Churn Prediction Model for Apps

The Problem with Fixing Churn After It Happens

Every Monday we'd pull the weekly retention report. Users who churned last week. Win-back campaigns sent. Discount codes deployed. Response rates: 3–7%. The users were already gone.

We were fighting churn retrospectively — responding to an event that had already happened rather than preventing it. The entire retention strategy was essentially a post-mortem dressed up as engagement.

By the time a user churns, you've already lost. The window to intervene is 10–14 days before the churn event — when behavior starts to degrade but the user is still reachable and still cares.

The goal shifted: stop measuring churn. Start predicting it. Specifically, predict it 14 days before it happens — when there's still something we can do about it.

What We Observed

The 14-Day Degradation Signal

We pulled 2 years of churned user behavioral data and mapped it backwards from the churn date. The pattern was remarkably consistent: users who churned showed measurable behavioral degradation starting 14 days out.

Days Before Churn	Avg Session Frequency Change	Avg Session Depth Change	Transaction Rate Change
D-28 to D-21	−8%	−5%	−3%
D-21 to D-14	−22%	−18%	−14%
D-14 to D-7	−41%	−35%	−28%
D-7 to D-0 (churn)	−67%	−52%	−47%

The sharpest degradation happened between D-21 and D-14. That 7-day window was both the earliest detectable signal and the last viable intervention window. Before D-21: too early, too noisy. After D-7: too late.

The question was whether the D-21 to D-14 behavioral changes were specific enough to be predictive — or whether they were just statistical noise that looked like a pattern in retrospect.

Behavioral Fingerprint

What Pre-Churn Users Actually Did

We labeled 8 months of churned users (N=4,200) and matched them against active users for the same period. Then we looked at behavioral differences starting at D-28 before churn.

Behavioral Signal	Pre-Churn Users (D-14)	Retained Users (same period)	Signal Strength
Session frequency drop >30%	71%	12%	High
No return within 7 days of last session	68%	18%	High
Core feature usage stopped	61%	9%	Very High
Last 3 sessions: browse only, no action	54%	14%	High
Support ticket OR complaint in D-30	38%	8%	Moderate
Comparison of competitor features viewed	29%	4%	High

Core feature usage stopping was the strongest individual predictor — 61% of users who churned had stopped using the core feature 14+ days before churning. Most teams interpret this as "disengagement." It's actually the decision phase of leaving.

The Model

From Retrospective to Predictive

The conceptual shift was simple to articulate, harder to execute: treat churn not as an event to respond to, but as a probability to be managed continuously. Every user, every day, has a churn probability. The job is to monitor that probability and intervene when it crosses a threshold.

A user doesn't churn suddenly. They decide slowly, then leave quickly. The decision phase is when you can still change the outcome. The leaving phase is where you lose the argument.

The model we needed wasn't a churn classifier — it was a churn probability monitor. The output wasn't "will this user churn?" but "what's the probability this user churns in the next 14 days, and is that probability increasing?"

Model Design

The Churn Risk Monitoring System

We built in four components, each necessary. The hardest was the intervention playbook — getting the model right was actually the easier half.

Daily Churn Probability Scoring

Gradient boosting model scoring every active user daily. Output: 0–100 churn probability for the next 14 days. Input: 22 behavioral features, rolling 28-day windows. Model retrained weekly with previous week's actual churn outcomes.

Risk Velocity Tracking

Not just the score — the rate of change. A user at 40% churn probability for 2 weeks is stable. A user who went from 20% to 55% in 3 days is in active degradation. Velocity triggers different (faster) interventions than level alone.

Segmented Intervention Playbook

Three risk tiers: Yellow (40–60%), Orange (60–75%), Red (>75%). Each tier gets a different intervention: Yellow = product nudge; Orange = personalized email with relevant feature highlight; Red = human outreach within 24 hours.

Intervention Outcome Tracking

Every intervention tagged with user ID, risk score at trigger, and intervention type. Churn outcomes tracked per intervention type × risk tier. Playbook updated quarterly based on what's actually working.

Risk Tier	Churn Probability	Intervention	Response Rate	Churn Prevention Rate
Yellow	40–60%	Personalized in-app nudge	34%	61%
Orange	60–75%	Email with relevant case study or feature	28%	44%
Red	75%+	Human outreach within 24 hours	51%	38%

Technical Design

How It Runs Day-to-Day

Daily Scoring Pipeline

→Nightly batch job: feature extraction for all active users in last 30 days
→22 behavioral features per user: session frequency, depth, feature usage vectors, support interactions
→Model inference run at 2am IST; risk scores available by 6am for intervention queue

Intervention Queue

→Yellow tier: in-app nudges queued for next active session
→Orange tier: personalized email generated and sent within 2 hours of morning score run
→Red tier: ticket created in CRM, assigned to customer success rep with context note

Human Outreach System

→Red tier users: automated context card sent to CS rep — shows user's score history, last 3 sessions, what feature they stopped using
→CS rep playbook: first goal is understanding (why the withdrawal?), not retention offer
→Outcome logged: retained, partially retained (downgraded), or churned despite intervention

Rollout

90 Days of Iteration

Month 1 · Model Training

Labeling + Feature Selection

→ Model live in staging. Daily scoring pipeline running but not yet triggering interventions.

✓4,200 churned users labeled manually from 8 months of data
✓22 features selected from initial 60 candidates by importance score
✓Model validation: AUC 0.81 at 14-day prediction horizon

→ Model live in staging. Daily scoring pipeline running but not yet triggering interventions.

Month 2 · Playbook Design

Intervention Testing

→ Intervention playbook drafted. First 200 at-risk users saved.

