We propose a pair of summary measures for the predictive power of a prediction
function based on a regression model. The regression model can be linear
or nonlinear, parametric, semi-parametric, or nonparametric, and correctly
specified or mis-specified. The first measure, R-squared, is an extension of
the classical R-squared statistic for a linear model, quantifying the prediction
function's ability to capture the variability of the response. The second
measure, L-squared, quantifies the prediction function's bias for predicting the
mean regression function. When used together, they give a complete summary of
the predictive power of a prediction function. Please refer to Gang Li and Xiaoyan Wang (2016)