Stripe·Data Scientist·Technical Phone Screen·Senior

SeniorPrefer not to say

Jun 2026Remote

Summary

Stripe DS interview with a single deep-dive case on predicting merchant subscription adoption. The whole thing was one long multi-part ML design question and it went places I did not fully anticipate.

Questions Asked (1)

Design a production-ready classification system to predict which non-subscription merchants will adopt Stripe's Subscription product within the next 60 days, covering labeling strategy, leakage prevention, feature engineering, modeling choices, evaluation, and post-deployment monitoring.

System DesignProduct Analytics & MetricsTechnical Trade-offs

Author's notes

This was basically six questions wrapped in one and I did not realize how deep it went until I was already three sub-parts in.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business objective and success metrics, then walk through the ML lifecycle: labeling, feature engineering with strict leakage prevention, model selection, evaluation, and monitoring. Emphasize production readiness by discussing data pipelines, retraining, and how to handle class imbalance and temporal drift.

Pro tip: Use a time-based split for validation and test sets to mimic real deployment, and consider using survival analysis or a sliding window approach to handle the 60-day horizon and censoring. Also, align the model's output with business actions (e.g., targeted outreach) to demonstrate impact.

1. Define Objective and Labeling Strategy

Clarify the prediction goal: identify non-subscription merchants likely to adopt Subscriptions within 60 days. Define the label as whether a merchant adopts within 60 days of a reference date, and handle censoring by excluding merchants with insufficient observation time or using survival analysis.

2. Feature Engineering and Leakage Prevention

Engineer features from merchant attributes, transaction history, product usage, and engagement metrics. Prevent leakage by ensuring all features are computed using only data available before the prediction date, and avoid using future information or target-derived features.

3. Model Selection and Training

Choose models robust to class imbalance and temporal effects, such as gradient boosted trees (XGBoost, LightGBM) or logistic regression with regularization. Use time-based cross-validation and consider survival models if censoring is significant.

4. Evaluation and Threshold Selection

Evaluate using metrics like AUC-ROC, precision-recall AUC, and lift at top-k. Select an operating threshold based on business costs (e.g., cost of outreach vs. conversion value) and validate on a holdout set from a later time period.

5. Deployment and Monitoring

Deploy the model with a pipeline for batch or real-time scoring. Monitor input drift, output drift, and performance metrics (e.g., precision@k) over time. Set up alerts for degradation and plan for periodic retraining with new data.

Key Points to Mention

Temporal validation to prevent leakage and account for concept drift
Handling class imbalance with techniques like SMOTE, class weights, or focal loss
Feature importance and interpretability for business stakeholders
Business impact metrics (e.g., incremental conversions, ROI) and how to measure them
Monitoring data quality, feature drift, and model performance post-deployment
Retraining strategy and feedback loops to incorporate new adoption data

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.