What This Chart Shows
This chart shows how cumulative marketing lift builds over time and how certain that estimate is. The main line is the day-by-day sum of attributed lift. The shaded bands represent uncertainty: they widen on days where the model’s internal “reconciliation” is less precise (i.e., where daily attributed lift doesn’t match the observed KPI change as closely). You see both the level of attributed impact and where that estimate is more or less reliable over the period.
Key Questions This Chart Helps Answer
- How does cumulative marketing lift evolve day by day?
- On which days is attribution more or less precise?
- Is attribution uncertainty stable over time or does it spike in certain periods?
Axes, Metrics, and Units
Element | Description |
|---|---|
| X-axis (daily view) | Simulation Day: Day index 1 through T (number of days in the analysis window). |
| Y-axis (daily view) | Cumulative Marketing Lift (KPI units): Sum of daily attributed lift from day 1 to each day. |
| Solid line | Cumulative Lift: Day-by-day cumulative sum of daily KPI delta (Full KPI − Base KPI). |
| Light shaded band | 90% credible interval (P5–P95): ±1.645 σ per day, then cumulated. σ = daily reconciliation error. |
| Darker shaded band | 50% credible interval (P25–P75): ±0.675 σ per day, then cumulated. |
| Zero baseline | Horizontal dashed line labeled "Baseline (no marketing)" at zero cumulative lift. |
| High-precision note | When reconciliation error is < 1% of lift, an annotation explains that bands are floored at 5% so they remain visible. |
| X-axis (ensemble fallback) | Quantile: P5, P25, P50, P75, P95, Mean. |
| Y-axis (ensemble fallback) | Lift: Scenario-level lift (KPI units) for each quantile/mean. |
| Bars (ensemble fallback) | One bar per quantile/mean from sim_plan_ensemble_runs. |
All daily-view values are modeled.. Ensemble fallback shows estimated lift distribution across scenarios.
Control Options Reference
Control | Meaning |
|---|---|
| KPI Classification | Filter by KPI type so the chart uses the selected target(s). |
How to Interpret the Results
Daily fan chart:
- Solid line: Best estimate of cumulative lift; use for the expected trajectory.
- Narrow bands: Days where reconciliation is good; higher confidence in that day’s contribution.
- Wide bands: Days where the model fits the KPI less well; more uncertainty; worth checking data or one-off events.
- Band width over time: Widening in a period suggests more uncertainty; stable narrow bands suggest consistent, precise attribution.
- Zero baseline: Cumulative lift is incremental above “no marketing”; negative cumulative lift would mean the model attributes a net negative effect over that span.
- Conservative planning: Use the lower edge of the 90% band (P5) for downside scenarios; use the central line for the expected case.
Ensemble bar chart (when only ensemble data exists):
- Bars show the distribution of lift across scenarios (P5, P25, P50, P75, P95, Mean). Use for spread and central tendency when no daily HCF series is available.
Practical Applications for Marketers
Application | How to use this chart |
|---|---|
| Budget and planning | Use the central line for expected lift and the lower band for conservative forecasts. |
| Quality checks | Treat widening bands as a signal to check data quality, gaps, or one-off events in that period. |
| Stakeholder communication | Show that attribution comes with a range, not a single number, and that precision varies by day. |
| Period prioritization | Focus reporting and decisions on periods where bands are narrow (higher confidence). |
| Model trust | Consistently narrow bands indicate the HCF ledger reconciles well; wide bands indicate more uncertainty. |
Common Mistakes & Misinterpretations
Mistake | Why it is a problem | How to avoid |
|---|---|---|
| Treating the band as "error" in the sense of mistake | Bands reflect uncertainty (precision), not that the model is wrong. Narrow bands mean the model fits the data well on that day. | Interpret bands as "plausible range," not "error bars" in a causal sense. |
| Ignoring band width | Wide bands in a period mean less reliable attribution there; decisions based only on the line can be overconfident. | Always consider band width when judging reliability of lift in a given period. |
| Comparing raw cumulative lift across different date ranges or KPIs | Cumulative lift depends on length and KPI; comparing different windows or targets is misleading. | Compare within the same KPI and same window length, or use normalized/rate metrics. |
| Assuming bands are from repeated experiments | Bands are from the reconciliation error (model fit), not from A/B test or bootstrap. | Use for model-based uncertainty, not for experimental variability. |
| Over-reacting to a single wide day | One noisy day can widen the band; it may be data or one-off. | Look at patterns over several days before concluding a systematic issue. |
Caveats & Considerations
- Data dependency: Requires HCF daily outputs. If neither daily KPI nor ensemble data is available, the chart shows: "Uncertainty analysis requires HCF attribution data. Run the pipeline to compute daily KPI and reconciliation error."
- Reconciliation error as uncertainty: Bands are based on how well daily attributed lift explains the KPI change. They are a model-fit uncertainty proxy, not full causal uncertainty.
- Band floor: When reconciliation error is very small (< 1% of lift), sigma is floored at 5% of |delta| so bands stay visible; an on-chart note explains this.
- Cumulative nature: Bands are built from daily σ then cumulated; early wide days can make later cumulative bands wide even if later days are precise.
- KPI filtering: If you use KPI classification/hierarchy, the chart reflects the selected target(s) only; aggregate and per-target views can differ.