Slides: slides.mitchel...
Source materials: github.com/mit...
Large collections of related time series are commonly structured with aggregation constraints, whereby each series possesses various attributes that identify their relation to other series. These attributes typically relate to what is being measured, such as product categories or store locations for the sales of a product over time. When there exists many attributes for time series data, the number of series in the collection quickly becomes unmanageable with disproportionately many uninformative disaggregated series. This presents many problems for forecasting, since producing many forecasts can be computationally infeasible and the forecast accuracy for aggregated series of interest can worsen.
To overcome these problems I propose using time series features to identify noisy, uninformative, or otherwise unwanted series and leveraging the graph structure from topic 1 to safely remove them while preserving coherency constraints. Pruning series from the bottom of the structure would result in graph coherency constraints since a common bottom level is no longer present. Various control points are possible, including specification of features, thresholds, and coherent pruning rules to produce a reduced set of coherent series for forecasting. Pruning subgraphs of time series from the collection can substantially reduce the number of series to forecast, while retaining most of the information. This helps limit the computational complexity of forecasting, while improving forecast accuracy for aggregated series due to reduced model misspecification in more disaggregated series.
Негізгі бет ISF2024: Feature based graph pruning for improved forecast reconciliation
Пікірлер