FinTSBridge: A New Evaluation Suite for Real-world Financial Prediction with Advanced Time Series Models

Yanlong Wang, Jian Xu, TianTian Gao, Hongkang Zhang, Shao-Lun Huang, Danny Dongning Sun, Xiao-Ping Zhang · 2025

Read FinTSBridge: A New Evaluation Suite for Real-world Financial Prediction with Advanced Time Series Models by Yanlong Wang, Jian Xu, TianTian Gao, Hongkang Zhang, Shao-Lun Huang, Danny Dongning Sun, Xiao-Ping Zhang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-06928.txt.

Introduction

Yanlong Wang and co-authors build FinTSBridge, a test suite that grades forecasting models on real finance tasks, in a paper from 2025. Many new time series models shine on generic data yet go untested on markets. Prices follow their own rules, and plain error scores miss whether a model tracks market moves. The suite was built to bridge that gap.

It holds three data sets drawn from finance. More than ten recent models were run on them. New scores were added to catch how well a model's path matches the market's path over time. Tasks mirror real jobs in asset pricing and risk work. The results show which models hold up when the data turn jagged and driven by news. The authors hope the suite steers model design toward real use.

Catalog record

CREATORYanlong Wang, Jian Xu, TianTian Gao, Hongkang Zhang, Shao-Lun Huang, Danny Dongning Sun, Xiao-Ping Zhang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE86 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-06928

