Long-term forecasting of hourly retail customer flow on intermittent time series with multiple seasonality

In this study, we address a demanding time series forecasting problem that deals simultaneously with the following: (1) intermittent time series, (2) multi-step ahead forecasting, (3) time series with multiple seasonal periods, and (4) performance measures for model selection across multiple time series. Current literature deals with these types of problems separately, and no study has dealt with all these characteristics simultaneously. To fill this knowledge gap, we begin by reviewing all the necessary existing literature relevant to this case study with the goal of proposing a framework capable of achieving adequate forecast accuracy for such a complex problem. Several adaptions and innovations have been conducted, which are marked as contributions to the literature. Specifically, we proposed a weighted average forecast combination of many cutting-edge models based on their out-of-sample performance. To gather strong evidence that our ensemble model works in practice, we undertook a large-scale study across 98 time series, rigorously assessed with unbiased performance measures, where a week seasonal naïve was set as a benchmark. The results demonstrate that the proposed ensemble model achieves eye-catching forecasting accuracy.