Predicting corporate failure is an increasingly important topic in the world of economics today. This paper, with the help of the credit ranking company Syna AB, aims to investigate a few different statistic strategies to do just that. This task is accomplished by using a data material of 250 000 Swedish companies divided into two subsamples. The first sample is used to develop the model and the second as a validation sample.
The model developing sample holds 160 000 companies divided into five different subgroups based on size and age. Binary logistic regression and probit regression were chosen for the analysis, with insolvency or not as the dependent variable. To find out which variables to include in the analysis different kinds of univariate tests are used. After the initial screening, a test of multicollinearity is also performed. Finally stepwise logistic regression is applied to lower the number of variables further.
In describing the different characteristics of each company a total of 78 variables are used as possible predictors in the model making process. Models based on logistic regression, probit regression and stepwise logistic regression alone are assessed by looking at a number of different measures. The models are compared by size, goodness of fit, significance of the parameter estimates and discriminating ability. Of the original 78 predictors about 10-25 remains in the final models created for each subgroup.
The result from the analysis implies that the most useful models are achieved through two of the three methods adopted. Stepwise logistic regression alone produces the smallest models with significant parameter estimates and high discrimination ability. Probit regression, on the other hand, manages to create models with the highest discrimination ability with the drawback of containing some insignificant parameter estimates and more variables than the previous. Unfortunately the Hosmer & Lemeshow??s goodness of fit-test indicates that model fit is uncertain or poor for all models. But since other measures are satisfactory the regression models should be useful for further analysis. Because the main purpose with this paper is to make models with high predictive ability the probit regression models are chosen as the final ones.
The validation displays that in overall the final models are valid when tested...
