11
2.3 The Design of the EWT......................................................................................... 12
2.3.1 A self-assessment: Inbound model.................................................................... 12
2.3.2 An advisory agency combined with qualified volunteers:
Outbound model.................................................................................................. 14
2.4 What Should Be Measured? ............................................................................... 16
2.4.1 Liquidity test.......................................................................................................... 16
2.4.2 Balance sheet test................................................................................................ 16
2.5 The Importance of Confidentiality..................................................................... 17
3. ARTICLE 3 OF THE EU DIRECTIVE.................................................................................. 18
3.1 The Scope of Article 3.......................................................................................... 19
3.2 Minimalistic versus Comprehensive Approach................................................ 20
3.3 Potential Risks Involved in an EWT..................................................................... 21
3.4 Interaction between the EWT and the Estonian Insolvency Framework...... 22
3.4.1 Applying for restructuring procedures.............................................................. 22
3.4.2 Directors’ liabilities............................................................................................... 22
4. OVERVIEW OF THE EWT (OUTBOUND MODEL)........................................................... 24
5. SOURCES OF DATA.......................................................................................................... 27
5.1 Financial Information........................................................................................... 28
5.1.1 Description............................................................................................................ 28
5.1.2 Availability............................................................................................................. 29
5.1.3 Challenges............................................................................................................. 30
5.1.4 Suggested indicators........................................................................................... 31
5.2 Credit History ....................................................................................................... 32
5.2.1 Description............................................................................................................ 32
5.2.2 Availability............................................................................................................. 34
5.2.3 Challenges............................................................................................................. 34
5.3 Tax Information.................................................................................................... 35
5.3.1 Description............................................................................................................ 35
5.3.2 Availability............................................................................................................. 36
5.3.3 Challenges............................................................................................................. 36
5.4 Governance .......................................................................................................... 37
5.4.1 Description............................................................................................................ 37
5.4.2 Availability............................................................................................................. 37
5.4.3 Challenges............................................................................................................. 37
5.5 Macroeconomic Information ............................................................................. 38
5.5.1 Description............................................................................................................ 38
5.5.2 Availability............................................................................................................. 38
5.5.3 Challenges............................................................................................................. 38
6. THE MODEL/ALGORITHM FOR THE EWT...................................................................... 39
6.1 What Exactly Should Be Measured? .................................................................. 40
6.1.1 The what................................................................................................................ 40
6.1.2 The when............................................................................................................... 41
6.1.3 The how................................................................................................................. 42
6.1.4 The process........................................................................................................... 42
6.2 Measuring likelihood of insolvency: Traditional solutions versus Artificial
Intelligence............................................................................................................ 44
6.2.1 A traditional single statistical model based on regression analysis:
Probability of default estimation under the Basel framework....................... 44
6.2.2 A more advanced ensemble learning model supported by
machine learning.................................................................................................. 48
6.3 Overview of Other Models.................................................................................. 52
6.3.1 Models used by other EWTs................................................................................ 52
6.3.2 Other models........................................................................................................ 53
6.4 Review of Options................................................................................................ 55
6.4.1 Option 1: Machine learning tool......................................................................... 55
6.4.2 Option 2: A traditional single statistical model based on
logistic regression................................................................................................ 58
6.4.3 Main solutions’ opportunities and drawbacks.................................................. 59
6.5 Institutional Aspects............................................................................................ 60
6.5.1 Model ownership.................................................................................................. 60
6.5.2 Platform development......................................................................................... 61
6.5.3 Access to information.......................................................................................... 63
7. THE PROCESS OF ALERTING DEBTORS......................................................................... 64
7.1 Self-Help................................................................................................................ 65
7.2 Submitting a Notification.................................................................................... 65
7.3 Network Responsible for Delivering Advice...................................................... 67
MAIN REFERENCES.............................................................................................................. 70
ANNEX A: IMPLEMENTATION OPTIONS: MACHINE LEARNING SOLUTION.................. 72
ANNEX B: IMPLEMENTATION OPTIONS: TRADITIONALSINGLE STATISTICAL MODEL
BASED ON REGRESSION ANALYSIS................................................................................... 77
4
Estonia / Early Warning Tools
Technical Note / May 2020
GLOSSARY
AI Artificial Intelligence
AIS Association for Information Systems
API Application Programming Interface
BA Bankruptcy Act
CIT Corporate Income Tax
EAD Exposure at Default
EBA European Banking Authority
EBF European Banking Federation
EBIT Earnings before Interest and Tax
EBITDA Earnings before Interest, Tax, Depreciation, and Amortization
EC European Commission
ECB European Central Bank
EL Expected Loss
ETL Enrich, Transform, and Load
EU European Union
EWE Early Warning Europe
GAAP Generally Accepted Accounting Principles
GDP Gross Domestic Product
IAS International Accounting Standard
IFRS International Financial Reporting Standards
IP Insolvency Practitioner (in Restructuring)
IRB Internal Rating Based
KNN K-Nearest Neighbors
LGD Loss Given Default
ML Machine Learning
MoE Ministry of Economic Affairs and Communications
MoF Ministry of Finance
MoJ Ministry of Justice
MSME Micro, Small, and Medium Enterprise
NACE Nomenclature statistique des Activités économiques dans la
Communité Européenne
5
Estonia / Early Warning Tools
Technical Note / May 2020
NBE National Bank of Estonia
NPL Nonperforming Loan
ODF Observed Default Frequency
PD Probability of Default
RA Reorganization Act
RIK Centre of Registers and Information Systems
ROD Return on Debt
ROI Return on Investment
SLA Service Level Agreement
SME Small and Medium Enterprise
SRSS Structural Reforms Support Service
TCB Tax and Customs Board
TRIM Targeted Review of Internal Models
VAT Value-Added Tax
WBG World Bank Group
XBRL Extensible Business Reporting Language
6
Estonia / Early Warning Tools
Technical Note / March 2020
ACKNOWLEDGMENT
The report has been prepared by the World Bank Group (WBG) at the request of
the Estonian Ministry of Economic Affairs and Communications (MoE) as part of
a technical cooperation project funded by the Directorate General for Structural
Reform Support (DG REFORM) of the European Commission (EC). The objective of
the report is to assist Estonian authorities in the design of an early warning tool
(EWT) to identify and assist debtors in potential financial distress, in line with the
recommendations of the European Union (EU) Directive on Preventive Restructur-
ings (the Directive).
For the preparation of the report, the WBG team1 consulted with law and account-
ing firms, financial institutions, public agencies, and other stakeholders inter-
viewed during visits to Tallinn on March 18-22 and July 8-12, 2019, and January
21-24, 2020. The feedback obtained during these visits was essential to under-
standing the context and providing the recommendations included in the report.
The team is extremely grateful to the stakeholders and authorities of Estonia for
the cooperation and input shared with the team in the elaboration of the report.
1 Composed of Raquel Letelier (Senior Financial Sector Specialist, Task Team Leader), Fabrizio Fraboni (Lead
Financial Specialist), Fernando Dancausa (Insolvency Specialist), and Sandy Shandro (Senior Consultant and
Insolvency Specialist), supported by Fabio Tarantini (Credit Reporting Specialist), under the guidance of and
with comments from Mahesh Uttamchandani (Practice Manager).
1. EXECUTIVE SUMMARY
8
Estonia / Early Warning Tools
Technical Note / May 2020
The recently enacted EU Directive urges EU member states to set up EWTs that
detect financial distress as early as possible, encouraging businesses to take early
action to avoid insolvency. The focus of the EWTs promoted by the EU Directive
is to alerts debtors (rather than creditors or public authorities, which has been
traditionally the case), and this shift has important implications for the design and
potential uses of these tools. This is a welcomed development that will surely con-
The focus of the EWTs
promoted by the EU tribute further to the creation of a rescue culture across the EU, thereby benefiting
Directive is to alerts businesses, economic activity, and employment. Although the Directive imposes
debtors this important obligation on all member states, it provides little guidance on how
to implement EWTs, leaving each member state plenty of discretion to design a
system that best fits its needs. The report explores the available options and rec-
ommends a way forward to implement an EWT in Estonia.
Estonia seems to meet many of the necessary preconditions to successfully de-
velop a sophisticated EWT: a highly digitized economy, institutions committed to
delivering high-quality public services, and a restructuring culture that mostly re-
lies on amicable solutions that avoid court involvement whenever possible. At the
same time, Estonia faces unique challenges in the implementation of this reform,
including a relatively small economy where a strong business reputation is key, the
prevalence of micro, small, and medium enterprises (MSMEs), and the lack of pos-
itive credit information that may compensate businesses’ prior defaults. Despite
these challenges, implementing an EWT is an excellent opportunity for Estonia to
innovate and design a system that could eventually be replicated across the EU.
When discussing how to design an EWT, the report conceptually relies on three
pillars: (1) the existence and availability of data that should be used by the EWT; (2)
Despite challenges, the algorithm that can indicate companies in distress; and (3) the institutional ar-
implementing an EWT rangements surrounding the EWT, including governance of the agency in charge
is an excellent oppor- of the EWT’s operations and the process of notification of distressed businesses.
tunity for Estonia
On the first pillar (data), the report concludes that Estonia counts on multiple
sources of data that form the foundation for an effective system. Specifically, an
impressive amount of tax payment information history is available and publicly
accessible from the tax authority. Advanced credit information and credit scorings
are also widely available, although at a cost, from local credit bureaus. Businesses’
financial information is also available from the Centre of Registers and Informa-
tion Systems (RIK), although 95 percent of Estonian companies are exempt from
submitting full financial statements (including a cash flow statement), which poses
a significant challenge.
9
Estonia / Early Warning Tools
Technical Note / May 2020
On the second pillar (the algorithm), the report provides the Estonian authorities
with two options. The first option is a machine learning model based on the gradi-
ent boosting technique. The second option is a statistical model based on multivar-
iate logistic regression analysis. Each option has advantages and disadvantages,
but both solutions would allow Estonia to fully comply with the Directive. The re-
port has a slight preference for the machine learning model, especially in case pos-
itive credit information becomes available in Estonia soon. Incorporation of these
Both solutions would
data would greatly increase the model’s accuracy, and therefore it is advisable that
allow Estonia to fully
comply with the such incorporation be possible, an option not afforded by the statistical model.
Directive
On the third pillar (institutional arrangements), the report recommends that RIK
play a predominant role regardless of the option chosen to design the algorithm
under the second pillar. In that respect, RIK should hold ownership of the EWT
and be the only point of entry for businesses interested in accessing information
about the EWT. This approach would ensure options for providing notice: self-as-
sessment by the companies and notification by a third party. Responsibility for
developing the platform could be shared by CreditInfo (as sole owner of credit in-
formation in Estonia and holder of the information required to notify companies)
and RIK, both of which must remain involved in all cases. Finally, the report rec-
ommends that RIK be responsible for notifying the businesses in distress. Upon
their notification, businesses should also be informed that they may benefit from
advisory support.
2. INITIAL
CONSIDERATIONS
11
Estonia / Early Warning Tools
Technical Note / March 2020
2.1 DEFINITION An EWT helps businesses detect financial difficulties so that they can be addressed
in a timely manner. Within this broad definition, EWTs can adopt many different
forms, ranging from purely internal control systems (involving corporate bodies
like the supervisory board or corporate auditors), to external control systems that
rely on the intervention of third parties, either due to their professional expertise
or because they are better informed about the debtors’ affairs.
The literature on this topic seems to be excessively permissive, allowing the term
“EWT”’ to be used for any action, system, or event that may help detect, directly or
indirectly, the existence of financial distress2. This approach is perhaps misleading,
as it implies that the organization of events, establishment of advisory services, or
even information campaigns would amount to the establishment of an EWT. The
report, however, adopts a narrower approach and considers EWTs as those sys-
tems that, after analyzing the detailed circumstances of a specific business, notify
the management of such business of the existence of potential problems, urging
management to take early action.
2.2 PURPOSE It is well established that the earlier a debtor can detect its financial difficulties,
the higher the probability of avoiding insolvency3. Similarly, if the viability of a
business is permanently impaired, the more orderly and efficient the liquidation
process will be.
Until recently, EWTs were typically designed to alert creditors and public authorities
to the upcoming distress of corporate and, in some occasions, supervised debt-
ors. Specifically, financial institutions in the EU have the obligation to implement
adequate internal procedures to identify and manage potential nonperforming
clients at a very early stage4. Public authorities, including financial sector super-
visors, have also developed EWTs to prevent the failure of systemic entities and
preserve financial stability.
However, the focus of the EWTs promoted by the EU Directive is on the debtor,
and this shift has important implications for the design and potential uses of these
tools. Among other differences with the Directive’s predecessors, one can high-
light the enhanced need to rely on publicly available data, the lack of sophistica-
tion of the ultimate beneficiaries of the alert, or the natural aversion by businesses
(and micro, small and medium enterprises MSMEs in particular) to acknowledge
the existence of financial distress.
Although the focus of the tool is to serve all debtors that engage in economic
activities, EWTs are expected to be particularly useful to SMEs. This is because,
when facing financial difficulties, SMEs often do not have the necessary resources
to cope with high restructuring costs, including hiring advisors that can prevent
2 See, for example, “Bankruptcy and Second Chance for Honest Bankrupt Entrepreneurs”, ECORYS (2014).
3 See, among others, “Business Dynamics: Start-Ups, Business Transfers and Bankruptcy”, EC (2011).
4 See “Guidance to EU banks on Non-Performing Loans”, ECB (2017).
12
Estonia / Early Warning Tools
Technical Note / May 2020
or mitigate the effects of insolvency. Considering the limited size of the Estonian
economy (where SMEs are predominant), the report bears this in mind as it ex-
plores variations of design and implementation of EWTs.
2.3 THE DESIGN OF THE EWT Statistical and technical features (discussed in section 6) aside, EWTs can be based
on one of the two approaches discussed later. These approaches are not mutually
exclusive, and the Estonian authorities may decide that the adoption of both ap-
proaches, mutually reinforcing, is the best option for Estonia (funding available to
introduce either or both options will be a factor). Naturally, the accuracy of these
approaches will vary depending on the amount and quality of data provided, a
critical issue discussed in section 5.
2.3.1 A SELF-ASSESSMENT: In this approach, the EWT encourages entrepreneurs (and entrepreneurs’ inter-
INBOUND MODEL nal corporate bodies) to assess their own situation. The assessment is the first
line of defense: before the involvement of third parties takes place, the system
prefers to place the responsibility for detecting financial distress on the ultimate
beneficiaries of such information, including shareholders, business supervisory
boards, and employees’ representatives. The underlying principle is that it is in
the debtor’s own interest to detect these difficulties as soon as they arise. When
potential distress is not detected by management (the main corporate body re-
sponsible for doing so), then other bodies within the business should alert or seek
an explanation from management as to the severity of the potential distress. Such
assessment can be conducted by the business’s own corporate bodies, whenever
access to the required information is possible, or by a public agency, which would
conduct the assessment at the request of the business.
There are different ways to design an EWT. The simplest approach is delegating
the analysis to a corporate body—the approach implemented in France, as ex-
plained in Box 15.
Box 1: Alert Procedure in France
The French alert procedure (procédure d’alerte) was introduced in 1985 and is part of the insolvency provisions
included in the French Commercial Code. It refers to several alternatives that the internal bodies of a business can
use to raise an alarm about the potential distress of the business. The procedure can begin as soon as the facts
about the struggling business are available. External stakeholders can launch the alert as well.
Among internal bodies, the council of workers, or business partners, has the authority to start the alert procedure.
As soon as any evidence suggesting the potential distress of the debtor is discovered, the council of workers can
contact the management of the business and request a formal explanation for the existence and severity of the
distress. Management has an obligation to respond; if this obligation is not complied with, then the council of
workers can prepare a report and send it to the company’s auditor or, if an auditor has not been appointed, to the
President of the Commercial Court. In the case of a partnership, non-managing partners may raise the presence
and severity of distress with a managing partner.
5 For more information on this model, see the Labour Code of France, Article L.2323-78, as well as Law No.
85-98 of January 25, 1985, on the Recovery and Judicial Liquidation of Companies.
13
Estonia / Early Warning Tools
Technical Note / May 2020
A similar approach, also implemented in France6, is to delegate the analysis to the
business’s statutory auditors, who are required to formally alert management to
the circumstances that endanger the business as a going concern. Once this infor-
mation is received, management has the obligation to act as early as possible; the
timing would depend on the standard set by the insolvency law. This alternative
approach, however, would only be useful to those companies that have appointed
an auditor, which are estimated to be about 6,000 in Estonia (approximately 4
percent of total companies).
A variation of this approach consists of an interface offered by a public agency,
typically the one in charge of registering businesses’ financial statements on an
annual basis. In this approach, the entrepreneur voluntarily requests a “health
check” by the agency once a year, and provides the information required to per-
form the assessment. Next, the EWT automatically calculates the likelihood of the
business’s becoming insolvent based on a set of indicators considered to be of
key importance. Subsequently, the EWT relays its findings to the debtor’s manage-
ment, who would again be encouraged to seek advisory support without delay.
This approach has already been implemented in Estonia, although only in large
companies, as discussed in the following box.
Box 2: Alert from the companies registry in Estonia
In Estonia, all businesses are required to file their annual financial statements with RIK, which assesses whether
the company’s net assets meet the minimum threshold established by law (see section 176 of the Estonian Com-
mercial Code). If the net assets fall below the legal threshold, RIK notifies the company’s management and urges
them to find a solution to address the financial problems. At that stage, the managers have a responsibility to file
for bankruptcy, or improve the company’s balance sheet.
If after delivery of the notification the company fails to communicate to RIK how management will restore its cap-
ital to the legally required level, then the court register may issue a formal petition to for management to do so. If
after that formal request the company still fails to provide an explanation, then the company can be fined. There
are different ratios of fines, starting from EUR 300 and up to EUR 3200 but might go higher and in some cases the
fines can be imposed repeatedly.
One shortcoming specific to this approach is that it fully depends on the willing-
ness of the entrepreneur to conduct the self-assessment (and that those entrepre-
neurs who either do not anticipate a situation of financial distress or who simply
are not aware of the existence of the EWT will not benefit from its capabilities). In
addition, the self-assessment approach will naturally end when the entrepreneur
receives the results generated by the interface. No support is offered to the entre-
preneur even if it has been alerted that its financial condition is compromised. A
similar approach has been adopted in Germany, where no EWT exists.
6 This is the approach adopted in France with respect to the largest companies only (société anonyme). See
Article L234-1 of the French Commercial Code. The implementation of an alerting tool in Italy also follows
this approach.
14
Estonia / Early Warning Tools
Technical Note / May 2020
Box 3: Voluntary support to companies in distress in Germany
Germany does not have a full-fledged EWT with a predictive tool designed to notify companies experiencing dis-
tress. A nonprofit organization, Team-U, supports entrepreneurs facing difficulties. The model is inbound, as en-
trepreneurs themselves must seek help from Team-U in order to receive the support they need.
