Prior Mismatch in Tabular Foundation Models and How to Fix It

Tabular foundation models such as TabPFN are increasingly used on financial time series.There they are overconfident, and they lose skill at market regime breaks. We show thatthis failure is shared by three TabPFN generations (TabPFN-2, -3 and -3.5), and we traceboth parts of it to the synthetic prior the models were trained on. Overconfidence followsfrom one property of that prior: how much noise there is in the answers. In a controlledexperiment with small prior-fitted networks, a model trained on cleaner tasks than it is testedon is overconfident (temperature 2.70). A model trained on noisier tasks is underconfident(0.38), and matched noise gives calibrated predictions (1.02 and 0.98). To repair skill aswell as confidence, the prior must also behave like the market. We build a realistic marketprior: simulated stocks with volatility clustering, regime breaks and market-level noise, runthrough the real feature pipeline and fitted to 20 calibration stocks kept apart from the testset. We fine-tune TabPFN-2 (7M parameters) and TabPFN-3.5 (220M) on it and run bothon 20 test stocks and 7 currencies never used to build the prior. Both models reach thesame result. Their temperature falls from about 3.6 to 1.15–1.38, so their confidence comesclose to how often they are right. On stocks, break-day AUC rises from about 0.62 to 0.76,and break-day skill turns positive, while logistic regression and the base models stay near orbelow zero. With the same inputs, the fine-tuned models also beat a volatility-persistencerule, models trained on real data and recalibrated base models. They are also better innormal periods (AUC increase from about 0.65 to 0.70 on stocks, with smaller gains oncurrencies). Overconfidence and the loss of skill at regime breaks are therefore properties ofthe prior, not of the architecture, and a prior matched to the domain repairs both.