How can we leverage the predictive power of Machine Learning to estimate the cross-section of expected stock returns without losing all economic intuition in a black box?

I combine Machine Learning with Demand System Asset Pricing to infer expected returns from portfolio holdings of financial institutions. In particular, I train neural networks to predict demand for stocks. Then, I compute expected excess demand (EXD). Finally, expected market clearing implies that EXD predicts stock returns with a coefficient that equals the inverse aggregate demand elasticity.

This relationship holds empirically. First, EXD is a strong predictor of returns with a coefficient that implies a demand elasticity of 1.6, consistent with the index effects literature. Second, EXD absorbs anomalies related to liquidity and trading. Third, a long short trading strategy using EXD-sorted portfolios returns an annual alpha of 20% at a Sharpe ratio of 1.4. Overall, EXD performs on par with the direct neural net prediction, but uses only a sparse set of anomaly variables and remains economically interpretable.