Problem
The Times funnel has unregistered, registered and subscribed users. Each request from a non-subscriber can end in a registration wall, a paywall or free access. Each choice moves several KPIs in different directions: subscriptions, registrations, engagement and bounce rate. Blocking too much hurts content discovery and SEO. The previous monthly batch model (Dynamic Meter) set a free-article quota per registered user and could not use real-time or article data.
Example from the article: a user who will probably register but not subscribe gets a registration wall. A user unlikely to do either but likely to engage gets free access.
Solution in brief
Train one causal predictor per KPI on always-on randomized data, with the action as a feature. At request time, combine the predicted KPIs per action with weights and pick the best action. Refit the weights daily by backtesting on recent randomized data so wall rates match business targets.
Questions this case answers
- How to choose one of several actions per request in real time when several competing objectives must be traded off under fixed action-rate limits
- How to keep a causal prescriptive policy within business-set action rates as traffic shifts day to day
- How to decide per request whether to show a registration wall, a paywall or free access?
- How to balance subscription, registration and engagement in one decision policy?
Tools and methods
- S-learner on RCT data — one supervised predictor per objective, with the action as a feature, gives counterfactual predictions per action.
- Weighted multi-objective policy — a convex combination of the predicted objectives picks the action.
- U-NSGA-III — derivative-free optimizer that finds the weights matching target action rates.
- Inverse probability weighting (Hájek estimation) — forecasts the effect of changing a business target.
- Always-on RCT and A/B tests — benchmark for the policy and validation of model changes.
Google or a one-time email code. No password.