Predicting Enterprise Contract Renewal After a Quality Incident

Context

A video-conferencing provider experienced a spike in call disconnects. You need to design a practical modeling approach to predict whether an enterprise customer will renew at their next contract decision point so the company can prioritize retention actions.

Task

Propose a modeling and decisioning plan that covers the following:

1. Problem Framing
   • Define the label (renewal within the next contract period) and the prediction horizon (when predictions are made relative to renewal date).
2. Features
   • Specify and precisely define features such as:
     • Disconnection rate per 1,000 minutes
     • Percent of meetings affected
     • Time-to-resolution (e.g., MTTR)
     • SLA breaches
     • Support ticket volume
     • NPS/CSAT
     • Active seats and usage intensity
     • Industry
     • Account tenure
     • Price/discounts
3. Modeling Choices
   • Discuss when logistic regression is preferable to gradient-boosted trees or deep models (interpretability, small-N, linear signal, sparse features, latency constraints).
   • Discuss when more complex models are justified.
   • Optionally consider monotonic constraints.
4. Data Hygiene
   • Prevent leakage (e.g., avoid features created post-renewal or after the prediction cutoff).
   • Address class imbalance.
   • Consider survival analysis vs. binary classification for censored data.
5. Evaluation
   • Go beyond AUROC: include PR-AUC for rare churn, calibration curves/Brier score, cohort- and time-split backtests, and stability under distribution shift (pre/post incident).
   • Compare rank-based actioning vs. calibrated thresholding.
6. Decisioning and Actions
   • Translate scores into actions with cost–benefit thresholds, expected ROI, and capacity constraints.
7. Minimal Viable Feature Set (MVFS)
   • Provide a minimal feature subset and justify why it’s sufficient to launch.