Context & Problem

The company was seeing rising churn but had no clear visibility into why customers were leaving. Revenue was being lost, but the root causes were buried in fragmented data across billing, support, and usage systems.

Without a structured way to identify at-risk customers early, the retention team was playing whack-a-mole - reacting to cancellations instead of preventing them.

Research & Discovery

• Audited existing data sources: billing records, support tickets, feature usage logs, NPS scores
• Interviewed customer success team to understand qualitative churn signals they observed
• Mapped the customer lifecycle to identify critical "danger zones" where churn risk spikes

Solution & Approach

1. Data Consolidation

Unified data from 3 sources (billing, support, usage) into a single analysis-ready dataset using SQL joins and Python preprocessing.

2. Exploratory Data Analysis

Identified statistical patterns: tenure vs. churn correlation, support ticket frequency as a leading indicator, feature adoption depths.

3. Customer Segmentation

Built customer segments based on behavior patterns. High-risk segment: low feature adoption + high support contact + first 90 days.

4. Dashboard & Strategy

Power BI dashboards with real-time KPIs, segment health scores, and automated alerts for high-risk accounts.

Implementation

The analysis pipeline prioritized speed-to-insight over complexity:

# Customer risk scoring approach
import pandas as pd
from sklearn.preprocessing import StandardScaler
Feature engineering for churn prediction
features = ['tenure_days', 'support_tickets_30d',
            'feature_adoption_score', 'days_since_last_login',
            'billing_issues_count']
Normalize and weight features
scaler = StandardScaler()
risk_scores = scaler.fit_transform(df[features])
Segment: High risk = top 20% risk score + tenure < 90 days
df['risk_segment'] = np.where(
    (df['risk_score'] > df['risk_score'].quantile(0.8)) &
    (df['tenure_days'] < 90),
    'High Risk', 'Standard'
)

Outcome & Metrics

• Before: No Visibility - No way to identify at-risk customers
• After: ~15% - Potential churn reduction
• Dashboards: 3 - KPI, Segments, Alerts

Learnings

What Worked

Starting with qualitative input from the CS team before touching data ensured I was looking for the right signals. The best features in the model came from their observations, not from statistical fishing.

What I'd Change

Would build a proper ML prediction model rather than a rule-based risk score. The heuristic approach was fast but ceiling-limited. A gradient boosted model could capture non-linear churn patterns.