The rationale behind this approach is that (1) help from a neutral counterpart, rather than a public authority, may
make entrepreneurs feel more comfortable with questioning their own financial status and receiving advice; and
(2) a self-assessment based on qualitative information is considered more effective in identifying the reasons
for distress. The support to distressed companies is provided through a network of volunteer consultants, often
ex-entrepreneurs who experienced financial difficulties during their professional life.
2.3.2 AN ADVISORY AGENCY A more ambitious approach involves a public or private agency that, at the mini-
COMBINED WITH QUALIFIED mum, is responsible for (1) designing and operating the EWT; as well as (2) notify-
VOLUNTEERS: OUTBOUND ing debtors that their business might be facing difficulties in the short term. The
MODEL EWT could also include optional advisory services that would assist with restruc-
turing the business, or any other beneficial measures.
The main advantages of this model are that it can accommodate a gradual ap-
proach and it allows for a series of steps to remedy the distress situation. In addi-
tion, a public or private advisory agency is better placed and, most likely, has more
adequate resources to design and operate a sophisticated system that comprises
all types of data. They include not only financial information about the entrepre-
neur, but also payment history, governance, and macroeconomic data.
Belgium and Denmark are the two countries within the EU that have recently im-
plemented models based on this approach. A summary of the main features ob-
served in the inbound and outbound models is provided in section 4. Both models
are also compared in Table 4.
Box 4: The Danish Business Authority and EWE in Denmark
Early Warning is a public program run by the Danish Business Authority (DBA) in Copenhagen under the Ministry
for Business and Growth. The program is largely composed of two different and well-defined stages: (1) a first
stage, involving the identification of distressed companies based on an artificial intelligence model operated by
the DBA; and (2) a sophisticated network of consultants and volunteers who provide advice to distressed compa-
nies, run by Early Warnings Europe (EWE).
The DBA model is based on accessible accounting data, which are free in Denmark, and calculates probabilities
of default of distressed businesses. The model uses parameters such as income statement, gross profit, result of
ordinary operations, profit for the year, equity development, and short-term and long-term debt. The results sug-
gested by the model are a basis, not a substitute, for subsequent individual, qualitative processing of the model’s
results; the processing is performed by EWE.
After an EWE expert has validated the model’s findings, a consultant or a volunteer (see section 7 for more in-
formation) approaches the management of the distressed company and offers advice on reconducting the busi-
ness’s financial situation. The first action taken by the consultant is an initial screening, which involves a complete
15
Estonia / Early Warning Tools
Technical Note / May 2020
overview of the circumstances, including financial, economic, and often private issues of the distressed business.
EWE consultants are based across the country in regional development centers, making it easier to reach strug-
gling entrepreneurs in person.
If a company can be saved, the consultant assigns one of the organization’s 100 volunteer advisors to support a
turnaround of the company. It has been reported that approximately 40 percent of assisted companies are fol-
lowed by a volunteer advisor for a longer period. A volunteer develops an action plan and is the point of reference
for the company owner throughout the restructuring process.
Box 5: The chamber for companies in financial difficulty in Belgium
The court for enterprises in Belgium has a separate department—the Chamber for Companies in Financial Difficul-
ty—that is responsible for collecting and analyzing information from various sources that help identify struggling
companies. The Chamber has two primary goals: (1) to alert companies before they turn insolvent; and (2) to
protect creditors’ interests by liquidating companies with no chance of survival.
Investigations to determine which companies are at risk of bankruptcy are carried out primarily by lay judges within
the Chamber for Companies in Financial Difficulties. The Chamber relies on several indicators to find companies
in distress. It receives information on other court rulings, for example, related to debt payments. In particular, the
Chamber is informed if companies cease to make rent payments or social security and tax contributions. In addition,
Graydon (a credit information company) also delivers to the Chamber a list of companies that may be at risk. To make
this determination, Graydon uses 13 indicators to identify companies in danger of insolvency (discussed in section 6).
Once Graydon sends the list of companies at risk, the Chamber determines whether to invite company managers
for an interview to obtain additional information. Appearance in court is mandatory. Managers must appear in
court in person; they may be accompanied by other persons of their choice, including counsel and accountants.
Interviews are conducted by lay judges who, in addition to working within the Chamber, often also have active
careers as accountants or bankers. Therefore, they are specialized in insolvency and can consult the court on what
to do with each case.
The interview usually begins with a question on whether the company is a going concern; if not, it is often too late
to attempt restructuring. The Chamber then investigates what would happen in the case of bankruptcy. Judges
may request more details from or additional interviews with the manager or other company employees, and
eventually prepare a report summarizing the situation.
Following the report, judges can make one of three decisions:
1. The Chamber may initiate bankruptcy by sending the case to the public prosecutor. Most cases fall into this
category.
2. The Chamber may appoint an administrator who has a maximum of four months to complete restructuring.
Otherwise, bankruptcy or liquidation proceedings are initiated. Administrators are seldom appointed, because
few companies have a chance of survival.
3. The Chamber also has the power to liquidate companies. Almost all courts for enterprises apply the new liqui-
dation procedure as a priority for companies that no longer engage in any real activity because they are unable
to repay debts but remain registered.
16
Estonia / Early Warning Tools
Technical Note / May 2020
2.4 WHAT SHOULD BE
MEASURED?
2.4.1 LIQUIDITY TEST Insolvency occurs when a company cannot meet its contractual obligations as
they become due. This has been widely referred to as the “liquidity test” for in-
solvency. It generally requires that the debtor has ceased making payments and
will not have enough cash flow to service its existing obligations in the ordinary
course of business. From this perspective, the objective of the EWT would be to
identify entrepreneurs that have not reached that stage yet, but that may reach
it unless certain actions are taken without delay. Indicators that may be used to
reach that conclusion include the debtor’s failure to pay rent, taxes, salaries, em-
ployee benefits, trade accounts payable, and other essential business costs.
The liquidity test is recognized in Estonia as the entry criterion for insolvency pro-
ceedings, both under the bankruptcy procedure7 and under the reorganization
procedure8. The difference between the standard applicable to each of these
procedures is that the state of insolvency must be imminent but not have been
reached yet in the reorganization procedure, thereby encouraging the prevention
restructuring.
Considering its importance and its wide implementation in Estonia, it is advisable
that any EWT potentially developed in Estonia focus on the liquidity test as the trig-
ger of financial distress. Sources of data that can be used by the tool to determine
whether this test has been met include (1) financial information, and the cash flow
statement more specifically; (2) payment history, to the extent this information is
available and any consent from the debtor that is necessary is obtained; and (3)
payments to the Tax and Customs Board (TCB), which are public and are currently
being used by a tool recently launched by the Estonian Ministry of Finance (MoF).
The key challenge that the EWT will face when determining business distress is
how early can that status be anticipated, and with what degree of accuracy. This
issue has already been highlighted and is discussed in detail in section 6.
2.4.2 BALANCE SHEET TEST When designing an EWT, a shortcoming traditionally associated with the liquidity
test needs to be considered: a debtor’s inability to pay its debts as they become
due may point to a solely temporary cash flow or liquidity problem for a business
that is otherwise viable9. If the EWT deems the liquidity test as the sole indication
of distress, the tool may naturally identify “false positives,” which, considering the
negative implications the notification may have, should be avoided at all costs. For
this reason, it is desirable that any EWT designed for this project look beyond the
liquidity test and include other variables that do not look exclusively at the cash
flow or payment history of a company to determine its financial status.
7 Article 1.2 of the Bankruptcy Act reads: “A debtor is insolvent if the debtor is unable to satisfy the claims of
the creditors and such inability, due to the debtor’s financial situation, is not temporary.”
8 Article 7.2 of the Reorganization Act reads: “A reorganization application of an undertaking shall set out an
explanation of the undertaking regarding the reasons for the economic difficulties and shall substantiate that:
1) the undertaking is likely to become insolvent in the future;
2) the enterprise requires reorganization;
3) the sustainable management of the enterprise is likely after the reorganization.”
9 Competition in modern markets may compel market participants to accept even losses on a temporary
basis in order to become competitive and maintain or gain market share.
17
Estonia / Early Warning Tools
Technical Note / May 2020
To complement the liquidity test, indicators that focus on the “balance sheet test,”
which centers on excess of liabilities over assets as an indication of financial dis-
tress, are an advisable option. Specific indicators that may serve this purpose are
largely based on the debtor’s balance sheet information and may include the debt
to EBITDA (a measure of a company’s overall financial stability) or debt to sales
(these are discussed in further detail in section 5.1.4).
Of note, in Estonia, the Commercial Code uses the balance sheet test when declar-
ing corporate directors personally liable for damages caused to creditors when
the business is irremediably insolvent, and they have failed to take any actions to
mitigate this situation. In this standard, insolvency is understood as the moment
when corporate liabilities exceed the corporate assets and directors have failed
their duty to file for bankruptcy10.
2.5 THE IMPORTANCE OF The information collected and handled by the EWT, regardless of the final form or
CONFIDENTIALITY structure adopted, will of course be of a commercially sensitive nature, confiden-
tial, and perhaps even subject to obligations owed to other persons. In a relatively
small market like Estonia, the potential spread of rumors involving the potentially
compromised financial status of a debtor can have disastrous consequences. For
this reason, it is desirable that an EWT include provisions to protect all information
collected by the tool, as well as any analysis provided by the tool, from abuse by
creditors or competitors that may take advantage of it in any way. There should
also be a legal obligation for EWT staff to use information received or produced by
the tool only for the purposes of the debtor at hand.
This rule may accept certain exceptions, like the provision of information in the
context of criminal proceedings, for example.
10 Article 146.1 of the Commercial Code reads: “A claim for payment of compensation to a private limited
company for damage specified in subsection (1) of this section may also be submitted by a creditor of the
private limited company if the assets of the private limited company are not sufficient to satisfy the claims
of the creditor.”
3. ARTICLE 3 OF THE
EU DIRECTIVE
19
Estonia / Early Warning Tools
Technical Note / March 2020
3.1 THE SCOPE OF Before considering the development of an EWT in Estonia, it is important to have
ARTICLE 3 a clear understanding of what Article 3 requires, and of the limits of its intended
application. Article 3 provides in relevant part as follows:
1. Member States shall ensure that debtors have access to one or more
clear and transparent early warning tools which can detect circum-
stances that could give rise to a likelihood of insolvency and can
signal to them the need to act without delay.
2. Early warning tools may include the following:
(a) alert mechanisms when the debtor has not made certain types
of payments;
(b) advisory services provided by public or private organizations;
(c) incentives under national law for third parties with relevant infor-
mation about the debtor, such as accountants, tax and social se-
curity authorities, to flag to the debtor a negative development.
3. Member States shall ensure that debtors and employees’ represen-
tatives have access to relevant and up-to-date information about the
availability of early warning tools as well as of the procedures and
measures concerning restructuring and discharge of debt.
4. Member States shall ensure that information on access to early warn-
ing tools is publicly available online and that, in particular for SMEs, it
is easily accessible and presented in a user-friendly manner.
5. Member States may provide support to employees’ representatives
for the assessment of the economic situation of the debtor.
The following observations follow from the wording of Article 3 (see section 3.4 for
a discussion on how these observations may apply in Estonia):
• The time at which an EWT is to be made available is before the debtor faces any
“likelihood of insolvency,” which is the standard applicable under Article 4.1 of
the Directive.11 This implies that the debtor should be alerted, at least in theory,
before he is under any obligation to open a restructuring procedure.
• It follows that additional duties on directors required under Article 19 of the
Directive , which are to apply where there “is a likelihood of insolvency,” should
not be triggered solely based on any EWT notification or other contact with the
debtor under Article 3.
11 See also Recital 71 of the Directive and also paragraph 2.19 of the WB Report “Preventive restructuring in
Estonia” of July 2019.
20
Estonia / Early Warning Tools
Technical Note / May 2020
• EWTs are clearly intended to be for the benefit of the debtor and no one else.
The intention of the Directive is to make it possible to save the “success stories”
(and to liquidate promptly the failures) by giving some form of advance indi-
cation that financial problems requiring a formal restructuring might occur in
the future if early steps are not taken. This is consistent with another principal
aim of the Directive: to encourage a spirit of entrepreneurialism within the EU
as well as both reasonable risk-taking by honest debtors and a more generous
attitude toward forgiveness of failure.12
• There may be a risk that employees, whose interests and wishes may not al-
ways coincide with those of the debtor, may not stand by to permit the debtor
to try to find long-term solutions for the business, and might seek to intervene
to force their views on the debtor. In this regard, note that Article 3(5) provides
that a member state “may” provide for support to employees’ representatives
for the assessment of the economic situation of the debtor. If a policy decision
is made to provide such support, then consideration should be given to the fea-
sibility of covering off any liability risk in legislation. Otherwise, in the absence
of their consent, debtors are entitled to assume that they, not the member
state or any agency involved in the operation of the EWT, control the informa-
tion flow to employees and/or their representatives.
3.2 MINIMALISTIC VERSUS Apart from the examples given in Article 3 itself, the Directive offers little guid-
COMPREHENSIVE ance on how member states should comply with the requirement to give debtors
access to “one or more” EWTs. There is, however, plenty of scope in the Directive
APPROACH for each member state to do as much as it considers appropriate to its needs and
institutional capacity.
There are many factors to be considered before deciding on the approach to com-
pliance that is best for Estonia. The agencies in charge of designing and operating
such EWTs need to exercise great care in order to minimize liability risk as well as
the prospect of unintended damaging consequences.
If a policy decision is taken to do the minimum necessary, compliance would be
quite straightforward. For example, since it is already the case that a late payment
of VAT is something recorded by the TCB, a notification sent to the debtor warn-
ing of the need to remedy the situation would amount to compliance with the
Directive. Indeed, Recital 22 of the Directive states that EWTs could take the form
of “alert mechanisms” that indicate when the debtor has missed a payment, such
as taxes or social security contributions, and leaves it to member states to decide
whether the public or the private sector should develop the notification tool.
There are two advantages in taking a straightforward approach that is solely based
on information on missed payments. First, it avoids the confidentiality problems
12 See Recitals 1, 3, 5, and 7.
21
Estonia / Early Warning Tools
Technical Note / May 2020
relating to the basis of notification inherent in a more elaborate EWT (discussed
later) because the VAT position of a debtor is already a matter of public record
in Estonia: anyone dealing with the debtor can ascertain the position. Second, it
minimizes the risk of the potential liability of the agency in charge of operating the
EWT for having brought about or contributed to a company’s insolvency by virtue
of the inappropriate or inaccurate functioning of the EWT.
However, there are also disadvantages in adopting the minimalistic approach.
While reducing the chances of false positives and minimizing liability risk, this
approach also delays the identification of a distressed business. The delay may
reach back to the stage when the business had already missed a payment, which
may not be considered early enough and is certainly later than the standard tar-
geted by the Directive. In this sense, it could be argued that the Directive falls
into a contradiction. A more comprehensive approach would seem justified if the
EWT is expected to “detect circumstances that could give rise to a likelihood of
insolvency.” If this standard is to be met, the EWT needs to take into consideration
additional information, including financial or economic information, to come to
such conclusion.
Several related risks need to be considered in adopting a more comprehensive
approach that regards any sign of financial problems a trigger for issuing an alert.
Perhaps the most important one is that “financial problems” can only rarely be
fully explained by financial indicators. And, more pertinently, the extent to which
other factors can explain such problems is often unknown. For example, there
may be many reasons that a financial indicator shows signs of deterioration: the
departure of employees or negative rumors in the marketplace or a changed
management structure each may be a valid explanation for poor financial perfor-
mance. However, it is practically impossible for an EWT to establish a link between
these variables.
3.3 POTENTIAL RISKS If and to the extent that an EWT is triggered by an event that turns out not to have
INVOLVED IN AN EWT been a “real problem,” then the tool is generating false positives and therefore is
not functioning optimally. The consequences may be serious:
• The spirit of entrepreneurship could be damaged, as debtors may be less in-
clined to take reasonable commercial risks if they are afraid of receiving noti-
fication of financial distress. This is particularly true in smaller economies and
smaller populations. The EWT must not, therefore, seek to eliminate what one
might call the “freedom to fail” on the part of an honest debtor. Since the EWT
is intended to be accessed by the debtor for its own benefit, there is a persua-
sive argument that if the debtor so chooses, it should be possible to ignore a
notification or reject it as wrong under the circumstances;
• The cost of credit could increase, as financial creditors will almost certainly
make it a condition of credit that the fact of any notification to the debtor be
disclosed; and
22
Estonia / Early Warning Tools
Technical Note / May 2020
• The risk that parties other than the debtor will learn of a notification (already
significant in the context of Estonia) will increase as the number of notifications
to the debtor increases, with creditors and competitors taking advantage in
whatever ways they can.
The materialization of these risks may imply the potential liability of the member
state for damages caused to the debtor. Indeed, although Recital 22 of the Di-
rective states that the Directive itself “should not impose any liability on Member
States for potential damage incurred through restructuring procedures which are
triggered by … early warning tools,” that does not mean that the risk of liability
does not exist. Such a liability could arise either (1) as a result of a loss of confi-
dentiality of information essential for the profitable operation of the debtor’s busi-
ness; or (2) as a result of analytical errors made by the EWT in the assessment of
problems facing the business, which, in turn, caused the debtor to enact changes
that made the situation worse, or that contributed to or caused a later insolvency.
Suggestions of solutions that effectively mitigate the risks described here are de-
tailed in following sections13.
3.4 INTERACTION BETWEEN
THE EWT AND THE
ESTONIAN INSOLVENCY
FRAMEWORK
3.4.1 APPLYING FOR Any EWT developed under Article 3 will form an integral part of the Estonian re-
RESTRUCTURING structuring regime. It is important, therefore, to know where, exactly, EWTs fit
PROCEDURES within that system. Section 7(2)1) of the Reorganization Act (RA) requires a debtor
applying to open a restructuring proceeding to substantiate that the debtor is
“likely to become insolvent in the future.” In contrast, Article 3 moves the timeline
to an earlier time: EWT(s) should be designed to “detect circumstances that could
give rise to a likelihood of insolvency.” Importantly, this may, in many cases, be
before the debtor is eligible to open a restructuring procedure under the current
RA or under the RA, as it may be amended to comply with the Directive.14 In this
sense, Article 3 should be viewed as innovative.
3.4.2 DIRECTORS’ LIABILITIES As stated, it follows that the duties of directors required under Article 19 of the
Directive , which are to apply where there “is a likelihood of insolvency,” should
not, therefore, be triggered solely based on any EWT notification. Whether those
duties apply or not will depend in each case on the financial position of the debtor
when the EWT is accessed by the debtor (or when a notification is made to the
debtor, if that is the model chosen). For example, if, at the time that an EWT is
13 See Section 5 on sources of data and data privacy-related issues, Section 6 on EWT modeling, and Section 7
on debtors’ alerting process.
14 See too Recital 71 of the Directive and also paragraph 2.19 of the WBG Insolvency Report of July 2019.
23
Estonia / Early Warning Tools
Technical Note / May 2020
accessed, the debtor is not, in fact, facing a likelihood of insolvency (because the
financial or other difficulty can be resolved quickly and simply, or for some other
reason), then the debtor should not incur liability, at some future date, for having
failed to take into account the interests of those mentioned in Article 19 of the
Directive; at that time of the early warning, the likelihood of insolvency test did not
apply to the debtor. That is not to say that the debtor might not incur such liability
from and after some later date; however, the key point is that the actual “time”
at which these additional Article 19 duties apply is not automatically going to be
when the debtor accessed the EWT. A court assessing the conduct of the debtor
that has accessed an EWT and later become insolvent must, therefore, consider
every case with great care.
