Market Microstructure Meets Deep Learning: Forecasting and Asset Allocation with PIN Signals

This study examines whether informed trading signals derived from the Probability of Informed Trading (PIN) model can systematically enhance predictive modeling and investment strategies. Using Brazilian equities as a representative emerging market, we integrate PIN-based features into machine learning frameworks and portfolio allocations. Results reveal a clear divergence: PIN inputs often deteriorate return forecast accuracy—particularly in deep learning models—while select configurations, such as XGBoost and deep Q-learning with lagged PIN features, yield statistically significant improvements in directional accuracy of up to 12.8%. More notably, when used for asset allocation, PIN consistently enhances risk-adjusted returns. Portfolios constructed using short-window PIN estimates achieve Sharpe ratios above 1.9 on average and outperform market benchmarks by a wide margin, particularly under high-intensity and directional PIN-based strategies. These findings suggest that the informational value of PIN resides more in signal filtering for investment than in improving forecast magnitudes. Our analysis challenges the assumption that informational variables uniformly improve model performance and highlights the importance of contextual deployment. By bridging predictive modeling with microstructure-based signals, the study contributes to both financial machine learning and market microstructure literature, offering practical guidance on deploying informed trading metrics in data-driven investment workflows.