Time series and forecasting · insight-1233
Let a damped trend compete with a linear trend whenever automatic forecasts run beyond a few periods
- Evidence type
- documented fact
- Difficulty
- easy
- Roles
- Data scientist, Data analyst
- Industries
- retail, aviation, cross-industry
Summary: situation, action, reason, boundary
Situation
A monthly forecast for 1,000 route-level series is produced automatically, and a linear-trend Holt model projects 24 months of continued growth on routes that recently grew.
Action
Include damped-trend variants among the candidates, estimate the damping parameter with the other parameters, check that it is not pinned at a boundary, and compare candidates at the longest horizon over rolling origins.
Reason
FPP3 notes that Holt's linear method tends to over-forecast at long horizons and that damped methods have been very successful, being arguably the most popular single methods for automatic forecasting of many series.
Boundary
Large collections of series with moderate trends where nobody inspects each fit.
Symptom or risk
A constant trend runs on indefinitely, so long-horizon forecasts for routes whose growth is slowing are systematically too high.
Exceptions
Series with structurally constant growth, such as a scheduled build-out, may favour an undamped trend.
Trade-offs
Damping can under-forecast growth that really does continue.
Pitfalls
Fixing the damping value by hand to make the flattening visible, or accepting a value so close to one that it is indistinguishable from no damping.
Lessons and concepts
Claims and their sources
FPP3 reports that Holt's linear trend method tends to over-forecast, especially at longer horizons, and that damped trend methods have proved very successful, arguably the most popular individual methods for automatic forecasting across many series.