24
Estonia / Early Warning Tools
Technical Note / May 2020
4. OVERVIEW OF THE EWT
(OUTBOUND MODEL)
25
Estonia / Early Warning Tools
Technical Note / May 2020
As anticipated, in section 2 the report recommends the introduction of an early
warning tool in Estonia based on the “outbound” model, which conceptually con-
sists of three main blocks, illustrated in Figure 1:
1. A set of data providers supplying data to regularly feed the predictive model;
2. A platform manager who runs the technical infrastructure in charge of (1) gath-
ering and integrating the data as well as of (2) elaborating a risk matrix that
leads to a list of businesses at risk; and
3. A notification and counseling entity providing advisory services to the notified
companies.
Figure 1: The Three Blocks of the EWT
CREDIT
CREDIT REGISTRY
DATA
TAX AGGREGATOR
BUSINESS COUNSELING
REGISTRY
TAX DATA
BIZ DATA
MODELING
BIZ REGISTRY
DATA
PROCESSING
The main phases of the EWT’s operations are:
• Data providers periodically supply information through a reporting portal. In-
formation is normalized, verified15, and uploaded to a database (staging area).
• A data aggregation procedure centralizes data flows from different data
providers.
15 Data providers may receive notifications if information provided doesn’t comply with established validation
rules and may be asked to resubmit the reporting flows.
26
Estonia / Early Warning Tools
Technical Note / May 2020
• A calculation engine (see the different methodologies described later) elabo-
rates risk scores (for example, likelihood of insolvency) for companies reported
in the data flows. Companies with a likelihood of insolvency among an estab-
lished threshold are included in a list.
• At least for the first two years after implementation, the model output (for
example, a list of companies in distress) is subject to validation by a team of
experts prior to company notification16.
• A platform manager available through a notification service (or both the plat-
form manager and the counseling service, as described in section 7) contacts
companies at risk and informs them that advisory support may be available to
define appropriate strategies for facing the situation.
Table 1 synthetizes the different options to implement and operationalize the
components of the EWT. The different options and their related variations with the
associated pros and cons are described in more granular details in sections 5–7.
Table 1: Overview of Main Implementation and Operationalization Options
Phase and topic Options available
Data suppliers RIK, credit bureau, tax authority, central bank
Platform ownership (e.g,. who operates the E.g., RIK (business register)
platform)
Platform development Option 1: Internal development with external support
Option 2: External development (e.g., CreditInfo)
Platform management Option 1: Managed internally
Option 2: Outsourced management (CreditInfo)
Model methodology Option 1: Artificial intelligence (machine learning)
Option 2: Statistical modeling
Option 3: Combined approach17
Notification service Option 1: Notification from platform manager
Option 2: Combined approach18
Counseling service19 Option 1: Based on hired staff and a network of professionals available upon demand
Option 2: Based on hired staff and a network of volunteers
Option 3: All volunteers working on a non-remunerated basis
16 For model fine-tuning purposes. As described, avoiding false positives should be considered a key priority.
17 For example, customization of an existing statistical model as a “quick win” while a machine learning tool is
being implemented.
18 Notification from the platform manager and counseling service based on the likelihood of insolvency.
19 See Section 7.3 for the detailed suggestion on the most suitable organization to provide counseling
services.
27
Estonia / Early Warning Tools
Technical Note / May 2020
5. SOURCES OF DATA
28
Estonia / Early Warning Tools
Technical Note / May 2020
5.1 FINANCIAL INFORMATION
5.1.1 DESCRIPTION All Estonian legal entities must submit annual reports to RIK during the six months
following the end of their year. For most Estonian companies, the deadline is June
30. The annual report is meant to give an overview of the enterprise’s financial po-
sition and financial performance, and it must be prepared according to Estonian
Generally Accepted Accounting Principles (Estonian GAAP) or under International
Financial Reporting Standards (IFRS)20. The report’s information is the cornerstone
of any potential financial analysis to be conducted by the EWT.
Since 2016 Estonian companies have been divided into four categories for the
purposes of their financial reporting:
1. Micro enterprises, for which the annual report consists of a balance sheet and an
income statement, and which may add notes to the accounts;
2. Small enterprises, which must submit the same information as micro enterprises
as well as a management report on the company’s operations and significant
events of the financial year; and
3. Medium enterprises, which must include in their submission a balance sheet and
income statement, and notes to the accounts as well as a cash flow-statement
and statement of change in owner equity; and
4. Large enterprises, which must submit the same information as medium
companies.
RIK provided a breakdown of the number of companies allocated to each category
that submitted financial information in 2017 and 2018, summarized in Figure 2.
Figure 2: Submission of Annual Accounts in Estonia, By Type
90
Thousands
80
70
60
50
40
30
20
10
0
2017 Total: 136k companies 2018 Total: 131k companies
Micro-sized entity Small-sized entity Medium-sized entity Big-sized entities
20 The annual report (including main reports and annexes) needs to be compiled and submitted in the e-ac-
counting system: http://www.rik.ee/en/company-registration-portal.
29
Estonia / Early Warning Tools
Technical Note / May 2020
The following observations can be drawn from Figure 2’s data, which have import-
ant consequences for the design of the EWT:
• Most companies in Estonia are considered micro-size or small-size. Together
they numbered 130,000 in 2017 and 126,000 in 2018, or approximately 96 per-
cent of all companies in Estonia in 2018. As explained earlier, these companies
submit only limited financial information in their annual reports.
• About 6,200 companies (2017) and 5,000 companies (2018) were considered
medium-size or large-size companies, and therefore needed to include a cash
flow statement in their annual reports. These companies represent less than 5
percent of companies engaged in economic activities in Estonia (see bullet on
condominium associations) and less than 1 percent of companies obliged to
submit full financial information.
• Only legal entities must submit their annual report to RIK. Self-entrepreneurs
and individuals engaged in economic activities submit their financial informa-
tion to TCB only. They number approximately 30,000.
• Conversely, legal entities that do not engage in economic activities, such as
condominium associations, have an obligation to submit their annual report
to RIK. These entities are estimated at approximately 20,000. Although these
debtors are not supposed to be covered by the EWT, they may compromise the
analysis of data provided.
• A significant number of companies do not submit any financial information to
RIK at all, despite the obligation to do so. These companies amount to approxi-
mately 20,000 per year, and typically they are deleted from the RIK system after
two or three years of inactivity.
5.1.2 AVAILABILITY RIK has developed an electronic tool for submitting annual accounts. Although
companies have the option to submit their accounts in signed pdfs, most accounts
are submitted using the Extensible Business Reporting Language (XBRL) Annual
Reporting System, which is accessible through the Company Registration Portal.
The system allows digital signatures as well as electronic auditing. As a result of
these technological incentives, about 99 percent of the companies registered in
Estonia use the e-Annual Reporting System, which makes financial information
machine readable. These features have positive implications for the development
of an EWT, as explained later.
Information submitted to RIK is public and accessible to any interested party, al-
though access is not free. Financial information on a specific company is made
accessible for 1 euro per request. The fee is for RIK’s legal obligation to be a fi-
nancially self-sustainable institution. A commercial agreement currently exists be-
tween RIK and CreditInfo Estonia to share information of all incorporated compa-
nies. In fact, CreditInfo purchases in bulk all financial and governance information
about businesses. Updates to the database are made in real time.
30
Estonia / Early Warning Tools
Technical Note / May 2020
It should also be noted that, following the standards set by the General Data Pro-
tection Regulation (GDPR), RIK plays the role of “data processor,” in charge of the
elaboration of information controlled by the courts (data controller)21.
5.1.3 CHALLENGES The following obstacles should be considered when using financial information
on a company in Estonia to design its EWT:
• Limitations in scope: The financial analysis performed by the EWT will be fairly
limited, as 95 percent of Estonian companies submit only minimal financial in-
formation22. For many companies (about 50,000), no financial analysis will be
possible, as there are no data available, either because the entities are self-en-
trepreneurs or because they simply do not submit any financial information.
In addition, self-entrepreneurs (approximately 30,000) will not be captured by
the EWT.
• Outdated data: The annual reports submitted will typically be dated as of the
end of the calendar year and become available only as of July of the following
year. This implies that by the time the financial statements become available to
the EWT, the records reported will be 12–18 months old. If financial problems
are already apparent in such out-of-date information, it is highly probable that
the debtor has not implemented any measures to reverse the financial distress,
and therefore, insolvency is inevitable. This important shortcoming should be
analyzed in the context of the EU Directive, which emphasizes the word early in
EWT. From this perspective, it can be argued that an EWT that is designed solely
on financial information cannot be considered an “early” warning tool.
• Potential inaccuracies: The quality of annual reports will vary greatly depending
on the expertise and diligence of the managers in charge of preparing them.
There will necessarily be cases of underreporting the financial distress of a
company, and in these cases, the EWT will not be able to identify potential prob-
lems based solely on financial information. RIK has reported that only about
6,000 companies audit their annual accounts.
21 The standards are set by Article 4 of the GDPR: a data processor is defined as “the natural or legal person,
public authority, agency or other body which, alone or jointly with others, determines the purposes and
means of the processing of personal data.” A data controller is defined as “a natural or legal person, public
authority, agency or other body which processes personal data on behalf of the controller.”
22 This is due to a restrictive transposition of Directive 2013/34/EU (Accounting Directive, setting harmonized
reporting standards across EU member states), aimed at minimizing the reporting burden of Estonian
businesses. For a comparative analysis of Accounting Directive transposition across the EU, see https://
pdfs.semanticscholar.org/245f/1f5e2aaa6d2bc1a3f3460f33f2a99e0cbc0b.pdf and https://theaccountant.
org.mt/transposition-of-the-accounting-directive/.
31
Estonia / Early Warning Tools
Technical Note / May 2020
5.1.4 SUGGESTED INDICATORS There are myriad indicators of early financial distress. To identify the indicators
most accurate for Estonian companies, one should first determine those areas
where business distress significantly impacts financial statements. Consequently,
financial difficulties normally result in:
• Barely sustainable financial debt;
• Weakened capital adequacy;
• Low liquid return on assets;
• Decreased available short-term liquidity; and
• Higher share of tax and social security debt on assets.
Considering the challenges explained earlier and the fact that most Estonian com-
panies do not report cash flow statements, the following indicators are proposed
as the most effective for Estonian companies:
• Financial debt sustainability: to be measured as the percentage of financial debts
on sales. This indicator aims at monitoring business sustainability. An increase
or decrease in sales should not be considered in isolation, but only as relative
to the levels of corporate debt. If sales increase less than financial debt, then
the business risk increases. This indicator is often referred to as return on debt
(ROD) and it’s a key variable for calculating return on investment (ROI). A similar
indicator that is widely used to track debt sustainability is debt to EBITDA.
• Capital adequacy: to be measured as net worth on total debts. As a company
receives the funds it needs from stakeholders (net worth), voluntary funding
(financial institutions), and non-voluntary funding (for example, the tax author-
ity), a decrease of capital adequacy indicates a shift of company control from
stockholders to stakeholders. In other words, the company depends more on
third-party resources than on its own resources. This indicator is strictly cor-
related with the leverage ratio (total assets/net worth), which in turn is a com-
ponent of the Dupont triangle23 and Altman’s Z-score24.
• Liquid return on assets: to be measured as cash flow on assets. Although other
indicators track asset profitability (see, for example, ROI and asset turnover),
cash flow becomes paramount for SMEs25 as they try (typically with difficulty) to
rapidly obtain from the market the liquidity necessary to meet financial obliga-
tions or to resolve emergencies. Therefore, a decreased share of cash flow on
assets indicates an increased business risk for the company.
• Liquidity: to be measured as short-term assets on short-term liabilities (com-
monly referred to as current ratio). Although the indicator helps to monitor
how well the company is managed (for example, when short-term liabilities
grow less than assets), some components of the liquid assets (for example,
23 ROE = ROA * Leverage.
24 See section 6.3.2.
25 About 95% of businesses in Estonia are SMEs.
32
Estonia / Early Warning Tools
Technical Note / May 2020
credits, inventory) also should be monitored to ensure that they move in har-
mony with the increase (or decrease) of sales.
• Tax and social security debt: to be measured as a percentage of the debt on
assets. Getting funding from TCB can be cheaper and less burdensome than
applying for a new loan or a loan extension. Nevertheless, an increase of tax
debt without proper planning can be disastrous for a business. As described
earlier, TCB can become an unaware funder, and delaying tax (or social secu-
rity) payments is usually considered one of the earliest signs of a company’s
distress.
All of these indicators can be calculated from all types of report formats included
in the balance sheet and income statement currently used in Estonia. If the cash
flow statement is not available26, cash flow should be calculated from the balance
sheet variations and income statement (direct or indirect method) information.
Obviously, a company can be vulnerable based on the results of more than one
indicator. The challenge for the EWT is to determine the extent to which a deteri-
oration in one (or several) of the proposed indicators constitutes a deterioration
severe enough to justify the submission of an alert. Naturally, a worsening of all
five indicators usually indicates that a company should be considered in distress.
However, deterioration typically will be observed based on one or two indicators
only. The definition of the thresholds and criteria for making that judgment are
explained in section 6.
5.2 CREDIT HISTORY
5.2.1 DESCRIPTION Credit information sharing in Estonia is on a voluntary basis. The information
shared by financial sector operators and other non-regulated entities27 is limited
to negative events, including delayed payments from companies and private per-
sons. Such information service is provided to data contributors by the Negative
Registry28, which was established in 2001 and is operated by the privately owned
company CreditInfo Estonia, which acts solely as data processor as far as such
service is concerned. CreditInfo Estonia is a subsidiary of CreditInfo Group, which
is an Iceland-based company operating in more than 20 countries.
It has been reported that some changes in the legal framework regulating credit
information sharing and the status of the Negative Registry are under discussion.
One the key transformations being considered is the creation of a clear framework
for positive information sharing, which would include, among other data, data
related to direct and indirect exposures and the payment histories of borrowers
26 See section 4.13. A cash flow statement is submitted by medium to large businesses (about 5% of regis-
tered and active businesses in Estonia).
27 Telecom operators and most fuel companies post information about payment defaults.
28 Maksehäireregister.
33
Estonia / Early Warning Tools
Technical Note / May 2020
and guarantors. Should these legal changes materialize in the short term, the
information sharing market in Estonia would be brought in line with internation-
ally recognized best practices, thereby significantly improving the quality of risk
management and decision making.
Although these legal changes remain under consideration, a broad array of ser-
vices is provided by local operators to support the credit granting process and
credit risk management. In this space, CreditInfo is Estonia’s leading institution,
allowing access to the broadest dataset and offering in parallel a series of sophis-
ticated, value-added services. For the purposes of the EWT, CreditInfo has devel-
oped a rating system supporting lenders’ assessment of borrowers’ creditworthi-
ness*, which could be of great use in identifying companies in distress. Here is
some key information about the rating system:
• The rating calculates a probability of default within a 12-month horizon.
• The rating methodology is based on an expert model complemented by a de-
cision tree.
• The rating calculates a probability of default based on information extract-
ed from the Negative Registry, RIK, annual statements, TCB, and public
announcements.
• The rating calculates 10 rating grades (from AAA to O).
• The rating is calculated from three types of scores, covering financial, econom-
ic, and payment performance.
The Negative Registry service is structured as follows:
• Credit reporting is provided on a real-time basis by all domestic lenders and
other users.
• A credit is classified as “in default” when payment is delayed for more than 45
days from the day after the term of payment, and for outstanding amounts
above 30 euro.
• Information is provided on both legal entities and individuals.
• Payment defaults are recorded within the last 24 months from the reporting
date.
• Borrowers have the right to access their data on the credit bureau, submit
claims, and ask for data rectifications.
• Data quality standards of credit information are generally optimal29.
• Reporting channels among different information providers are already in place
to gather complete sets of information, provide comprehensive credit reports,
and calculate credit scores30.
* Which is used by lenders not only at loan’s approval, but also throughout the loan cycle.
29 Also considering the low quantity of claims and rectifications.
30 For example, CreditInfo has reporting channels to gather information from RIK, TCB, and other sources
(public announcements).
34
Estonia / Early Warning Tools
Technical Note / May 2020
It should also be noted that under the current regulation in Estonia31, negative
credit information sharing does not require borrower consent. In fact, no legal
restrictions exist on providing legal information on legal entities32. In the case of
private persons, the matter is regulated by Article 5 of the General Data Protection
Regulation (GDPR)33 and Paragraph 10 of the Personal Data Protection Act, where
the purpose of assessment of the data subject’s creditworthiness or any other
similar purposes are cited as permittable uses of data.
5.2.2 AVAILABILITY Current negative information can be made available to third parties upon estab-
lishing a commercial agreement, and therefore could be used by the EWT if such
an option was allowed in Estonia. However, if positive information also was includ-
ed in the ratings, then this would require borrower consent. In this case, a specific
regime for the EWT’s use of the data might be defined under the “permissible
purpose principle.34” A careful evaluation should be conducted when the legal
framework is formalized.
Given the state-of-the-art ICT standards adopted by CreditInfo, from a purely
technical standpoint, no relevant issues emerge for credit information sharing.
5.2.3 CHALLENGES One of the options the report explores is to employ the CreditInfo credit ratings
as a main source of data used by the EWT to identify companies in distress (see
section 4). The most relevant challenge presented by this option is the limited
credit information currently shared in Estonia: negative information has restricted
predictive power, as it only highlights existing difficulties. On the other hand, pos-
itive credit information plays a more important role in the analysis of borrowers’
ability to repay existing debts.
Another challenge of using this information is that it is privately owned by Credit-
Info, which implies that the EWT would have to purchase it. Considering that the
scope of the EWT would be to cover all companies that engage in economic activ-
ities in Estonia (about 130,000), the resources required to implement this option
would be significant.
31 The current regulatory framework may be subject to change, considering the new legal framework.
32 If incorrect data are published, then the company may seek protection of its rights in court.
33 Principles relating to processing personal data.
34 See Art. 5 of the EU General Data Protection Regulation (GDPR), which establishes the principles relating to
processing personal data.
35
Estonia / Early Warning Tools
Technical Note / May 2020
5.3 TAX INFORMATION
5.3.1 DESCRIPTION Because distressed enterprises typically delay their tax and social security pay-
ments first, TCB may recognize financial difficulties before other creditors do. For
this reason, the detailed data TCB collects, especially when combined with those
from the sources mentioned earlier, could be extremely useful in preventing fi-
nancial distress (indeed, the Directive refers to incentives to public creditors, in-
cluding tax authorities, as an example of EWTs). The data fields most beneficial for
the purposes of the report are:
• Submitted tax declarations: under the Taxation Act, all companies that have
made payments subject to tax calculations are obligated to submit tax decla-
rations to TCB. Like tax payments, the timely submission of tax declarations is
also an indicator of distress, as one can anticipate that if the tax declarations
are not submitted on time, tax payments will not be on time either. The delay
between the declaration and payment due dates varies depending on the tax,
and this needs to be considered by the EWT35. Tax declarations that are VAT
related are the most valuable to the EWT. Property tax and income tax are pay-
able annually in Estonia, which limits their value to the EWT.
