Prediction of variable forecast horizons with artificial neural networks by embedding the temporal resolution warping

Abstract

Artificial neural networks (ANNs) have achieved many successes in time series forecasting. The shortcomings of them are a fixed forecast horizon and an increasing inaccuracy for multi-step forecast techniques to extend the forecast horizon. We embed temporal resolution warping into an ANN to provide a dynamic forecast horizon, excluding multi-step forecasts. The ANN is improved to recognize different representations of patterns by mapping spacial frequencies to new frequencies according to their relevance in time. We demonstrate the drastically improvement in forecast accuracy on different datasets. In comparison to the multi-step approach, we achieve a constant accuracy for extending the forecast horizon.