🎤 Podcast

You can now listen to the accompanying podcast here: https://soundcloud.com/eghbal-rahimikia/revisiting-time-series-foundation-models-in-finance

🆕 GitHub Model Loading Support (NEW)

All models can now be loaded directly from GitHub. The repository includes utilities and setup instructions. 🔗 https://github.com/DeepIntoStreams/TSFM_Finance

🚀 TSFMs Release

We are pleased to introduce FinText-TSFM, a comprehensive suite of time series foundation models (TSFMs) with 613 models pre-trained for quantitative finance. This release accompanies the paper : Re(Visiting) Time Series Foundation Models in Finance by Eghbal Rahimikia, Hao Ni, and Weiguan Wang (2025).

💡 Key Highlights

• Finance-Native Pre-training:
  Models are pre-trained from scratch on large-scale financial time series datasets — including daily excess returns across 89 markets and over 2 billion observations — to ensure full temporal and domain alignment.

• Bias-Free Design:
  Pre-training strictly follows a chronological expanding-window setup, avoiding any look-ahead bias or information leakage.
  Each variation includes 23 separately pre-trained models, corresponding to each year from 2000 to 2023, with data starting in 1990.

• Model Families:
  This release includes variants of Chronos and TimesFM architectures adapted for financial time series:

  • Chronos-Tiny (8M) / Mini (20M) / Small (46M)
  • TimesFM-8M / 20M

• Model Collections:

  • U.S.: Covers U.S. market-wide excess returns from 2000 to 2023, with one pre-trained model per year.
  • Global: Covers excess returns across 94 global markets from 2000 to 2023, with one pre-trained model for each year.
  • Augmented: Extends the global data with augmented factors from 2000 to 2023, with one pre-trained model for each year.
  • The remaining 253 pre-trained models are available for download via the FinText.ai Portal. These include models pre-trained with varying hyperparameter configurations for extended experimentation and performance comparison.

• Performance Insights:
  Our findings show that off-the-shelf TSFMs underperform in zero-shot forecasting, while finance-pretrained models achieve large gains in both predictive accuracy and portfolio performance.

• Evaluation Scope:
  Models are benchmarked across U.S. and seven international markets, using rolling windows of 5, 21, 252, and 512 days, with over 18 million out-of-sample forecasts spanning 22 years (2001–2023) of daily excess returns, evaluated at both the statistical and economic performance levels.

🧠 Technical Overview

• Architecture: Transformer-based TSFMs (Chronos & TimesFM)
• Compute: 50,000 GPU hours on NVIDIA GH200 Grace Hopper clusters

📚 Citation

Please cite the accompanying paper if you use these models:

Re(Visiting) Time Series Foundation Models in Finance.
Rahimikia, Eghbal; Ni, Hao; Wang, Weiguan.
SSRN: [https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5770562)

🔋 Acknowledgments

This project was made possible through computational and institutional support from:

• UK Research and Innovation (UKRI)
• Isambard-AI National AI Research Resource (AIRR)
• Alliance Manchester Business School (AMBS), University of Manchester
• N8 Centre of Excellence in Computationally Intensive Research (N8 CIR)
• The University of Manchester (Research IT & Computational Shared Facility)
• University College London (UCL)
• The Alan Turing Institute
• Shanghai University

---