Lilly Tech Systems
Data & Features for Churn Prediction
The best churn models are built on rich, multi-dimensional data. Learn which behavioral signals, usage patterns, and customer health indicators are most predictive of attrition and how to engineer features that capture early warning signs.
Key Feature Categories
| Category | Examples | Predictive Power |
|---|---|---|
| Usage Patterns | Login frequency, feature adoption, session duration | Very High |
| Engagement Trends | Week-over-week usage change, declining activity | Very High |
| Support Interactions | Ticket volume, complaint severity, resolution satisfaction | High |
| Contract Details | Contract length, renewal date proximity, payment history | High |
| Customer Profile | Industry, company size, plan type, customer tenure | Medium |
| Sentiment Signals | NPS scores, survey responses, social media mentions | Medium |
Feature Engineering for Churn
- Trend Features: Calculate the slope of key metrics over 7, 14, 30, and 90 day windows. A declining login trend is far more predictive than a single snapshot of login count.
- Ratio Features: Compare current period usage to historical averages (e.g., this month's logins vs. average monthly logins). Values below 0.5 are strong churn signals.
- Recency Features: Days since last login, last purchase, last support interaction, and last feature usage. Increasing recency is a key early indicator.
- Engagement Depth: Track breadth of feature usage, not just frequency. Customers using only 1-2 features are more at risk than those deeply integrated across the product.
- Support Sentiment: Apply NLP to support tickets to extract frustration levels, competitive mentions, and cancellation intent signals.
The Customer Health Score
Healthy (80-100)
Active usage, growing engagement, positive support interactions, on-time payments, and expanding feature adoption. Low intervention needed.
At Risk (50-79)
Declining engagement trends, reduced feature usage, or recent negative support experiences. Proactive outreach recommended.
Warning (25-49)
Significant usage drops, multiple support complaints, missed payments, or competitive evaluation signals. Urgent intervention required.
Critical (0-24)
Near-zero engagement, cancellation inquiries, or explicit dissatisfaction signals. Executive-level save attempt may be warranted.