I have a data set contains 5400 healthy banks and 470 failed banks during 2009 to 2014.
First, what is the major criteria to divide the data to training set and validation set. Is there a specific percentage? or an equation?
Second, I was thinking to use the T test to determine the most important explanatory variables that differentiate between failed and healthy banks and then apply LRA, MDA, and the Dynamic Hazard model using all variables and the significant variables after T test.
Finally, after dividing data to training set, validation set, the whole variables and the significant variables after T test. I will have four different data sets.
Which results should I adopt?
Training set (to construct your model) is recommended to be 80% of the data set, whereas validation set (to test your model) is 20%.
Hi Ahmed
Typically the purpose of the univariate t-tests on the individual variables is to cut down to a short list (which will still be lengthy) of variables to consider for your forecast modelling. So long as the model development sample and holdout sample have equal proportions of surviving versus failed banks (5400/5870) and are randomly chosen for each of the 2 samples you should be okay. Alma's suggestion to use an 80:20 split is reasonable. The US Federal Reserve System CAMELS approach uses logit. Nevertheless, you could conduct robustness tests as well using the other techniques such as MDA, Cox Proportional Hazard Model, probit, et cetera. Also, you may want to consider dummy year variables or split up your time period as you are mixing together years that are considered part of the financial crisis (2009-2010) with post-financial crisis (2011-2014).
Raymond A. K. Cox
Thompson Rivers University
Hi Ahmed
i agree with Alma on the 80/20.
Also you could add country dummies to control for unobserved factors. Or if you have diffrent types of banks you can dummy them out. Finally if you are not going to splite the dataset, then I would recommend you to conduct structural break test.
I hope this helps
Faisal