• Paid taxes and tax arrears are the most useful tax data field. Any differences
between the amount of tax declared and the amount actually paid are an ob-
jective sign of distress. This is particularly revealing in the case of VAT taxes
and labor taxes, which are typically the tax payments that an entrepreneur in
trouble misses first. The fact that they are due quarterly and monthly, respec-
tively, allows the EWT to monitor performance of the business multiple times a
year. Delinquent payments are not the only useful information here: evolution
of the paid amounts is also powerful information. The sudden decrease in the
amount of taxes declared and paid indicates potential problems in the activity
of a company.
• Total tax debt and interests, showing the total owed by a company to the tax
authority, is a clear indication of liquidity problems of a distressed debtor.
• Postponed tax debt and interests is the result of a distressed debtor’s success-
ful negotiation with the tax authority to reschedule payment dates, increasing
the likelihood of full payment. The mere existence of postponed tax debt is
an objective indicator of financial problems, which should prompt the EWT to
monitor these taxpayers. The trends associated with the postponed debt are
also a key input for valuable information: the increase of the postponed debt
implies a deterioration of the debtor’s financial condition. Conversely, the de-
crease in postponed amounts, together with on-time payments of the new tax
obligations as they accrue, is a sign of increased creditworthiness.
35 For example, in the case of income tax for natural persons, the declaration is due in March and the pay-
ment is due in June. Focusing on the delay in submitting the declaration may allow the EWT to gain three
valuable months.
36
Estonia / Early Warning Tools
Technical Note / May 2020
5.3.2 AVAILABILITY Under Estonian law36, tax secrecy does not apply to any of these data fields as
potential indicators of economic distress, although some exceptions apply to tax
arrears, as explained in the following sections. This means that the EWT may be
able to use all of this tax information without the need to obtain prior consent
from the taxpayer to use its data.
Representatives of TCB have expressed that these data, stored in a single data-
base, remain available electronically for use by the EWT. The data are updated
daily to reflect taxpayer declarations and payments. The legal framework and the
format and storage determined by the tax authority significantly increase the po-
tential uses of tax information.
5.3.3 CHALLENGES Considering the detail and easy access of tax information in Estonia, this source
of data should be one of the main pillars on which the EWT is built. The following
caveats, however, should be borne in mind when designing the EWT, as they may
limit the effectiveness of tax information when predicting potential distress:
• Information obtained from the corporate income tax (CIT), which in Estonia is
only accrued when dividends are distributed, is limited. In fact, payment of low
amounts under this tax may be unrelated to the economic performance of the
company, since the directors may decide not to distribute the profit earned.
This is unlike most EU member states, where the accrual of this tax takes place
upon earning profit.
• Information collected by TCB on the amounts of unpaid taxes unfortunately is
not categorized by tax type. This means that it will not be possible for the EWT
to ascertain whether the amounts due by a potentially distressed debtor cor-
respond to VAT payments or CIT, which may prompt different interpretations.
• The VAT payment, probably the most efficient indicator as likely the first pay-
ment to be missed, is only imposed on companies with a turnover greater than
16,000 euros. In practice, this may leave out SMEs with a turnover less than this
threshold.
• Information on tax arrears assessed by TCB typically does not become public
until the determination has been communicated to the taxpayer and the ob-
jections or appeals to the decision, if any, have become final. This implies that,
often, the information on tax arrears does not become available until several
months after the first arrear.
36 See article 27 of the Estonian Taxation Act.
37
Estonia / Early Warning Tools
Technical Note / May 2020
5.4 GOVERNANCE
5.4.1 DESCRIPTION One of the most predictive sources of data when assessing the potential insolven-
cy of a business is the financial status of those parties having legal or economic
connections to the business. Specifically, assessing whether the directors and/or
shareholders of a given business are in financial difficulty has been identified by
other models implemented in Europe as the most relevant factor when predicting
financial distress.
There are different options for integrating the analysis of these data into the EWT.
A first option is to conduct a full analysis, using the algorithm described in section
6, to those related parties identified as potentially relevant. This may, however, be
overly ambitious and not always possible, especially considering that the directors
and shareholders of a distressed business may be natural persons on which no
financial information is available. To mitigate this challenge, a simpler approach—
tracking whether those related parties have filed for insolvency, or whether there
are any judgments issued against their assets for the non-payment of debts—may
be used.
The governance information that may potentially be used by the EWT would there-
fore consist of two different sets of data: (1) the identity of the director(s) and the
ultimate beneficial owner(s) of a given company; and (2) whether they have been
declared insolvent and/or whether there are any judgments rendered in civil or
commercial procedures against their assets.
5.4.2 AVAILABILITY In Estonia, data on the identity of the ultimate beneficial owner seem to be col-
lected by the MoF, although RIK operates as the data holder and manager of the
database. During the missions to Tallinn, the team was informed that while this
information is not readily available, RIK is in possession of it and would be able to
make it available relatively easily. In fact, some of it is already being used by the
tax authority and CreditInfo in their systems.
5.4.3 CHALLENGES The only critical aspect that emerges in the case of any enforcement or bankrupt-
cy order issued against company managers, shareholders, or beneficial owners
is that this information is only managed by the court, and is published anony-
mously, and therefore cannot be stored in a central database. Based on this,
information on risk deriving from businesses’ related parties may be limited to
creditworthiness.
38
Estonia / Early Warning Tools
Technical Note / May 2020
5.5 MACROECONOMIC
INFORMATION
5.5.1 DESCRIPTION Macroeconomic information should play an important role when implementing
an EWT. In fact, improving or worsening trends of the economic cycle should be
reflected in the number of businesses in distress. These variables should there-
fore be taken into consideration when validating the models and establishing rel-
evant thresholds, allowing the identification of businesses in distress. The most
relevant macroeconomic variables are:
• GDP growth;
• Consumer price index (inflation);
• Unemployment rate;
• Income and wages;
• Reference interest rates;
• Balance of trade; and
• Sector-specific statistics37.
5.5.2 AVAILABILITY The information is generally available and can be retrieved from various public
sources, such as Statistics Estonia, the Bank of Estonia (central bank), and the Eu-
ropean Central Bank38.
5.5.3 CHALLENGES Although there is no relevant challenge regarding data gathering, relevant ca-
pacity is needed for capturing country specificities, infra-sector correlations, and
impact on sectorial trends in the economic cycle of the EWT.
Even though solutions based on artificial intelligence usually perform auto-fitting
adapting model outcomes to changing macroeconomic contexts, statistical model
validation considering these factors is typically more complex.
37 Agriculture, manufacturing, and services contribute to GDP and sector interlinkages.
38 ECB Statistical Data Warehouse, macroeconomic and sectorial statistics, see https://www.rik.ee/en/
international/e-business-register
39
Estonia / Early Warning Tools
Technical Note / May 2020
6. THE MODEL/ALGORITHM
FOR THE EWT
40
Estonia / Early Warning Tools
Technical Note / May 2020
6.1 WHAT EXACTLY SHOULD
BE MEASURED?
As explained in previous sections, the Directive aims at shifting authorities’ focus
to a prognostic approach rather than just performing damage control measures:
the concept of “early intervention” is a key pillar of applying the Directive more
generally. This has important implications for the EWT’s foundation.
6.1.1 THE WHAT Identifying a clear definition of prospective insolvency is perhaps the most com-
plex task of implementing the EWT. To do so, the report suggests a “prudential
approach” based on the observation of a significant increase in credit risk to de-
tect the emerging difficulties before they occur39. To achieve this objective, the
starting point should be the liquidity test (see subsection 2.4.1), complemented by
the balance sheet test to avoid the identification of false positives (for example,
companies experiencing temporary, and not structural, liquidity issues).
To track the liquidity test, the EWT should take into consideration the businesses’
capacity to pay the following obligations as they become due:
• Debts to suppliers;
• Debt service;
• Tax and VAT; and
• Payments to employees and social security.
Key questions that immediately arise are how to track an increase in credit risk
based on the ability to pay these debts, and of course, what indicators could be
used to signal those early difficulties (see subsection 5.1.4). A subsequent ques-
tion is: what distress severity must be reached to trigger an alert and lead to the
objective perception that the risks identified may give rise to insolvency? After
careful consideration of the sources of data available, the report suggests the
alert triggers described in Table 2.
Table 2: Financial Obligations, Data Sources, and Prospective Insolvency Indicators
Topic Data Source Main Perspective Insolvency Indicators (Alert Triggers) Rationale
Increased payable days, increased gap between receivable and Increased number of days needed to pay suppliers may
payable days40, unsubmitted financial statements highlight increasing liquidity issues. Increased time to
receive payments may highlight internal difficulties (debt
Financial collection) or decreased debtors’ creditworthiness.
reporting
Debts to Cash flow41 (decrease), sales (decrease), EBITDA or sales/financial All of these ratios provide key indications of business
suppliers debt (decrease), leverage (increase), capital adequacy (decrease), sustainability and capacity to repay debts.
current ratio (decrease), quick ratio (decrease)
Tribunals and Court notifications, debt rescheduling agreements, public an- All of these indicators are strong alerts of decreased
other sources nouncements referred to company, shareholders, beneficial own- overall debt repayment capacity.
ers, managers’ or owners’ involvement in bankruptcy procedures
39 For the purpose of this definition, it should be noted that a similar approach has been identified in the IFRS9 Standard. In view of applying prevention rather
than damage control, financial institutions are mandated to apply additional provisioning (e.g., based on a lifetime instead of 12 months of expected losses) to
those exposures where a significant increase of credit risk has taken place.
40 Estimation of payable and receivable days can be done through balance sheet items used to calculate the current ratio (payables and receivables) and sales and
cost of raw materials (or cost of goods sold) in the income statement.
41 Estimation of cash flow (for all companies not mandated to report the cash flow statement) can be done using the balance sheet items, yearly variations, and
income statement at the reporting date (direct or indirect method).
41
Estonia / Early Warning Tools
Technical Note / May 2020
Topic Data Source Main Perspective Insolvency Indicators (Alert Triggers) Rationale
Financial Cash flow (decrease), sales (decrease), EBITDA or sales/financial All of these ratios provide key indications of business
reporting debt, leverage, capital adequacy, current ratio, quick ratio, cash sustainability and capacity to repay debts.
ratio, decrease of client base, decreased number of suppliers,
unsubmitted financial statements
Payment performance (e.g., negative information) of companies, Emerging loan repayment difficulties of companies,
shareholders, beneficial owners and warrantors owners and ultimate beneficial owners, warrantors are
highly predictive alerts of perspective insolvency.
Debt
service Payment behavior and credit exposure (positive information): de- All of these indicators can significantly help to predict
creased number of credit facilities, decreased number of reporting future difficulties. These indicators will only be
Credit bureau institutions, rejected applications, increased number of queries, available when positive credit information
decreased credit limits, increased use of credit lines, rejected sharing is operational in Estonia.
renewals of credit facilities, switched financing from short-term and
self-liquidating credit facilities to term loans, debt rescheduling. All
of the above indicators can be referred to companies, sharehold-
ers, beneficial owners and warrantors.
Financial Cash flow (decrease), sales (decrease), EBITDA, sales/financial All of these ratios provide key indications of business
reporting debt, leverage, capital adequacy, Current ratio, quick ratio, cash sustainability and capacity to repay tax and VAT debt.
ratio, decrease of client base, decreased number of suppliers,
Tax and unsubmitted financial statements
VAT
Tax and Unsubmitted tax declarations, emerging tax arrears, increased
Customs tax debt and interest, postponed tax debt and interest (company,
Board beneficial, owners, shareholders)
All of these indicators are highly predictive and provide
Staff Same Same financial indicators, increase of social security debt, unpaid alerts on company capacity to face obligations.
social information social security debt (company, beneficial, owners, shareholders)
security sources as tax
and VAT
6.1.2 THE WHEN There is no indication in the Directive of how “early” those risks of insolvency
should be identified. It is therefore understood that each member state should be
able to determine the time horizon applied by the EWT, and such determination
must carefully manage the trade-off between prediction accuracy and effectivity
of the intervention.
Within that margin of discretion, it is reasonable to assume that the goal of the
EWT should be to identify those companies that are likely to face difficulties in the
following six to 12 months. This is an adequate time horizon for several reasons:
• Although projections within a time frame of less than six months may be more
accurate, they most probably identify hardly recoverable situations (given the
little time available).
• Although projections of more than 12 months may ensure more effective in-
terventions (given the significant time available), many concurrent factors ma-
terializing within this time frame may have an adverse impact on prediction
accuracy (changed context, availability of new data, and so on) and therefore
increase the possibility of false negatives.
42
Estonia / Early Warning Tools
Technical Note / May 2020
6.1.3 THE HOW Based on relevant case studies, methodologies to predict the likelihood of insol-
vency could be based either on predefined statistical models42 or on a more so-
phisticated and flexible approach grounded in artificial intelligence and machine
learning43.
The detailed description of both methodologies and of the main advantages and
drawbacks deriving from their application is explained in Annex A (which explores
a possible machine learning-based solution) and Annex B (which explores a pre-
defined statistical modeling).
6.1.4 THE PROCESS The model implementation process consists of three main phases: (1) model de-
sign; (2) implementation; and (3) operationalization. A detailed description of the
aims, tasks, and main deliverables expected from each phase is provided in Annex
A (machine learning-based solution) and Annex B (statistical model).
Regarding model design, based on the complexity of the definition and the wide
range of eligible alert indicators, the model implementation process should be
based on two main streams of activity that appear in both approaches (AI and
statistical modeling):
• Definition of a clear perimeter of variables, ratios, and thresholds by a work-
ing group44, for which a retrospective analysis of bankrupt or forcefully deleted
companies may be used.
• Once the perimeter has been defined, a quantitative analysis should validate
the preceding definition through statistic methodologies45, confirming that the
range of indicators under consideration accurately separates healthy compa-
nies from distressed companies.
A blended qualitative and quantitative approach combining these factors is the
best option to apply, as:
- The working group can provide overall indications (suggestions about data and
indicators in the report might be taken into consideration) and key inputs on
the perimeter to be identified.
42 See for example the scoring model implementation by private credit bureaus and internal rating modeling
implementation by financial institutions.
43 See for example the machine learning-based solution implemented by EWE in Denmark.
44 Based on the experiences observed in relevant case studies, the working group should be structured as
follows: commercial register, Statistics Estonia, Minister of Economy, central bank, significant financial
institutions, CreditInfo (and eventually other credit information providers), TCB, chamber of commerce
(and eventually other business associations), eventually chartered accountants’ association. Vertical groups’
roundtables should also be held to identify the relevant sets of information and alerts for each type of
data (e.g., financial ratios, credit data, tax and social security data, governance and other companies’ data,
macroeconomic information).
45 For these purposes, insolvency rates should be calculated for different clusters, and the movements of
insolvency rates may be analyzed through a transition matrix.
43
Estonia / Early Warning Tools
Technical Note / May 2020
- The quantitative and sampling analysis will provide key inputs to confirm the
definitions or highlight the need of further fine-tuning to enable better perim-
eter identification.
After the design phase, the working group will complete the implementation and
operationalization tasks46:
• Establish a project governance framework (including a definition of roles and
tasks of entities involved, and an implementation roadmap.
• Establish a data governance framework (including current and prospective
information needs, a definition of data sources and information providers,
elaboration of a data dictionary, a definition of relevant acceptance based on
data quality, key performance indicators, key risk indicators47, and potential
redundancies).
• Plan, define, and implement impact studies and pilot tests.
• Plan, define, implement, and support the definition and implementation of ac-
ceptance tests before EWT go-live.
• Plan, define, implement, and support the definition of a communication strate-
gy before EWT go-live.
• Plan the EWT’s operationalization phase and oversee the first period48 after
platform go-live (EWT’s consolidation phase).
Roles and tasks of the working group may vary depending on the definition of the
institutional arrangements surrounding the EWT.
46 See for example the European Banking Federation’s study on the Internal Rating Based Models implemen-
tation process (see http://www.ebf-fbe.eu/uploads/Internal_Rating_Based_models_in_Europe.pdf).
47 For a reference data governance model, see https://www.researchgate.net/publication/44939125_A_Mod-
el_for_Data_Governance_-_Organising_Accountabilities_for_Data_Quality_Management
48 Normally from six months to one year.
44
Estonia / Early Warning Tools
Technical Note / May 2020
6.2 MEASURING LIKELIHOOD
OF INSOLVENCY:
TRADITIONAL SOLUTIONS This Section describes most common approaches towards credit risk modeling
being the likelihood of insolvency a key element when measuring a debtor’s credit
VERSUS ARTIFICIAL risk49. Naturally, a significant number of models exist to measure counterpart’s
INTELLIGENCE credit risk, and the choice of the optimal modelling solution mainly depends on:
a) the type of counterpart (e.g. models to measure corporate credit risk, mort-
gage loans credit risk, retail credit risk), and b) the type of statistical and techno-
logical approach applying to credit risk measurement (ranging from univariate
correlation to highly complex neural network deep learning based on Artificial
Intelligence).
Consequently, a wide literature exists to describe all these models. It is not the
purpose of this Report to describe all main credit risk measurement models, and
some main references have been included in the Main References Section. Section
6.3 also provides a broad overview of other models that may also apply to the
context.
Given the specificities of the context and of the tool, the Team’s opinion is that two
main solutions can be a valid reference to be considered by Estonian authorities,
as follows:
- A traditional single statistical model based on regression analysis, drawing its
inspiration from the Basel framework compliant calculation of a Probability of
Default (PD),
- A more advanced ensemble learning model supported by a machine learning
technique.
Both models rely on statistical forecasting and are based on a statistical regres-
sion that uses historical data to project future impact on dependent variables,
mainly the likelihood of future insolvency. The main implications, advantages and
disadvantages deriving from each of the solutions mentioned above are summa-
rized in the following Sections.
6.2.1 A TRADITIONAL SINGLE Scoring systems developed under the Basel framework to determine credit expo-
STATISTICAL MODEL BASED sures’ credit risk play a prominent role among statistical models. In fact, since the
ON REGRESSION ANALYSIS: beginning of the millennium, national regulators agreed on identifying two ap-
PROBABILITY OF DEFAULT proaches to credit risk measurement: a “standardized approach” and an “internal
ESTIMATION UNDER THE rating based (IRB) approach” to be adopted by financial sector operators.
BASEL FRAMEWORK
Among most relevant benefits of such methodologies, the Basel framework pro-
motes a level playing field where financial operators apply well-described validat-
ed scoring models, resulting in fair cost of credit, lowered adverse selection, and
sound risk management. In fact, proactive and effective risk management results
49 Credit risk can be defined as the current or perspective risk of a debtor failing to meet its obligations in
accordance with agreed terms.
