Iter-Transformer: Transformer-based Multi-step Multivariate Time Series Forecasting Using Iterative Strategy

In the financial sector, stock price forecasting has always been a critical and challenging task. Accurate stock price prediction can support financial institutions in risk management and asset allocation, and help investors make more informed decisions. Traditional methods have limitations in dealing with nonlinear relationships and complex multidimensional data, and due to Transformer's powerful capabilities, it has been gradually explored and successfully applied to solve this task. However, in multi-step prediction tasks, Transformer directly generates predictions for multiple time steps through the self-attention mechanism and global dependency, which often leads to a significant reduction in prediction accuracy. In this paper, an iterative prediction strategy is proposed. It utilizes the iterative idea to decompose multi-step prediction into recursive single-step prediction tasks. We conducted comparative experiments on several stock market indices around the world. The results show that Iter-Transformer's prediction results generally outperforms the traditional Transformer, and also improves to varying degrees compared to other models such as Transformer variants.