47% returns. Sharpe 6.8. Foundation models aren’t ready — until you retrain them from scratch.

A new paper benchmarks large-scale Time Series Foundation Models (TSFMs) across 34 years, 94 countries, and 2 billion+ return observations.

The result?

• Off-the-shelf TSFMs underperform CatBoost,

• Fine-tuning helps a little,

• But pre-training from scratch on financial data yields up to 5.4 Sharpe,

• And when you add synthetic data + JKP factors + tuning... → TSFMs hit 41.9% return, 6.78 Sharpe.

Still not plug-and-play — but close.