Learning Time-Varying Forecast Combinations

Combining forecasts has been demonstrated as a robust solution to noisy data, structural breaks, unstable forecasters and shifting environmental dynamics. In practice, sophisticated combination methods have failed to consistently outperform the mean over multiple horizons, pools of varying forecasters and different endogenous variables. This paper addresses the challenge to “develop methods better geared to the intermittent and evolving nature of predictive relations”, noted in Stock and Watson (2001), by proposing an adaptive non-parametric “meta” approach that provides a time-varying hedge against the performance of the mean for any selected forecast combination approach. This approach arguably solves the so-called “Forecast Combination Puzzle” using a meta-algorithm that adaptively hedges weights between the mean and a specific forecast combination algorithm or pool of forecasters augmented with one or more forecast combination algorithms. Theoretical performance bounds are reported and empirical performance is evaluated on the seven country macroeconomic output and inflation dataset introduced in Stock and Watson (2001) as well as the Euro-area Survey of Professional Forecasters