Behavioral propensity
The propensity model is a gradient-boosted ensemble trained on tens of millions of historical conversion events across the platform's industry portfolio. Inputs are decay-weighted behavioral features, demographic context, psychographic overlay (where licensed), and identity-graph confidence. The model is retrained on a rolling 90-day window, with separate per-industry heads to capture vertical-specific decision dynamics. Outputs are raw probability estimates in [0,1] which are then calibrated.
Identity graph confidence
Behavioral signals are useless without high-confidence linkage to the right individual or household. The identity graph uses a probabilistic record-linkage approach (Fellegi-Sunter framework, panel-calibrated) over devices, addresses, hashed contact records, and behavioral fingerprints. Every linkage carries an explicit confidence value; downstream scoring weights signal contribution by linkage confidence. The graph's accuracy is benchmarked quarterly against a labeled ground-truth panel at 97.4% as of the v4.7 release.
Decay-aware signal weighting
Behavioral signal value decays non-linearly from the moment of observation. For high-velocity categories (luxury travel, automotive), half-life can be 48–72 hours; for B2B SaaS purchase intent, 14 days. Decay curves are fit per (industry × signal class) using a Hawkes-process formulation that captures self-excitation: a new signal of the same class extends the usable life of prior signals from the same individual. Stale signals are not discarded — they are down-weighted and retained for retrospective modeling.
Demographic and psychographic overlay
Behavior alone is not enough for high-consideration verticals. Demographic features (household composition, wealth indicators, life-stage) and psychographic overlays (values, attitudes, motivations) materially improve probability calibration, especially in healthcare, finance, luxury travel, and political. Overlays are merged at score time, not at training time, so they can be enabled per customer based on contractual data access.
Calibration
A raw 0.7 model output is not the same as a calibrated 70% probability. The platform applies isotonic regression on a held-out panel per industry and publishes calibration curves on request. Calibration quality is monitored by Brier score; drift triggers re-calibration without full model retraining. Every score that ships to customers carries a confidence interval reflecting both model uncertainty and identity-graph linkage uncertainty.
Model governance
Models are versioned (v4.7 in production at time of writing). Enterprise customers receive quarterly model governance reviews covering: model lineage, training data composition, calibration drift, signal-class contribution, and fairness analyses where applicable. No model is promoted to production without passing a panel calibration regression and an adversarial signal-integrity audit.
Signal half-life — production model
Predictive cohort vs. cold list
Citations
- · Fellegi, I., & Sunter, A. — A Theory for Record Linkage. JASA, 1969.
- · Liniger, T. — Multivariate Hawkes processes for self-exciting event data. ETH Zürich, 2009.
- · Brier, G. W. — Verification of forecasts expressed in terms of probability. Monthly Weather Review, 1950.
- · Niculescu-Mizil, A., & Caruana, R. — Predicting good probabilities with supervised learning. ICML, 2005.
- · Kahneman, D., & Tversky, A. — Prospect theory: An analysis of decision under risk. Econometrica, 1979.