The Label Horizon Paradox: Rethinking Supervision Targets in Financial Forecasting

While deep learning has revolutionized financial forecasting through sophisticated architectures, the design of the supervision signal itself is rarely scrutinized. We challenge the canonical assumption that training labels must strictly mirror inference targets, uncovering the Label Horizon Paradox: the optimal supervision signal often deviates from the prediction goal, shifting across intermediate horizons governed by market dynamics. We theoretically ground this phenomenon in a dynamic signal-noise trade-off, demonstrating that generalization hinges on the competition between marginal signal realization and noise accumulation. To operationalize this insight, we propose a bi-level optimization framework that autonomously identifies the optimal proxy label within a single training run. Extensive experiments on large-scale financial datasets demonstrate consistent improvements over conventional baselines, thereby opening new avenues for label-centric research in financial forecasting.

Many stock-forecasting models are trained with a simple assumption: if you want to predict tomorrow’s return, you should also train the model using tomorrow’s return as the label. We found this “same target, same label” rule often fails in practice, because tomorrow’s return can be dominated by short-term randomness, so the model ends up learning noise. Our key idea is to decouple the prediction goal from the training label. We still evaluate the model on the original goal (e.g., predicting next-day return), but during training we let the model choose a different horizon as its label—one that carries a cleaner, more learnable signal. We design an end-to-end method that automatically searches for this best label horizon while the model trains, instead of running many expensive trials by hand. In experiments across multiple model types and markets, this label choice reliably improves results. More broadly, our work argues for a label-centric view of financial prediction: in noisy domains, choosing what to learn from can matter as much as choosing which model to use.