Moscow, December 11, 2018 — Business Systems Consult. Credit risk assessment can be performed on a collective or individual basis. One of the approaches to individual assessment is to apply a scoring card to assessment of borrowers. However, this operation is often mistaken.

A Scoring Card: What Is It

A scoring card is a relationship which links the borrower’s characteristics measured at present with the chances to fall in default in a future. One of the possible specifications is as follows:

here xk represent borrower’s characteristics (k = 1, ..., K), while coefficients βk explain how the values of these characteristics lead to default chances. The default processes are not deterministic, and a random variable ε (with zero average) models the indeterminacy in default of the specific borrower.

For a corporate borrower, its characteristics used in the formula above may be constructed basing on the last financial report. The examples include sales margin and other similar coefficients, ratios representing the balance sheet structure, turnovers. The direct usage of these characteristics is not a good practice. The better way to identify a scoring card is to bin the values of characteristics in intervals and to replace the values by the weights of evidence of the corresponding intervals. These procedures lie beyond the subject of this article; some specialised books are recommended.

A special focus should be made on β0 coefficient. In reality, the target event for the identification of a scoring card may not be a default event but something else. The essence of this event depends on the data availability. The scoring card is rarely built on the data of the specific bank or leasing company, such entity may not have enough data. But the external data sources may not contain default data (where default is defined as an event of more than 90 days of past due). The target event may be an event of bankruptcy or license breakdown.

The other intuition behind the target event change is the rarity of defaults. They can be treated as a sugar in a cup of tea. To investigate the reasons for defaults, sometimes additional defaults from the other datasets are added (as if additional sugar to the tea to the concentration when sugar can not be dissolved). But after the research, the natural frequency of defaults should be restored. In this example, the replacement event is a default in the other dataset.

The main requirement is the relationship between the initial target event and the event used in scoring card calibration. These events should be correlated, the variables describing them should be instrumental to each other.

It might seem that the replacement of the target event is a sign of impossibility to build a correct scoring model. Luckily, it is not truth. It is proved as a theorem, to get back to the initial target event the only β0 coefficient should be recalculated. Continuing the analogy to a cup of tea, upon research of the solid sugar, it should be dissolved back to the concentration in a cup of tea.

A Typical Error in Scoring Card Calibration in IFRS 9

At this point in IFRS 9 implementation, a temptation emerges. The result of the fall is the estimators for probabilities of default that completely do not satisfy the requirements of IFRS 9. Moreover, these estimators are the rape of reason. The temptation is to calibrate the scoring card to default rate observed in the entity.

What is default rate in a credit portfolio? It is a frequency of defaults during the following year subject to the condition that a borrower is not in past due at the start of the period. Therefore, some methodologists think, if default rates for 2012, 2013, 2014, 2015, 2016, 2017 are calculated, they may be linked to some macroeconomic parameters (like GDP growth, price inflation, unemployment et al), and then the forecast of a default rate for the following year can be calculated subject to an widely accepted but actually random macroeconomic scenario (for example, some government scenario). In this way, these methodologists believe, they obtain a point-in-time scoring card.

Another article discusses the issue whether IFRS 9 actually requires to apply point-in-time scoring card, as well as how the random (although well discussed and widely accepted) scenario breaks IFRS 9. The remainder of the article will focus on default rate forecasting for IFRS 9.

First of all, the default rate time series is not stationary. For example, if the observation time period starts from January 01, and if (in another calculation) the start of observation period is on April 01, the default rates are significantly different (see figure). Further discussion of this method may be stopped here.

Second, IFRS 9 requires that all credit risk models have to satisfy all the conditions formulated in the Standard. The model of the borrowers’ defaults is not an exemption. Moreover, common sense (or experience of credit portfolio analysis) shows that these requirements are extremely well-based, economically justified and correspond to the real processes of credit losses maturation. The realisation of credit losses, for example, default rate, is a result of the combination of various independent factors described in the Standard:

•     various deals which resulted in default rate have various initial terms (cf. pp. B5.5.10, B5.5.5(e))

•     and various remaining term to maturity (B5.5.17 (a, e, p));

•     credit quality of the specific deals (B5.5.4, B5.5.5(b), 5.5.9);

•     efficiency of collection procedures (5.5.17 (b), B5.5.29).

Even if default rate measured this way has some correlation with macroeconomic characteristics, this will be a spurious correlation and it will be an example of the first-rate blunder.

Third, building a relationship between default rate and macroeconomics basing on 6 points lies well beyond minimal requirements to quality of econometric research. Due to the lack of data, the IFRS 9 requirements of reasonability and supportability are not satisfied. Back testing out-of-the-sample is impossible.

The Cost of Such Errors

Is it possible to satisfy auditors with such an error? Unfortunately, yes, although the adjustment procedures for the models will vary from year to year.

Does this error carry any problems for an entity? Yes, it does, because beyond explicit IFRS 9 violations this error makes the objective estimation of credit risk level impossible. Moreover, if an auditor is not ready to live with the conflict of interest, this error breaks all the methodologies of flexible credit allowance management.