Re(Visiting) Time Series Foundation Models in Finance

Battle Count: 216

Highly relevant to quantitative trading. The paper directly addresses daily excess return forecasting for cross-sectional stock selection, portfolio construction (long-short decile portfolios), and risk-adjusted performance evaluation. Key findings include: ensemble models (CatBoost) remain the best benchmark with Sharpe ratios up to 6.79; TSFMs pre-trained from scratch on financial data achieve Sharpe ratios up to 5.42-7.69 (with augmentation); transaction costs significantly impact all strategies; TSFMs show slower performance degradation over time. The paper provides actionable guidance on when and how to use TSFMs in trading contexts.

High implementation complexity due to: (1) large-scale data processing (~2 billion observations across 94 countries), (2) GPU-intensive pre-training of Transformer architectures (50,000 GPU hours total), (3) multiple model variants and training regimes requiring careful hyperparameter tuning, (4) complex data cleaning and harmonization across CRSP and Compustat sources, (5) expanding-window estimation with 23 yearly models per configuration, (6) need for specialized infrastructure (NVIDIA GH200 GPUs with 96GB HBM). However, pre-trained models are available on Hugging Face for inference.

Models are publicly available through FinText.ai portal and Hugging Face repository (https://huggingface.co/FinText). The paper provides detailed descriptions of data cleaning, model architectures, hyperparameter settings, and evaluation procedures. Computational resources are documented (50,000 GPU hours, 25,000 CPU hours). However, the full dataset (CRSP, Compustat, JKP factors) requires institutional access. Random number generators are initialized for consistency.