We use machine learning to capture nonlinearities and interactions in the relation between trades and holdings of multiple market participants and future stock returns. Our predictor yields a long-short portfolio with significant out-of-sample alpha, forecasts firm fundamentals, and assigns stocks on the right side of most anomalies. Predictability is stronger for smaller or illiquid stocks and stocks with lower analyst coverage or higher idiosyncratic volatility. A factor model based on our predictor achieves higher Sharpe ratio than existing models. Our findings suggest that incorporating nonlinear interactions between trades and holdings of various participants reveals valuable information for price discovery.