Meta-learning for return prediction in shifting market regimes

We propose a meta-learning framework for cross-sectional return prediction that adapts to regime-dependent dynamics. Instead of learning a fixed mapping from features to returns, we condition our model forecasts on recent feature-return relationships. This allows it to adjust to evolving market states without explicit regime labels or frequent re-estimation. We implement the framework with a Transformer-based Bayesian predictor, the Financial Prior-data Fitted Network (FinPFN), and evaluate it on daily Chinese A-shares and monthly U.S. equities. During regime changes, proxied by large volatility shifts, our method significantly outperforms benchmarks, offering a practical tool for dynamic return prediction.