Customer Churn Drivers: A Simulated Example

Suppose a telco lost a large block of customers last quarter. Customer success says it's pricing. Product blames service quality. Finance points to payment friction. Everyone has a theory. Nobody has data.

Here's the problem: without proper telco customer churn drivers analysis, you're flying blind. You might throw retention budget at the wrong levers, subsidizing customers who weren't leaving anyway or missing the structural issues pushing your best customers to competitors. Churn driver analysis answers one question: which factors actually move the churn needle, and by how much?

Before we draw conclusions, let's check the experimental design. This isn't an A/B test. You're analyzing observational customer data. That means you can identify correlates of churn and build predictive models, but causal claims require careful thinking. Contract type correlates with churn, yes. But does month-to-month cause churn, or do churn-prone customers self-select into flexible contracts? We'll come back to this.

This article walks through a dual-method approach: logistic regression for interpretable coefficients plus random forest for non-linear variable importance. We'll build a simulated customer set where plan type has no effect on churn, show how a gap still appears, and then walk through how to read each kind of churn driver output. By the end, you'll know how to tell a lever from a bystander.

Worked Example: A Plan Gap With No Plan Effect (Simulated Data)

Question: In this made-up set, how much higher is churn for flex plans than for annual plans?

Inputs: Made-up data, seed 2718. The script makes 1,000 people. Half are at high risk. Those at high risk have a 0.75 chance of a flex plan. Those at low risk have a 0.25 chance. The chance of churn is 0.60 for those at high risk and 0.15 for those at low risk. The script builds in no effect of plan type on churn. The whole gap comes from which customers pick which plan.

Method: Count who left in each group. Divide by the group size to get each rate. Then take the flex rate minus the annual rate. The script also gives a 95% range for the gap, which shows how much the gap may shift by chance.

import math
import random

SEED = 2718
N = 1000
rng = random.Random(SEED)
rows = []

for _ in range(N):
    high_risk = rng.random() < 0.5
    p_flexible = 0.75 if high_risk else 0.25
    flexible = rng.random() < p_flexible
    p_churn = 0.60 if high_risk else 0.15
    churn = rng.random() < p_churn
    rows.append((flexible, churn))

def summarize(group_value):
    group = [churn for flexible, churn in rows if flexible == group_value]
    events = sum(group)
    rate = events / len(group)
    se = math.sqrt(rate * (1 - rate) / len(group))
    return len(group), events, rate, se

n_f, e_f, r_f, se_f = summarize(True)
n_a, e_a, r_a, se_a = summarize(False)
diff = r_f - r_a
se_diff = math.sqrt(se_f**2 + se_a**2)
lo, hi = diff - 1.96 * se_diff, diff + 1.96 * se_diff

print(f"Simulated data, seed {SEED}")
print(f"Customers: {N}")
print(
    "Inputs: high-risk share 0.50; flexible contract chance 0.75 for "
    "high-risk and 0.25 for low-risk; churn chance 0.60 for high-risk "
    "and 0.15 for low-risk"
)
print(f"Flexible group: {e_f}/{n_f} churned ({r_f:.3f})")
print(f"Annual group: {e_a}/{n_a} churned ({r_a:.3f})")
print(f"Observed rate gap (flexible minus annual): {diff:.3f}")
print(f"Approximate 95% confidence interval for gap: [{lo:.3f}, {hi:.3f}]")

Printed output:

Simulated data, seed 2718
Customers: 1000
Inputs: high-risk share 0.50; flexible contract chance 0.75 for high-risk and 0.25 for low-risk; churn chance 0.60 for high-risk and 0.15 for low-risk
Flexible group: 235/488 churned (0.482)
Annual group: 132/512 churned (0.258)
Observed rate gap (flexible minus annual): 0.224
Approximate 95% confidence interval for gap: [0.165, 0.282]

Result: The script builds in no effect of plan type. The whole gap comes from which customers pick which plan. Flex: 235 of 488 left (0.482). Annual: 132 of 512 left (0.258). The rate gap was 0.224, close to the 0.225 the inputs imply (0.4875 for flex minus 0.2625 for annual). The 95% range of 0.165 to 0.282 describes only the observed gap. It does not say plan type matters, because the plan effect built into the script is zero.

