Impact of Multistep Forecasting Strategies on Recurrent Neural Networks Performance for Short and Long Horizons

Forecasting is one of the most important tasks in temporal data mining. Actually, most of forecasting applications require performing multistep forecasts. Thus, what is the appropriate multistep forecasting strategy to use with each model? To answer this question, this study evaluates the impact of multistep forecasting strategies on the performance and the stability of recurrent neural networks models (SRN, LSTM and GRU) for short term and long term horizons. This comparison is based on well-defined benchmarking univariate time series (deterministic, stochastic and chaotic), whose properties are known and challenged their modeling. The results of this study proved that direct and hybrid strategies are computationally expensive than multiple outputs and recursive strategies. Further, we had come out with the following guidelines: For SRN and LSTM models, multiple outputs strategy is suggested for short term forecasting for all types of time series. While for long term forecasting, multiple outputs is suggested for stochastic time series, however, direct strategy for deterministic data and hybrid strategy for chaotic data. On the other hand, for GRU model, multiple outputs strategy is suggested for short and long horizons for all data types, except for the case of long term forecasts with chaotic data, where direct or hybrid are preferred. Thereupon, Switching-based forecasting system is suggested as a relevant alternative to build up accurate predictions for each model.