CMStatistics 2017: Start Registration
View Submission - CFE
A1622
Title: Improving the classification accuracy of the logit model: The case of US bank failures Authors:  Francesco Audrino - University of St Gallen (Switzerland)
Alexander Kostrov - University of St. Gallen (Switzerland) [presenting]
Juan-Pablo Ortega - University St. Gallen (Switzerland)
Abstract: A MIDAS-style weighting scheme is introduced for constructing flow predictors in financial studies. MIDAS is introduced in the context of the logit model and we then address the issue of classification accuracy evaluation in the presence of severe class imbalance in the data. A `risk group' approach is presented as a better alternative compared to the standard indicators of the classification accuracy, which are misleading for imbalanced data. Re-weighting of observations in the log-likelihood function is suggested to mitigate the class-imbalance problem, where the cross-validation selects an optimal weight for the rare class. In the empirical part of the study, we apply these innovations to improve a well-established logit model for predicting US bank failures in 2004-2016. The gain in forecasting accuracy is confirmed to be both statistically and economically significant. In our setting, MIDAS-style weights characterize the relationship between the probability of individual US bank failures and past values of an explanatory flow variable.