Training LSTMS with circular-shift epochs for accurate event forecasting in imbalanced time series

Abstract

The use of LSTM models for forecasting has increased substantially because they have proven to yield excellent performance in numerous applications involving time series consisting of normal observations. However, LSTM models tend to perform poorly in forecasting extreme events in imbalanced time series due to the lack of representative extreme event data during training. This study introduces a breakthrough Circular-Shift Epoch Training (CSET) strategy for LSTMs that makes the accurate forecasting of extreme events no more difficult than predicting events in normal time series with similar accuracies. CSET is (a) specifically designed to preserve the natural ordering of observations in epochs during training without any attempt to balance extreme and normal observations, (b) formulated for both univariate and multivariate time series forecasting, (c) applicable to scaler and vector forecasting for the multivariate case and can handle upward, downward, and mixed extremes without any modification, and (d) universal because it can be applied to train LSTMs to forecast events in normal time series or in imbalanced time series in exactly the same manner. Most importantly, the forecasting of normal time series and imbalanced time series do not have to be treated as separate problems. The results from an extensive set of experiments involving forecasting events in simulated and real imbalanced time series confirm that LSTM models trained using CSET dramatically outperform the standard fixed-epoch trained baseline models in (a) forecasting both extreme and normal events and (b) classifying extreme events. Given the ubiquitous use of LSTMs for forecasting, the accurate forecasting of extreme events through the incorporation of the innovative CSET strategy can have a major impact in decreasing the disastrous consequences of dangerous extreme events.