Re(Visiting) Time Series Foundation Models in Finance

Abstract

Financial time series forecasting is central to trading, portfolio optimization, and risk management, yet it remains challenging due to noisy, non-stationary, and heterogeneous data. Recent advances in time series foundation models (TSFMs), inspired by large language models, offer a new paradigm for learning generalizable temporal representations from large and diverse datasets. This paper presents the first comprehensive empirical study of TSFMs in global financial markets. Using a large-scale dataset of daily excess returns across diverse markets, we evaluate zero-shot inference, fine-tuning, and pre-training from scratch against strong benchmark models. We find that off-the-shelf pre-trained TSFMs perform poorly in zero-shot and fine-tuning settings, whereas models pre-trained from scratch on financial data achieve substantial forecasting and economic improvements, underscoring the value of domain-specific adaptation. Increasing the dataset size, incorporating synthetic data augmentation, and applying hyperparameter tuning further enhance performance.

Summary

This paper investigates the effectiveness of Time Series Foundation Models (TSFMs) in financial time series forecasting, a challenging area due to the noisy, non-stationary, and heterogeneous nature of financial data. The authors conduct a comprehensive empirical study using a large-scale dataset of daily excess returns from global financial markets spanning 34 years across 94 countries. They evaluate TSFMs under three regimes: zero-shot inference, fine-tuning, and pre-training from scratch, comparing their performance against strong benchmark models, including linear models, ensemble models (CatBoost, XGBoost, LightGBM), and neural networks. The key finding is that off-the-shelf pre-trained TSFMs perform poorly in zero-shot and fine-tuning settings. However, TSFMs pre-trained from scratch on financial data achieve substantial improvements in both forecasting accuracy and economic performance. Furthermore, increasing dataset size, incorporating synthetic data augmentation, and applying hyperparameter tuning further enhance the performance of these domain-specific TSFMs. The best-performing CatBoost model achieves an annualized return of 46.50% and a Sharpe ratio of 6.79 using a window size of 252. These results highlight the importance of domain-specific adaptation for TSFMs in finance. The authors make their models available through FinText.ai and Hugging Face.

Key Insights

• •Off-the-shelf TSFMs perform poorly in zero-shot forecasting of daily excess returns, underperforming benchmark ensemble models (e.g., CatBoost, LightGBM). For example, Chronos (large) model with 512 past excess returns generates out-of-sample R^2 of -1.37% and directional accuracy just above 51%, while the TimesFM (500M) model attains R^2 of -2.80% and directional accuracy just below 50%.
• •Fine-tuning pre-trained TSFMs on financial data yields limited improvements and fails to close the performance gap with benchmarks.
• •TSFMs pre-trained from scratch on financial data achieve substantial gains in both predictive accuracy and portfolio performance. For example, considering the Chronos (small) model, its R^2 for window size 5 increases substantially to -3.18% from -77.07%, and for window size 512, it increases to -0.59% from -1.27%.
• •Expanding the training universe from U.S. to global markets yields mixed results, improving the linear model's R^2 by 0.43-0.60% but weakening directional accuracies and portfolio performance for most models.
• •Combining TSFMs with financial factors and synthetic data augmentation leads to consistent improvements in both statistical and economic outcomes. The Chronos (small) model achieves the highest directional accuracy of 51.74%, compared with 51.16% obtained by CatBoost using a window size of 512.
• •Hyperparameter tuning is critical for the performance of TSFMs; with appropriate optimization, TSFMs can outperform benchmark models even without scaling the dataset.
• •Both TSFMs and benchmark models experience gradual degradation in portfolio performance over time, but the decline is markedly slower and less severe for TSFMs.

Practical Implications

• •The research suggests that practitioners should prioritize pre-training TSFMs on domain-specific financial data rather than relying on off-the-shelf models or fine-tuning generic models.
• •Quantitative analysts and portfolio managers can potentially leverage domain-adapted TSFMs to improve financial forecasting and portfolio construction, leading to enhanced risk-adjusted returns.
• •The open-source release of the models encourages further research and development of TSFMs in finance, allowing researchers to build upon the findings and explore new applications.
• •The insights into the importance of data scaling, synthetic data augmentation, and hyperparameter tuning can guide future research on improving the performance of TSFMs.
• •Future research can explore the application of TSFMs to other financial tasks, such as risk management, algorithmic trading, and derivative pricing.

Links & Resources

Authors

Cite This Paper