Predictive Analytics Template for PMs

Not necessarily. Simple models (logistic regression) can be built by an analyst with SQL and Python experience. Many product analytics tools (Amplitude, Mixpanel) now offer built-in predictive features that require no coding. The [churn prediction template](/templates/churn-prediction-template) provides a rules-based alternative that any PM can implement without a data scientist. However, for complex models with many features and large datasets, a data scientist significantly improves model quality and reliability.

At minimum, 6 months of data with at least 500 positive examples (e.g., 500 users who actually upgraded). More data improves model stability. For rare events (base rate <2%), you may need 12-18 months of data to get enough positive examples. The key constraint is not total data volume but the number of positive examples in your training set. A model trained on 50 positive examples will be unreliable regardless of how many negative examples you have.

A/B testing measures the causal impact of a specific change. Predictive analytics forecasts outcomes based on observed patterns without necessarily establishing causation. They are complementary: use predictive analytics to identify who to target, then A/B test the intervention to measure its causal effect. For example, a propensity model identifies users likely to upgrade, and an A/B test measures whether a targeted email actually increases their upgrade rate.

Focus on three things: (1) What the model predicts and how accurately (in business terms, not statistical jargon). (2) The top 3-5 features driving predictions (e.g., "Users who hit the paywall 5+ times and invited team members are 4x more likely to upgrade"). (3) The action we take based on predictions and the expected business impact. Skip the algorithm details unless asked. Stakeholders care about "does it work?" and "what do we do with it?" not "how does gradient boosting work?"

Retrain monthly if your product changes frequently (new features, pricing changes, market shifts). Retrain quarterly if your product and market are stable. Monitor the key metric (AUC-ROC or calibration accuracy) weekly and retrain immediately if it drops below your threshold. The most common reason for model degradation is "concept drift": the relationship between features and outcomes changes because of product or market changes.