Hi,

that depends to a very large extent on what you are trying to achieve and what is your model type. In particular:

- are you trying to get a result which is a probability of a failure (default) within an year, or within lets say 3 years in advance? If you are trying to forecast N years ahead then it is unlikely to be able to do this reliably by separately studying each year.

- what is the type of your model - are you trying to forecast specific ratios or characteristics of the bank with time series, or are you just trying to find the causes (ex post fact) for the failures?

Of course there are other things that will drive your decision, but these two should be clear in advance.

Best regards,

 Stan

I feel analyzing bank failure each year separately will give useful information.

But it depend on the objective/s of the study.   First decide the objective/s of the study then automatically everything will be clear. 

Hi Ahmed,

Are you trying to develop a predictive model based on historical data or determine the likelihood of failure of a bank at a point of time? If you wish to determine likelihood of failure at a point of time, you may use Altman Z-Score that requires you to use financial ratios. But if your objective is to develop a predictive model, you may need to employ a binary logistic regression model where you have firm-level and industry-level data to identify factors that best predict failure. 

You should do an analysis of covariance to see whether pooling the data across years is warranted. I think you should also distinguish the probability of insolvency which is a precondition for failure from the decision to close the bank.This will help you to see the resource constraints that limit failures in bad economic times.

Thank you for these fruitful answers. 

I have another question related to the same topic

How can I test the prediction accuracy of three or four statistical models. For Instance, If i'm going to use the Logistic regression analysis, Multivariate Discriminant analysis and The Dynamic Hazard model. how can I test the prediction accuracy of each model 

for example, X model predicts 90% of failure while model Y predicts 85% and so on ....

If you need "n" years of data to use model X and Y, you can determine model accuracy if you have a string of data that comprises of any number of years that is higher than n (t>n).

X(n)=x, Y(n)=y, RV(n)=rv, Model X error = (x-rv)/rv, Model Y error=(y-rv)/rv

You can repeat this (t-n) times and determine a string of errors. Use an average error and done.

Hi Ahmed,

It is rare to see an answer from one of the leading authorities in an area. I had the pleasure of hearing Ed Kane decades ago and still recall being impressed with his work on failures of financial institutions.

With that said, I am not sure I can contribute much; however, the following ideas might add some things to consider.

It might make a difference if you are trying to forecast the timing of failure. Also, it may depend on your tolerance of Type I and Type II errors, which would seem to change when the focus is a single year compared with several years as a group.

I hope this helps.

Rob