京都大学 大学院経済学研究科・経済学部

This paper develops a fund-level mixture hidden Markov model (MHMM) to study stock–bond allocation ability among Balanced Funds in the CRSP mutual fund database. The model combines time-invariant latent ability classes with time-varying allocation regimes and links fund returns to an allocation-correctness measure based on lagged stock–bond tilts and subsequent relative benchmark performance. Using 256,038 fund-month observations from 1,893 funds, the estimated two-class, three-state MHMM identifies clear latent class separation, strong regime persistence, and meaningful heterogeneity in how correct allocation signals translate into returns. One class exhibits consistently positive return responses to correct stock–bond allocation, while the other shows mixed or negative responses. Out-of-sample tests confirm that the estimated ability structure has predictive content: high-ability funds outperform low-ability funds, ability rankings sort future returns monotonically, and the MHMM outperforms simpler benchmark models on predictive log score.