||In this paper, we forecast stock market crashes by innovative EEMD-based neural network approach with stock market indices. The concept of EMD is to decompose nonlinear and nonstationary time series signal for obtaining instantaneous frequency data. In order to predict stock crashes for early-warning proposes, this study uses Ensemble Empirical Mode Decomposition to decompose historical prices into several frequency component and a residual trend, and then take these components as inputs in neural network. |
The result of our research shows that the prediction model is significant by better forecasting performances. In seven sample countries, if we take a look on the forecasting result of the out of sample that including crisis periods and non-crisis periods, the lowest accuracy is 70.8% for Dow Jones Average Industrial and the highest is 99.1% of Australia Index. If we only take a look of the forecasting result of the crisis period alone, the lowest accuracy is 33% for Hang Seng Index, and the highest is 100% for Singapore STI Index, UK FTSE All-Share Index, and Australia ASX Index. Compare the forecasting performance with a logit model, the forecasting accuracy of the logit model in crisis period is much lower than the EEMD-based neural network approach. Finally, we explain the factors explaining the various price components, and found a significant relationship between low frequency price component and microeconomic variables, while mid-frequency components of IMFs can be explained by investor sentiment.