N-BEATS-RNN: Deep learning for time series forecasting

Abstract

This work presents N-BEATS-RNN, an extended version of an existing ensemble of deep learning networks for time series forecasting, N-BEATS. We apply a state-of-the-art Neural Architecture Search, based on a fast and efficient weight-sharing search, to solve for an ideal Recurrent Neural Network architecture to be added to N-BEATS. We evaluated the proposed N-BEATS-RNN architecture in the widely-known M4 competition dataset, which contains 100,000 time series from a variety of sources. N-BEATS-RNN achieves comparable results to N-BEATS and the M4 competition winner while employing solely 108 models, as compared to the original 2,160 models employed by N-BEATS, when composing its final ensemble of forecasts. Thus, N-BEATS-RNN's biggest contribution is in its training time reduction, which is in the order of 9x compared with the original ensembles in N-BEATS.