Memory augmented coherent probabilistic forecasts for hierarchically related time series

Coherent probabilistic forecasting, which involves predicting multivariate time series with hierarchical aggregation, is vital for accurate decision-making in industrial applications. Recent advancements have emphasized end-to-end approaches that simultaneously learn from all-time series in the hierarchy while integrating the reconciliation step into a single trainable model. However, prior studies often neglect the distinct characteristics of time series at different hierarchical levels. For instance, time series at the lower hierarchy levels are typically sparse and lack the high-level patterns evident in aggregated levels. Notably, in a hierarchical time series, the aggregated series (parent) is the sum of its lower-level series (child), thereby retaining intrinsic time series properties. Motivated by these observations, we propose hierarchical time series forecasting through Hierarchy-Aware Memory Network (HAMN). HAMN employs a hierarchy-aware memory module that stores representations of aggregated series at each hierarchical level in an external memory module and retrieves these representations to augment features of the sparse lower-level series. Evaluation across four public datasets demonstrates improvements of 6.4 %-20.9 % over state-of-the-art baselines in terms of the CRPS metric.