Forecast combinations: An over 50-year review

Forecast combinations have flourished remarkably in the forecasting community and, recent years, become part of mainstream research and activities. Combining multiple forecasts produced for a target time series is now widely used to improve accuracy through integration information gleaned from different sources, thereby avoiding need identify single “best” forecast. Combination schemes evolved simple combination methods without estimation sophisticated techniques involving time-varying weights, nonlinear combinations, correlations among components, cross-learning. They include combining point probabilistic forecasts. This paper provides an up-to-date review extensive literature on forecast reference available open-source software implementations. We discuss potential limitations various highlight how these ideas developed over time. Some crucial issues concerning utility are also surveyed. Finally, we conclude with current gaps insights future research.