✓Yellow tier in-app nudges: A/B tested 6 message variants
✓Orange tier email: personalized vs generic offer — personalized 2.8× better click rate
✓Red tier: manual human outreach for first 6 weeks before systematizing

→ Intervention playbook drafted. First 200 at-risk users saved.

Month 3 · Full Rollout

System Running at Scale

→ Monthly churn rate: 8.4% → 5.1%. Significant drop in first 60 days of full rollout.

✓All 3 tiers automated (except Red — human loop preserved)
✓Weekly model refresh established
✓Churn prevention dashboard live for CS team

→ Monthly churn rate: 8.4% → 5.1%. Significant drop in first 60 days of full rollout.

Results

What Changed

−39%

Monthly Churn Rate

8.4% → 5.1%

51%

Red Tier Response Rate

human outreach

₹2.1Cr

Annual Revenue Retained

from prevented churn

14 days

Prediction Horizon

before churn event

The most surprising finding: Yellow tier (40–60% churn probability) had the highest prevention rate at 61%. These users hadn't made up their mind yet — a well-timed intervention shifted the outcome. The users most at risk (Red tier) were harder to save because the decision was further along.

Learnings

The Non-Obvious Findings

Velocity matters more than level for urgent interventions

A user at 70% churn probability who's been there for 2 weeks is more stable than a user who jumped from 30% to 65% in 3 days. The second user is actively making the decision to leave. Score velocity should trigger faster intervention than score level alone.

The earliest intervention window has the highest success rate

Yellow tier (40–60%) had 61% prevention rate. Red tier (75%+) had 38%. Intervention earlier in the decision process — when users are ambivalent, not committed — is significantly more effective.

Human outreach beats automation for the highest-risk users

For Red tier users, automated messages felt tone-deaf. What worked was a CS rep who had context (what the user stopped doing, when, what they'd previously valued) and led with curiosity rather than retention offers.

Core feature abandonment is the single most predictive signal

Session frequency drops are visible but noisy — people go on vacation. Core feature abandonment is specific: a user who stops doing the thing they came for is making a statement about the product's value to them.

Failures

Where We Got It Wrong First

✕ Initial model had no velocity component

Why it broke

First version: daily score, static. A user at 65% for 3 weeks and a user who jumped to 65% in 2 days got the same intervention. Urgency was identical; it shouldn't have been.

What fixed it

Added 7-day score velocity as a feature. Users with velocity > 15-point increase per week flagged as urgent. Red tier expanded to include high-velocity users even below the 75% threshold.

✕ Yellow tier nudge was too generic

Why it broke

'We noticed you haven't been as active lately. Here's what's new.' Ignored by 73% of users. Too vague to be compelling, too generic to be personal.

What fixed it

Nudge personalized to the specific feature the user had stopped using. 'You haven't checked your rate alerts recently — here's a rate change that matches your saved comparison.' Click rate 4.2× better.

Future

What's Still Left to Build

Churn Root Cause Model

Current model predicts who will churn. Missing: why. A secondary classifier identifying whether the churn signal is price sensitivity, feature gap, or competitive. Different root causes need different interventions.

Re-engagement Scoring

Churned users who come back are underserved by the current system. A re-engagement model that identifies which churned users have elevated return probability, and what triggered them to reconsider.

Product-Level Churn Attribution

Connect churn signals to specific product features or flows. Which parts of the product are consistently present in the journeys of churning users? Inform product roadmap with churn data, not just engagement metrics.

Churn doesn't happen on the day someone cancels. It happens when they quietly start looking for reasons not to stay. The 14-day window is where the decision actually lives.

Using ML to Reduce CAC by 31% via Smart Audience Segmentation

Machine Learning

Personalizing Insurance Offers Using ML-Driven User Clustering

Machine Learning

Building a Real-Time Intent Scoring Engine for Lead Conversion

9 min read

Machine Learning

How We Built an AI-Powered Recommendation Engine for Financial Products

12 min read

Machine Learning

How We Used ML to Predict Insurance Buyers (50,000+ Data Points)

12 min read

All Case Studies

Browse all 30 breakdowns

Work With Me

Let's build something together

Connect on LinkedIn

Follow for more insights

Where to go next

Consumer App Growth Consultant
The consulting work this sits inside, with the rest of the evidence behind it.
Day 30 retention benchmark, fintech & banking
Published figures, traced to their sources — and where the sources disagree.
Day 7 retention benchmark, e-commerce & retail
Published figures, traced to their sources — and where the sources disagree.
Product Growth Score
Thirteen questions that locate which part of the funnel is actually constraining growth.

Predicting User Drop-Off Before It Happens (Churn Prevention Model)

Results at a glance

The Problem with Fixing Churn After It Happens

The 14-Day Degradation Signal

What Pre-Churn Users Actually Did

From Retrospective to Predictive

The Churn Risk Monitoring System

How It Runs Day-to-Day

90 Days of Iteration

What Changed

The Non-Obvious Findings

Velocity matters more than level for urgent interventions

The earliest intervention window has the highest success rate

Human outreach beats automation for the highest-risk users

Core feature abandonment is the single most predictive signal

Where We Got It Wrong First

What's Still Left to Build

You might also like

Where to go next

The Problem with Fixing Churn After It Happens

The 14-Day Degradation Signal

What Pre-Churn Users Actually Did

From Retrospective to Predictive

The Churn Risk Monitoring System

How It Runs Day-to-Day

90 Days of Iteration

What Changed

The Non-Obvious Findings

Velocity matters more than level for urgent interventions

The earliest intervention window has the highest success rate

Human outreach beats automation for the highest-risk users

Core feature abandonment is the single most predictive signal

Where We Got It Wrong First

What's Still Left to Build