Forecast Collapse in Time-Series Foundation Models

“Forecast Collapse in Time-Series Foundation Models” examines why models such as TimesFM and Chronos can produce nearly flat forecasts for difficult financial targets. On Finance1K, an hourly panel covering 1,000 US equities and 28,510 time steps, mean-squared-error training predicts stock returns at only 0.027 of target amplitude, with cross-sectional information coefficient, or IC, of 0.046. The paper identifies two mechanisms: low predictability shrinks the amplitude of calibrated point forecasts, while per-series objectives fail to identify the cross-series coupling required for ranking assets. Optimizing IC directly raises return-ranking performance to 0.124 but produces forecasts at 25.260 times the target amplitude, revealing a severe calibration-ranking tradeoff. The proposed CalibRank objective combines MSE with differentiable per-timestamp Pearson correlation; on the controlled backbone it reaches IC 0.126 and raw amplitude 1.836, nearly tripling IC while remaining close to the target scale. Across 12 forecasting architectures, CalibRank improves IC for every model, showing that the failure is not specific to a Transformer design. A matched volume experiment provides a negative control: because volume is more predictable, MSE forecasts retain substantial scale and ranking quality. Across 97 GIFT-Eval configurations, raw amplitude also tracks achieved predictability, with correlations of 0.88 for Chronos and 0.87 for TimesFM. The central implication is that time-series benchmarks must evaluate per-series accuracy together with amplitude and the cross-sectional structure used by downstream decisions.