Robust Multivariate Time Series Forecasting with Deep Reconstruction

An intriguing observation regarding the development of deep learning-based multivariate time series (MTS) forecasting models over the past few years is that improved forecasting accuracy does not necessarily correlate with an increase in the number of model parameters. This work reveals that imposing irrational inductive biases on intraseries and interseries relations is one of the dominant factors that is responsible for the failure of heavy forecasting networks. Thus, heavy forecasting networks suffer from more aggravated overfitting problems than lightweight forecasting networks do, leading to their worse performance. In contrast, MFDR, a novel MTS forecasting network with deep reconstruction, is proposed in this work. MFDR reconstructs and forecasts MTSs in parallel to align the distributions of prediction sequences with those of previous observations. Moreover, MFDR adopts wavelets to hierarchically and completely extract intraseries relations. In addition, a novel Cucconi attention mechanism is proposed herein to extract interseries relations; thus, the problem of misalignment among different series can be alleviated. Therefore, MFDR can achieve superb and robust MTS forecasting performance. Extensive experiments conducted with six baselines and eight benchmarks demonstrate the state-of-the-art performance attained by MFDR under various settings and circumstances.