Expected Credit Losses: Future Looking AI Model – Simplified
Expected Credit Losses is a very important area of risk management. It is even more important if you are working for a financial institution. Financial Institutions like banks work with large sums of credit and it is not always “healthy”. Although every application is thoroughly reviewed before acceptance, a large number of customers still tend to default.
These institutions must therefore have additional measures to predict and avoid such losses. With the introduction of IFRS 9: Financial Instruments by International Accounting Standards Board (IASB) and now required by SECP to be implemented, Pakistani banks are in process of developing ECL models.
However we observed that some institutions out there are designing models based on past data on product/pool level and applying the percentage of default to the future product/pool. They may fulfill the regulator’s requirement, but can “5% of car loans in rural areas default” identify exactly or nearly who are those that are going to default so that something could be done? Guess not.
Therefore, IFRS/SECP or not, banks must keep themselves secure and instead of implementing such measure out of fear of regulators, they must do it for the right reasons i.e. increasing profits and looking for shareholders’ interests.
There are many approaches to calculate expected credit losses but here we are going to suggest an artificial intelligence (AI) based model. AI has been radically improved/matured over time and now there are tools available for students and professionals to experiment with AI like tensorflow.
As most of the professional bankers are not very familiar with computer technology, the objective of this article is to make this so simple that anyone could understand how it would work.
Our proposed model is based on two types of data:
- Past : Financials, Credit history etc
- Future : Forecasts such as economic growth, weather (i.e. floods), real estate etc
So it is a futuristic approach of identifying credit risk. AI models are very flexible. For example an AI model can directly tell you the amount of allowance required to be provided for, without even calculating the actual figure of expected losses or probability of default etc.
Suppose you are a bank with 250 existing customers and you already know how much provision you had required for each of them. On the basis of that data you want to create provision against loans of your fifty (50) new customers. Or you have the data of 300 customers but you want to test the model based on first 250 customers’ data.
The first step would be to normalize the data. That is, to convert the data in a form that the maximum value in a column is 1 and minimum is 0.
Here we used Tensorflow. We created a single hidden layer, relu activated model which is the most basic and simple ANN structure out there. There are just eight (8) inputshapes and one (1) unit. Iterations/passes are left to the users as the model is live and can be started and stopped in browser at anytime. Unless browser is refreshed, it always starts from where it is stopped.
The objective of Artificial Neural Network (ANN) is to train the model on data provided. In this case the first 250 cases. Data of both inputs and outputs has been provided to the model and when training is started, model starts learning from data. There are two lines in line charts. One is the actual data and the other is trained data. When trained data and actual data are near identical, your model is trained well enough to predict the data for which it does not know the outputs of, in this case the last 50 cases.
You can visit the following page for live model : http://watchdog.saber.pk/ai/ecl/
The result of our live run is pretty satisfactory.
I hope you enjoyed the research. In case of any question, please do contact.Tags: AI, ANN, Artificial Intelligence, Artificial Neural Networks, Credit Risk, ECL, Expected Credit Loss, IFRS 9, Risk Management