A multi-asset allocation study based on a time-varying hidden Markov regime-switching model

Traditional linear predictive regression models perform poorly in out-of-sample stock return predictions. One competing hypothesis for this result is that structural changes in the financial markets introduce model instability. This paper constructs a two-state multi-asset timevarying regime switching (TVTP-RS) model to investigate industry stock return predictability. In this model, the time-varying industry expected returns are driven by economic variables, but the relation between them may change due to shifts in market states, where market states are unobservable and follow a Markov chain with time-varying transition probabilities. In-sample estimation results reveal that jointly utilizing information from industry returns and economic variables can effectively identify latent market states, and the relation between economic variables and expected returns indeed depends on market states — The coefficients of some economic variables reverse in different market states. Out-of-sample industry return predictions and industry allocation strategies based on this model consistently outperform benchmark models and linear predictive regression models. This study provides new evidence of the time-varying relation between economic variables and industry expected returns and demonstrates that considering the impact of market state transitions can effectively reduce the model instabilities. The proposed TVTP-RS model also offers a reliable solution for industry rotation strategies in practice.