45
Estonia / Early Warning Tools
Technical Note / May 2020
in more predictive models, which in turn allows for less internal capital spending
when granting loans (that is, less capital is needed for the same amount of loans
disbursed).
Most relevant downsides lie in the complex and burdensome implementation and
validation activities (validation of an internal statistical model can be costly and
may take up to one year for systemically important financial institutions). More-
over, model risk can be relevant in situations of steady mutation of macroeco-
nomic context, when models should show not only robustness, but also a quick
adaptivity to reflect the changed context.
Both approaches are based on the measurement of Expected Losses, as follows:
EL=EAD x PD x LGD x M
Whereas:
EAD = exposure at default (for example, the expected amount of the expo-
sure when the event of default is expected to happen)
PD = probability of default (for example, the probability – a value from 0 to
1 – of a default happening in over a 12-month time horizon)
LGD = loss given default (for example, the loss amount after deducting
the recoverable amount of eventual risk mitigation instruments, such as
collaterals)
M = maturity (for example, the expected maturity of the credit instrument)
In IRB modeling, the PD is the backbone of credit risk modeling under the Basel
framework. For the purpose of PD calculation, statistical models (often based on
multivariate regression analysis50) for the different portfolio segments are imple-
mented and undergo a periodic validation process by both financial institutions
and the supervisory authority51.
The 4 main steps of PD calculation under the Basel framework are as follows52:
1. Definition of default: The definition of default used by banks to compute their
PDs follows the definition of default of Basel 3, in some cases complemented by
guidance from national authorities. However, in addition, the entities may in-
ternally decide on the definition of cure periods or amount of threshold, which
have an impact on the number of defaults and therefore on the final PD.
50 As the name implies, multivariate regression is a technique that estimates a single regression model with
more than one outcome variable. When there is more than one predictor variable in a multivariate regres-
sion model, the model is a multivariate multiple regression
51 See for example the ECB Targeted Review of Internal Models (TRIM)
52 See EBF’s “Study on Internal Rating Based models in Europe” http://www.ebf-fbe.eu/uploads/Internal_Rat-
ing_Based_models_in_Europe.pdf
46
Estonia / Early Warning Tools
Technical Note / May 2020
2. Build-up of data series/data samples: This is critical for the construction of the
models and for the calibration of the PD. At present, it is a wide-spread practice
to use different time periods for model construction and for model calibration
purposes (e.g. data series for calibration purposes can be more limited). Most
of the banks use time series of up to 10 years for modelling purposes, trying to
capture at least 2 economic cycles. In general, data series including recession
periods tend to be more conservative. Broadly considered, portfolio character-
istics and available data (quality and historical length) limit modelling choices.
In this regard, as a general practice in most jurisdictions, regulators require a
minimum number of years for the analysis, which is different across countries,
and determine exclusions of some historical data – for example, the period
2009-10 – or the inclusion of some assumptions. Another element to be taken
into account is the weighting method for the different years included in the
sample. When calibrating, most banks give the same weight to all the available
years with the rest following alternative approaches.
3. Application of a scoring model: Regarding the scoring model, some banks have
developed models for new clients, existing clients distinguishing between those
with enough financial information from those without, clients with some form
of delinquency and employees. At the time Basel II was implemented in Europe,
regulators admitted different types of scoring models. For retail portfolios, it
is very common to find behavioral and application models, which are used to
assign a score to the obligor and/or the operation. In general, banks develop
both types of models; whereas the behavioral models use most recent informa-
tion and therefore are considered more procyclical, the application models only
take into account information at one point in time. Regarding the behavioral
approach, the common steps usually undertaken for PD calculation are briefly
described below:
- Univariate analysis, aimed at identifying single variables with high correla-
tion with the future default event (e.g., predictors),
- Multivariate analysis, aimed at shortlisting the list of variables identified by
the univariate analysis, eliminating redundancies, and identifying the opti-
mal combination of predictors,
- Logistic regression, to identify subsets of stratified population of borrowers
with homogeneous features (for example: sectors) and assign the weights of
predictors in each subset.
4. PD calibration: Once a score is given to an operation/borrower, the following
step is the assignment of PD to the score. This process could be broken down
in two stages:
- The computation of Observed Default Frequencies (ODF) and
- The transformation of ODF into PD, taking into account the cycle adjustments.
47
Estonia / Early Warning Tools
Technical Note / May 2020
Model calibration mainly depends on how ODFs are computed (simple average,
average by operation or by borrower, the treatment of withdrawn operations,
etc.) and on the different methodologies for cycle adjustment, which could be due
to either the choice of the financial entity or to regulatory requirements.
Regarding the first step, most of the banks compute PD as simple averages, but in
some cases, they choose a customer weighted method.
As was mentioned previously, most banks have behavioral scorecards, which
are by nature more pro-cyclical. However, banks consider the nature of their PD
models as mixed or hybrid mainly due to the way they compute their PD, which
is linked to the second step of the calibration process. In relation to the cycle ad-
justment, banks make use of different methodologies when computing the cycle
adjustment. Among the most common methodologies used by the banks are the
following:
- Variable scalar methodologies.
- Add-on capital requirement or increase of PD.
- Usage of longer historic data series.
Lastly, another key topic that affects PD estimation is frequency of recalibration.
Most of the banks adopting Basel compliant PD modelling recalibrate annually or
quarterly, but there are other institutions that do it with a lesser frequency.
A typical example of PD calibration is provided below53:
Figure 3: Scoring based on PD calibration
Description Rating Grade Mark Rating Grade Numeral PD min Average PD PD max
Enterprises with lowest probability of default A0 1 0.00% 1.11% 1.40%
Enterprises with very low probability of default A1 2 1.40% 2.04% 2.72%
Enterprises with low probability of default A2 3 2.72% 3.05% 3.34%
Enterprises with increased probability of default A3 4 3.34% 4.36% 5.54%
Enterprises with below-average probability of default A4 5 5.54% 6.81% 8.32%
Enterprises with above-average probability of default A5 6 8.32% 9.15% 10.11%
Enterprises with average probability of default A6 7 10.11% 12.48% 15.33%
Enterprises with above-average probability of default A7 8 15.33% 18.04% 21.49%
Enterprises with high probability of default A8 9 21.49% 38.18% 99.99%
Enterprises with very high probability of default A9 10
Enterprises in default, classified in A90 B1 11
Enterprises in default, classified in B1 B2 12 100%
Enterprises in default, classified in B2 B3 13
Enterprises in default, classified in B3 CC 14
53 National Bank of Croatia “Behavioral model of assessment of Probability of Default of Non-fi-
nancial corporations” (2018), https://www.hnb.hr/documents/20182/2259329/w-056.
pdf/05593eb1-da2d-4cc1-8b8b-33a2b6d5afc9
48
Estonia / Early Warning Tools
Technical Note / May 2020
6.2.2 A MORE ADVANCED The evolution of artificial intelligence and data analytics together with the in-
ENSEMBLE LEARNING MODEL creased technological capacity of managing large volumes of data have had a
SUPPORTED BY MACHINE significant impact on the evolution of credit risk management and counterparts’
LEARNING insolvency risk measurement techniques54.
Among the new data intelligence techniques, machine learning is commonly re-
ferred to as a set of methods that can automatically detect patterns in data, and
then use the identified patterns to predict future data or to perform other kinds of
decision making under uncertainty.
Pattern detection using machine learning is not based on coding or predefined
statistical algorithms, which allows for implementing highly flexible and cost-ef-
fective models.
Two main approaches to machine learning are used, namely:
1. Supervised machine learning (often based on structured data). This approach is
normally used to identify predefined samples, such as companies in distress;
and
2. Unsupervised machine learning (often based on unstructured data). This tech-
nique is normally used for the elaboration of large amounts of data to discover
unpredicted trends or still unidentified classifications (for example, customer
behaviors, customer segments, marketing operations targets). In other words,
users may not know exactly what to detect when applying this technique in
respect to the available sets of data.
A brief and simplified explanation of how techniques based on machine learning
work is provided here:
Figure 4: The Machine Learning Workflow
DATA TRAINING MODEL
SERVE
PREDICTIONS
ANSWER
QUESTIONS
54 On this matter, see the BIS “How do machine learning and non-traditional data affect credit scoring? New
evidence from a Chinese Fintech firm” (2019) https://www.bis.org/publ/work834.htm
49
Estonia / Early Warning Tools
Technical Note / May 2020
In a nutshell, machine learning algorithms are used to learn the structure of the
data, and then can be used for prediction, description, and any other purposes.
The next figure describes what machine learning is basically about: data intelli-
gence and pattern identification through structured or unstructured data.
Figure 5: The Purpose of Machine Learning
The most commonly used machine learning techniques for pattern recognition
are55:
• Linear regression;
• Logistic regression;
• K-nearest neighbors (KNN);
• Decision tree;
• Random forest; and
• Gradient boosting.
Selecting machine learning techniques often depends on the expected outcome
(for example, the question to be answered) and on the choice of acceptable trade-
off between model performance, predictivity, and interpretability. A brief expla-
nation of advantages and disadvantages of each of these techniques is provided
in Table 3.
55 To have an overall view of ML techniques, see:
https://scikit-learn.org/stable/supervised_learning.html#supervised-learning
50
Estonia / Early Warning Tools
Technical Note / May 2020
Table 3: A Brief Explanation of the Main ML Techniques56, Advantages and Downsides
ML Technique Definition Main Advantages Main Downsides
Linear regression The linear regression algorithm is used to Fast to implement, easy to understand, Linear regression performs poorly
see a relationship between the predictor less prone to overfitting when there are non-linear relation-
variables and explanatory variable. ships. It is hard to use on complex
data sets.
Logistic regression It is commonly used to predict the Fast to implement, easy to understand, Algorithm performs poorly where
probability of an event occurring. less prone to overfitting there are multiple or non-linear
decision boundaries or capturing
complex relationships.
K-nearest neighbors It is an algorithm that classifies objects Good for large data, it learns well on Hard to find the right K-value, bad
based on its closest training example in complex patterns, it can detect linear for higher dimensional data, and
its featured space. The “k” in “k-nearest or non-linear distributed data, and is requires a lot of computation when
neighbors” refers to the number of robust to noisy training data. fitting larger amounts of features.
nearest neighbors the model will use for
its prediction.
Decision tree A decision tree is made up of nodes. Each Algorithm is very easy to understand, It loses its predictive power by not
node represents a question about the robust to outliers, and can work on collecting other overlapping features.
data, and the branches from each node non-linear relationships. It is a complex tree that does not
represent the possible answers. generalize well to future data, hard to
interpret. Duplication is possible.
Random forest Grouping algorithm averaging numerous Accuracy, high predictive power Poor intelligibility
decision trees together (ensemble
learning model)
Gradient boosting Ensemble method training many models High precision rate (e.g., performing A balance of computational cost and
and taking a consensus of their prediction well when minimizing false positives). accuracy is a concern. Interpretability
to transform a weak model into a strong It’s one of the best off-the-shelf can be difficult.
one. Good for tabular data, with a limited algorithms used for great accuracy with
number of variables. decent run time and memory usage.
Among criteria used to evaluate the suitability of a machine learning technique,
the most commonly used are:
• Predictive accuracy (performance);
• Speed (how fast does it take for the model to deploy);
• Scalability (can the model handle large datasets);
• Robustness (how well does the model handle outliers or missing values); and
• Interpretability (is the model easy to understand).
Advantages of machine learning techniques—cost-effectiveness, flexibility, and
predictivity—are important aspects to consider. These benefits are a direct conse-
quence of an approach that doesn’t need any coding or pre-defined algorithms57
to identify connections among dependent and independent variables in the
datasets.
56 Cells in yellow highlight the optimal machine learning technique identified by the Team for the imple-
mentation of the EWT.
57 See for example the punctual identification of correlations between predictors and default described in
the previous Section.
51
Estonia / Early Warning Tools
Technical Note / May 2020
Consequently, machine learning facilitates the syncing of models with changed
contexts, which in turn represents a significant issue for predefined and coded
algorithms of statistical models58.
Among most relevant disadvantages, application of machine learning—and more
in general, artificial intelligence techniques—to datasets can lead to unwanted
consequences. In fact, totally open data intelligence can detect patterns or behav-
iors that are specific to certain segments of the population (or customers, compa-
nies, and so on). This in turn may lead to the identification of personal information
like race, gender, religion, political views, and so on, and eventually to some form
of discriminatory scoring/evaluations59. Therefore, responsive data governance
(for example, identification of a minimum relevant set of information that may
reasonably lead to pattern identification) and solid model risk management to be
applied by expert data scientists are preconditions of avoiding unwanted conse-
quences deriving from the application of these techniques.
The second barrier when it comes to applying machine learning techniques is the
concept of “explainability,” which refers in IT jargon to the “the black box issue.”
Explainability of the outcomes is key to all stakeholders involved, such as borrow-
ers, lenders, and regulators.
Explainability is the act of presenting textual or visual artifacts that provide quali-
tative understanding of the relationship between the instance’s components (for
example, application data, words in text, regions in an image) and the model’s
prediction.
Identifying effective solutions to make these relationships fully intelligible by us-
ers of platforms based on machine learning can often be a complex task.
Lastly, another potential barrier when applying machine learning techniques is
overfitting, defined as a significant decrease of model predictivity when shifting
from training to test datasets60 (e.g. limited capacity of the model to identify gen-
eral patterns regardless the data observed).
Normally, mitigators of the risk of overfitting are a) Focus on dataset quality (to
minimize the “noise” effect, e.g. presence of inconsistent information or outliers)
and b) Choice of an appropriate machine learning technique (ensemble learning
techniques61 based on several models normally tend to minimize overfitting).
Naturally, trade-offs between mitigation of risk of overfitting and all other factors
(for example: interpretability) should be taken into consideration when identifying
the appropriate machine learning technique that will fit for the purpose.
58 BIS (Bank of International Settlements) identified ineffective model risk management and poor statistical
model predictivity as enabling conditions of the 2007-2008 global financial crisis (see Rakesh Mohan: Global
financial crisis: causes, impact, policy responses and lessons https://www.bis.org/review/r090506d.pdf).
59 For credit risk measurement techniques, this is often referred to as “creditworthiness by association.” For
example, a poorly managed model risk may result in lower credit scores (or higher probability of default)
for companies established in a given neighborhood or zone of a country.
60 See Elite data science “Overfitting in machine learning” for a comprehensive description of overfitting and
how to prevent it. https://elitedatascience.com/overfitting-in-machine-learning.
61 See Cha Zang and Yunqian Ma, “Ensemble machine learning techniques: methods and applications” (2019).
52
Estonia / Early Warning Tools
Technical Note / May 2020
6.3 OVERVIEW OF OTHER
MODELS
6.3.1 MODELS USED BY OTHER The report has briefly analyzed the tools and the distressed companies’ assistance
EWTS process in Belgium and Denmark. One of the clearest lessons learned from this
process is that there is no “one size fits all” tool, and every model works well to
support the need in its own context. Relevant specificities of the Belgian and Dan-
ish alerting tools are described in Table 4.
The team also analyzed the EWT system in Germany, although detailed analysis of
this system was discarded, because it did not rely on an alerting tool but rather on
an inbound model62 and soft information63.
Table 4: Early Warning Tools in Belgium and Denmark: A Comparative Analysis
Topic Denmark Belgium
Platform ownership Public consortium 64
Private
Platform development Internal (Danish Business Authority) External65
Platform scope All registered companies (sole proprietorships are All registered companies
excluded)
Definition of “distressed company” Discretional definition (counseling service’s judgment) Discretional definition (judges’ decision on shortlist-
ed companies receiving notifications)66
Data used by the tool Financial information in business register A wide range of information including financial
information, tax, VAT, social security arrears, seized
payments, payment behavior, court sentences, and
bankruptcy procedures
Information processing and data Machine learning (x-gradient boosting technique) Expert statistical model based on a decision tree and
intelligence calculation of a probability of insolvency
Output List of potentially distressed companies (filtered by List of potentially distressed companies (filtered by
technical assistance providers) judges)
Type of notification and recipients Companies are contacted by assistance providers67 A letter is sent by the tribunal to summon companies
Platform cost and funding Roughly 500.000 EUR for tools implementation68 Development dates back many years (established in
the early 2000s) and costs are unknown (privately
funded69).
62 E.g., companies contact a local NGO (Team-U), providing several types of support to companies in distress,
such as bankruptcy and a second chance.
63 Mostly based on a status check, e.g., a questionnaire available on Team-U’s website.
64 Danish Business Authority and Early Warning Europe.
65 The tool aimed at identifying lists of companies at risk was implemented by a private credit reporting
agency (Graydon) more than 15 years ago.
66 However, some pre-established criteria exist, such as (1) when financial statements are not submitted (past
two reporting dates); and (2) when company managers have previously undergone bankruptcy or insolven-
cy proceedings.
67 EW Denmark regional hubs.
68 In 2016 Early Warning Europe (part of the consortium owning the EWT in Denmark) benefited from 3.8
million EUR funding from DG GROW and EASME under the European Commission COSME program. The
funding covered the 75% of expenses for a three-year support of a European network for early warning
and for support of enterprises and second starters (see https://ec.europa.eu/easme/sites/easme-site/files/
documents/call_text_eweu_2016.pdf).
69 A trade agreement exists between the private credit reporting operator and national authorities: costs for
EWT development and maintenance are borne by the credit reporting agency, which in turn has free access
to relevant information that is used for its core business (court sentences, notifications, information from
various registries, etc.).
53
Estonia / Early Warning Tools
Technical Note / May 2020
6.3.2 OTHER MODELS Professor Edward Altman in 1968 pioneered the application of discriminant anal-
ysis to the prediction of corporate bankruptcy with what he called the “Z-score.”
Altman’s model is one of the first statistical models based on multivariate analysis
(linear discriminant analysis) aimed at identifying companies in distress.
Using accounting data on healthy and bankrupt companies, Altman calculated
the financial ratios used by accountants and analysts to assess the solvency of
business firms, and obtained the following “discriminant function” to distinguish
healthy companies from those with a high probability of bankruptcy:
Z-score = 0.717 x X1+0.847 x X2+3.107 x X3+0.420 x X4+0.998 x X5
Whereas:
X1 = % working capital / total assets
X2 = % retained earnings / total assets
X3 = % EBIT70 / total assets
X4 = market value of shares / total Assets
X5 = sales / total assets
Altman’s statistical study showed that the Z-score was usually distributed for both
groups and the respective means of the distributions was 4.14 for healthy compa-
nies and 0.15 for those that were bankrupt.
Thus, by obtaining the Z-score for a company, one can decide to which group it
is more likely to belong (healthy or bankrupt) by performing the corresponding
statistical tests of hypothesis.
This comment underlines one of the main characteristics of all statistical tech-
niques, and discriminant analysis is no exception; namely, that absolute certainty
as to which group a company belongs is impossible. One can only place a com-
pany in one group or the other with a certain probability or confidence level. This
means that mistakes can and will be made, but this is true of any method.
The important thing here is that in discriminant analysis, as opposed to more tra-
ditional techniques, the actual probability of making a mistake and the associated
costs can be estimated. Figure 6 illustrates the concept.
