Predictive gains from forecast combinations using time-varying model weights

Several frequentist and Bayesian model averaging schemes, including a new one that simultaneously allows for parameter uncertainty, model uncertainty and time varying model weights, are compared in terms of forecast accuracy over a set of simulation experiments. Artificial data are generated, characterized by low predictability, structural instability, and fat tails, which is typical for many financial-economic time series. Sensitivity of results with respect to misspecification of the number of included predictors and the number of included models is explored. Given the set up of our experiments, time varying model weight schemes outperform other averaging schemes in terms of predictive gains both when the correlation among individual forecasts is low and the underlying data generating process is subject to structural locations shifts. In an empirical application using returns on the S&P 500 index, time varying model weights provide improved forecasts with substantial economic gains in an investment strategy including transaction costs.