Limit: This is made-up data. People at high risk were more apt to pick flex plans, and that mix led to the gap. The result does not show that a plan change would cut churn. The script runs one group comparison only; logistic regression and random forest, covered below, are shown as ways to read churn driver output, not run here.

Why You Need Two Methods, Not One

Most churn analyses pick one method and stop there. That's a mistake. Logistic regression and random forest answer different questions, and you need both.

Logistic regression gives you directional effects with statistical significance. Each coefficient is a change in the log-odds of churn: if this variable increases by one unit, holding the others constant, the odds of churning multiply by exp(coefficient). That is a change in odds, not a fixed multiplier on churn probability; how much the probability moves depends on the customer's starting risk. You get p-values, confidence intervals, and interpretable odds ratios. The output is a ranked list of drivers with effect sizes. This is what you show your CFO when she asks "which lever has the biggest ROI?"

But logistic regression assumes linear relationships and additive effects. Real customer behavior is messier. A variable might matter only for certain segments or interact with other features in non-obvious ways.

Random forest captures those non-linearities. It builds hundreds of decision trees, each trained on a random subset of data and features. Variable importance here is measured by mean decrease in Gini impurity: how much, summed across all trees, the splits on that variable make the churned and retained groups purer. (Scrambling a variable and measuring how much worse predictions get is a different measure, called permutation importance.) Gini importance tends to favour continuous variables with many possible split points, so read it alongside the logistic results. Strobl, Boulesteix, Zeileis and Hothorn published a study of bias in random forest importance measures in 2007. High importance means the variable helps split customers into churned vs. retained groups, regardless of whether the relationship is linear.

Here's how to use both: if a variable ranks high in random forest but shows weak logistic regression coefficients, you're looking at non-linear effects or interactions. Check your cross-tabs and heatmaps. If both methods agree a variable matters, you have convergent evidence. That driver is robust across modeling assumptions.

What Real Churn Data Usually Holds

Telco churn records usually contain demographics, account information, services subscribed, contract terms, payment method, and a binary churn indicator (1 = churned, 0 = retained). This is the kind of data you'd export from your CRM or billing system. The example above uses none of it; it only shows what a plan gap can look like when the plan does nothing.

Typical variables include:

• Contract type: Month-to-month, one year, or two year
• Tenure: Months as a customer
• Internet service: DSL, Fiber optic, or None
• Payment method: Electronic check, mailed check, bank transfer, or credit card
• Monthly charges: Current monthly bill amount
• Total charges: Lifetime revenue from customer
• Add-on services: Online security, tech support, streaming TV, etc.

The churn variable is binary-encoded (1/0). Both logistic regression and random forest consume this directly.

Did you randomize? Usually not: this is observational data. Customers choose their own contract types, payment methods, and service bundles. That means selection bias is baked in. High-risk customers may self-select into month-to-month contracts because they're uncertain about staying. The correlation can be real while the causal arrow is ambiguous, exactly as in the simulated example. Flag actionable vs. descriptive drivers as you go.

Reading Churn Rate by Contract Type

A first cut is the churn rate for each contract type. A large spread between month-to-month and longer contracts is a link, not an effect. Whether it's a psychological sunk cost effect, switching friction, or adverse selection (risky customers avoid long contracts), the rates alone cannot say which. A gap between two groups is not the effect of converting a customer: only a randomized test can tell you how much of that gap a conversion would remove.

