Skip to content

Interview Questions — Customer Churn Analysis

These are the questions an interviewer might ask after you present this project. Prepare answers for all of them.


[Beginner] What is customer churn and why do companies care about it?

Show answer

Churn is when a customer stops using a product or service. Companies care because: (1) acquiring a new customer costs 5-7× more than retaining an existing one; (2) churned customers represent permanent lost revenue; (3) high churn signals product or service problems.

For a subscription business, even a 2% monthly churn = 22% annual churn — the business must replace nearly a quarter of its customers every year just to stand still.


[Beginner] How did you calculate churn rate in this project?

Show answer

For this dataset, the churn rate is a cross-sectional snapshot:

Churn Rate = Number of churned customers / Total customers
           = 1869 / 7043 = 26.5%

In a production setting, you'd calculate it as a period metric:

Monthly Churn Rate = Customers who churned in month M / Customers at the start of month M

This dataset doesn't have dates, so we can't compute the true time-based rate — a limitation I'd note when presenting.


[Mid-level] You found that fiber optic customers churn at 42% — how would you determine if this is a pricing problem or a quality problem?

Show answer

I'd segment the analysis further:

Price signal: do fiber customers who pay the highest charges churn more than lower-paying fiber customers? If yes, price sensitivity may be a factor.

Quality signal: do fiber customers who contacted support multiple times churn more? I'd need to join with the support ticket table: JOIN support_tickets ON customerID. If churners had 2+ unresolved tickets in their final 30 days, quality is the likely cause.

Competitor signal: did churn increase around competitor promotions? Needs external market data.

A practical first step: a post-churn survey. Ask recently churned fiber customers why they left. Qualitative data is often faster than building the cross-join.


[Mid-level] Walk me through the churn risk scoring logic you built.

Show answer

The churn risk score is a weighted sum of the features most correlated with churn in the analysis:

Feature Weight Rationale
Month-to-month contract +30 Highest churn rate (42.7%)
Tenure < 12 months +20 New customers churn most
Fiber optic service +15 42% churn vs 19% DSL
Monthly charge > £70 +10 Correlated with fiber + high expectations
0 add-ons +15 No sticky services → easy to leave
Manual payment +10 Lower commitment signal

This is a heuristic model, not a trained ML classifier. It's interpretable and can be built entirely in SQL. A proper churn model would be logistic regression or gradient boosting — but this is a good first approximation that the business can use immediately.

Limitation: I haven't validated the weights statistically. I would calibrate these against actual churn data — do high-scoring customers actually churn at higher rates than low-scoring ones?


[Mid-level] What other data sources would improve this analysis?

Show answer
  1. Support tickets — customers who churn often contact support 30-60 days before leaving. Number of tickets, resolution time, and ticket category are powerful predictors.

  2. Product usage logs — how often does the customer log in, use the app, consume data? Declining usage often precedes churn.

  3. Billing history — late payments are a leading indicator. Customers who consistently pay late are more likely to churn.

  4. Competitor pricing — knowing when competitor promotions launched helps explain churn spikes.

  5. Marketing touchpoints — was this customer recently emailed? Did they open the retention campaign? Response to retention efforts is itself a predictive signal.

  6. NPS/CSAT scores — customers who rate satisfaction 6/10 or below are at risk. This is a direct early warning signal.


[Senior] The retention team has budget to contact 200 customers this month. You have 847 at-risk customers. How do you prioritise?

Show answer

I'd prioritise by expected value of retention:

Expected Value = P(churn | no intervention) × P(retain | intervention) × CLV

Practical approach: 1. Prioritise by revenue at risk, not just churn probability. A 60% churn-risk customer paying £120/month (£72 expected MRR at risk) outranks a 80% churn-risk customer paying £25/month (£20 expected MRR at risk).

  1. Within same revenue band, prioritise by win-back probability. Customers who contacted support about a specific issue (fixable) are more likely to be retained than customers who said "switching to competitor" in their cancellation survey.

  2. Use propensity scoring. If we have historical retention campaign data, build a model: given a customer profile, what is P(retain | outreach)?

Segmentation for the 200 contacts: - Top 50: >£100/month MRR, score 70+, fixable issue → call - Next 100: £50-100/month, score 50+, discount eligible → call + offer - Next 50: £30-50/month, score 60+ → email + discount


[Senior] This dataset shows 26.5% churn — much higher than typical telecom. A stakeholder argues the model won't generalise to TelcoMax's real customers. How do you respond?

Show answer

The stakeholder is right to raise this. This is a dataset bias concern.

Valid limitations: - The IBM dataset was created for educational purposes and may not reflect any real company's distribution - Our risk score was trained and evaluated on the same dataset — no out-of-sample validation - TelcoMax's product mix, pricing, and customer demographics may differ significantly

How I would address it: 1. Validate feature direction on TelcoMax data. I expect the same features to matter (contract type, tenure, add-ons) even if the exact weights differ. Run the same SQL queries on TelcoMax's real data and verify the directional relationships hold.

  1. Recalibrate scores. Using TelcoMax's historical churn data, fit a logistic regression to re-estimate the weights for each feature.

  2. Track model lift. After using the risk score to prioritise retention outreach for one quarter, compare: did customers with high risk scores actually churn more than low-risk customers? This is the real validation.

The educational dataset teaches methodology. Production deployment requires validation on real data.


← Insights and Recommendations · ← Project Overview · Next Project: Sales Dashboard →