Asset pricing with complexity

Machine learning methods for big data trade off bias for precision in prediction. To understand the implications for financial markets, I formulate an asymmetric information model where agents optimally choose a biased estimator. The model identifies a novel cost of complexity that arises endogenously and makes it optimal to ignore costless signals – limiting the value of big data in finance. Complexity drives a wedge between the private and social value of data lowering price informativeness. Short-term price reversals do not identify liquidity demand and there is return predictability in in-sample tests not driven by priced risk or behavioral biases. Return predictability persists out-of-sample when machine learning methods applied in the test are more advanced than investors’ predictions, especially when signals are included that investors ignore