Query:Probit vs Logit in Credit Models like RiskCalc

deepnarayan · August 17th, 2005, 3:31 pm

Risk Calc(by Moody's ) uses financial statement information and builds a probit model. The output of the probit model is then mapped to actual default probabilities as observed in the portfolio.Would the output of this whole exercise be very significantly different if instead of Probit model a Logistic regression(Logit) model was run?Would it be advisable to map the output of this logistic regression to actual default rates(for respective baskets that may be created)?

KL · August 18th, 2005, 10:22 am

I think RIskCalc uses a non-linear logit - which is really many univariate-like logit regressions and then combined. the probit is a normal dist and logit is a logistic dist. the logistic dist has slighltly fatter tails than the normal distribution.As for the output - there is no practical difference in the coefficients, Tstats may be slightly different. But really in practice it doesn;t ,matter.Logits are used because they are more conevnient to compute although with modern day computers they matter not.What do you mean by mapping to buckets instead of portfolio ??