Abstract:
We propose a new no-change benchmark to evaluate forecasts of series that are temporally aggregated. The new benchmark is the last high-frequency observation and reflects the null hypothesis that the underlying series, rather than the aggregated series, is unpredictable. Under the random walk null hypothesis, using the last high-frequency observation improves the mean squared prediction errors of the no-change forecast constructed from average monthly or quarterly data by up to 45 percent. We apply this insight to forecasts of the real price of crude oil and show that a new benchmark that relies on monthly closing prices dominates the conventional no-change forecast in terms of forecast accuracy. Although model-based forecasts also improve when models are estimated using closing prices, only the futures-based forecast significantly outperforms the new benchmark. Introducing a more suitable benchmark changes the assessments of different forecasting approaches and of the general predictability of real oil prices.