The Label Horizon Paradox: Rethinking Supervision Targets in Financial Forecasting
A new research paper introduces the "Label Horizon Paradox," challenging the standard practice of using direct inference targets as training labels in financial forecasting. The authors propose that optimal supervision signals often deviate from prediction goals, shifting across intermediate horizons based on market dynamics. They developed a bi-level optimization framework to identify these optimal proxy labels, demonstrating improved performance on large-scale financial datasets. AI
IMPACT Introduces a novel theoretical framework for optimizing AI model training in financial forecasting, potentially improving accuracy and generalization.