Putting VAR forecasts of the real price of crude oil to the test

This study reevaluates crude oil price forecasts from state-of-the-art VAR models (Baumeister et al., 2022). Unlike Baumeister et al., who use the average-price no-change forecast, we employ the end-of-period no-change forecast, corresponding to the traditional random walk hypothesis. VAR forecasts do not significantly outperform the random walk for horizons under one year. The average-price benchmark systematically biases the Diebold–Mariano test statistic, affecting inference on forecast improvements up to 18 months. Similar biases are observed for alternative forecast criteria. The fact that naive benchmark choice alters inference even at extended horizons is relevant for all forecasts targeting averaged series.