Modeling the behavior of retail borrowers of the Bank

On February 14 at the research seminar “Empirical Researches of Bank Activity” School of finance were presented the results of the study Kosarev V.R. postgraduate, School of finance, HSE faculty of Economics: Modeling the behavior of retail borrowers of the Bank

Postgraduate of School of Finance Kosarev Vladislav answered the questions about the content of the study

What input data do you use to predict the borrower's payment behavior?

To predict payment behavior, you need fairly detailed data about the loan, namely the number of payments the client makes in repayment of his loan.

What is the percentage of reliability (accuracy) of prediction of the borrower's payment behavior?

Now the accuracy of prediction is quite modest-45%, such results do not suit anyone, but due to the use of more suitable modelling techniques (finite mixture regression models), they can be improved. In my opinion, the error should be reduced to 20 percent or less.

You offer several ways to apply a prediction of the borrower's payment behavior, and in what direction in your opinion there are the greatest prospects of development? And why?

The first method I told you about - this is an "on the forehead" simulation of the number of payments. There are obvious advantages in it - knowing the number of payments you can get anything you like, incl. accurately calculate the expected cash flow. The disadvantage here is that for the modeling of such a value, standard econometric techniques are not suitable because the distribution is multimodal.

The second is the modeling of the probability that a loan will be paid 0, 1, 2, 3, and so on. In it from the point of view of econometric modeling the situation is ambiguous. On the one hand, it is sufficient to use multinomial regression, but on the other hand it is not clear how much of the payment is limited - how many probabilities should 10 or 50 be modeled?