Customer Churn Analysis

Problem

Correlation vs. actual prediction

Churn was already known: 28.9% of customers, $20.5M in lost lifetime value. What wasn't known was which customers were about to leave, or which factors actually predicted it rather than just correlated with it by coincidence.

What the analysis needed to find: which behavioral signals reliably predict churn, and can they be turned into a segmentation model precise enough to prioritize a retention budget?

Method

Dataset overview and cleanup

50,000 customer records, 25 behavioral, transactional, and demographic features, spanning 8 countries and 4 signup cohorts (Q1 through Q4).

Before analysis, the dataset needed real cleanup: nearly 49,000 missing values scattered across columns, an age field topping out at 200, and negative purchase amounts. Each issue got a documented rule rather than a blanket drop, since missingness wasn't random and dropping rows would have quietly skewed the churn rate itself.

Ruling out geography and time

Churn held at 27–30% across every one of the 8 countries and every signup quarter. Customers with 3+ years of tenure churned at about the same rate (29%) as new signups. A regional, seasonal, or loyalty problem would have shown up as variation here — it didn't, so those were ruled out.

Looking at behavior instead

Three engineered features carried the signal: cart abandonment rate, engagement score (a weighted composite of login frequency, session duration, and wishlist activity, normalized 0–1), and purchase recency (days since last purchase).

Building the risk segments

These three features fed a rule-based Risk Segment (Low, Medium, High), cross-tabulated against Customer Tier (Low, Mid, High Value, by lifetime-value percentile), then validated in PostgreSQL across three joined tables (customers, engagement_metrics, churn_summary) with eight analysis queries.

SELECT c.customer_id,
       c.country,
       c.customer_tier,
       e.risk_segment,
       ROUND(c.lifetime_value::NUMERIC, 2) AS lifetime_value,
       ROUND(e.engagement_score::NUMERIC, 4) AS engagement_score,
       ROUND(e.cart_abandonment_rate::NUMERIC, 4) AS cart_abandonment_rate,
       c.days_since_last_purchase
FROM customers c
JOIN engagement_metrics e ON c.customer_id = e.customer_id
WHERE c.customer_tier = 'High Value'
  AND e.risk_segment = 'High Risk'
  AND c.churned = 0
ORDER BY c.lifetime_value DESC
LIMIT 20

Findings

# Engagement Score: weighted composite, login frequency matters most
engagement_scaled = MinMaxScaler().fit_transform(df[['Login_Frequency', 'Session_Duration_Avg', 'Wishlist_Items']].fillna(0))
df['Engagement_Score'] = (0.4 * engagement_scaled[:, 0] + 0.35 * engagement_scaled[:, 1] + 0.25 * engagement_scaled[:, 2]).round(4)
 
# Risk Score: 3 binary flags on cart abandonment, engagement, and recency
df['Cart_Risk'] = (df['Cart_Abandonment_Rate'] > df['Cart_Abandonment_Rate'].quantile(0.66)).astype(int)
df['Engagement_Risk'] = (df['Engagement_Score'] < df['Engagement_Score'].quantile(0.33)).astype(int)
df['Recency_Risk'] = (df['Days_Since_Last_Purchase'] > df['Days_Since_Last_Purchase'].quantile(0.66)).astype(int)
df['Risk_Score'] = df['Cart_Risk'] + df['Engagement_Risk'] + df['Recency_Risk']
 
def assign_risk(score):
    if score >= 2: return 'High Risk'
    elif score == 1: return 'Medium Risk'
    else: return 'Low Risk'
 
df['Risk_Segment'] = df['Risk_Score'].apply(assign_risk)

Recommendations

Five prioritized actions, each sized against the data.

Limitations and Next Steps

The dataset is a snapshot, not a live feed, so the risk model needs to be re-run monthly to stay accurate. It also has no campaign cost data, so the ROI estimates lean on industry-standard conversion benchmarks rather than this business's actual numbers.

Correlation isn't causation, either: cart abandonment predicts churn, but the dataset can't say why customers are abandoning carts. The natural next step is a controlled test, holding an intervention against a control group of High Risk customers, to move from "this predicts churn" to "this fixes it."

Takeaways