Predicting Publisher Churn

The problem

Indiegraf serves 180+ independent publishers on a subscription model. Losing a publisher is costly, and without a system to flag risk early, the pattern only becomes visible after it’s too late to act.

Approach

• Data: account-level usage and behavioral signals collected across the publisher base
• Model: Random Forest, chosen for its ability to handle mixed feature types and surface which signals matter most, without requiring heavy feature engineering upfront
• Scope: 250 publisher accounts

Tools

Python, pandas, scikit-learn for modeling and evaluation. Feature exploration and validation in Jupyter notebooks.

Result

• AUC: 0.81 — in plain terms, given one at-risk account and one healthy account, the model correctly ranks the at-risk one higher about 81% of the time
• Behavioral signals (engagement patterns, usage trends) were the strongest predictors, stronger than firmographic data like publisher size or tenure
• Data quality gaps capped accuracy further. That’s a finding on its own: it points to specific gaps worth closing before the next iteration, not just a limitation to note in passing

What this demonstrates

Going from a business problem to a working, evaluated model. Not the algorithm choice on its own, but the judgment behind what to build, what data to trust, and what the result actually means for the business.