A foundation model is not a foundation
Measured evaluation of Google's TimesFM 3.0 against 118,857 bars of Nasdaq futures tape, five sessions, strictly causal and session confined.
Direction accuracy on price was exactly 50.0%, and error magnitude came in 7.4% worse than repeat-last-value. Coarser bars did not help. A purpose-built mean-reverting target designed to suit the model did not help. Covariates, the supported mechanism for side information, made forecasts worse at every horizon.
A probability signal built from the returned quantiles was anti-informative when most confident: 24.7% realized against a 34.3% base rate.
The cause is architectural. The model is stateless, has no text interface, and cannot be instructed. It maps a numeric window to a numeric window. It cannot hold a regime, carry state, or condition one signal on another, which is precisely the shape of whatever structure survives in a liquid market.
Credit where due: 98.7% better than naive on a seasonal series with perfect direction accuracy, zero-shot. Roughly 1.3 GB of VRAM, about 8% of a mid-range consumer card, 7.9 ms per series at batch 128. And on a random walk it performed slightly worse than doing nothing, which is the correct behaviour, not a defect.
The word "foundation" describes a training and transfer regime, not a level of competence, and it sets an expectation the architecture was never built to meet.
https://michaelhairetis.medium.com/a-foundation-model-is-not-a-foundation-a3edf3383be9