Time series and forecasting · insight-1250
Include lagged promotion or advertising effects and pick the lag count by AICc on a common sample
- Evidence type
- documented fact
- Difficulty
- medium
- Roles
- Data scientist, Data analyst
- Industries
- retail, e-commerce, telecom
Summary: situation, action, reason, boundary
Situation
Weekly sales respond to a campaign over several weeks, but the model contains only this week's campaign flag.
Action
Add the current value and a few lags of the campaign-intensity variable in a regression with ARIMA errors, select the lag count and error orders by AICc, drop the first observations from every candidate so all use the same sample, then refit the winner on everything.
Reason
Advertising can influence sales for months after a campaign, and AICc comparisons are fair only when every model is fitted to the same data.
Boundary
Promotions, advertising bursts and price changes whose effect is spread over time.
Symptom or risk
Effects that arrive later are attributed to trend or noise, the immediate effect is overstated, and the post-campaign weeks look like unexplained misses.
Exceptions
If successive campaigns overlap heavily, lagged terms are almost collinear and effects are not separable.
Trade-offs
Each extra lag costs an estimated coefficient; use a small candidate set.
Pitfalls
Comparing AICc across candidates fitted to different numbers of observations.
Lessons and concepts
Claims and their sources
FPP3 includes the current and several lagged values of advertising as predictors in a regression with ARIMA errors, selects the lag count by AICc, and removes the first observations from all candidates so they are compared on the same estimation period.