Mini Case Study — Customer Churn EDA

This case study runs the full EDA workflow on a synthetic customer dataset. Everything is generated in code — no external files needed. Work through each step yourself before reading the commentary.

The business question: Which customers are most at risk of churning, and what signals predict it?

---

The Dataset

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns

rng = np.random.default_rng(42)
n = 500

# Simulate customer data with realistic patterns
tenure      = rng.integers(1, 60, n)               # months as a customer
age         = rng.integers(20, 65, n)
segment     = rng.choice(["Basic", "Premium"], n, p=[0.65, 0.35])
region      = rng.choice(["North", "South", "East", "West"], n)
calls       = rng.integers(0, 12, n)               # support calls last 6 months
monthly_fee = np.where(segment == "Premium",
                        rng.uniform(800, 1500, n),
                        rng.uniform(200, 600, n))

# Churn probability depends on tenure, calls, and segment
churn_prob = (
    0.5
    - tenure * 0.008          # longer tenure → lower churn
    + calls  * 0.06            # more calls → higher churn
    - (segment == "Premium") * 0.15
    + rng.normal(0, 0.05, n)
)
churn_prob = np.clip(churn_prob, 0.02, 0.95)
churn = (rng.uniform(size=n) < churn_prob).astype(int)

df = pd.DataFrame({
    "customer_id":  range(1001, 1001 + n),
    "age":          age,
    "tenure_months": tenure,
    "segment":      segment,
    "region":       region,
    "support_calls": calls,
    "monthly_fee":  monthly_fee.round(2),
    "churn":        churn        # 1 = churned, 0 = retained
})

# Inject realistic messiness
df.loc[rng.choice(df.index, 18, replace=False), "age"]          = np.nan
df.loc[rng.choice(df.index,  9, replace=False), "support_calls"] = np.nan
df.loc[rng.choice(df.index,  4, replace=False), "region"]        = None

print(df.shape)
print(df.dtypes)

---

Step 1 — First Contact

print(df.head(8))
print(f"\nRows: {df.shape[0]}, Columns: {df.shape[1]}")
print(f"Churn rate: {df['churn'].mean():.1%}")

What each row represents: one customer. Target variable: churn (1 = left, 0 = stayed).

---

Step 2 — Data Quality Audit

# Missing values
missing = pd.DataFrame({
    "count": df.isna().sum(),
    "pct":   (df.isna().mean() * 100).round(1)
}).query("count > 0")
print(missing)
# Output:
#                count  pct
# age               18  3.6
# support_calls      9  1.8
# region             4  0.8

# Duplicates
print(f"Duplicate customer_ids: {df['customer_id'].duplicated().sum()}")  # 0
print(f"Fully duplicate rows:   {df.duplicated().sum()}")                 # 0

Decisions:

• age: 3.6% missing. Impute with median; add age_missing indicator flag.
• support_calls: 1.8% missing. Missing likely means zero calls — impute with 0.
• region: 0.8% missing (4 rows). Impute with "Unknown" category.

df["age_missing"] = df["age"].isna().astype(int)
df["age"] = df["age"].fillna(df["age"].median())
df["support_calls"] = df["support_calls"].fillna(0)
df["region"] = df["region"].fillna("Unknown")

print("Remaining NaN values:", df.isna().sum().sum())  # 0

---

Step 3 — Univariate Analysis

Numeric Features

print(df[["age", "tenure_months", "support_calls", "monthly_fee"]].describe(
    percentiles=[0.1, 0.25, 0.5, 0.75, 0.9]
))

fig, axes = plt.subplots(2, 2, figsize=(12, 8))
cols = ["age", "tenure_months", "support_calls", "monthly_fee"]

for ax, col in zip(axes.flatten(), cols):
    sns.histplot(df[col], kde=True, ax=ax, color="#0D9488")
    ax.set_title(col.replace("_", " ").title())

plt.suptitle("Univariate Distributions", fontsize=13, y=1.01)
plt.tight_layout()
plt.savefig("univariate.png")

Findings:

• tenure_months: roughly uniform (1–60). No concentration near cancellation dates.
• support_calls: right-skewed. Most customers make 0–3 calls; a tail makes 8–12.
• monthly_fee: bimodal — two clear peaks matching the Basic and Premium segments.
• age: approximately normal, centred around 42.

Categorical Features

for col in ["segment", "region"]:
    vc = df[col].value_counts()
    print(f"\n{col}:\n{vc}")
    print(f"  Churn rate: {df.groupby(col)['churn'].mean().round(3)}")

---

Step 4 — Bivariate Analysis (Feature vs Target)

Numeric Features vs Churn

fig, axes = plt.subplots(1, 4, figsize=(16, 5))
cols = ["age", "tenure_months", "support_calls", "monthly_fee"]

for ax, col in zip(axes, cols):
    sns.boxplot(data=df, x="churn", y=col, ax=ax,
                palette={0: "#0D9488", 1: "#F59E0B"})
    ax.set_title(col.replace("_", " ").title())
    ax.set_xlabel("Churned")

plt.suptitle("Feature Distributions: Retained (0) vs Churned (1)", y=1.02)
plt.tight_layout()
plt.savefig("bivariate_target.png")

Findings:

for col in ["age", "tenure_months", "support_calls", "monthly_fee"]:
    retained = df[df["churn"] == 0][col].mean()
    churned  = df[df["churn"] == 1][col].mean()
    print(f"{col:20s}  retained={retained:7.1f}  churned={churned:7.1f}  diff={churned-retained:+.1f}")

Expected output pattern: - tenure_months: churned customers have ~15 months; retained have ~35. Strong negative relationship. - support_calls: churned customers average ~5 calls; retained ~2. Strong positive relationship. - monthly_fee: churned customers pay less (mostly Basic segment). - age: minimal difference — probably not a strong predictor.

