Decoding financial trading decisions in the wild with physiological foundation models and LLMs

Biosignal foundation models are evaluated almost entirely on curated laboratory and clinical recordings. In the real world, signals are noisy and behavior is confounded with context, so effects are expected to be small, and what decides whether neurotechnology works outside the lab is whether pretraining delivers that edge robustly, and whether it survives alongside the behavioral context a deployed system would already have. We record scalp EEG and single-lead ECG from professional day traders during live trading, and ask whether the physiology in the moments before a trader places an order can be told apart from quiet moments from the same trader on the same day. Across $n=18$ traders and 16,872 examples, a linear probe on frozen CBraMod embeddings reaches 0.594 AUROC in a window ending five seconds before the order, against 0.536 for per-channel band power and 0.537 for a Conformer trained end-to-end on the same data. The advantage over band power is significant in every pre-event window and persists when frontal channels are removed and when events with uncertain order times are discarded, conditions under which handcrafted decoding degrades and, in the second case, falls to chance. In the cardiac modality the frozen model matches handcrafted heart-rate features on mean accuracy but is $3.3\times$ more consistent across traders. Supplying the same embeddings to a frozen LLM together with the trader's recent order history raises AUROC from 0.690 with context alone to 0.774. The physiological edge is small in isolation, but adds to, rather than dissolves into, strong behavioral context. Pretrained physiological representations thus recover a weak anticipatory signal that neither hand-designed features nor a task-trained network reach, and one that becomes useful once combined with the context a deployed system already has.