Generalized random forests (GRFs) are a flexible, nonparametric statistical estimation method that extends traditional random forests to estimate any quantity of interest identified through local moment equations. Unlike classical kernel-based methods, GRFs use adaptive weighting derived from a forest to capture heterogeneity in the target quantity. The method involves growing a forest of trees, where each tree contributes a weighted set of nearby training examples, and solving a plug-in version of the estimating equation using these weights. This approach allows for efficient, computationally stable estimation of quantities such as conditional means, quantiles, and heterogeneous treatment effects via instrumental variables. GRFs are built on the idea of adaptive nearest neighbor estimation, where the forest's structure is used to determine the weights for each test point. This weighting function is derived from the fraction of trees in which an observation appears in the same leaf as the target value. The method is designed to handle high-dimensional data and avoids the curse of dimensionality by using a forest-based adaptive weighting scheme. The algorithm is computationally efficient and can be implemented using existing tree software, such as the ranger package in R. Theoretical analysis shows that GRFs are consistent and asymptotically Gaussian, with valid confidence intervals. The method is applicable to a wide range of statistical tasks, including nonparametric quantile regression, conditional average partial effect estimation, and heterogeneous treatment effect estimation. The approach is robust to model misspecification and can be used in various settings where local moment conditions are relevant. GRFs are implemented in the grf package for R and C++, and the method is supported by a large sample theory that establishes asymptotic normality of the estimates. The algorithm uses a gradient-based splitting scheme to maximize heterogeneity in the target functional, leading to more accurate and stable estimates. The method is also compatible with subsampling and honesty techniques to ensure consistency and reduce overfitting. Overall, GRFs provide a flexible, efficient, and theoretically sound approach to nonparametric estimation and inference.Generalized random forests (GRFs) are a flexible, nonparametric statistical estimation method that extends traditional random forests to estimate any quantity of interest identified through local moment equations. Unlike classical kernel-based methods, GRFs use adaptive weighting derived from a forest to capture heterogeneity in the target quantity. The method involves growing a forest of trees, where each tree contributes a weighted set of nearby training examples, and solving a plug-in version of the estimating equation using these weights. This approach allows for efficient, computationally stable estimation of quantities such as conditional means, quantiles, and heterogeneous treatment effects via instrumental variables. GRFs are built on the idea of adaptive nearest neighbor estimation, where the forest's structure is used to determine the weights for each test point. This weighting function is derived from the fraction of trees in which an observation appears in the same leaf as the target value. The method is designed to handle high-dimensional data and avoids the curse of dimensionality by using a forest-based adaptive weighting scheme. The algorithm is computationally efficient and can be implemented using existing tree software, such as the ranger package in R. Theoretical analysis shows that GRFs are consistent and asymptotically Gaussian, with valid confidence intervals. The method is applicable to a wide range of statistical tasks, including nonparametric quantile regression, conditional average partial effect estimation, and heterogeneous treatment effect estimation. The approach is robust to model misspecification and can be used in various settings where local moment conditions are relevant. GRFs are implemented in the grf package for R and C++, and the method is supported by a large sample theory that establishes asymptotic normality of the estimates. The algorithm uses a gradient-based splitting scheme to maximize heterogeneity in the target functional, leading to more accurate and stable estimates. The method is also compatible with subsampling and honesty techniques to ensure consistency and reduce overfitting. Overall, GRFs provide a flexible, efficient, and theoretically sound approach to nonparametric estimation and inference.