Multi-Task Transformer for Stock Market Trend Prediction

This paper presents a novel stock market prediction method by taking transformers’ advantages in analyzing the sequential data. The previous techniques usually tend to learn/understand the pattern of the market by analyzing the historical market data, while those patterns are very complex and implicit. To learn these patterns effectively, we cope with this challenge by leveraging deep neural models, i.e., transformers. We employ transformers to predict the stock trend. Since this kind of deep learning model needs a massive amount of data to be trained, the data paneling approach is hired to extend the dataset. Also, the multi-task technique is utilized to reduce the optimization searching space, which causing to speeding up the coverage and finding relatively optimal parameters and consequently improved accuracy. Note that the method of labeling the trend which is used in this paper is financially meaningful and more practical. Our proposed method has been evaluated on the real-world stock market, specifically the Iranian stock market, and the results confirm its effectiveness.