Speaking in general terms, discriminant analysis identifies a distribution of scores
for “good” debtors and another for “bad” debtors. Notice that the shaded area in
Figure 6 is ambiguous, in the sense that there are small probabilities of compara-
ble orders of magnitude, and that debtors with scores in this region can belong to
one group or the other.
70 EBIT=Earnings Before Interest and Taxes.
54
Estonia / Early Warning Tools
Technical Note / May 2020
This points to the well-known “type I” and “type II” errors in statistics, which in the
language of credit scoring means:
1. Placing the debtor in the healthy group when it really belongs in the bad one
(type I error); or
2. Placing the debtor in the bad group when it is good (type II error).
Thus, when a score falls in the fuzzy area, in the decision of which group to place
the debtor, there are two elements to consider; namely, the probabilities of
the debtor belonging to one group or the other, and the costs associated with
misclassification.
Figure 6: Altman’s Discriminant Analysis
Number of firms
(Frequency)
Bankrupt Group Healthy Group
“Z-score”
0.15 4.14
Type I and II Errors
Lastly, other commonly used credit risk modeling techniques are Merton/Black
and Scholes, Moody’s modelling, KMV and Credit Metrics. However, these models
are generally more focused on corporate credit risk (as described, large corpo-
rates only cover a small share of the market in Estonia, and the EWT should be
mostly focused on identifying MSMEs facing difficulties).
Therefore, reference on relevant literature on these models can be found in the
Main References Section.
55
Estonia / Early Warning Tools
Technical Note / May 2020
6.4 REVIEW OF OPTIONS
This section has described three models from three different waves of innovation
of statistical modeling:
1. The first wave, based on multivariate analysis (for example, Altman’s model71);
2. The second wave, based on logistic regression (for example, internal scoring
methodologies under the Basel framework, based on the calculation of a prob-
ability of a default event)72; and
3. The third and more recent wave, based on ensemble learning techniques (sup-
ported by machine learning73) and AI in general.
After analyzing the specificities of the Estonian context and relevant benchmark
models, the report identifies two possible EWT implementation solutions74:
1. A machine learning tool based on ensemble learning, gathering businesses’
financial, tax and social security, governance, and credit information (current
credit payments defaults). Some macroeconomic indicators may also be used
by the model.
2. A traditional single statistical model based on regression analysis, gathering
businesses’ financial, tax and social security, governance, and credit informa-
tion (current credit payments defaults).
A brief outline in terms of rationale, data used, modeling, project governance and
project phases is described in the following sections for both suggested options.
Main solutions’ potential opportunities and drawbacks are also described to facil-
itate final decision making by national authorities.
6.4.1 OPTION 1: MACHINE a. Rationale behind the adoption of the suggested solution
LEARNING TOOL
Adoption of a machine learning based approach supports the implementation of
a flexible, cost-effective, and highly predictive solution.
Availability of a relevant quantity of data, as discussed in section 5, has the po-
tential to support the setup of training and test environments, identifying prob-
abilities of distress, false positives and negatives (for example, type I and type II
errors), and achieving a high predictive capacity.
Finally, additional predictive information (for example, company credit histories
and ultimate beneficial owners) can be easily integrated into the model when
available to further enhance model predictivity.
71 See section 6.3.
72 See section 6.2.1.
73 See section 6.2.2.
74 The report considers that while the Z-score may be used as a data point both for sampling purposes and
model implementation, both machine learning and logistical regression may better suit the purpose of the
implementation of an EWT.
56
Estonia / Early Warning Tools
Technical Note / May 2020
Data used
Table 5 describes the main data points used by the model (and the report strongly
suggests that the punctual definition of the data model be made by the working
group)75. Highlighted cells refer to positive information (not currently available).
Table 5: Main Data Points Used by the Machine Learning Tool
Data Classification Main Data Points Data Source and Reporting
Information Provider Frequency
Financial data Date of submission of financial statements, annual variation of all RIK Yearly
balance sheet, income, cash flow statements items, historical trends of
liquidity, debt sustainability, capital adequacy, leverage, profitability and
productivity ratios (including receivable, payable, and inventory days)
Tax, VAT, and social For companies and their (1) subsidiaries and parent companies; (2) TCB Monthly
security data company and ultimate beneficiary owners; and (3) shareholders:
submitted and unsubmitted tax declarations, tax arrears, tax debt and
interest, postponed tax debt and interest (company, company owners,
warrantors), postponed social security debt
Payment performance For companies and their (1) subsidiaries and parent companies; (2) Negative credit registry Daily
(payment default, e.g., company and ultimate beneficiary owners; (3) shareholders; and (4)
negative information) warrantors: dates the debts were incurred and settled; range of amount
at the origin; and number and outstanding amount of arrears
Credit histories (e.g. For companies and their (1) subsidiaries and parent companies; (2) Credit bureau Daily
credit positive data) company and ultimate beneficiary owners; (3) shareholders; and (4)
warrantors: number and disbursed amount of credit facilities, number
of reporting institutions, number of accepted and rejected applications,
number of queries, amount and variation of credit limits, use of credit
lines, renewals of credit facilities, and rescheduled debts
Company activity, Business name, register code, legal form and date of registration RIK One-off and updates
structure, legal form, (number of years in business), company headquarters (address), regis-
governance tered fixed capital, business type, status (registered, under liquidation,
receivership, etc.), company representatives, company shareholders,
subsidiaries and associates, beneficial owners, regular representation
rights, parent company’s main sphere of business
Other information Court notifications, debt rescheduling agreements, public announce- Various data sources76 Daily
ments referred to companies, shareholders, beneficial owners, manag-
ers’ or owners’ involvement in bankruptcy procedures
Macroeconomic data GDP, price index (inflation), unemployment rate, income and wages, Bank of Estonia, Quarterly
reference interest rates, balance of trade, sector-specific statistics (in Statistics Estonia,
particular: agriculture, manufacture, and services contribution to GDP ECB statistical data
and interlinkages) warehouse
75 See section 6.1.3.
76 Collection of this information is already made by CreditInfo.
57
Estonia / Early Warning Tools
Technical Note / May 2020
b. The importance of positive credit information
Following many international case studies, positive information of credit histo-
ries has a strong predictive power. In fact, many financial sector operators and
supervisory authorities globally use positive credit data to capture trigger events
identifying portfolio deterioration77, such as:
• Progressive stress of overdrafts (for example, the amount used to get more
and more close to the authorized limit);
• Swapping revolving or self-liquidating facilities with term loans; and
• Decrease of reporting institutions of a debtor, and so on.
Granularity of information and its multi-faceting (for example, type of loan, pur-
pose, maturity, payment schedule, performance, status) typically enables detailed
pattern identification and risk portfolio segmentation. In international risk man-
agement practice, this information is used for:
• The setup and validation of IRB modeling;
• Portfolio staging classification under IFRS9;
• Support to portfolio segmentation according to the accounting standard;
• Ongoing evaluation and monitoring of provisioning policies;
• Setup of alert systems for early identification of portfolio segments to undergo
strict periodic monitoring; and
• Ongoing estimation of internal capital to face expected losses.
In general, highly granular credit information exploited by machine learning can
help identify patterns of deterioration earlier than can financial ratios (represent-
ing a snapshot of the company situation made available with a significant delay)
and tax or credit payment defaults.
c. Data modeling
Like the Danish model, the data model should be based on the gradient boosting
technique.
This technique is an ensemble learning method based on the aggregation of sev-
eral different models to enhance predictivity and minimize the risk of overfitting.
The machine therefore develops a model, determines the errors produced, and
then develops a new optimized model. The machine iterates the model optimiza-
tion operation until an acceptable level of performance is achieved. The detailed
machine learning tool implementation methodology is described in Annex A.
77 See the International Committee of Credit Reporting (ICCR), chaired by the World Bank, publication “The
Role of Credit Reporting in Supporting Financial Sector Regulation and Supervision” http://pubdocs.world-
bank.org/en/954571479312890728/CR-2016-role-credit-reporting-in-supporting-financial-regulation.pdf.
58
Estonia / Early Warning Tools
Technical Note / May 2020
6.4.2 OPTION 2: A TRADITIONAL a. Rationale behind the adoption of the suggested solution
SINGLE STATISTICAL
MODEL BASED ON LOGISTIC According to many case studies, the implementation of a statistical model based
REGRESSION on logistic regression analysis can provide alerts based on a confidence score78
through clearly defined algorithms and achieving acceptable predictivity and
interpretability.
Development of such a model is typically a standardized activity79, with a relatively
low project risk. The option of building the model upon existing companies’ risk
evaluation algorithms may be taken into consideration by national authorities.
b. Data used
Table 6 describes the main data points used by the model. The report strongly
suggests that the punctual definition of the data model be made by a working
group80 to incorporate the specificities of the local context.
Table 6: Main Data Points Used by the Statistical Model
Data Classification Main Data Points Data Source and Reporting
Information Provider Frequency
Financial data Date of submission of financial statements, annual variation of all RIK Yearly
balance sheets, income, cash flow statement items, historical trends of
liquidity, debt sustainability, capital adequacy, leverage, profitability and
productivity ratios (including receivable, payable, and inventory days),
sectorial indicators, insolvency rate81
Credit information Dates the debts were incurred and settled, range of amount at the Negative credit registry Daily
origin, number and outstanding amount of arrears
Tax, VAT, and social Submitted and unsubmitted tax declarations, tax arrears, tax debt and TCB Monthly
security data interest, postponed tax debt and interest (company, company owners,
warrantors), postponed social security debt
Company activity, Business name, register code, legal form and date of registration RIK One-off and updates
structure, legal form, (number of years in business), company headquarters (address), regis-
governance, sectorial tered fixed capital, business type, status (registered, under liquidation,
data receivership, etc.), company representatives, company shareholders,
subsidiaries and associates, beneficial owners, regular representation
rights, parent company’s main sphere of business
Other information Court notifications, debt rescheduling agreements, public announce- RIK One-off and updates
ments referred to companies, shareholders, beneficial owners, manag-
ers’ or owners’ involvement in bankruptcy procedures
Macroeconomic GDP, price index (inflation), unemployment rate, income and wages, Bank of Estonia, Statics Quarterly
indicators reference interest rates, balance of trade, sector-specific statistics (in Estonia, ECB statistical
particular: agriculture, manufacture, and services contribution to GDP data warehouse
and interlinkages)
78 E.g., the likelihood of an insolvency event taking place within the next six to 12 to 24 months.
79 Main steps of the modelling process are described in Section 6.2. and Annex B
80 See section 6.1.3.
81 Data described earlier are used by the model in the training phase to perform univariate and multivariate
analysis to define the shortlist of relevant trigger events.
59
Estonia / Early Warning Tools
Technical Note / May 2020
Highlighted cells refer to a set of macroeconomic data to be tracked for periodic
statistical model’s calibration, validation and fine-tuning activities. However, mac-
roeconomic data are not part of the dataset used for the statistical model. Annex
B provides a detailed description of the modeling approach implementation, main
project phases, and statistical model project governance.
6.4.3 MAIN SOLUTIONS’ OPPORTU- A recap of the most relevant opportunities and drawbacks deriving from the
NITIES AND DRAWBACKS adoption of solution 1 versus 2 are described in Table 7:
Table 7: Opportunities and Drawbacks Deriving from Highlighted Solutions
Solution 1: Machine Learning Tool Solution 2: Traditional Single Statistical Model
Main opportunities 1. Introduction of an innovative tool; model is adaptive 1. Apparently easier and less costly implementation compared to
and can easily evolve and leverage additional solution 1.
predictive information at low cost. 2. Model may be built upon existing algorithms (see, e.g., credit score
2. Data intelligence and pattern recognition through or expert model of TCB tool).
model training enable applying one model to all
clusters and types of company and to automatically
sync the model with changed macroeconomic context.
3. Availability of positive credit information has potential
to maximize the effectivity of a machine learning
tool (machine learning tools performs better with
significant volumes of granular data, e.g., payment
behaviors of companies, company owners, and
ultimate beneficial owners, warrantors).
Main drawbacks 1. Complex data integration given the high quantity of 1. Rigidity of the model. Availability of additional information to be used
data. in the future will require the development of a new statistical model.
2. Choice of institutional solution may be conditioned by 2. Ongoing model syncing process with changed context is needed.
limitations in data sharing82. 3. Model may be less predictive in case of lack of significant infor-
3. Interpretation of main risk drivers for a given mation. Additional datasets and models may be needed for some
confidence score can be complex. company clusters.
4. Data governance and model risk should be carefully 4. Calculation of benchmark ratios and historic ratio trends (if not yet
handled to avoid incurring in identifying patterns go- available) may be a burdensome activity.
ing beyond the purpose of the platform and entering 5. Shortlisting country-specific indicators and identifying the optimal
in the private sphere of the entrepreneurs. model may be complex and subject to risk of analysts’ bias.
Based on these considerations, although both solutions are valid and will enable
the compliance with Article 3 of the Restructuring Directive, the opinion of the
report is that decision making about the most suitable approach should consider:
a) Whether future EWT enhancements are expected to include additional gran-
ular data in the future (that is, if the scope of the tool is expected to evolve in
terms of data covered) and b) Most relevant EWT’s requirements to comply with
(e.g. main needs of platform, taking into consideration the possible tradeoffs:
minimization of false positives vs. interpretability, model predictivity vs. ease of
implementation).
82 E.g., borrowers’ consent is needed for sharing positive credit information with third parties.
60
Estonia / Early Warning Tools
Technical Note / May 2020
If behavioral information (e.g. positive credit histories) is expected to support the
EWT in the near future and model predictivity and minimization of false positives
are considered as top priorities, then adopting solution 1 may be the preferred
option for Estonian Authorities.
In addition, a mixed approach may also be selected by national authorities to
further mitigate the main drawbacks of both solutions: implementation of a tra-
ditional single statistical model83 in the short term as a “quick win” and buildup of
a machine learning tool to leverage the power of granular data in the medium to
long term.
6.5 INSTITUTIONAL ASPECTS Regarding institutional aspects, Table 8 describes the main suggestions, which
are further expanded in the following subsections.
Table 8: Institutional Aspects
Solution 1: Machine Learning Tool Solution 2: Traditional Single Statistical Model
Ownership RIK
Platform Possible development solutions: Possible development solutions:
development
1. Internal development with support by a specialist 1. Internal development by RIK
(hosting by RIK) 2. Outsourced development by a specialist (hosting by RIK)
2. External development by CreditIinfo (hosting by 3. Outsourced development by CreditInfo (hosting by CreditInfo)
CreditInfo)
4. Outsourced development by a third party (hosting by RIK)
3. External development by CreditInfo. (hosting by RIK)
5. Outsourced development by TCB (hosting by TCB)
6. Outsourced development by TCB (hosting by RIK)
Access to Main types of access to the list of distressed companies are described here (non-mutually alternative):
information by
companies
1. Free access to RIK via APIs when accessing https://www.eesti.ee after login (access to basic information)
2. Free access to RIK via notification service https://www.rik.ee/en/e-business-registry/notification-service (access to basic information)
3. Counseling service (full information is disclosed)
6.5.1 MODEL OWNERSHIP As described earlier, the report suggests public ownership of the EWT. Under this
assumption, ownership by RIK seems the best option for the EWT, as RIK:
1. Is the official counterpart when incorporating companies and a financial infor-
mation sharing process; and
2. Already has a notification service.
83 Or customization of an existing statistical model to fit with the EWT’s purposes.
61
Estonia / Early Warning Tools
Technical Note / May 2020
6.5.2 PLATFORM DEVELOPMENT Solution 1 (machine learning tool)
Table 9 describes the main advantages and drawbacks deriving from this option.
Table 9: Opportunities and Drawbacks Deriving from
Highlighted Solutions (ML Tool)
Options Main Opportunities Main Drawbacks
Option 1: Platform • Opportunity to enhance internal • Data integration process (retrieving
development—internal capacities dealing with AI data from credit bureau, TCB, and
development with other sources) may be complex,
support by a specialist expensive, and burdensome.
• Necessity to establish proper service
level agreements (SLAs)
• In case future use of positive credit
information is planned, consent by
borrowers will be required.
Option 2: Platform • Opportunity to enhance internal • Higher budget costs than Option 1
development—external capacities dealing with AI • Necessity to establish proper SLAs
development by
• No data integration needed,
CreditInfo
as all necessary information is
already stored by CreditInfo (in-
cluding negative information).
• No additional data processing
cost
• CreditInfo is a global firm with
relevant skill and track record
handling AI.
• Easier and faster integration
of positive credit information
when available
Option 3: Platform • Opportunity to enhance internal • Necessity to establish proper SLAs
development — capacities dealing with AI • Credit information and tax/VAT/
external development
• No data integration needed, as social security transfer protocols to
by CreditInfo (hosting
most necessary information is be defined
by RIK)
already stored by CreditInfo (in- • In case future use of positive credit
cluding negative information). information is planned, consent by
• CreditInfo is a global firm with borrowers will be required.
relevant skill and familiarity
with AI-related topics.
62
Estonia / Early Warning Tools
Technical Note / May 2020
Solution 2 (traditional single statistical model)
Table 10 describes the main advantages and drawbacks deriving from this option.
Table 10: Opportunities and Drawbacks Deriving from Highlighted Solutions (traditional single statistical Model)
Options Main Opportunities Main Drawbacks
Option 1: Platform • Full project ownership • Relevant engagement in terms of human resources for
development—internal data integration and model implementation
• Lower expected budget costs
development by RIK
• Credit and tax/VAT/social security information transfer
protocols to be defined
• In case future use of positive credit information is
planned, consent by borrowers will be required.
Option 2: Platform • Few internal resources engaged in the data integration and • Higher budget costs than Option 1
development—internal model development process • Necessity to establish proper SLAs
development with
support by a specialist • Credit and tax/VAT/social security information transfer
protocols to be defined
• In case future use of positive credit information is
planned, consent by borrowers will be required.
Option 3: Platform • The project will benefit from economies of means, as most • Higher budget costs than Option 1
development—External necessary information is already stored by CreditIinfo (including • Necessity to establish proper SLAs
development by credit negative information, financial information, tax and social
CreditInfo (hosting by security payment history). • Necessity to implement an interface to export lists of
CreditInfo) companies in distress
• CreditInfo has the skill and capacity to implement an alert
system. A credit scoring model is already available and may be • In case future use of positive credit information is
customized to achieve the specific purpose of the EWT. planned, consent by borrowers will be required.
• No additional costs for data processing
Option 4: Platform • Credit Info has the skill and capacity to implement an alert • Higher budget costs than Option 1
development—external system. • Necessity to establish proper SLAs
development by
• A credit scoring model and a range of additional services84 are • Credit and tax/VAT/social security information transfer
CreditInfo (hosting by
already available and may be customized to achieve the specific protocols to be defined
RIK)
purpose of the EWT.
• In case future use of positive credit information is
planned, consent by borrowers will be required.
Option 5: Platform • The project will benefit from some economies of means, as some • Higher budget costs than Option 1
development—external necessary information is already stored by TCB (e.g., company • Necessity to establish proper SLAs
development by TCB information and tax payments).
