Abstract
This study empirically evaluates the predictive power of an econometric and the machine learning
method in forecasting financial crises in emerging and developed markets as early warning
systems. These methods were developed for both emerging and developed countries over a short
horizon of two quarters and a long horizon of four quarters for two different types of crises, which
are stock and currency crises. A different specification for each crisis variable is proposed to
evaluate how it enables prediction of new crisis onsets. The results indicated that the multinomial
logit model has a high forecasting power relative to the random forest in both regions over the long
term, particularly in emerging economies. Overall, crises detection in both models was relatively
more accurate in emerging economies, and this can be attributed to the existing high volatility in
the region due to its vulnerabilities such as weaker institutions, higher exposure to external shocks,
volatile capital flows, and higher levels of indebtedness. These factors make it relatively easier to
identify early warning signals in emerging economies.