0
$\begingroup$

I'm using neural network for a binary classification problem of bankruptcy prediction using patternnet function in MATLAB, so i have probability of bankruptcy for out sample (final report). I'm searching for a method to combine probability of bankruptcy and size ( for example total assets) of companies. I want have biggest companies with biggest bankruptcy probability so i need a hybrid method to combine these features. (the probability is probability of being in bankrupt class for a specific firm). How can i do that?

$\endgroup$
  • $\begingroup$ What is your target variable in your training data? Do you have a set of companies, for which you know, whether they went bankrupt or not, or you are trying to model expert judgement, that larger companies are more likely to go bankrupt? In the first case you would have discrete target variable, e.g. 0=non bankrupt, 1=bankrupt. In the second case, you would construct the target variable yourself, e.g. for company A the probability that it WILL go bankrupt is 0.7 (whereas it is not known from the data, whether it is bankrupt). Which case are you dealing with? $\endgroup$ – inzl Aug 28 '14 at 13:16
  • $\begingroup$ Thank you for your comment. My input variables are financial ratios and variables and my target variable is bankrupt (1), otherwise (0). After training neural network (or SVM model) and inserting out sample data (new year data), now I can have probabilities of bankruptcy besides binary output ( 0 and 1). Now i want a hybrid method to sort data. I don't want only sorting based on size of firm or probability. Suppose that you want help bankrupt companies. You want know relatively biggest companies with higher probability. Someone recommended me to use PCA for this but i don't know how. $\endgroup$ – user2991243 Aug 28 '14 at 13:21
1
$\begingroup$

It seems that you want to model something like expected loss, the probability of loss times the loss if bankrupt.

I would not modify the bankruptcy probabilities output by the model. I would try to quantify loss if bankrupt, e.g. number of employees losing jobs, or the amount of bank loan, then multiply if by the probability output by the predictive model, and then rank from the highest expected loss to the lowest.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ That's a good idea but if i want rank data based on size, How can i do that? $\endgroup$ – user2991243 Aug 28 '14 at 13:36
  • $\begingroup$ What do you mean "rank data based on size"? $\endgroup$ – inzl Aug 28 '14 at 15:58
  • $\begingroup$ As I mentioned in main question, if i want bigger companies based on log(total assets) besides higher probability of bankruptcy, what should i do? $\endgroup$ – user2991243 Aug 28 '14 at 16:00
  • $\begingroup$ I may be still misunderstanding the problem, but if you want to rank companies based on size independently of their bankruptcy probability, why can't you just rank them based on log(total assets). $\endgroup$ – inzl Aug 28 '14 at 18:52

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.