Econometrics Seminar Series - Mitchell O’Hara-Wild (Monash University)

Title: Graphs for forecasting many related time series

Abstract: Forecasting in practice rarely involves a single time series. Organisations forecast large collections of related series, such as product sales across stores and regions, which requires software that handles many series and models consistently and produces probabilistic forecasts that can be evaluated, combined, and compared. The fable R package addresses this by treating forecasting as a workflow of composable stages designed with tidy data analysis principles. The relationships between related time series are commonly encoded with aggregation constraint matrices, which forecast reconciliation uses to produce coherent and more accurate forecasts. Directed acyclic graphs (DAGs) offer a more general representation that make large collections of series easier to explore and manipulate. As series are disaggregated by many attributes, their number grows rapidly, increasing computational cost and degrading accuracy through noisy, uninformative, or misspecified series. Using time series features and graph search algorithms, uninformative subgraphs can be efficiently pruned while preserving coherency, substantially reducing the number of series to forecast while retaining most of the information. This seminar begins with the software design of fable for forecasting at scale, then introduces graph-based forecast reconciliation and feature-based graph pruning for accurate forecasting at scale.