Forecast Collapse in Time-Series Foundation Models
The paper says flat forecasts can be a calibrated response to low-predictability targets, but that breaks stock ranking.
The authors found the collapse while forecasting hourly returns for 1,000 US equities; the same setup did not show the issue for trading volume. They tested TSFMs, 12 deep-learning models, and 97 benchmark configurations, tying the failure to target predictability and per-series training objectives. Their proposed objective, CalibRank, nearly tripled cross-sectional correlation on Finance1K while keeping forecast amplitude close to the target. HF Daily Papers' note
The authors found the collapse while forecasting hourly returns for 1,000 US equities; the same setup did not show the issue for trading volume. They tested TSFMs, 12 deep-learning models, and 97 benchmark configurations, tying the failure to target predictability and per-series training objectives. Their proposed objective, CalibRank, nearly tripled cross-sectional correlation on Finance1K while keeping forecast amplitude close to the target. HF Daily Papers' note
score 5