Fixing Forecast Collapse in Time-Series Models

Summary & Methodology Analysis

Time-series foundation models often encounter forecast collapse, a failure mode where the model outputs nearly flat predictions during hourly return forecasting. This limits the model's utility for downstream tasks like ranking stocks, as the point forecasts lose their discriminative power. The authors identified that this issue stems from a combination of low predictability, which constrains the amplitude of calibrated point forecasts, and the reliance on per-series objectives that fail to account for relationships between different assets in the dataset.

To address these limitations, the paper introduces CalibRank, an objective function designed to improve forecasting performance by explicitly balancing two competing goals: minimizing squared errors and maximizing cross-sectional correlation. By optimizing for these concurrently, the method ensures that the model preserves the relative ranking of financial assets while maintaining accurate forecast amplitudes. This replaces or supplements the standard per-series optimization approaches that typically struggle to identify the underlying structural relationships in cross-series data.

Evaluation on the Finance1K dataset demonstrated that CalibRank consistently improves cross-sectional correlation across all tested models. The paper notes that conventional evaluation metrics are insufficient for detecting these issues because they focus on individual series metrics, which often obscure broader structural failures across the entire dataset. While the paper provides performance gains, it does not specify the latency impact or the computational resource requirements for training or inference with the CalibRank objective.