Mixed‐frequency predictive regressions with parameter learning

We explore the performance of mixed-frequency predictive regressions for stock returns from perspective a Bayesian investor. develop constrained parameter learning approach sequential estimation allowing belief revisions. Empirically, we find that models improve predictability, not only because combination predictors with different frequencies but also due to preservation high-frequency features such as time-varying volatility. Temporally aggregated misspecify evolution frequency volatility dynamics, resulting in poor timing and worse portfolio than specification. These results highlight importance preserving potential nature regressions.