Re(Visiting) Time Series Foundation Models in Finance
Eghbal Rahimikia, Hao Ni, Weiguan Wang
TL;DR
This study delivers the first rigorous empirical assessment of time series foundation models (TSFMs) in global finance, contrasting zero-shot, fine-tuned, and finance-domain pre-training. It finds that generic pre-trained TSFMs underperform established benchmarks, while finance-native pre-training on large, domain-specific data substantially boosts predictive and economic performance, especially for longer input windows. Data scaling to global datasets and augmentation with factors (JKP) or synthetic series further enhances outcomes, though gains are sensitive to hyperparameters and come with high computational costs. The results highlight that TSFMs hold promise for finance when tailored data, task-aligned pre-training, and careful tuning are combined, and they emphasize the ongoing trade-off between predictive power and real-world trading frictions.
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.