This paper evaluates machine learning methods for empirical asset pricing, focusing on measuring asset risk premiums. The authors compare various machine learning techniques, including trees and neural networks, against traditional regression-based strategies. They find that machine learning methods significantly outperform existing approaches, with some strategies doubling the performance of leading regression-based methods. The best-performing methods are trees and neural networks, which capture nonlinear interactions among predictors that other methods miss. All methods agree on the same set of dominant predictive signals, including variations on momentum, liquidity, and volatility. The study uses a large dataset of nearly 30,000 individual stocks over 60 years (1957-2016), with 94 characteristics for each stock, interactions with eight aggregate time-series variables, and 74 industry sector dummy variables. The authors find that machine learning methods can improve the empirical understanding of asset returns by creating return forecasting models that dominate traditional approaches. They also show that machine learning methods can provide substantial economic gains, with portfolio strategies based on neural network forecasts achieving higher Sharpe ratios than traditional methods. The paper discusses the advantages of machine learning in asset pricing, including its ability to handle a large number of predictor variables and complex functional forms. It also addresses the challenges of overfitting and the need for regularization techniques to improve out-of-sample performance. The study highlights the importance of variable selection and dimension reduction in high-dimensional settings. The authors compare various machine learning methods, including linear regression, generalized linear models, dimension reduction techniques (PCR and PLS), regression trees, and neural networks. They find that nonlinear methods, such as trees and neural networks, significantly improve return prediction. The paper also discusses the economic implications of these findings, showing that machine learning can lead to substantial gains in portfolio performance. The study concludes that machine learning has great potential for improving risk premium measurement, which is fundamentally a problem of prediction. It highlights the importance of accurately modeling the conditional expectation of future returns given available information. The paper also discusses the limitations of traditional methods in handling large numbers of predictor variables and the need for more advanced statistical tools in machine learning to overcome these limitations. The authors argue that machine learning methods can provide more accurate and robust predictions of asset returns, leading to better investment decisions.This paper evaluates machine learning methods for empirical asset pricing, focusing on measuring asset risk premiums. The authors compare various machine learning techniques, including trees and neural networks, against traditional regression-based strategies. They find that machine learning methods significantly outperform existing approaches, with some strategies doubling the performance of leading regression-based methods. The best-performing methods are trees and neural networks, which capture nonlinear interactions among predictors that other methods miss. All methods agree on the same set of dominant predictive signals, including variations on momentum, liquidity, and volatility. The study uses a large dataset of nearly 30,000 individual stocks over 60 years (1957-2016), with 94 characteristics for each stock, interactions with eight aggregate time-series variables, and 74 industry sector dummy variables. The authors find that machine learning methods can improve the empirical understanding of asset returns by creating return forecasting models that dominate traditional approaches. They also show that machine learning methods can provide substantial economic gains, with portfolio strategies based on neural network forecasts achieving higher Sharpe ratios than traditional methods. The paper discusses the advantages of machine learning in asset pricing, including its ability to handle a large number of predictor variables and complex functional forms. It also addresses the challenges of overfitting and the need for regularization techniques to improve out-of-sample performance. The study highlights the importance of variable selection and dimension reduction in high-dimensional settings. The authors compare various machine learning methods, including linear regression, generalized linear models, dimension reduction techniques (PCR and PLS), regression trees, and neural networks. They find that nonlinear methods, such as trees and neural networks, significantly improve return prediction. The paper also discusses the economic implications of these findings, showing that machine learning can lead to substantial gains in portfolio performance. The study concludes that machine learning has great potential for improving risk premium measurement, which is fundamentally a problem of prediction. It highlights the importance of accurately modeling the conditional expectation of future returns given available information. The paper also discusses the limitations of traditional methods in handling large numbers of predictor variables and the need for more advanced statistical tools in machine learning to overcome these limitations. The authors argue that machine learning methods can provide more accurate and robust predictions of asset returns, leading to better investment decisions.