Bankruptcy prediction : a comparative study on logit and neural networks.

Abstract

This study extends the research of earlier studies on bankruptcy prediction. Logit models are built using different independent variables to predict the probability of bankruptcy in the oil and gas industry. Neural network models are then built to predict bankruptcies using the same independent variables used in the logit models. Each model is then ranked on the rate of prediction error on an outside sample using a cross-validation method.

The study will encompass six sets of data and two estimating methods. The six sets of data are from the oil and gas industry, and these sets are: (1) Accrual based ratios. (2) Accrual based ratios adjusted by interest rates and oil prices (real accrual ratios). (3) Cash flow based ratios. (4) Cash flow based ratios adjusted by interest rates and oil prices (real cash flow ratios). (5) Real accrual ratios with economic variables. (6) Real cash flow based ratios with economic variables. The economic variables are the interest rate and oil price. The two estimating methods are: (1) Logit model. (2) Neural networks.

The main finding of this study is that all of the models that were estimated with neural networks outperformed all of the models that were estimated with logit. This finding agrees with the findings of other studies that compared the two methods. The ability of neural networks to generalize and their freedom from the data characteristic and estimation assumptions that must be present for other estimation techniques to perform well, are the main reasons why neural networks outperformed logit models.