An Empirical Comparison of Early Warning Bank Failure Models Using McNemar's Test
32 Pages Posted: 30 Jan 2019
Date Written: December 15, 2018
In a field of study as robust as bank failure, many researchers investigate the same question: What factors increase the likelihood of bank failure? Models built to answer this question are highly accurate and span decades of research. What has yet to be answered is whether the models are different. This paper uses McNemar’s Test on out of sample predictions to show with statistical significance that discordant errors (where one model is correct and the other is wrong) provide the answer. This paper uses Cole and White (2012) as a baseline model to which two other papers (Martin (1977) and DeYoung and Torna (2013)) are compared. We find that the Martin model is highly accurate on its own, but in a different way than Cole and White. Additionally, we find that the DeYoung and Torna model is statistically different from Cole and White, but an interesting trade-off in accuracy occurs. Finally, we compare a hybrid model which relies on variables of interest from all three papers and find that the new approach is dissimilar and more accurate. The testing methodology is useful for all predictive models which benefit from a vast source of prior research and serves as a tool for researchers to determine if their new predictive model is different in a meaningful way, which may help future researchers decide how best to contribute to their field of research.
Keywords: bank, bank failure, CAMELS, FDIC, McNemar’s Test, early warning system
JEL Classification: G17, G21, G28
Suggested Citation: Suggested Citation