Reproduce “The Label Horizon Paradox: Rethinking Supervision Targets in Financial Forecasting” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chen-Hui Song, Shuoling Liu, Liyuan Chen
The authors propose selecting the return horizon used to train a stock forecaster while evaluating predictions against a fixed final horizon. A bi-level method learns horizon weights using validation against that target before retraining the predictor. On daily next-day forecasting with an LSTM on CSI 500, test IC rises from 0.0845 to 0.1029; results vary by market setting and horizon.
Key insights
- The "Label Horizon Paradox" reveals that the optimal supervision signal for financial forecasting is often an intermediate time horizon, not the final prediction target, driven by a dynamic trade-off between information gain and noise accumulation.
- A theoretical framework, extending Arbitrage Pricing Theory, quantifies how generalization performance is governed by the relative rates of marginal signal realization and noise accumulation.
- The optimal label horizon is dynamic and cannot be predetermined, necessitating an adaptive mechanism to capture its equilibrium where marginal information gain balances marginal noise penalty.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.