Title: FinTSB: A Comprehensive and Practical Benchmark for Financial Time Series Forecasting

Abstract: Financial time series (FinTS) document the decision-making processes influenced by human cognition, preserving historical data that can be utilized to develop profitable investment strategies. Consequently, this field has garnered significant interest from the research community, leading to the proposal of numerous methods built upon diverse architectural backbones. Despite this progress, current evaluations in the domain are hampered by three systemic shortcomings: first, a failure to encompass the full range of stock movement patterns present in dynamic markets, creating a diversity gap; second, a lack of standardized assessment protocols that compromises the reliability of performance comparisons across different studies, resulting in a standardization deficit; and third, the oversight of essential market structural elements, which yields inflated performance metrics that lack real-world applicability, known as the real-world mismatch. To address these challenges, we introduce FinTSB, a robust and practical benchmark designed for financial time series forecasting (FinTSF). To enhance data variety, we classify movement patterns into four distinct categories, apply tokenization and preprocessing techniques, and evaluate data quality using specific sequence characteristics. To mitigate biases arising from disparate evaluation settings, we establish standardized metrics across three dimensions and develop a user-friendly, lightweight pipeline that integrates methods from various backbones. Furthermore, to faithfully replicate real-world trading environments and support practical deployment, we rigorously incorporate various regulatory constraints, such as transaction fees. Finally, we perform extensive experiments using FinTSB to derive key insights that assist in selecting appropriate models under different market conditions. In summary, FinTSB offers researchers a new and holistic platform for advancing and assessing FinTSF methodologies. The code is available at https://github.com/TongjiFinLab/FinTSB.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC