Time series and forecasting · insight-1241
Treat cross-classified series as grouped, not as an arbitrary single tree
- Evidence type
- documented fact
- Difficulty
- medium
- Roles
- Data scientist, Data analyst
- Industries
- aviation, retail, logistics
Summary: situation, action, reason, boundary
Situation
Seats sold are available by route, cabin and channel, and an analyst forces them into route, then cabin, then channel to use a hierarchical method, although channel totals across routes are also reported.
Action
Define the bottom-level series and a summing matrix that includes every aggregate that is reported, then reconcile with a method that handles several aggregation paths. Evaluate accuracy at each reported aggregate after reconciliation.
Reason
Attributes that cross rather than nest have no unique tree. FPP3 describes such grouped series as having several sets of aggregation constraints that must hold at once.
Boundary
Data cut by several attributes such as route and cabin, product and region, or store and channel.
Symptom or risk
Only one of the aggregates that stakeholders use is made coherent; the others still disagree with the bottom level.
Exceptions
If only one aggregation path feeds decisions, a plain hierarchy is enough.
Trade-offs
More aggregates mean a larger summing matrix and more series to reconcile.
Pitfalls
Reporting aggregates outside the chosen tree by summing the bottom level, which hides incoherence.
Example
Hypothetical: 3 routes crossed with 2 cabins give 6 bottom series and 12 series in total (1 network, 3 route, 2 cabin, 6 bottom).
Lessons and concepts
Claims and their sources
FPP3 describes grouped time series as attributes that cross rather than nest, with no unique hierarchy, so the summing matrix must encode several sets of aggregation constraints at once.
The HierarchicalForecast library lists bottom-up, top-down, middle-out and MinTrace reconciliation among its methods.