Marketing Mix Model — Causal Explainer Agent

An agentic system that interprets Marketing Mix Model outputs, selects appropriate causal methods (DiD, Synthetic Control, BSTS), and generates natural-language explanations of promotional ROI.

Context: Built to automate the explanation workflow for MMM results. Analysts spend significant time translating Bayesian posterior distributions into business recommendations — this agent handles the first pass.

flowchart TB
    U[User Query] --> R[Router Agent]
    R --> |ROI question| ROI[ROI Explainer]
    R --> |Causal question| CI[Causal Inference Engine]
    R --> |Budget question| BO[Budget Optimizer]
    R --> |Data question| SQL[Text-to-SQL Agent]
    ROI --> S[Synthesizer]
    CI --> S
    BO --> S
    SQL --> S
    S --> E[Executive Summary]

Loading

🎯 Problem

Marketing analysts need to answer causal questions about promotional effectiveness: "Did the digital campaign cause the sales lift, or was it seasonal?" Standard MMM provides correlational ROI, but stakeholders need causal reasoning with appropriate uncertainty quantification.

🧮 Mathematical Foundation

Causal Inference: Potential Outcomes Framework

$$\text{ATE} = \mathbb{E}[Y_i(1) - Y_i(0)] = \mathbb{E}[Y_i(1)] - \mathbb{E}[Y_i(0)]$$

where $Y_i(1)$ = outcome under treatment, $Y_i(0)$ = outcome under control.

Synthetic Control Method

For treated unit $j_0$ at time $T_0$:

$$\hat{Y}_{j_0,t}^{(0)} = \sum_{j \neq j_0} w_j Y_{j,t}, \quad \text{s.t.} \sum w_j = 1, w_j \geq 0$$

Treatment effect: $\tau_t = Y_{j_0,t} - \hat{Y}_{j_0,t}^{(0)}$ for $t > T_0$

Geo-Lift Analysis (Bayesian Structural Time Series)

$$y_t = \mu_t + x_t \beta + \epsilon_t, \quad \mu_{t+1} = \mu_t + \delta_t + \eta_t$$

Posterior predictive distribution for the counterfactual:

$$p(\tilde{y}_t \mid \text{data}) = \int p(\tilde{y}_t \mid \theta) p(\theta \mid \text{data}) , d\theta$$

Granger Causality (Time Series)

$$y_t = \alpha + \sum_{i=1}^{p} \beta_i y_{t-i} + \sum_{i=1}^{p} \gamma_i x_{t-i} + \epsilon_t$$

Test: $H_0: \gamma_1 = \gamma_2 = \ldots = \gamma_p = 0$ (x does not Granger-cause y)

DAG-Based Causal Reasoning

$$P(Y \mid \text{do}(X=x)) = \sum_z P(Y \mid X=x, Z=z) P(Z=z)$$

The backdoor adjustment formula — distinguishing $P(Y \mid X)$ (observation) from $P(Y \mid \text{do}(X))$ (intervention).

Information-Theoretic Channel Attribution

$$I(X_i; Y) = \sum_{x,y} p(x,y) \log \frac{p(x,y)}{p(x)p(y)}$$

Mutual information between channel spend $X_i$ and outcome $Y$ provides a non-parametric attribution measure.

🏥 Enterprise Pharma Application

This repo directly automates my causal inference workflow in enterprise pharma settings:

Enterprise Work	Agent Implementation
Geo-lift analysis for TV campaigns	CausalInferenceEngine with synthetic control
"Did email cause the Rx lift?"	DAG-based causal reasoning with do-calculus
Budget reallocation justification	Counterfactual simulation with CIs
PoC causal inference with NBE team	Bayesian structural time series
PyMC-Marketing MMM interpretation	ROI Explainer with posterior reasoning

Key insight: In enterprise pharma, I PoC'd geo-lift analysis applying causal inference methods. This agent automates the "explain why" step that analysts spend hours on.

🚀 Quickstart

git clone https://github.com/fab-admasu/mmm-causal-explainer-agent.git
cd mmm-causal-explainer-agent
pip install -r requirements.txt

# Generate synthetic causal scenarios
python scripts/generate_causal_scenarios.py --n_scenarios 200

# Fine-tune causal reasoning adapter
python scripts/train_causal_sft.py --config configs/causal_config.yaml

# Run agent demo
python scripts/run_agent.py --query "Did our TV campaign cause the 8% Rx lift for Cardivex?"

📊 Evaluation

Metric	Base LFM2.5	+ Causal SFT	+ Tool-Use
Causal reasoning accuracy	35%	72%	84%
Correct uncertainty hedging	28%	68%	81%
Appropriate method selection	22%	65%	78%
Actionable recommendations	45%	76%	88%
Hallucination rate	18%	5%	3%

License

MIT

📸 Visual Tour

---