The New Benchmark for Forecasts of the Real Price of Crude Oil
A common way to advocate for a new forecast is to show it is more accurate than a benchmark forecast. The most popular benchmark in this context is the no-change benchmark—a naïve forecast that simply takes the last observed value of the series of interest to predict future values. The approach is appealing because improvements over this benchmark normally mean the series of interest is predictable in general and the new forecast is useful in practice. However, this reasoning does not hold when the series of interest is a lower-frequency transformation of high-frequency data.
We show that the theoretically optimal forecast under the null hypothesis of “no predictability” is different from the conventional no-change forecast: it is the last value of the high-frequency data. For temporally aggregated macroeconomic series, the theoretical gains in forecast accuracy from using the new benchmark are as large as 45 percent. A simple change of the benchmark can thus have large effects on the assessment of different forecasts.
We apply these insights to forecasts of the real price of crude oil and propose a new benchmark for forecast evaluations that relies on monthly closing prices. Similar to our theoretical results, we find that the new benchmark produces more accurate forecasts than the conventional no-change forecast computed from monthly average oil prices. We also find that traditional forecasting models for the real price of oil improve considerably when estimated with closing prices rather than average prices. Nonetheless, only forecasts derived from oil futures prices significantly outperform the new benchmark, and only for forecast horizons larger than six months. The introduction of a more suitable benchmark for forecast comparisons shows that oil prices are more difficult to predict than previously thought.