Early warning system for financial crisis: application of random forest

Abstract

The study identifies important variables in detecting the likely occurrence of a financial crisis 1 to 3 years from its onset . We do this by implementing random forest on Macroeconomic Historical time series data set for 16 developed countries from 1870-2016. By comparing the misclassification error for logistic regression to that obtained for random forest, we show that random forest outperforms logistic regression under the out-of-sample setting for long historical macroeconomic data set. Using the SMOTE technique, we show that minimising class imbalance in the data set improves the performance of random forest. The results show that important variables for detecting a financial crisis 1 to 3 years from its onset vary from country to country. Some similarities are however also observed. Credit and money price variables for instance emerge as very important predictors across a number of countries.