Empirical Results

A. Logistic and Stepwise Logistic Models

Tables 3A-3B [ PDF 212.8KB | 13 pages ] report the results of the first logistic regression, incorporating all available financial ratios in domestic bankruptcy prediction models for Japan and Indonesia and a cross-country model taking advantage of information from both. Tables 4A-4C [ PDF 212.8KB | 13 pages ] report the results of the second step, where we have used factor analysis to narrow the range of variables used in the prediction model considerably. In all tables, rather than the coefficient estimates, we report the log of the odds ratio, which is derived from the coefficient estimates⁶ and represents the increased odds of bankruptcy for each unit increase in the independent variable.

In the stepwise regressions, for both domestic prediction models, the behavior of loans, in particular the ratio of loans to deposits or loans to equity, are significant indicators of bankruptcy. In Indonesia the ratio of loans to total assets and nonperforming loans are also significant indicators of bankruptcy. This is perhaps not surprising, as troubled banks may increase lending in the face of financial difficulty as a way of bringing in revenue and this lending may in fact tend to go to riskier borrowers who can pay higher interest rates. For the domestic Japanese model, the fact that OTA (ratio of other securities to assets) and ROA (return on assets) enter positively is contrary to our expectations but neither odds ratio is significantly different from 0⁷. In the case of ROA, this may be signaling increasing risk, requiring higher return on assets.

The cross-country stepwise regression results, reported in Table 4C [ PDF 212.8KB | 13 pages ], also suggest that loan behavior is very significant in predicting bankruptcy. The loan to equity ratio and ratio of non-performing loans for Indonesian banks enter statistically significantly in the cross-country model. The ratio of securities to total assets and for Indonesian banks and equity to total assets for both Indonesian and Japanese banks (STA and ETA) also enter significantly positive, and the odds ratio for STA is particularly large. This is contrary to our expectations, but may reflect depositor flight prior to bankruptcy, which would reduce short-term liabilities, thereby increasing the ratio or equity to assets in the short-run, as longer-term assets were fixed. The same phenomenon may be reflected in the very large positive odds ratio on capital to deposit ratios for Japanese banks.

B. Goodness of Fit

Goodness of fit test for all three bank failure prediction models – Japan, Indonesia and the Cross-Country Model – display good fit with actual observed bankruptcy (Table 5 [ PDF 212.8KB | 13 pages ]). Using the Hosmer and Lemeshow (2000) test⁸, we cannot reject the null hypothesis that there is no difference between observed bankruptcy and bankruptcy predicted by our model.

C. Predictive Power

Table 6 [ PDF 212.8KB | 13 pages ] reports the predictive power of all three models. All three models correctly classified over 90% of the outcomes. Negative predictive power, the probability of a bank surviving given that our model had classified it as a survivor, was highest for the model using only Japanese data, at 99.47%, but negative predictive power was fairly high in all cases, 97% for the crosscountry model and 93% for the Indonesian model. Perhaps more important to regulators is positive predictive power, the probability of a bank actually going bankrupt given that the model classifies it as such. Positive predictive power was highest for the cross-country model at 32.41%, and significantly lower for the two domestic country models at around 22% for Japan and 26% for Indonesia.

Type I and type II error are also reported in Table 6 [ PDF 212.8KB | 13 pages ]. Type I error, the percentage of surviving banks that were incorrectly predicted to fail by the model, was below 5% in all cases, and lowest at 0.20% for the domestic Japan model. Perhaps more significant for regulators, type II error, the percentage of failed banks that were incorrectly predicted to survive by the prediction model, was substantially higher even for the cross-country model, which at 51% displayed the lowest type II error.

D. Sensitivity and Specificity

Table 6 [ PDF 212.8KB | 13 pages ] and graphs 1a to 1c [ PDF 212.8KB | 13 pages ] display the sensitivity and specificity of the prediction model. Specificity, the fraction of observed survivals that are correctly classified by the model, was fairly high – over 95% - for all the models, and highest for Japan at 99%, followed by the Indonesian model at 97%. Sensitivity, the fraction of observed bankruptcies that are correctly classified, was significantly lower for all three models and for the two domestic models was even below 10% (9.5% for Japan and 8.7% for Indonesia. Of the three, the cross-country model performed best on sensitivity, with 49% of observed bankruptcies correctly classified.

Graphs 1A [ PDF 212.8KB | 13 pages ], 1B [ PDF 212.8KB | 13 pages ], and 1C [ PDF 212.8KB | 13 pages ] display graphically the ROC curve, the trade-off between sensitivity and 1-specificity as the cutoff point is varied between 0 and 1. A model with no predictive power would display a straight 45 degree line (50% of the graph beneath the curve) and in general the more bowed the line is and the larger the area beneath the curve, the better the performance of the model. Our cross-country model displays the best performance, with 92% of the area under the ROC curve. This is higher than the single country curves for either Japan (88%) or Indonesia (78%)

Download this Discussion Paper [ PDF 286.6KB| 22 pages ].

Comment(s)

There are [0] comment(s) for this entry. Post a comment.