Explanation: Assisted vs Exogenous Baseline

A core AMMM design choice is how to represent baseline dynamics (trend and seasonality) in MMM.

Two Baseline Modes

1) Exogenous baseline (fully endogenous estimation)

Baseline terms are engineered from time index and external controls (for example Fourier harmonics and holiday indicators), then estimated jointly with media effects.

Benefits:

• no target-derived regressor leakage,
• clearer separation between structural assumptions and observed outcome.

Risk:

• higher collinearity in finite samples,
• weaker numerical behaviour when trend/seasonality bases overlap with media patterns.

2) Assisted baseline (Prophet-derived components)

AMMM can run Prophet decomposition and inject components (for example trend, yearly, weekly, holidays) into controls.

Configuration keys include:

• prophet.trend
• prophet.yearly_seasonality
• prophet.weekly_seasonality
• prophet.include_holidays

Benefits:

• practical stabilisation of baseline representation,
• reduced pressure from high-VIF explicit seasonal bases,
• improved operational robustness in many real business datasets.

Risk:

• prophet.trend: true introduces target-derived trend information into regressors,
• this can attenuate channel coefficients and shift attribution towards baseline,
• interpretation should therefore be explicitly associational.

AMMM Position

AMMM treats both modes as legitimate when assumptions are documented and sensitivity is performed.

Recommended protocol:

1. Fit with prophet.trend: true (assisted baseline).
2. Fit with prophet.trend: false (seasonality/holidays without assisted trend).
3. Compare diagnostics and downstream conclusions (contributions, ROI rankings, optimisation recommendations).
4. If conclusions are highly unstable, report the instability rather than forcing a single narrative.

Reporting Standard

When assisted baseline is used:

• state the choice explicitly,
• state that attribution is associational under this specification,
• avoid causal language unless separate identification evidence exists.

Practical Failure Signals

Assisted baseline may be unnecessary or harmful if:

• no meaningful trend is visible,
• trend component dominates decomposition with implausible media suppression,
• calibration and residual diagnostics worsen.

In such cases, prefer exogenous baseline specification and re-check diagnostics.

Relationship to Workflow

This baseline decision should be treated as an upstream modelling choice that influences every downstream stage, especially:

• 40_decomposition/
• 60_response_curves/
• 70_optimisation/
• 80_interpretation/

See Methodology and Workflow Stages for gate context.

Causal Caveat

Neither baseline mode guarantees causal identification. The choice controls bias–variance and interpretability trade-offs under model assumptions.