How we built it: Smart Retries

Kiran Chandran23 Jan 2024

Problem

Twenty-five percent of lapsed subscriptions are lost purely because a payment failed (involuntary churn). Causes include insufficient funds, an expired card, new card details and technical faults. The best retry time depends on the cause: almost immediately for a technical fault, the next day for a low balance, about a week for an expired card. Many businesses skip the problem, often because building it in-house is slow and complex. Stripe says recovered subscriptions continue for seven more months on average.

Example from the article: a debit-card payment fails for insufficient funds. The retry window should then reach at least the customer's next payment period.

Solution in brief

Train a model on more than 500 customer, business, payment, calendar and billing attributes to predict the best retry time. Because the best retry is often days away, a slow stacked ensemble can replace a fast single model. Add customer-set retry rules and a default window from network data.

Questions this case answers
  • How to predict the best future moment to repeat a failed action when the right delay depends on the failure cause, the actor and the calendar
  • How to expose manual limits on an automated timing model while keeping already scheduled predictions stable
  • How to choose the best time to retry a failed subscription payment so that revenue is recovered
  • How to reduce involuntary churn caused by payment failures

Tools and methods

  • Stacked ensemble from Auto-ML — replaces the earlier XGBoost model; weak base models feed a stacker and ensemble selection
  • XGBoost — the previous production model, strong on tabular data
  • Sentence Transformers — embed text and numeric features into a shared space inside the stack
Continue reading — it's free

Google or a one-time email code. No password.