A dual-stage advanced deep learning algorithm for long-term and long-sequence prediction for multivariate financial time series

Long-term and long-sequence prediction for multivariate time series (MTS) data is an important and challenging problem. The key to solving this problem is learning the long-term dependency period in a time series. Different attention-based models like RNN variants and transformers are actively researched to provide short-term and long-term predictions for a time series. These attention -based models consider time series as sequential data. They are not designed to learn the different characteristics (trends, seasonality, and irregularity) of a time series. The complex temporal patterns of data prohibit these models from finding a reliable period of dependency. This paper proposes a dual-stage advanced deep learning framework (DST2V-TRANSFORMER), which exploits moving average smoothing, the percentage change in data, and time embedding to improve transformers for time series forecasting. In the first stage, data is smoothened using the moving average method, and to learn the periodic and non-periodic temporal patterns in time series, a vector representation for time is added as a feature to the data. This preprocessed data is passed through a transformer to learn temporal correlation and find the reliable period of dependency. The DST2V-TRANSFORMER achieves state-of-the-art efficiency and accuracy, with at least a 49% relative improvement in error on seven MTS datasets from the financial domain.(c) 2022 Elsevier B.V. All rights reserved.