The people behind the numbers.
Customer lifetime value, segmentation, churn, propensity, and forecasting — in one model that knows which customers your media is actually reaching. Reconciled to MMM elasticities so customer-level forecasts and channel-level decompositions tell a coherent story.
One graph, four modelling layers.
BTYD-family models — Pareto/NBD, BG/NBD with Gamma-Gamma — implemented via PyMC-Marketing. Bayesian uncertainty, hierarchical pooling across segments, the canonical method since 1987.
HDBSCAN as default — density-based, doesn't require choosing k, robust to non-spherical clusters. RFM as the interpretable floor. Longitudinal segmentation tracks customers across cohorts.
LightGBM workhorse classifiers with isotonic / Platt calibration. Reliability diagrams shown by default — the calibration matters more than the AUC. Survival analysis when time-to-churn matters.
Why is this customer at risk? Surface the specific behaviours and product relationships that drove the score. Not a black-box prediction — an auditable inference.
Forward-looking scenario simulation reconciled to MMM elasticities via MinT (Hyndman et al. 2011). When you change the media plan, the customer-level forecast updates coherently.
Send a segment to MMM as a covariate, to Experiments as an audience definition, or out via the public API for activation. Same segment object, every direction.
Canonical methods, current implementations.
The methods are mostly four decades old. The implementations are not.