Forecasting automobile gasoline demand in Australia using machine learning-based regression

We use a variant of machine learning (ML) to forecast Australia's automobile gasoline demand within an autoregressive and structural model. By comparing the outputs of various model specifications, we find that training set selection plays an important role in forecasting accuracy. More specifically, however, the performance of training sets starting within identified systematic patterns is relatively worse, and the impact on forecast errors is substantial. We explain these systematic variations in machine learning performance, and explore the intuition behind the ‘black-box’ with the support of economic theory. An important finding is that these time points coincide with structural changes in Australia's economy. By examining the out-of-sample forecasts, the model's external validity can be demonstrated under normal situations; however, its forecasting performance is somewhat unsatisfactory under event-driven uncertainty, which calls on future research to develop alternative models to depict the characteristics of rare and extreme events in an ex-ante manner.