Prediction of Early Warning Crises by a Hidden Markov Model with Covariates

We propose an early warning system for financial crisis prediction, which is tailored to longitudinal data with missing values and time-varying covariates. The proposed system is based on a hidden Markov model that includes selected time-varying economic drivers and the lagged response variable, thus relaxing the local independence assumption. Partially missing outcomes at a given time are considered under the missing-at-random assumption, and partially missing values on the covariates are accounted for by dummy indicators. We study in-sample and out-of-sample model performance in terms of forecasting, considering an application related to country-level financial crises.