This paper investigates why Transformer-based models often fail in financial time-series forecasting despite their success in structured domains like electricity or traffic. The authors show that under squared loss, the Bayes-optimal predictor for a martingale price process is trivial: a flat trajectory for prices or zero for returns. When the conditional mean is degenerate, increasing model expressivity does not reduce bias but instead amplifies noise through interpolation, leading to higher prediction variance. The paper formalizes this by comparing a simple linear model with a rich class containing interpolating predictors, proving that the latter yields strictly larger expected squared error. Empirical experiments on high-frequency EUR/USD data confirm that a PatchTST model produces larger errors than a linear benchmark across most forecasting windows, supporting the theoretical claim that the collapse is inherent to the learning objective rather than a result of optimization or data limitations.