FEC-Real: Enhancing Financial Time Series Task with a Hybrid Encoder

The application of mathematical and computational techniques in financial investment has emerged as a prominent area of research, leading to the development of various tasks including factor mining, stock prediction, and analysis of financial statements. In this work, we particularly focus on the task of predicting the future trend for stocks. In existing fintech research, different transformer-based models have been explored for predicting future stock trend. This study is motivated by the need for a more efficient network architecture that can enhance the interpretation of real-time data. However, transformer based models are not always efficient for real world high speed trading data. To address this, we specifically explore the effectiveness of Kolmogorov Arnold Network (KAN) for financial time series model. We propose a KAN based encoder (FTS2K) which utilizes both KAN and transformer architecture to predict future stock price movements. The empirical results show that our proposed Encoder improves the average accuracy by 2.62%. Our approach consistently outperforms in four datasets (i.e. China A Daily, China A Min, China Futures Min, Dow 12 Daily), achieving superior results in both ACC and Top-100 ACC metrics. Additionally, we validate the versatility of the FTS2K encoder by applying it to a portfolio management task on a public dataset, where it achieves improved cumulative return across multiple reinforcement learning baselines.