Is contract type actionable? Yes and no. You can't force customers into long contracts, but you can design incentives. Offer a discount for annual commitment. Waive activation fees for two-year contracts. Remove month-to-month as the default option in signup flows. These are testable interventions: run a proper A/B test with randomized treatment assignment to measure causal impact.

One caution: month-to-month customers may be fundamentally different from annual customers. They value flexibility, may have shorter expected tenure, or face income uncertainty. Converting them to annual contracts helps only if the contract itself causally reduces churn. If you lock in high-risk customers, you may just delay churn and increase early termination fee disputes. This is where churn driver analysis meets customer segmentation: not all retention levers work for all segments.

Reading Churn Rate by Internet Service Type

Compare churn rates across service types. If a premium product shows the highest churn, that is a red flag worth checking, since a premium tier should have better speeds and reliability.

Two hypotheses: (1) Pricing is too high relative to perceived value, driving price-sensitive customers to churn when competitors undercut you. (2) Service quality isn't meeting expectations: latency, outages, or support issues push customers to leave despite higher speeds. Both are testable.

Check your NPS or CSAT scores by service type. If premium customers report lower satisfaction, you have a quality problem. If satisfaction is comparable but churn is still high, you have a pricing or competitive pressure problem. Run a win-back survey on churned customers: "Why did you leave?" If most cite price, you know where to intervene.

Here's a retention experiment: randomly assign a subset of premium-tier customers to a loyalty discount. Measure churn over the next 90 days vs. a control group at full price. If churn drops clearly, price elasticity is high and retention discounts may pay off. If churn stays flat, discounting won't solve your problem. You need to fix the product.

Reading Churn Rate by Payment Method

Compare churn rates across payment methods. One possible reason a manual method shows more churn: it requires an active payment each month, which creates a monthly decision point ("Should I pay this bill or cancel?"). Automatic payment methods (bank transfer, credit card) remove the decision, and inertia may keep marginal customers subscribed. The rates alone cannot confirm this.

If it holds, this is an actionable driver. Migrate manual payers to automatic payments. Offer a discount for switching to auto-pay. Run an email campaign: "Never miss a payment: switch to auto-pay and save." Measure conversion rates and churn impact in a randomized experiment.

One subtlety: manual payers may be financially constrained or prefer payment control. If they can't afford automatic charges, forcing them into auto-pay increases failed payment rates and involuntary churn. Segment by payment history: customers with consistent on-time manual payments are good auto-pay candidates. Customers with frequent late payments need a different intervention: maybe payment plan flexibility or downgrade options.

Reading Logistic Regression Coefficients

A logistic regression ranks drivers by log-odds coefficients. Positive coefficients increase churn probability; negative coefficients decrease it. Statistical significance is marked; read it together with the interval and the size of the effect, not alone.

Each categorical coefficient is relative to a reference group, such as month-to-month for contract. Always check which group is the reference before you read the sign.

A negative coefficient on tenure means each additional month as a customer lowers the odds of churning, holding the other variables constant. A coefficient on a contract term compares that term with the reference group in the same way.

What's your sample size? Check the model output. With few churn events, treat borderline-significant results with skepticism. Is the analysis adequately powered? If your study has low power to detect a meaningful effect, you're likely missing true drivers or flagging false positives.

How to use these coefficients: exp(coefficient) gives you the odds ratio. With an illustrative coefficient of -1.0 for two-year contracts, two-year customers would have exp(-1.0) = 0.37 times the odds of churning of month-to-month customers, holding other variables constant. Put the other way, month-to-month customers would have about 2.7 times the odds of two-year customers. A large association like that is worth acting on if the variable is something you can influence.

Reading Random Forest Variable Importance

The random forest importance plot ranks variables by mean decrease in Gini impurity. Higher bars = more important for splitting customers into churned vs. retained groups. This captures non-linear effects logistic regression might miss.

Read the ranking with some care. Variables that are continuous, such as tenure, monthly charges or total charges, offer many possible split points and so tend to rank high. Lifetime spending is tied closely to tenure, so much of its signal overlaps with tenure.

