Regime signaling techniques for non-stationary time series forecasting

Abstract

An accuracy-based signaling technique as an alternative to a statistics-based signaling fordetecting changes in a time series distribution is proposed. Three different forecasting scenariosare considered for deciding whether to reuse historically successful neural network modelsor retrain new ones when a change in the distribution is signaled. The validity of the proposedtechnique is evaluated in the context of either low-noise or high-noise, non-stationarytime series obtained as a mixture ...