Categorical Features vs Churn

for col in ["segment", "region"]:
    ct = pd.crosstab(df[col], df["churn"], normalize="index") * 100
    ct.columns = ["Retained %", "Churned %"]
    print(f"\n{col}:\n{ct.round(1)}")

Findings:

• segment: Basic churns at ~38%; Premium churns at ~20%. Segment is a strong signal.
• region: Churn rates similar across regions (~29–33%). Region is probably weak.

---

Step 5 — Correlation Analysis

numeric_feats = ["age", "tenure_months", "support_calls", "monthly_fee", "age_missing"]
corr = df[numeric_feats + ["churn"]].corr()

fig, ax = plt.subplots(figsize=(7, 6))
mask = np.triu(np.ones_like(corr, dtype=bool))
sns.heatmap(corr, mask=mask, annot=True, fmt=".2f",
            cmap="coolwarm", center=0, ax=ax)
ax.set_title("Correlation Matrix incl. Target")
plt.tight_layout()
plt.savefig("correlation.png")

Key correlations with churn:

• tenure_months: strong negative (~-0.55). Longer-tenure customers churn less.
• support_calls: strong positive (~+0.50). More calls = higher churn.
• monthly_fee: moderate negative (~-0.30). Higher fee customers (Premium) churn less.
• age: near zero. Not predictive on its own.

---

Step 6 — Outlier Check

def iqr_bounds(s):
    q1, q3 = s.quantile(0.25), s.quantile(0.75)
    iqr = q3 - q1
    return q1 - 1.5 * iqr, q3 + 1.5 * iqr

for col in ["support_calls", "monthly_fee", "tenure_months"]:
    lo, hi = iqr_bounds(df[col])
    n_out = ((df[col] < lo) | (df[col] > hi)).sum()
    print(f"{col}: {n_out} outliers (bounds: {lo:.1f} – {hi:.1f})")

No pathological outliers here — the ranges are consistent with the data generation. In a real dataset, always inspect the actual rows before deciding.

---

Written Insights

Using the Observation / Interpretation / Action format:

Insight 1 — Tenure is the dominant predictor Observation: Churned customers have an average tenure of ~15 months vs ~35 months for retained customers. Interpretation: New customers have not yet built switching costs or habitual use. The first year is the critical retention window. Action: Include tenure as a top feature. Consider a tenure_bucket feature (0–6, 7–12, 13–24, 24+) to capture non-linear effects.

Insight 2 — Support calls signal dissatisfaction Observation: Churned customers average ~5 support calls vs ~2 for retained. Customers with 8+ calls churn at over 70%. Interpretation: High contact volume reflects unresolved problems. By the time someone has called 5 times, they are already disengaged. Action: support_calls should be a primary feature. Consider an interaction feature: calls_per_tenure_month.

Insight 3 — Segment explains fee differences Observation: Premium customers churn at 20% vs 38% for Basic. Monthly fee is correlated with churn only because segment is — not because high-paying customers are more loyal per se. Interpretation: The product or service tier, not the price, drives retention. Premium customers likely have access to features that increase stickiness. Action: Use segment directly. Explore whether there are Premium features that could be offered to Basic customers to reduce their churn.

Insight 4 — Region is not predictive Observation: Churn rates across North/South/East/West are 29–33% — effectively uniform. Interpretation: Churn is not a regional phenomenon in this dataset. Local factors do not dominate. Action: Do not include region as a primary feature. May still capture marginal signal as an interaction.

Insight 5 — Age has negligible effect Observation: Pearson r between age and churn is near zero. Age distributions of churned vs retained customers are nearly identical. Interpretation: This product churns for behavioural reasons (low engagement, problems) not demographic ones. Action: Exclude age from the initial model. Revisit if the model is poor on a specific age subgroup.

---

Modelling Recommendations

1. Primary features: tenure_months, support_calls, segment, monthly_fee
2. Engineer: calls_per_tenure_month, tenure_bucket, binary high_calls flag (calls > 6)
3. Handle class imbalance: if churn rate is ~30%, use class_weight='balanced' in the model
4. Baseline model: Logistic Regression first — the relationships look roughly linear. Then compare against a tree-based model.
5. Metric to optimise: Recall on the churned class — it is more costly to miss a churning customer than to incorrectly flag a retained one.

---

What this case study demonstrated

A structured EDA — even on 500 rows of synthetic data — produces clear, actionable findings in under an hour. Two features (tenure_months and support_calls) explain most of the churn signal. That is exactly the kind of focus that makes modelling efficient.

---

06-eda-workflow | Week 02 — Machine Learning