Pretrained Time-Series Foundation Models for Financial Return Forecasting

Battle Count: 108

The paper is directly relevant to quantitative trading as it benchmarks practical forecasting tools for equity return prediction. However, the authors explicitly caution that gains over random-walk benchmarks are small and sparse, with only 2 of 10 cases showing statistically significant improvement. The paper does not evaluate trading performance, transaction costs, or portfolio construction. The main finding is that pretrained TSFMs reduce model-development costs in low-data settings but do not constitute reliable alpha generation engines. The theoretical framing (information-theoretic bounds, Kelly-Cover capacity) explicitly limits the achievable economic edge. The work is more relevant as a practical tooling benchmark than as a source of trading signals.

The benchmark involves 11 different model architectures spanning multiple families (MLP decomposition, transformers, quantized language models, KAN, decoder-only foundation models). Pretrained TSFMs require API access or specific checkpoint loading (TimeGPT, TimesFM-2.5, Moirai-2.0, Chronos, Chronos-2). Train-from-scratch baselines require per-ticker supervised training with MSE loss and early stopping. The rolling-origin evaluation protocol with 10 windows, equalized context of L=512, and Diebold-Mariano testing with Harvey-Leybourne-Newbold correction adds complexity. The theoretical framework spans PAC-Bayes bounds, information geometry, operator theory, and rough-path signatures, though these are interpretive rather than computational.

The paper provides a detailed reproducibility checklist (Appendix B) specifying data source (Yahoo Finance API), date range, return formulas, train/evaluation split, forecast horizon, rolling evaluation protocol, point forecast extraction method, primary baselines, and statistical test. However, the authors explicitly state that the paper does not constitute a complete executable replication package. Full computational reproducibility requires a companion code release with package versions, model checkpoints, API version identifiers, random seeds, optimizer settings, early-stopping split, rolling-origin calendar dates, and vendor-specific forecast parameters. No GitHub repository URL is provided.