Regime signaling techniques for non-stationary time series forecasting

An accuracy based signaling technique is proposed as an alternative to a statistics based signaling for detecting changes in a time series distribution. Three different forecasting scenarios are analyzed in order to decide whether to reuse historically successful neural network models or retrain new ones when a change in the distribution is signaled. The results obtained on low noise and high noise, non-stationary time series provide strong evidence in favor of the accuracy based signaling technique.