Where do the methods disagree? If a variable ranks low in the forest yet carries a large logistic coefficient, part of that gap can be the Gini bias toward continuous variables, and part may be that its effect overlaps with another variable, such as monthly charges.

Conversely, if a variable ranks high in the forest but is not significant in logistic regression, that suggests non-linear effects: maybe churn spikes above a certain price threshold, or price sensitivity varies by customer segment. Check your price elasticity curves by contract type or service tier.

How to interpret convergent evidence: when both methods rank a variable high, it is more likely to be robust across modeling assumptions. Variables that rank high in only one method deserve deeper investigation. They may represent non-linear effects or model artifacts.

Reading Churn Rate by Tenure and Contract Type

A heatmap of churn rate by tenure bucket and contract type can show an interaction. Look at whether early-tenure customers on flexible contracts churn at a higher rate than later-tenure ones, and whether customers on longer contracts stay low across all tenure buckets.

If the first year shows the highest churn for flexible-contract customers, the first year is make-or-break for them, and that is where retention budget should focus.

What interventions might work in the early tenure window? Onboarding quality, proactive support, and early-win moments. Send a welcome email sequence with setup tips. Flag customers who haven't activated key features early and offer live setup assistance. Monitor early usage: customers who don't reach a usage level you define are higher churn risk.

Run a randomized experiment: assign new month-to-month customers to either (A) standard onboarding or (B) enhanced onboarding with proactive outreach, setup calls, and usage monitoring. Measure 90-day churn. If enhanced onboarding lowers churn by enough to cover its cost, it's worth scaling.

If long-tenure month-to-month customers stabilize at a lower churn rate, they may be habitual customers who value flexibility but aren't actively shopping. Don't over-invest in retaining them if they're already sticky. Focus on the high-risk early cohorts.

How to Interpret Your Results

You've run the analysis. You have logistic regression coefficients, random forest importance scores, and cross-tabulated churn rates. Now what? Here's how to translate statistical output into retention strategy.

Step 1: Separate Actionable from Descriptive Drivers

Actionable drivers are variables you can change through intervention. Descriptive drivers identify high-risk segments but don't tell you what to do.

Actionable: Contract type (offer incentives to upgrade), payment method (migrate to auto-pay), add-on services (bundle or discount to increase attachment), onboarding quality (improve early experience).

Descriptive: Tenure (you can't make a new customer old), total charges (reflects tenure and pricing), demographics (age, location, hard to change).

Prioritize actionable drivers with large coefficients. Month-to-month contract and electronic check payment are both actionable and high-impact. These are your top retention levers.

Step 2: Calculate Expected Churn Reduction

Don't just look at statistical significance, estimate business impact. If you convert 1,000 month-to-month customers to one-year contracts, how much churn do you prevent?

With made-up numbers: month-to-month churn is 40%, one-year churn is 25%, an observed gap of 15 percentage points. If the whole gap were caused by the contract itself, expected prevented churn would be 1,000 × 0.15 = 150 customers. Treat that as an upper bound: some of the gap is self-selection, so the true effect of converting a customer is likely smaller. Multiply by average customer LTV to get a best-case retention ROI.

If LTV is $1,000 and it costs $100 in incentives to convert a customer to annual (e.g., one month free), the best-case result is: (150 saved customers × $1,000 LTV) - (1,000 conversions × $100 cost) = $150,000 - $100,000 = $50,000 net gain. The randomized test in Step 3 tells you how much of that is real.

Run this calculation for each actionable driver. Rank interventions by expected ROI. Test the top 3 in randomized experiments.