(hosting by TCB) • In case future use of positive credit information is
• TCB has already developed a tool aimed at identifying companies planned, consent by borrowers will be required
with tax payment issues. Such a tool may be customized to
achieve the specific purpose of the EWT. • Credit information transfer protocol to be defined.
Option 6: Platform • The project will benefit from some economies of means, as a • Necessity to establish proper SLAs
development—external data transfer channel is already operational between RIK and • Credit information transfer protocol to be defined
development by TCB TCB.
(hosting by RIK) • Tax and social security payment information transfer
• TCB has already developed a tool aimed at identifying companies protocol to be defined
with tax payment issues. Such a tool may be customized to
achieve the specific purpose of the EWT. • In case future use of positive credit information is
planned, consent by borrowers will be required.
84 Including Credit Opinion, calculating the probability of insolvency within 12 months.
63
Estonia / Early Warning Tools
Technical Note / May 2020
6.5.3 ACCESS TO INFORMATION Information calculated by the EWT should be made strictly confidential, fully ac-
cessible, and available for free to notified companies. Therefore, the best solution
seems to be to:
1. Prioritize e-notification using existing company credentials; and
2. Ensure that companies receive the notification in due time.
Given the world-class ICT standards of e-government and e-citizenship in Estonia,
using both the e-government portal and online business register notification ser-
vice seems to be the best option for accessing data about distressed companies.
When elaborating notifications, two different sets of information may be
calculated:
1. Basic information about the situation of distress and the probability of insol-
vency, as calculated by the model; and
2. Complete information about the situation of distress and the probability of in-
solvency, highlighting the areas of concern and the severity of each area and
showing critical indicators.
Only synthetic information on increasing the probability of distress should be dis-
played when submitting a notification for the first time. In fact, companies may
need technical assistance to fully understand key indicators and determinations
of the EWT. Access to a complete set of data without any advisory support may
result in a company owner’s underestimating the importance of a notification or
making uninformed decisions about the business’s future.
For this reason, RIK should:
1. Notify companies about the situation of distress, specifying that counseling is
available to address the difficulties;
2. Make the documentation of EWT methodologies available online (broad indica-
tions, to facilitate businesses’ understanding of key indicators).
3. Provide (RIK or platform manager) a first level of support (call center) to inform
companies of the purposes and the content of notifications.
4. Offer access to the complete information to those companies that acknowl-
edge the notification either by contacting the call center or confirming that they
have read and understand the tool’s methodology.
A RIK’s representative and the debtor must sign a nondisclosure agreement when
the debtor accepts counseling and before any access to EWT data.
64
Estonia / Early Warning Tools
Technical Note / May 2020
7. THE PROCESS OF
ALERTING DEBTORS
65
Estonia / Early Warning Tools
Technical Note / May 2020
7.1 SELF-HELP Regardless of the notification procedure selected (the main options are described
in the following section), debtors should have the right to access the results of
the model at any time in order to check their situation (for example, whether they
are on the list of notified companies). Information can be accessed either via the
“single window” or RIK portal, or via “my CreditInfo85,” depending on the model
implementation solution adopted.
7.2 SUBMITTING A When a notification is to be submitted by a third party, the notification procedure
NOTIFICATION can be based on one of two approaches:
1. E-notification to company administrators sent via RIK’s notification service. De-
pending on selected aspects of corporate law in Estonia that the report does
not cover, the notification may be extended to related parties, including share-
holders and members of the supervisory board. Automated emails should
contain an “acknowledge receipt.” If the email is not received or receipt is not
acknowledged after a certain period, a hard communication (a letter to a com-
pany’s official address as registered in the e-business register) should be sent
by the notification service.
2. Combined e-notification from RIK and contact by the counseling service for
companies identified with severity 1 (that is, a higher likelihood of becoming
insolvent). Companies identified with severity 2 (that is, a lower likelihood of
becoming insolvent) are notified only of the relevant likelihood of insolvency
and could contact the counseling service themselves.
Table 11 describes the main advantages and drawbacks deriving from the notifi-
cation options.
Regarding the frequency of notifications, two options are available:
1. Notifications should be sent periodically to all companies in distress (for ex-
ample, a company to which a distress solution has been reported receives au-
tomated notifications on subsequent reporting dates until the difficulties are
resolved or the business is liquidated; and
2. Notification to one company can occur only when the distress emerges, and
once within a given lapse of time (for example, up to 18–24 months, considered
a reasonable time to readdress a distressed situation86). The main reason for
this approach is twofold: (1) the EWT’s impact on companies’ business should
be minimal (for example, company managers and owners should be informed
only when significant anomalies emerge, avoiding false positives); and (2) dis-
tressed situations are not expected to be resolved within a short amount of
85 My Credit Info is the credit bureau’s portal disseminating all value-added services to the customers.
86 See for example the cure period indicated in European Banking Authority (EBA) guidance on management
of nonperforming and forborne exposures (https://eba.europa.eu/regulation-and-policy/credit-risk/
guidelines-on-management-of-non-performing-and-forborne-exposures).
66
Estonia / Early Warning Tools
Technical Note / May 2020
time, and multiple automated notifications deriving from the same unresolved
situation may result in unwanted stress for business owners/managers, and
should be avoided.
Final decision making on the notification process option and frequency should
be made by the working group during the design and operationalization phases.
Table 11: Opportunities and Drawbacks Deriving from Highlighted Solutions (Company Notification Procedure)
Options Main Opportunities Main Drawbacks
Option 1: e-notification • RIK is perceived as a neutral counterpart by companies. • Neutral communication based on data may not help support
from RIK company owners experiencing personal distress.
• Communication channel already established between
RIK and companies. • Company owners may fear consequences of the notification by
a public authority, and this may lead to unintended effects of
notifications.
Option 2: notification • Support by RIK and counseling service for different • Increased complexity of the model to identify different and
from RIK and counseling levels of likelihood of insolvency may help solve the homogeneous risk classes
service according to issues described in options 1 and 2 (e.g., low RIK ca- • Model’s reputational risk may increase significantly (e.g., risk
level of severity of pacity to deal with personal distress, lack of confidence of incurring false positives).
company distress of entrepreneurs toward a third party, and unknown
institution). • Increased complexity of company classification and notification
procedure (for example, how to manage infra-classification
• Counseling service focuses on situations where likeli- movement of a company from severity 1 to severity 2, and
hood of insolvency is more significant. vice-versa)
The notification workflow for the highlighted options is described in the follow-
ing figure:
Figure 7: Notification Process
CREDIT
CREDIT REGISTRY
DATA
TAX AGGREGATOR
BUSINESS COUNSELING
REGISTRY
TAX DATA
BIZ DATA
MODELING
BIZ REGISTRY
DATA
PROCESSING
67
Estonia / Early Warning Tools
Technical Note / May 2020
The notification process should be based on an incremental and blended ap-
proach, meaning that:
• After a pilot phase (to test the efficiency and effectiveness of the model) a “bulk
notification” should be sent from the EWT to all companies with a significant
probability of distress (as defined in previous sections). This phase is expect-
ed to demonstratively engage the platform manager, notification service, and
counseling support as (1) a significant number of businesses is expected to
receive notifications; and (2) businesses are expected to have little or no infor-
mation about the early warning tool87.
• In case of adoption of the one-off notification approach, after a first bulk notifi-
cation is sent, monthly updates of the model’s elaboration will identify only new
companies in distress (for example, companies that have not received notifica-
tions within the past 18–24 months).
• Following relevant case studies88, a sanity check by experts seems necessary
to validate the content of the notification list calculated by the EWT. In fact,
final validation of the list may ensure that model outcomes effectively identify
situations of distress, mitigating the risk of notification to false positives. Taking
into consideration that an incremental approach is suggested, the sanity check
(performed either by the platform owner or platform manager) should interact
with automated model outcomes in order to enable future notifications to com-
panies that have been classified as distressed by the model and then removed
from the list by the experts.
7.3 NETWORK RESPONSIBLE As mentioned earlier, the network responsible for delivering advice may adopt
FOR DELIVERING ADVICE an inbound model (where the distressed company requests assistance from the
counseling service) or an outbound model (where the counseling service ap-
proaches the distressed company and highlights the likelihood of distress and
offers assistance).
Even if both approaches have proved to be valid, based on cultural specificities
(for example, the overall willingness of public authorities to minimize the interfer-
ence with the private sector and entrepreneurial activity), a hybrid model may be
the best fit for Estonia. Therefore, companies receiving the notifications from RIK
may decide to contact the counseling service to receive support.
87 Like the information campaign that took place when XBRL reporting standards were adopted, a similar
activity (information to businesses through public events, seminars, business associations) is strongly
suggested to mitigate the risk of having an excessive burden managing requests of information from
businesses after the EWT’s go-live.
88 See for example Belgium and Denmark.
68
Estonia / Early Warning Tools
Technical Note / May 2020
For the purpose of identifying the entity that may deliver advice to distressed com-
panies, some key aspects to consider are:
• Target companies are expected to be SMEs, as roughly 95 percent of compa-
nies are SMEs in Estonia, and larger companies usually have easier access to
advisory services for support;
• A volunteering culture seems weak in Estonia (therefore an inbound model
with a network of volunteers, like the one in Germany, seems difficult to estab-
lish); and
• Given the combination of financial and personal difficulties that company own-
ers typically face when their business is in distress, trust in known counterparts
is critical.
Based on these considerations, a solution suggested by the team is that the Net-
work of County Development Centers89 could manage the advisory services.
This not-for-profit organization consists of about 160 employees (including nine
consultants, with experts in financial distress/insolvency) and is represented in
all 15 counties of Estonia. The entity supports approximately 5,000 entrepreneur
clients per year country wide by consulting on business ideas, the development
of business plans, access to grants, and offering mentoring and training services.
Based on opinions expressed during the WBG team’s missions in Estonia, the Net-
work of County Development Centers enjoys a good reputation within the private
sector.
Given the unlikelihood of establishing a network of volunteers, a more realistic ap-
proach for Estonia would be for (1) a company to contact the counseling service;
(2) the service to conduct a first screening to identify criticalities and company
needs; and (3) the service to contact a professional matter expert to provide finan-
cial advisory support to the company.
Although the activity of the counseling service may be free of charge, support by
the matter expert should be remunerated based on two distinct components:
1. An hourly fee; and
2. A success fee, when the objectives are met, and the company’s expectations
are satisfied.
The main advantages of this approach are:
• The company views the counseling intervention as neutral, and not aimed at
making a profit from its distress;
89 Maakondlike Arenduskeskuske Vorgustik (www.arenduskeskused.ee).
69
Estonia / Early Warning Tools
Technical Note / May 2020
• The company need only pay for the matter expert’s services (if it chooses addi-
tional support);
• The company perceives the value of the matter expert’s services because they
are not free; and
• Reliance on a network of matter experts is flexible, that is, based on the needs
expressed by the company.
The main drawbacks of this approach are:
• The company in distress may not be willing or able to remunerate an expert,
even based on success; and
• Due to company demand, a matter expert may not be available to the compa-
ny. This may have a serious impact for when financial distress should be tackled
rapidly.
Regardless of the hosting authority selected for company support, counseling
support should be given by a team of experts with a comprehensive professional
and personal skillset, ensured by the following activities:
• An intensive and ongoing counseling team capacity building90 process;
• Identification of the right mix of team skills both in terms of technical capacity
(for example, adequate skills to support all types of companies from different
industries) and personal capacity (for example, adequate skills to support peo-
ple living in distressed personal situations, empathy, communication skills); and
• Setup of an adequate organizational infrastructure (for example, a backbone
of full-time staff supported by a network of volunteers—ex-entrepreneurs who
have experienced failure or a second start in their working life—may be the
best option)91.
The identification of the most suitable counseling service for companies notified
by the EWT should start as soon as possible—and in parallel with the design of the
algorithm described in the annexes—as the adequate skillset and service infra-
structure implementation may require at least 12 months.
90 Based on its consolidated expertise, Early Warning Europe may support this training phase with a “train the
trainers” activity.
91 The suggestion is based on examples of company support in other countries (Denmark, Germany, and
Belgium).
70
Estonia / Early Warning Tools
Technical Note / May 2020
MAIN REFERENCES
Association for Information Systems (AIS). “A model for data governance—Organ-
ising accountabilities for data quality management.” (2018) https://www.research-
gate.net/publication/44939125_A_Model_for_Data_Governance_-_Organising_Ac-
countabilities_for_Data_Quality_Management.
Bank of International Settlements (BIS) “How do machine learning and non-tradi-
tional data affect credit scoring? New evidence from a Chinese Fintech firm” (2019)
https://www.bis.org/publ/work834.htm
Cha Zhang, Yunqian Ma, “Ensemble machine learning: methods and applications”
(2019)
Deutsche Bundesbank, “The Merton Model: A structural approach to default predic-
tion”, https://www.bot.or.th/Thai/FinancialInstitutions/Highlights/IFRS9_Doc/The%20
Merton%20Model.pdf
ECORYS. “Bankruptcy and Second Chance for Honest Bankrupt Entrepreneurs”
(2014). https://www.eea.gr/system/uploads/asset/data/9042/Bankruptcy_and_sec-
ond_chance_for_honest_bankrupt_entrepreneurs_FINAL_REPORT.pdf.
European Banking Authority (EBA). “Guidelines on management of non-perform-
ing and forborne exposures.” (2018) https://eba.europa.eu/regulation-and-policy/
credit-risk/guidelines-on-management-of-non-performing-and-forborne-exposures.
European Banking Federation (EBF). “Study on Internal Rating Based (IRB) Models
in Europe.” (2012) http://www.ebf-fbe.eu/uploads/Internal_Rating_Based_mod-
els_in_Europe.pdf.
European Central Bank (ECB) “Guidance to European Union (EU) Banks on Non-Per-
forming Loans” (2017) https://www.bankingsupervision.europa.eu/ecb/pub/pdf/
guidance_on_npl.en.pdf?b2b48eefa9972f0ca983c8b164b859ac.
European Commission “Business Dynamics: Start-Ups, Business Trans-
fers and Bankruptcy”(2011). https://ec.europa.eu/growth/content/
start-ups-business-transfers-and-bankruptcy-0_en.
EU Directive 2019/1023 of the European Parliament and Council “On preventive
restructuring frameworks, on discharge of debt and disqualifications, and on mea-
sures to increase the efficiency of procedures concerning restructuring, insolvency
and discharge of debt” (2019). https://eur-lex.europa.eu/legal-content/EN/TXT/
PDF/?uri=CELEX:32019L1023-.
EU Directive 2013/34, “Accounting Directive”(2013). https://eur-lex.europa.eu/
legal-content/EN/TXT/?uri=CELEX%3A32013L0034.
71
Estonia / Early Warning Tools
Technical Note / May 2020
EU Regulation 2016/679 of the European Parliament and of the Council “General
Data Protection Regulation”(2016). https://eur-lex.europa.e.u/legal-content/EN/TXT/
PDF/?Uri=CELEX:32016R0679.
International Committee of Credit Reporting (ICCR). “Facilitating SME financing
through improved credit reporting.” (2016) https://siteresources.worldbank.org/
EXTFINANCIALSECTOR/Resources/282884-1395933501015/Facilitating_SME_financ-
ing_through_CR_public_comments_web.pdf.
ICCR. “The role of credit reporting in supporting financial sector regu-
lation and banking supervision.” (2018) http://pubdocs.worldbank.org/
en/954571479312890728/CR-2016-role-credit-reporting-in-supporting-financial-reg-
ulation.pdf.
Journal of Business and Economic Development, “Default probability in Black
Scholes Formula: A Qualitative Study” (2017)
Michael Frenkel, Ulrich Hommel, Markus Rudolf, “Risk Management, challenge and
opportunity”, Springer (2017)
Moody’s analytics, “An overview of modelling credit portfolios”, https://www.
moodysanalytics.com/-/media/whitepaper/2013/2013-30-06-Overview-Model-
ing-Credit-Portfolios.pdf (2013)
National Bank of Croatia “Behavioural model of assessment of Probability
of Default of Non-financial corporations” (2018), https://www.hnb.hr/docu-
ments/20182/2259329/w-056.pdf/05593eb1-da2d-4cc1-8b8b-33a2b6d5afc9
World Bank, “Preventive Restructurings in Estonia” (2019)
72
Estonia / Early Warning Tools
Technical Note / May 2020
ANNEX A: IMPLEMENTATION OPTIONS:
MACHINE LEARNING SOLUTION
1. Model Implementation
The proposed solution based on machine learning uses an ensemble learning
method (gradient boosting technique), where predictive models on training data
are iteratively tested until a satisfactory model predictivity and performance are
achieved. The model optimization process is described in Figure 8.
Figure 8: Ensemble Learning Applied to Data Modeling
The sample data set consists in companies that have been in operation for the
past three years (the report discourages including startups in the EWT’s scope92).
A subset of this sample consists in “positive cases” (for example, bankrupt and
forcefully deleted companies. The second subset of the dataset (“negative cases”)
consists in businesses with no negative event in the past two years. These compa-
nies should undergo validation through sampling by information providers in the
working group. Eventual modifications or add-ons to the lists may be proposed.
Classification of positive and negative cases should be made using the following
information:
• Bankrupt and forcefully deleted companies from the commercial registry; and
• Negative cases from the commercial registry and credit bureau.
92 The reason for this suggestion is twofold: (1) startups typically benefit from different types of support
offered by public institutions and public-private partnerships (for example, seed funding, sandboxes, incu-
bators, technical assistance); and (2) including startups in the model may introduce “noise” (for example,
data being rendered meaningless given the high variance), with a potentially significant impact on the
model’s predictivity.
73
Estonia / Early Warning Tools
Technical Note / May 2020
Given the importance of the sampling for the subsequent model training and test
phases, the report strongly suggests applying a blended approach, for example:
• Agreement of the working group on a set of rules aimed at identifying positive
and negative cases in the sample (for example, earlier described rules, eventual
add-ons ensuring that the sample is highly homogeneous);
• A broad definition of the dataset based on the algorithms described earlier (or
other additional algorithms established by the working group); and
• Subsequent qualitative fine-tuning by the working group.
For modeling purposes, the report suggests considering selecting companies
within a time series of at least 24–36 months93 and having a subsequent 12 months
to eventually perform back-testing94.
Figure 9 describes the data sampling process:
Figure 9: Data Sampling Process
DATA
PROVIDERS 50%
random
selection
SET OF RULES SINGLE MODELS
AND MANUAL ∑ ∑ TEST SAMPLE
VALIDATION
POSITIVE TRAIN TEST
SAMPLE TRAINING TEST
SAMPLE SAMPLE SINGLE MODELS
TEST SAMPLE
50%
random
NEGATIVE TRAIN TEST selection
50%
random
selection
After defining the dataset (as a rule of thumb, a sample of a minimum 20,000 com-
panies should be defined, with an indicative event rate of 10 percent), a random
selection of 50 percent should be made to split the population for training and test
purposes (see Figure 9).
93 With a five-year data history considered the preferred option.
74
Estonia / Early Warning Tools
Technical Note / May 2020
Then, by using training cases, at least nine models should be created, out of which:
• Eight models to compare the pattern evolution of companies with positive
events from eight quarters to one quarter before event with companies with
negative events; and
• One model for fitting an ensemble model.
