“Not all customers are equally risky. Some churn quietly… others vanish the moment you send a survey.”

Welcome to the Cox Proportional Hazards model, a.k.a. the Sherlock Holmes of survival analysis — it investigates who’s more likely to churn and why, without needing to know exactly when.

🎯 The Core Idea¶

While the Kaplan–Meier curve tells you how long people survive, the Cox model tells you which features make them survive longer (or shorter).

In other words:

KM = “How loyal is everyone?” Cox = “Who’s the least loyal and why?”

🧮 The Magic Formula (Simplified)¶

[ h(t|x) = h_0(t) \times e^{(\beta_1 x_1 + \beta_2 x_2 + ... + \beta_n x_n)} ]

Where:

• (h(t|x)): hazard (risk of event) for a customer with features (x)

• (h_0(t)): baseline hazard (the average risk for everyone)

• (\beta_i): coefficients showing how risky each feature is

Basically:

The higher the hazard, the faster someone churns. A negative β means loyalty. A positive β means “brace for cancellation.”

💼 Example: Subscription Business Case¶

So if a customer:

• Has been around long enough,

• Pays less, and

• Rarely calls support — their hazard is basically “snoozing on autopay.”

🧪 Quick Python Example¶

from lifelines import CoxPHFitter
import pandas as pd

# Fake churn dataset
df = pd.DataFrame({
    'tenure': [10, 20, 5, 8, 15],
    'spend': [50, 80, 30, 40, 60],
    'calls': [1, 5, 2, 3, 1],
    'event': [1, 1, 0, 1, 0],
    'time': [30, 45, 60, 25, 80]
})

cox = CoxPHFitter()
cox.fit(df, duration_col='time', event_col='event')
cox.print_summary()

Interpret like a pro:

• exp(coef) > 1: increases risk of churn (⚠️ red flag)

• exp(coef) < 1: decreases risk (💚 loyal legend)

📊 Visualizing the Risk¶

cox.plot_partial_effects_on_outcome(covariates='spend', values=[30, 50, 80])

💡 The steeper the line → the faster churn happens as the variable increases. (Translation: “Stop charging your best customers more than your competitors.”)

🧩 Common Business Use Cases¶

🤓 Important Assumption: “Proportional Hazards”¶

The Cox model assumes the ratio of hazards between customers is constant over time. If it’s not, your model starts lying politely — like a sales forecast in December.

To check:

cox.check_assumptions(df, p_value_threshold=0.05)

If the check fails, you might need to:

• Add time-dependent covariates

• Split your model by time groups

• Or… pretend it passed and pray nobody asks

🧠 Practice Task¶

Try it out:

1. Load your company churn data.

2. Fit a CoxPHFitter.

3. Identify the top 3 “hazardous” features.

4. Present it to management as: “Here’s who’s most likely to leave us — and here’s how to stop them.”

🎭 TL;DR Summary¶

🧭 Next Stop¶

➡️ Customer Lifetime Value Modeling (CLV) — where we flip the script from “Who’s leaving?” to “How much are they worth before they do?”