Step 3: Design Proper Retention Experiments

Observational churn analysis identifies correlates. To prove causality, you need randomized experiments. Here's the protocol:

1. State your hypothesis: "Offering a 10% discount for annual commitment will reduce churn by at least 15 percentage points among month-to-month customers."
2. Calculate sample size: Use a power calculator. To detect a 15-point churn difference (for example, 40% down to 25%) with 80% power and α=0.05, the normal approximation gives about 150 customers per arm (300 total). Don't run underpowered tests.
3. Randomize assignment: Use a random number generator to assign customers to treatment (discount offer) or control (no offer). Check that treatment and control groups are balanced on key covariates (tenure, current charges, service type).
4. Set observation window: Measure churn over 90 days post-intervention. Too short and you miss delayed effects; too long and external factors confound results.
5. Analyze intent-to-treat: Compare churn rates for all customers assigned to treatment vs. control, regardless of whether treatment customers accepted the discount. This preserves randomization and gives you the causal effect of the offer.
6. Check statistical significance: Use a two-proportion z-test. If p < 0.05, you have evidence the intervention worked. If not, either the effect is smaller than expected or your sample size was too small.

Did you randomize? This is the question that separates correlation from causation. If you just offered discounts to high-risk customers (identified by propensity scores), you've introduced selection bias: high-risk customers churn more regardless of treatment. Randomization breaks the selection mechanism and isolates the causal effect.

Step 4: Monitor Driver Stability Over Time

Churn drivers aren't static. What mattered last quarter may not matter today. Re-run your analysis quarterly and track coefficient changes.

Set up monitoring alerts: if your overall churn rate moves by more than a margin you set in advance, re-run the driver analysis. A sudden churn spike may be driven by a new factor (e.g., competitor launched, pricing changed, service outage).

Track driver importance over time. If contract type was your #1 driver in Q1 but drops to #4 by Q3, something changed. Maybe you successfully shifted customers to annual contracts (reducing variance in that variable), or maybe a new driver (e.g., service quality issues) emerged. Investigate.

Working on this for your business? See the churn analysis tool.

Common Mistakes and How to Avoid Them

Mistake 1: Confusing Correlation with Causation

Tenure correlates negatively with churn. Does longer tenure cause lower churn, or do low-risk customers self-select into longer tenure? Both. You can't randomly assign tenure, so the causal arrow is ambiguous. Use tenure as a risk segmentation variable (new customers need more attention), but don't claim you can "increase tenure to reduce churn" without testing an intervention that actually keeps customers longer.

Mistake 2: Running Underpowered Experiments

You test a retention email on 50 customers and see 5% lower churn. Is that real or noise? Calculate power. With 25 customers per arm and a baseline churn rate of 40%, the normal approximation gives about 7% power to detect a 5-point effect. You'll miss a true effect of that size more than 90% of the time. Either increase sample size or widen your detection threshold. What's your minimum detectable effect before we start? Agree on this upfront.

Mistake 3: Ignoring Interaction Effects

Main effects tell you "on average, contract type matters." But maybe contract type only matters for high-price customers or only in the first 6 months. Check interactions in cross-tabs and heatmaps. If you find strong interactions, segment your retention strategy: don't offer the same intervention to all customers.

Mistake 4: Overfitting on Noise

Random forest can overfit, especially with small samples or many irrelevant features. If a variable ranks high in random forest but has no logical mechanism, it's probably spurious. Example: if "customer ID ends in 7" ranks high, that's noise. Use cross-validation on your random forest model to check out-of-sample performance. If importance scores shift dramatically between training and test sets, you're overfitting.

Mistake 5: Skipping the Holdout Check

For example, build your models on 70% of data and test on the remaining 30%. Do the same drivers emerge? Do predicted churn probabilities match observed churn rates in the holdout set? If your model performs well in-sample but fails out-of-sample, you've overfit. This is especially critical for random forest, which can memorize training data.

