Machine learning provides significant value for asset pricing

Machine learning can add significant value for investors who employ these predictive tools as an alternative to traditional regression analysis. A comprehensive study of equity prediction using machine learning shows a significant improvement in out of sample R-squared. 

This added predictive value is seen across a variety of machine learning techniques. Using machine learning also allows investors to identify the most informative predictive variables. Just as important, the more expansive list of predictors with different specifications and functional forms employed with machine learning allows for an improved understanding of the drivers for stock prediction. (See "Empirical Asset Pricing via Machine Learning")

The authors provide a good three-part definition of machine learning:

• A diverse collection of high dimensional models for flexible statistical prediction. 
• Regularized methods for model selection and overfit minimization to ensure high out of sample performance.
• Efficient algorithms for searching among a large set of model specifications to control computation costs.

Trees, (boosted regression trees and random forests), and neural nets clearly outperform linear regressions and allow for a broader set of predictive features. Machine learning, which focuses on nonlinear methods, needs not be overly complex to add value. Along with predictive gains, machine learning also provides clear economic gains through higher portfolio Sharpe ratios. The study shows that price trends, liquidity, and volatility measure prove to be the most effective predictors.

Machine learning, coupled with advances in data science, allow investors to generate better predictive models and improved methods for finding the key variables that drive stock performance. The cost of entry for using machine learning may be high, yet the greater predictive value makes this a useful investment.