Propensity Modeling (Ecommerce)

Propensity modeling in ecommerce is the use of supervised machine learning to assign each customer (and often each session or each order) a probability score for some future behavior of interest. The classic case is propensity-to-purchase, used for ad targeting. The margin-relevant cases are different: propensity-to-stack-coupons, propensity-to-abuse-returns, propensity-to-charge-back, propensity-to-churn-after-discount. These models are typically trained on historical data — past coupon redemption patterns, past return outcomes, past chargeback histories — and applied at a decision point where their score can shift policy. For example, a customer with high propensity-to-stack-coupons might receive a checkout-time policy that allows fewer concurrent discounts; a customer with low propensity-to-return might be approved for a margin-tight order that a high-propensity customer would be flagged for. The technical foundation is well-understood: gradient-boosted trees (XGBoost, LightGBM) work well at the typical commerce scale; deep models (transformers, GNNs) add value at enterprise scale where customer-graph signals matter. The harder problem is operations: training the model is straightforward; integrating it into a real-time checkout decision requires sub-100ms scoring latency, a feature-engineering pipeline that runs in real time on customer state, model-drift monitoring, and robust A/B-testing infrastructure to validate that policy differentiation actually helps margin without depressing conversion. For most mid-market merchants, propensity modeling is too complex to build in-house and the operational requirements push it into specialized platforms. The pattern that delivers value at mid-market scale is propensity-LITE: simple rules-of-thumb derived from the same historical data (customers who have stacked coupons in past 90 days are flagged; first-time customers without verified email are flagged), implemented in the policy engine. This captures 60-70% of the margin lift of full propensity modeling at 5% of the operational cost. Propensity modeling proper becomes worthwhile at the enterprise tier ($100M+ GMV) where the operational investment can be amortized over enough order volume.