Time-Series Models Face "Forecast Collapse" in Finance

When applying time-series foundation models (TSFMs) to forecast hourly returns for a large number of US equities, an unexpected issue termed "forecast collapse" has been observed. This phenomenon manifests as predictions becoming nearly flat and exhibiting poor performance in ranking stocks, as measured by cross-sectional correlation. Interestingly, this collapse largely disappears when the same models are used to forecast trading volume, suggesting a link to the predictability of the target variable. The study investigated forecast collapse across various TSFMs, deep-learning forecasting models, and benchmark configurations, concluding that it is strongly tied to the inherent predictability of the target data. Two primary causes were identified: low predictability inherently limits the amplitude of calibrated point forecasts, and the use of per-series optimization objectives fails to capture crucial cross-series structural information. These factors reveal a fundamental calibration-ranking tradeoff: optimizing for squared error tends to produce flat predictions, while directly optimizing for cross-sectional correlation can improve ranking but might inflate forecast amplitudes significantly. To mitigate this tradeoff, the researchers introduced CalibRank, a straightforward objective designed to balance both calibration and ranking. On the Finance1K dataset, CalibRank nearly tripled cross-sectional correlation while maintaining forecast amplitude close to the actual target values, and improved correlation across all tested models. These findings highlight a blind spot in conventional time-series evaluation, where per-series metrics can obscure failures in cross-series structure vital for downstream decision-making.