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

FinTSBridge: A New Evaluation Suite for Real-world Financial Prediction with Advanced Time Series Models Yanlong Wang, Jian Xu, TianTian Gao, Hongkang Zhang, Shao-Lun Huang, Danny Dongning Sun, Xiao-Ping Zhang Abstract Despite the growing attention to time series forecasting in recent years, many studies have proposed various solutions to address the challenges encountered in time series prediction, aiming to improve forecasting performance. However, effectively applying these time series forecasting models to the field of financial asset pricing remains a challenging issue. There is still a need for a bridge to connect cutting-edge time series forecasting models with financial asset pricing. To bridge this gap, we have undertaken the following efforts: 1) We constructed three datasets from the financial domain; 2) We selected over ten time series forecasting models from recent studies and validated their performance in financial time series; 3) We developed new metrics, msIC and msIR, in addition to MSE and MAE, to showcase the time series correlation captured by the models; 4) We designed financial-specific tasks for these three datasets and assessed the practical performance and application potential of these forecasting models in important financial problems. We hope the developed new evaluation suite, FinTSBridge, can provide valuable insights into the effectiveness and robustness of advanced forecasting models in finanical domains. Code can be found in https://anonymous.4open.science/r/FinTSBridge-0D6C . 1 Introduction Time series forecasting has become increasingly crucial in the financial domain, playing a critical role in decision-making processes related to asset pricing, risk management, and algorithmic trading ( Tang et al. 2022 ; Lim & Zohren 2021 ; Masini et al. 2023 ) . Accurate forecasts are essential for tasks such as stock index prediction, option pricing, and modeling cryptocurrency volatility, as they help optimize investment strategies and mitigate market risks ( Li et al. 2022 ; Doshi et al. 2024 ) . However, the non-stationary nature of financial markets-shaped by factors like geopolitical events and investor sentiment-adds significant complexity to these challenges ( Zheng et al. 2023 ) . Despite advancements in machine learning for time series analysis, transforming state-of-the-art models into actionable financial insights remains an ongoing obstacle ( De Prado 2018 ) . For example, while models may achieve low mean squared error (MSE) on synthetic datasets, their performance often deteriorates in real-world conditions due to unaccounted market dynamics, such as price-volume interactions and multi-scale volatility. Moreover, many studies focus on time series datasets with simplified statistical properties, such as electricity consumption or traffic flow, which typically exhibit strong stationarity and periodicity ( Chen et al. 2024 ) . Even commonly used financial datasets, such as historical exchange data, suffer from key limitations: daily resolution obscures intraday fluctuations, and limited variable coverage (e.g., the absence of derivatives metrics) fails to capture the full spectrum of multi-scale market interactions ( Bao et al. 2024 ) . Compounding these issues, traditional evaluation metrics like MSE and mean absolute error (MAE) prioritize point-wise accuracy while neglecting temporal correlations, a critical aspect in financial forecasting. For instance, our empirical results show that a naive model that predicts the last observed value can achieve competitive MSE but exhibits near-zero correlation with true market trends, making it ineffective for practical trading strategies. These shortcomings underscore the need for financial-specific benchmarks that reflect the complexity of real-world markets, along with evaluation frameworks that assess both predictive accuracy and economic utility. Figure 1 : The overall pipeline of FinTSBridge. First, three financial datasets are constructed and corresponding preprocessing is carried out. Then, a wide range of time series models are utilized for training. Through extensive task settings and comprehensive evaluations, the performance of the models is verified. Finally, investment strategies are formulated in real-world financial scenarios and the performance of these strategies is evaluated. To address these gaps, we introduce FinTSBridge, a comprehensive framework that connects state-of-the-art time series models with real-world financial time series and applications. As depicted in Figure 1 , our process begins with the pre-processing of financial data. We have developed a specialized method tailored to financial time series that enhances stationarity while preserving the interrelationships between variables, facilitating more effective model training. The data is then fed into the modeling phase, where we apply various AI-driven time series models to tackle diverse forecasting tasks. We design multi-perspective forecasting tasks and evaluate model performance using both traditional and novel metrics, complemented by financial metrics for assessing strategy performance. Depending on practical requirements, we assess models not only for predictive accuracy but also for their potential to support investment strategies through strategy simulation. This process aims to provide actionable decision support while deepening our understanding of financial data and processing techniques. By introducing new datasets, evaluation metrics, and forecasting tasks, FinTSBridge seeks to bridge the gap between advanced time series models and the complex challenges encountered in financial markets. Our work emphasizes the importance of a correlation-aware approach to financial time series forecasting, ultimately advancing the field of AI-driven finance. Our key contributions can be summarized as follows: • Three curated financial datasets are constructed to reflect diverse real-world market dynamics: – Global Stock Market Indices (GSMI): 20 indices (2005–2024), capturing cross-market dependencies and volatility regimes. – High-Frequency Option Metrics (OPTION): Minute-level implied volatility and Greeks for CSI 300 ETF options (2024), modeling intraday market microstructure. – Bitcoin Futures-Spot Dynamics (BTCF): Hourly price-volume sequences (2020–2024), capturing crypto market lead-lag effects. • The performance validation of over ten leading time series forecasting models from recent research works on financial time series is conducted to showcase their practical viability in this domain. • Two new evaluation metrics, msIC (mean sequential correlation) and msIR (correlation stability ratio), are developed as a complement to traditional metrics like MSE and MAE, to better capture temporal correlation and robustness under non-stationarity. • Financial-specific tasks over these datasets, such as index portfolio optimization, Timing trading and BTC futures long&short strategies, are further designed to assess the practical performance and application potential of forecasting models. 2 Related Work 2.1 Time-series Forecasting Models Recent advances in time-series forecasting have focused on enhancing model capabilities through multi-scale decomposition, attention mechanism optimization, lightweight architectures, and representation learning innovations. Decomposition-based methods remain foundational for handling non-stationary signals. For example, Autoformer ( Wu et al. 2021a ) pioneers autocorrelation-driven periodicity detection, replacing traditional moving averages with adaptive decomposition, while FEDformer ( Zhou et al. 2022 ) leverages Fourier-Wavelet hybrid spectral analysis to achieve multi-resolution frequency decomposition. Building on this, Non-stationary Transformers ( Liu et al. 2022 ) introduce dynamic stationarization modules that learn to normalize non-stationary inputs before applying attention mechanisms, significantly improving robustness to distribution shifts. These approaches highlight the synergy between frequency-domain analysis and temporal modeling. To address computational bottlenecks in Transformers, researchers have developed sparse attention variants: Informer ( Zhou et al. 2021 ) reduces quadratic complexity via probabilistic sparse attention with KL-divergence-based token selection, while Crossformer ( Zhang & Yan 2023 ) designs hierarchical cross-resolution attention to capture dependencies across temporal scales. Koopa ( Liu et al. 2023a ) bypasses attention entirely, proposing time-varying Koopman dynamic systems that model latent state evolution through linear operators. Concurrently, lightweight architectures challenge conventional wisdom-DLinear ( Zeng et al. 2023 ) demonstrates that decoupled linear projections for trend and residual components can outperform complex models,

Read the complete text