We postulate that return predictability is an intrinsic and time-varying asset characteristic potentially related to the cross-section of expected returns, instead of just an attribute of the chosen predictors or models. We develop a tree-based clustering method to gauge heterogeneous return predictability by grouping a panel of asset returns using high-dimensional asset characteristics and market-wide predictors. Our approach tells what types of assets exhibit greater return predictability under what market conditions, and empirically reveals substantial predictability heterogeneity in the U.S. equity market. Stocks with high earnings surprises, high earnings-to-price ratios, and low trading volumes exhibit the strongest predictability; predictability diminishes sharply with low market dividend yield but peaks with high dividend yield and low market liquidity. Out-of-sample, a new anomaly linked to investors' model misspecification easily generates monthly excess alphas exceeding 1\%, and investing in highly predictable clusters significantly outperforms conventional benchmarks with Sharpe ratios approaching 2.