Case Study

Glass Box Discovery in Business Intelligence

How Lumenais found a hidden risk cohort in 7,000 customers that standard AI models missed—and explained why.

4 cycles to converge
83.6% AUC
7,043 customers

The Problem

Predicting customer churn is easy; standard models like Random Forest do it every day. The problem is actionability.

A "Black Box" AI tells you who will leave (Risk Score: 0.89). It doesn't tell you why. If you don't know the why, you can't fix the product. You're just fighting symptoms.

We fed the classic IBM Telco Customer Churn dataset to Lumenais with a single objective: "Identify counter-intuitive drivers of churn that contradict standard pricing logic."

The Discovery

Objective: Find non-linear risk factors

Method: UFCT Learning Loop (Automated Feature Engineering)

Standard logic says: "Long-term customers are safe."

Lumenais proved this wrong. It autonomously engineered a new feature—TotalCharges_x_tenure—and found it was the single strongest predictor of churn (Coefficient: +0.886).

"Long-term customers with high lifetime spend are actually HIGHER risk than linear models predict. They hit a 'value ceiling'—a point where their high investment leads to higher expectations. If those expectations aren't met, they churn faster than new users."

This is a "Glass Box" discovery. It gives the business a specific lever: audit the experience of high-LTV vintage customers immediately. They are not safe; they are fragile.

Hidden Risk89% conf

The Loyalty Ceiling

Standard models assume 'High Tenure = Low Risk.' Lumenais discovered a non-linear interaction (TotalCharges × Tenure) revealing that long-term customers with high lifetime spend are actually HIGHER risk than expected. They hit a 'value ceiling' where loyalty breaks if service doesn't match their premium investment.

Key Driver86% conf

Price Sensitivity Inversion

For Fiber Optic users, 'MonthlyCharges' is a strong churn driver (+0.86 coefficient). However, the presence of 'TechSupport' acts as a value-anchor that significantly dampens this effect. High price is tolerated ONLY when accompanied by high-touch support.

Opportunity82% conf

The "Stuck" Senior Cohort

Senior Citizens on 1-year contracts show artificially low churn but high monthly charges. This isn't loyalty; it's inertia. This cohort is highly vulnerable to competitive offers that simplify billing, representing a proactive retention opportunity.

Validated Metric91% conf

Electronic Check Friction

PaymentMethod (specifically Electronic Check) emerged as a surprisingly strong independent predictor of churn, even after controlling for contract type. This suggests a user experience friction point in the payment process itself, separate from pricing or service quality.

Automated Science

Lumenais didn't just run a regression. It engineered its own features to find the signal. Here are the top predictors it discovered automatically:

PredictorCoefficientMeaning
TotalCharges × Tenure+0.886The "Loyalty Ceiling" Risk
MonthlyCharges+0.860Price Sensitivity
Tenure (Raw)-0.841Base Retention Effect
Contract (Long-term)-0.812Lock-in Effect

Note: The system converged in 4 cycles with an AUC of 0.836, outperforming the linear baseline.

Business Impact

Precision Retention

Stop spamming all "at-risk" users. Target the High-LTV / High-Tenure cohort with "VIP Service" offers, not just discounts. Discounts insult them; service saves them.

Pricing Strategy

Bundle Tech Support with high-tier Fiber plans by default. The data proves it's not an add-on; it's a churn vaccine for high-paying users.

Run Your Own Analysis

Upload your customer data. Find the hidden drivers of revenue and risk in minutes.