When Churn Driver Analysis Isn't Enough

Churn drivers tell you who churns and which variables correlate. They don't tell you:

• Why customers churn (their reasons, pain points, competitive alternatives). For that, run win-back surveys and exit interviews.
• Which retention offers work (discount vs. feature upgrade vs. service improvement). Test these in randomized experiments.
• Customer lifetime value by segment (retaining a $50/month customer is different from retaining a $200/month customer). Combine churn analysis with customer lifetime value modeling.
• Leading indicators of churn (usage drops, support tickets, payment failures). Build a churn early-warning system with behavioral signals, not just demographic features.

Use churn driver analysis as the foundation, then layer in qualitative research, experimentation, and predictive monitoring. The full retention stack combines all three.

From Insight to Action: Building a Retention Roadmap

You've identified your top churn drivers. Here's how to turn that into a 90-day retention roadmap:

1. Week 1-2: Prioritize interventions. Rank actionable drivers by estimated ROI. Pick the top 3. For each, draft a specific intervention (e.g., "Offer 1 month free for annual commitment upgrade").
2. Week 3-4: Design experiments. Write hypotheses, calculate required sample sizes, set up randomization infrastructure. Get stakeholder buy-in on running proper tests (no cherry-picking treatment groups).
3. Week 5-10: Run experiments. Launch treatment and control arms. Monitor enrollment, check balance, watch for implementation issues. Don't peek at results early: wait for the pre-specified observation window.
4. Week 11-12: Analyze results. Calculate intent-to-treat effects, check significance, estimate cost-benefit. If the experiment worked, plan scale-up. If it failed, diagnose why (wrong lever, insufficient dose, implementation failure).
5. Week 13+: Scale winners, iterate on losers. Roll out successful interventions to the full customer base. Re-test failed interventions with modifications (different offer amount, different customer segment, different messaging).

Quarterly, re-run the churn driver analysis. Track how driver rankings change as you roll out interventions. If contract type was your #1 driver but you successfully migrated 50% of month-to-month customers to annual, it may drop in importance, not because it stopped mattering, but because you reduced variance in that variable. That's success.

Your turn

Bring your own data and the question you actually need answered.

MCP AnalyticsData science by email Send us your data and your question, and we will reply. ds@mcpanalytics.ai

Frequently Asked Questions

Should I use logistic regression or random forest for churn driver analysis?

Use both. Logistic regression gives you interpretable coefficients with statistical significance tests, which is critical for understanding which features truly move churn probability and by how much. Random forest captures non-linear interactions and ranks variables by importance. Side-by-side comparison reveals where linear assumptions hold and where they break down. If both methods agree a variable matters, you have convergent evidence.

What's the minimum sample size I need for reliable churn driver analysis?

No single minimum fits every dataset. What you need depends on how many churn events you have, how many predictors you test, and how large the effects are. With few churn events for the number of predictors, estimates become unstable and an underpowered analysis can miss real drivers and flag noise as signal. Check how stable the results are on holdout data.

How do I know if a churn driver is actionable or just descriptive?

Ask: can we change this variable through intervention? Contract type is actionable. You can design incentives to shift customers from month-to-month to annual contracts. Tenure is descriptive. You can't make a 3-month customer into a 24-month customer. Focus retention spend on actionable drivers with large coefficients. Descriptive drivers help you identify high-risk segments but don't tell you what to do.

Why do some variables rank high in random forest but show weak effects in logistic regression?

Random forest captures non-linear effects and interactions that logistic regression misses. A variable may have weak main effects but strong conditional effects. It matters only for certain customer segments or in combination with other features. Check interaction terms in your cross-tabs. If tenure matters primarily for month-to-month customers, you'll see it in the heatmap even if the main effect is modest.

How often should I re-run churn driver analysis?

Quarterly for stable businesses, monthly if you're running active retention experiments or seeing market shifts. Churn drivers change as your product, pricing, and competitive environment evolve. Set up monitoring: if your overall churn rate moves by more than a margin you set in advance, re-run the analysis. What drove churn last quarter may not be the primary lever today.

Related Articles