The most significant impact on overall bank management is caused by the introduction of Expected Credit Loss approach to reflect credit risk in external accounting.
IFRS 9 requires the segmentation of financial assets based on similar credit risk characteristics. For each segment the expected credit loss needs to be calculated taking probability weighted macroeconomic scenarios into account. The calculation needs to consider the deterioration of credit quality since deal conclusion date. For this purpose, IFRS 9 differentiates in total 3 stages, taking 12-month or lifetime expected credit losses into account.
This approach shall ensure that at any time for all individual deals which belong to a specific segment a risk provision is available which covers the embedded credit risk.
The extend to which this risk provision matches with the required risk provision depends on several factors:
- Quality of macroeconomic parameters - Calibration
- Segmentation criteria which are used to identify deals with similar credit risk. Usually limited to few criteria because of limited processing capabilities. This results in a high risk due to distortions in the ECL calculation, since the credit risks of unequal risk objects are offset against each other and no allocation according to their origin is made. Some transactions are allocated too high credit risks, others too little.
- Mathematical approach
Too much risk provision will reduce the income in the income statement. This will reduce the capital in the balance sheet and finally the covering funds in legal reporting. Less covering funds finally limit the entity in doing new business which will have impact on future profit and loss.
Precision of prediction
The precision of Machine Learning-based predictions is done by comparison with other (classical) approaches such as Logit (Logistic Regression) by backtesting of predictions based on the models.
An example of this is the comparison for prepayment estimates (which has a direct impact on the EAD that is part of the ECL):
Comparison of out-of-sample pool-level predictions of the 5-layer neural network and the logistic regression model
A pool of 2 million mortgages is grouped into 2,000 portfolios by ordering loans according to the borrowers’ FICO score and then sequentially packaging every 1,000 loans into individual portfolios.
For each such portfolio, the figure shows the observed number of prepayments in the next 12 months on the x-axis and the predicted number of prepayments in the next 12 months from the two models, the 5-layer neural network and the logistic regression model, on the y-axis. The x = y line (in black) shows the ideal but hypothetical scenario under which the predicted and the observed number of prepayments coincide. It is seen that the predictions from the 5-layer neural network are much closer to this ideal line than those from the logistic regression model.
The use of Machine Learning helps to improve the quality of ECL parameters such as PD, LGD, EAD and macroeconomic parameters. Doing so the ECL on individual deal level will reflect best the future deterioration of credit quality.