CausalMMM: Learning Causal Structure for Marketing Mix Modeling

Online advertising relies on Marketing Mix Modeling (MMM) to predict Gross Merchandise Volume (GMV) and allocate budgets across channels, but traditional regression‑based MMM struggles with complex scenarios and assumes a fixed causal structure.

This work defines a new causal MMM problem that automatically discovers interpretable causal graphs from data and respects known marketing response patterns such as temporal decay and saturation.

The proposed CausalMMM model combines a Causal Relational Encoder, which uses a fully connected graph and Gumbel‑Softmax sampling to infer causal edges, with a Marketing Response Decoder that incorporates temporal and saturation modules to model carry‑over and diminishing returns effects.

Variational inference is employed to train the encoder‑decoder end‑to‑end, guaranteeing that learned causal relations correspond to Granger causality and scaling linearly with the number of stores.

Experiments on synthetic data with known causal graphs and real Alibaba advertising data compare CausalMMM against seven baselines, including Granger‑based causal discovery methods and traditional MMM models such as LSTM, Wide&Deep, and BTVC.

Results show that CausalMMM consistently outperforms baselines in causal structure learning (up to 5.7% improvement) and GMV prediction (best MSE at 7‑step horizon), and ablation studies confirm the contribution of each module.

A case study on a real beauty store reveals meaningful causal relationships between brand advertising, page views, and conversion channels, aligning with domain expert knowledge and marketing funnel theory.

In summary, CausalMMM advances MMM by jointly learning heterogeneous causal structures and realistic marketing response patterns, delivering superior predictive performance and actionable insights for advertisers.