For this purpose, a random selection (50 percent) on the test dataset is made to
perform testing of single models and testing of the ensemble model, whereas
all datasets (single models and ensemble model) contain the same number of
companies.
As described, the training dataset of the ensemble model uses the predictions of all
eight individual datasets to elaborate a final prediction and a final confidence score.
Figure 10 describes the main phases of the modeling process. Bidirectional arrows
indicate iterative procedures aimed at optimizing the model’s performance95.
Figure 10: Boosting Technique Modeling Process
CREDIT
TRAINED TRAINED ENSEMBLE
CREDIT MODEL (T-8) MODEL (T-1) MODEL
ET
L
REGISTRY
... ∑
TAX
ETL
TAX DATA
TRAINING
DATA SET
BIZ
L
ET
ETL
SINGLE MODELS SINGLE MODELS RAW LIST OF FINAL LIST OF
BIZ REGISTRY TEST DATASET TEST DATASET COMPANIES IN COMPANIES
DISTRESS WITH IN DISTRESS
CONFIDENCE
SCORE
MACRO
ECONOMIC ...
DATA
SCORING SCORING ENSEMBLE
DATA (T-6) DATA (T-1) SCORE
75
Estonia / Early Warning Tools
Technical Note / May 2020
As described in Figure 10, the outcome of the workflow is the list of companies in
distress. For a punctual indication of those companies, a confidence score (for exam-
ple, probability of being in a distressed situation) ranging from 0 to 1 is calculated.
The confidence score calculation is a key element in the workflow, as it allows
platform managers to fine-tune the process, minimizing the number of false pos-
itives and identifying the adequate levels of risk severity triggering notifications
based on how effective interventions after notifying the distressed situation are
(through an impact analysis).
2. Description of Main Project Phases
Table 12 briefly describes most relevant project phases, activities, actors involved,
and timeline (estimation of overall 15 months from platform design to go-live):
Table 12: Project Roadmap
Project Phase Broad Description of Activities Involved Counterparts Involved Estimated Timeline
Definition of project Project timeline, roles and tasks, and expected deliverables are Working group (informa- Six months
roadmap, governance, punctually established in this phase. Data points and sample dataset tion providers, platform
platform specifications, are also defined. developer, main financial
and sampling institutions, central bank,
chambers of commerce)
Platform Implementation should consist in the following four phases: Platform developer Three months
implementation
1. Developing code for data input (ETL), data review for missing data
and invalid values, data standardization, normalization, tokeniza-
tion, bucketization, definition of hyperparameters (preprocessing)
2. Training the model
3. Transforming input data into calculated properties, fostering
evaluation’s transparency
4. Elaborating a confidence score for every company and the final list
of companies
Platform validation Validation should consist in the following two phases: Working group Three months
1. Platform fine-tuning through working group’s validation of
distressed companies list
2. Back testing analysis to confirm predictivity performance
Pilot and go-live Pilot and go-live should consist in the following two phases: Platform developer and Three months
platform owner
1. Pilot test on a sample of new companies for final test of platform
and overall notification and support process effectivity
2. Go-live. Information is automatically made available and notified
through the established channels
76
Estonia / Early Warning Tools
Technical Note / May 2020
3. Project Governance
The WBG team’s suggestion for project governance is that separate entities
should have distinct goals to ensure both steering capacity and operational focus.
Therefore, a steering committee should perform project oversight and manage-
ment. The main participants in the steering committee should be the platform
owner, the platform developer, and the primary information providers.
The working group should provide vertical support to the different phases of the
tool’s implementation process:
• Data sampling and pre-processing;
• Training process;
• Validation; and
• Pilot and go-live.
In addition, independent advisory by a third party may help deal with unforeseen
issues, maximizing the effect of lessons learned in other countries implement-
ing similar tools. Independent advisory may also be provided by Early Warning
Europe.
A synthetic scheme of the team’s proposition regarding project governance is de-
scribed here:
Figure 11: Project Governance Arrangements
STEERING COMMITTEE
PLATFORM MANAGER ADVISOR
INFORMATION PLATFORM OTHER CENTRAL BANKS AND
PROVIDERS DEVELOPER INSTITUTIONS LENDERS’ REPRESENTATIVES
· Periodical working progress check · Periodical working progress check
· Quality checks and acceptance tests · Quality checks and acceptance tests
· Proposals for solutions of criticalities · Proposals for solutions of criticalities
OPERATIONAL TEAM
WORKING GROUP
1. SAMPLING 2. MODEL TRAINING 3. MODEL VALIDATION 4. PILOT AND GO-LIVE
77
Estonia / Early Warning Tools
Technical Note / May 2020
ANNEX B: IMPLEMENTATION OPTIONS:
TRADITIONALSINGLE STATISTICAL MODEL BASED ON
REGRESSION ANALYSIS
1. Model Implementation
The suggested approach based on logistic regression (or logit model) is common-
ly used to predict the probability of an event occurring96. Logistic regression is
an algorithm borrowed from statistics and uses a logic/sigmoid function (purple
formula) to transform its output, making it either 0 or 1.
The approach can then be extended to model several classes of events, and each
observation can be assigned a probability between 0 and 1. Using multivariate
analysis means acknowledging the complexity of reasons for which the likelihood
of insolvency can increase (for example, combining several different risk alerts—
each one with a risk weight—to calculate the confidence score).
A multitude of concurrent factors (for example, risk drivers) can be approached in
different ways in combining risk alerts.
When suggesting the optimal approach, the report takes into consideration cer-
tain aspects:
• Some risk alerts can be significantly more important than others (for example,
higher correlation with the likelihood of a future insolvency event); and
• Importance of minimizing the quantity of false positives97.
Based on these aspects, the report suggests a modeling approach structured
around a decision tree.
This approach allows (1) the definition of a clear workflow that identifies com-
panies with early signs of distress; as well as (2) the identification of the most
predictive signals, and therefore the measurement of the weight of each indicator
within the dataset.
Figure 12 gives an example of a decision tree. Country specificities are not con-
sidered when showing indicators in the decision tree. Preliminary correlation and
univariate analysis on financial ratios of a sample of Estonian companies may lead
to the identification of different ratios.
96 Most Internal Rating Based (IRB) systems complying with ECB regulations are implemented using logistic
regression.
97 See the WBG team’s considerations in section 2, and the importance of avoiding companies’ reputational
risk of being notified at all cost.
78
Estonia / Early Warning Tools
Technical Note / May 2020
Figure 12: EWT’s Decision Tree98
Negative or below regulatory
NET WORTH threshold EMERGING TAX, VAT, AND
OTHER PAYMENTS ARREARS
Positive or above regulatory
threshold
DEBT SERVICE True
COVERAGE RATIO (DSCR)
DSCR below LIKELIHOOD OF
threshold INSOLVENCY
DSCR above threshold,
unavailable, or unreliable
True
SECTORIAL ANALYSIS
Confidence score
Confidence score below cut-off value
below cut-off value OTHER RELEVANT ALERTS
(court notifications,
UNLIKELY INSOLVENCY debt rescheduling,
public announcements)
As anticipated, key indicators in the decision tree may reflect national specificities
and outcome of univariate and multivariate analysis.
A key phase of building up the model is the identification of early warning indica-
tors. As described in subsection 4.1.4, the ratio identification process should cover
at least the following topics:
• Debt sustainability;
• Capital adequacy;
• Liquidity;
• Liquid return on assets; and
• Tax and social security debt.
Alert triggers deriving from each of these areas should be based on the bench-
marking analysis of each company in its group of peers (for example, sectorial
analysis).
98 The example takes its inspiration from the Italian early warning system, whose adoption is still under-
way. The tool identifying a list of companies at risk in Belgium is also based on a decision tree (e.g., key predic-
tive indicators and statistical analysis measuring a probability of insolvency within 12 months).
79
Estonia / Early Warning Tools
Technical Note / May 2020
As described, selecting the ratio shortlist should be based on the following steps:
1. Statistical univariate analysis of each statistical indicator of the those areas (for
example, correlation of each risk driver with a future insolvency event); and
2. Multivariate analysis of shortlisted indicators aimed at identifying the optimal
combination of indicators identifying future distress.
As for point 1, indicators’ predictive capacity can be calculated using the following
ratios:
1. Gini99;
2. Gap between “good” and “distressed” companies’ ratios median values; and
3. Insolvency rates of indicator’s quantiles, focusing on its risk area (tail of indica-
tor’s distribution).
Examples of correlation analysis between some indicators and the probability
of a default event within a 12-month horizon to identify a shortlist of indicators
through univariate analysis are described here:
Figure 13: Correlation Analysis: Current Ratio100
6%
Default Frequency
5%
4%
3%
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Percentile
99 The Gini coefficient is one of the most popular metrics used by the financial industry for evaluating the
performance of credit score models. In practice, the Gini coefficient is a metric that indicates the model’s
discriminatory power, namely, the effectiveness of the model in differentiating between “bad” borrow-
ers, who will default in the future, and “good” borrowers, who won’t default in the future. See: https://
towardsdatascience.com/using-the-gini-coefficient-to-evaluate-the-performance-of-credit-score-models-
59fe13ef420 for more details on Gini and other most relevant credit scoring performance metrics (e.g. ROC
and AUC)
100 The indicator is calculated by CERVED (Italian balance sheet register) and clearly indicates a high correlation
between default frequencies and percentiles of the statistical distribution of the current ratio (e.g., percent
current assets/current liabilities): an increase of current ratio normally results in a decrease of insolvency
rates.
80
Estonia / Early Warning Tools
Technical Note / May 2020
Figure 14: Correlation Analysis: Percent Interest Costs / Sales
12%
11%
10%
9%
Default Frequency
8%
7%
6%
5%
4%
3%
2%
1%
0%
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Percentile
Other criteria to be used when selecting indicators are:
• Ease of calculation;
• Lack of computational problems;
• Interpretability;
• Minimum share of outliers; and
• Low correlation among shortlisted indicators.
Analysis aimed at shortlisting relevant alert indicators should be allowed at least
four years (the best option would be an entire economic cycle) before the insol-
vency event takes place.)
As a result, univariate analysis should lead to a shortlist of indicators, as follows
(example). Selected highlighted indicators have higher values of the accuracy ra-
tio101 (AR)102:
101 AR measures the overall predictive capacity of minimizing false positives and false negatives, e.g., the
discriminatory power in a classification model. Percent Debt/Sales and DSCR have the higher AR.
102 For more details about this table, see “Corporate credit risk modelling: Quantitative rating system and
probability of default estimation – Joao Eduardo Fernandes”: http://www.ratingexpress.net/content/JEF_Cor-
porateCreditRisk.pdf.
81
Estonia / Early Warning Tools
Technical Note / May 2020
Figure 15: Shortlisting Financial Indicators (example)
Type Name Definition Expected AR
Effect on PD
Net P&L / Assets [Net Profit & Loss] / [Total Assets] - 15.60%
Current Earnings / Assets [Current Earnings] / [Total Assets] - 16.60%
Profitability
Current Earnings and Depreciation / Turnover [Current Earnings + Depreciation] / [Turnover] - -1.00%
P&L / Assets [Net Profit & Loss + Depreciation + Provisions] / [Total Assets] - 5.80%
Gross Earnings / Production [EBT + Depreciation + Provisions] / [Production] - -3.20%
EBITDA / Production [EBITDA / Production] - -15.80%
Liquidity / Current Liabilities [Bank Deposits & Cash + Marketable Securities] / [Short- - 10.60%
Term Liabilities]
Liquidity
Current Ration [Current Assets] / [Short-Term Liabilities] - 9.20%
Liquidity / Assests [Bank Deposits & Cash + Marketable Securities] / [Total - 12.00%
Assets]
Equity / Assets [Equity] / [Total Assets] - 3.80%
Equity / Accounts Payable [Equity] / [Accounts Payable] - 4.40%
Leverage / Gearing
Bank Debt / Accounts Payable [Bank Debt] / [Accounts Payable] + 1.80%
Accounts Payable / Assets [Accounts Payable] / [Total Assets] + 5.20%
Liabilities / Assets [Total Liabilities] / [Total Assets] + 3.80%
Net Current Accounts Payable / Assets [Short-Term Accounts Payable - Bank Deposits & Cash] / + -0.40%
[Total Assets]
Gross Earnings / Liabilities [Current Earnings + Depreciation + Provisions] / [Total - 9.80%
Liabilities]
Debt Coverage
Debt Service Coverage [Current Earnings + Depreciation] / [Interest & Similar Costs] - 25.20%
P&L / L-T Liabilities [Net Profit & Loss + Depreciation + Provisions] / [Long-Term - 0.00%
Liabilities]
Operating Earnings / Debt Service [Operating Earnings] / [Interest & Similar Costs] - 16.60%
Interest Costs / Sales [Interest & Similar Costs] / [Turnover] + 40.20%
Activity
Inventories / Turnover [Inventories] / [Turnover] + 12.00%
Turnover / Assets [Turnover] / [Total Assets] - 20.40%
Produc- Productivity Ratio [Personnel Costs] / [Turnover] + 15.80%
tivity
Note: Turnover = Total Sales + Services Rendered
82
Estonia / Early Warning Tools
Technical Note / May 2020
Regarding point 2 (multivariate analysis of shortlisted indicators), identification of
alert threshold can be made through a top-down approach based on information
entropy103.
Box 6: Information Entropy
Information entropy is the average rate at which information is produced by a stochastic source of data. When the
data source produces a low-probability value (that is, when a low-probability event occurs), the event carries more
“information” (“surprisal”) than when the source data produce a high-probability value. In general terms, entropy
is considered the measure of disorder (or, conversely, the measure of purity) of available information (in this case,
shortlisted indicators). Being a logarithmic function, information entropy ranges from 0 (low level of disorder, or
high level of purity) to 1 (high level of disorder, or low level of purity).
This is a widely used approach that performs well both at identifying decision trees and at bucketing values of a
continuous variable. Application to a set of ratios allows the identification of cut-off values, dividing the population
in different homogenous groups: in this case “good” companies and “distressed” companies.
Analysis through information entropy should be performed on the overall sample
and then on segmented portfolios (sectors). In fact, different cut-off values may
characterize different sectors. Training sector-specific risk alerts can be funda-
mental to achieving acceptable model performance.
Sector-focused analysis should be aimed at minimizing the trade-off between in-
formation granularity and predictivity. In line with global regulatory standards,
the opinion of the report is that the classification complying with NACE standards
established by EU Regulation 1893/2006104 may well suit the purpose of the tool.
To enhance the effectivity of the tool, different benchmark models should be de-
fined based on expert analysis and on single ratios identified in the univariate and
correlation analysis.
Identified conceptually robust models should be compared with many alternative
models selected using ratios from the earlier listed areas of analysis through a
supervised combination analysis, to maximize the heterogeneity of signals and
full coverage of areas of concern.
As anticipated, as model accuracy is fundamental in general, minimizing the num-
ber of potential false positive companies receiving notifications should be a key
criterion when identifying the optimal model.
Once the optimal model has been defined in terms of sectors and ratios, insolven-
cy rates for companies falling within the established ranges of selected indicators
are calculated. Based on this, the model establishes critical thresholds for every
sector (see Table 13).
103 See https://towardsdatascience.com/entropy-how-decision-trees-make-decisions-2946b9c18c8.
104 The first level of the NACE hierarchy may be applied, e.g., single-digit classification. Considering more
detailed clusters may lead to statistically weak conclusions given the limited number of companies in Estonia.
83
Estonia / Early Warning Tools
Technical Note / May 2020
Table 13: Critical Values by Sector and Indicators (template)
RATIOS ALERT THRESHOLDS
NACE CLASSIFICATION DEBT CAPITAL SHORT-TERM SASH FLOW / TAX AND SOCIAL
SUSTAINABILITY ADEQUACY LIQUIDITY ASSETS SECURITY DEBT
A-Agriculture, forestry and fisheries
B+C+D
E
F41
F42+F43
G45+G46
G47+I56
H+I55
J+M+N
P+Q+R+S
Companies’ ranking in the statistical distribution of each indicator allows the
calculation of a weighted expected insolvency rate for every reporting date. In-
solvency rates above an established value (highlighting a significant likelihood of
insolvency within one year105) are the trigger of the alerting system (for example,
when the company is notified of the status calculated by the system).
Finally, for implementation and validation purposes, a comparison of historical
trends of estimated 12, 24, and 36 months calculated by the model and the overall
insolvency rates of companies should help sterilize the market effect and periodi-
cally evaluate the reliability of model outcomes.
105 E.g., the risk of migration toward insolvency is significantly higher than the median insolvency rate. The risk
of migration is calculated through a transition matrix.
84
Estonia / Early Warning Tools
Technical Note / May 2020
2. Description of Main Project Phases
The most relevant project phases, activities, actors involved, and timeline (estima-
tion of overall 15 months from platform design to go-live) are briefly described in
Table 14:
Table 14: Project Roadmap
Project Phase Broad Description of Activities Involved Counterparts Involved Estimated Timeline
Definition of project Project timeline, roles and tasks, and expected deliverables are Working group (informa- Six months
roadmap, governance, punctually established in this phase. Data points and sample dataset tion providers, platform
platform specifications, are also defined. Benchmark indicators, statistical distributions, and developer, main financial
and sampling insolvency rates for all clusters are calculated. institutions, central bank,
chambers of commerce)
Platform implementa- Implementation should consist in the following three phases: Platform developer Three months
tion and validation
1. Univariate analysis (indicators shortlist)
2. Multivariate analysis (model selection and weights)
3. Model validation and back testing
Pilot and go-live Pilot and go-live should consist in the following two phases: Platform developer and Three months
platform owner
1. Pilot test on a sample of new companies for final test of platform
and overall notification and support process effectivity
2. Go-live. Information is automatically made available and notified
through the established channels.
3. Project Governance
The WBG team’s suggestion for project governance is that separate entities
should have distinct goals in order to ensure both steering capacity and opera-
tional focus. Therefore, a steering committee should perform project oversight
and management. The main members of the committee would be the platform
owner, the platform developer, and the primary information providers.
The working group should provide support to the different phases of the tool’s
implementation process, through vertical working groups aimed at supporting:
• Data sampling;
• Model validation; and
• Pilot and go-live.
Figure 16 describes a synthetic scheme of the team’s proposition for project
governance.
85
Estonia / Early Warning Tools
Technical Note / May 2020
Figure 16: Project Governance Arrangements
STEERING COMMITTEE
PLATFORM MANAGER ADVISOR
INFORMATION PLATFORM OTHER CENTRAL BANKS AND
PROVIDERS DEVELOPER INSTITUTIONS LENDERS’ REPRESENTATIVES
· Periodical working progress check · Periodical working progress check
· Quality checks and acceptance tests · Quality checks and acceptance tests
· Proposals for solutions of criticalities · Proposals for solutions of criticalities
OPERATIONAL TEAM
WORKING GROUP
1. SAMPLING 2. MODEL TRAINING 3. MODEL VALIDATION 4. PILOT AND GO-LIVE
86
Estonia / Early Warning Tools
Technical Note / May 2020