Subscriber Churn Prediction
“Which subscribers will cancel?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
Which subscribers will cancel?
Streaming platforms lose 5-7% of subscribers monthly. Traditional churn models rely on usage decline, catching subscribers only after they've mentally checked out. They miss the graph signals: when a subscriber's social circle churns, when content in their preferred genres dries up, or when payment friction increases. For a platform with 30M subscribers at $12/month ARPU, reducing churn by 1 percentage point saves $43M annually.
Quick answer
Graph neural networks predict streaming subscriber churn 2-4 weeks before cancellation by detecting early signals that flat models miss: viewing sessions getting shorter, binge completion rates dropping, payment method failures, and social graph erosion (friends leaving the platform). For a 30M-subscriber platform, reducing churn by 1 percentage point saves $43M annually.
Approaches compared
4 ways to solve this problem
1. Usage threshold rules
Flag subscribers whose viewing time drops below a threshold (e.g., less than 2 hours in the past week). Simple and interpretable.
Best for
Quick-and-dirty churn flags. Works for obvious cases where viewing drops to near-zero.
Watch out for
Catches subscribers only after they have already mentally checked out. By the time usage crosses your threshold, retention intervention success rates drop below 10%. You are reacting, not predicting.
2. Logistic regression / XGBoost on aggregated features
Engineer features like 'days since last watch,' 'viewing hours trend,' and 'payment failures' into a flat table and train a classifier.
Best for
Solid baseline that captures the most predictive single-table signals. Easy to deploy and monitor.
Watch out for
Misses temporal sequences (declining-then-recovering patterns look different from steady decline) and cross-entity signals (when a subscriber's household co-viewers also disengage). Aggregated features destroy the time dimension.
3. Survival analysis (Cox proportional hazards)
Model time-to-churn using hazard functions. Naturally handles censoring (subscribers who haven't churned yet) and time-varying covariates.
Best for
Estimating when churn will happen, not just if. Good for planning retention campaign timing.
Watch out for
Assumes proportional hazards -- the relative risk of covariates stays constant over time. In streaming, the impact of a payment failure changes depending on viewing engagement level, violating this assumption.
4. KumoRFM (relational graph ML)
Connect subscribers, plans, watch history, payments, and devices into a temporal graph. The GNN learns early churn signals from the relational structure: viewing decay, payment friction, device abandonment, and social erosion.
Best for
Earliest detection window (2-4 weeks before cancellation). Captures compound signals like 'binge completion rate dropping + payment failure + household co-viewer churned' that no flat model can express.
Watch out for
Requires multi-table subscriber data with timestamps. Adds most value when you have behavioral depth (sessions, completion rates, device usage) beyond basic viewing hours.
Key metric: RelBench benchmark: relational models score 76.71 vs 62.44 for single-table baselines, translating to 2-4 weeks earlier churn detection.
Why relational data changes the answer
Churn is not a viewing-hours problem. The signals live across watch history (sessions getting shorter, completion rates declining for preferred genres), payments (failed charge, downgrade inquiry), devices (stopped using the Smart TV, only occasional mobile), and the social graph (household members disengaging, friends who recommended the platform have already cancelled). A flat feature table forces you to pre-aggregate all of this into static columns like 'avg_watch_hours_30d' and 'payment_failures_count,' destroying the temporal sequences that actually predict departure.
Relational models read these tables as a connected graph. They learn that Subscriber SUB102's churn risk is high not because of any single signal, but because of the compound pattern: watch time declined 72% month-over-month, a payment failed this billing cycle, the last active session was 9 days ago, and the subscriber has only one device (no household stickiness). On the RelBench benchmark, relational approaches score 76.71 vs 62.44 for single-table baselines. That gap translates to catching churners 2-4 weeks earlier, when retention interventions actually work.
Predicting churn from viewing hours is like a gym tracking only how often members swipe their badge. You catch the member who stopped coming, but you miss the one who used to do weights, spin class, and swimming but now only walks on the treadmill for 15 minutes. The relational model watches the full pattern of engagement across every touchpoint, detecting the slow disengagement before the member cancels.
How KumoRFM solves this
Graph-powered intelligence for media platforms
Kumo connects subscribers, plans, watch history, payments, and devices into a temporal graph. The GNN learns early churn signals: viewing sessions getting shorter, binge completion rates dropping, payment method failures, and social graph erosion (friends leaving the platform). PQL filters to active subscribers and predicts cancellation in the next 30 days, giving retention teams a 2-4 week intervention window.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
SUBSCRIBERS
| subscriber_id | plan | tenure_months | household_size |
|---|---|---|---|
| SUB101 | Premium | 18 | 3 |
| SUB102 | Standard | 4 | 1 |
| SUB103 | Premium | 24 | 4 |
PLANS
| plan_id | name | monthly_price | max_streams |
|---|---|---|---|
| PL01 | Standard | $9.99 | 2 |
| PL02 | Premium | $15.99 | 4 |
WATCH_HISTORY
| watch_id | subscriber_id | content_id | minutes_watched | timestamp |
|---|---|---|---|---|
| W6001 | SUB101 | MOV101 | 120 | 2025-02-28 |
| W6002 | SUB102 | SER201 | 15 | 2025-02-25 |
| W6003 | SUB103 | MOV305 | 95 | 2025-03-01 |
PAYMENTS
| payment_id | subscriber_id | amount | status | timestamp |
|---|---|---|---|---|
| PAY301 | SUB101 | $15.99 | Success | 2025-03-01 |
| PAY302 | SUB102 | $9.99 | Failed | 2025-03-01 |
| PAY303 | SUB103 | $15.99 | Success | 2025-03-01 |
DEVICES
| device_id | subscriber_id | type | last_active |
|---|---|---|---|
| D401 | SUB101 | Smart TV | 2025-03-01 |
| D402 | SUB102 | Mobile | 2025-02-20 |
| D403 | SUB103 | Smart TV | 2025-03-01 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT BOOL(SUBSCRIBERS.is_cancelled, 0, 30, days) FOR EACH SUBSCRIBERS.subscriber_id WHERE COUNT(WATCH_HISTORY.*, -30, 0, days) > 0
Prediction output
Every entity gets a score, updated continuously
| SUBSCRIBER_ID | PLAN | CHURN_PROB | RISK_TIER |
|---|---|---|---|
| SUB101 | Premium | 0.08 | Low |
| SUB102 | Standard | 0.81 | Critical |
| SUB103 | Premium | 0.15 | Medium |
Understand why
Every prediction includes feature attributions — no black boxes
Subscriber SUB102 -- Standard plan, 4-month tenure
Predicted: 81% churn probability (Critical)
Top contributing features
Watch time decline (30d vs prior 30d)
-72%
32% attribution
Payment failure in billing cycle
1 failed
24% attribution
Days since last active session
9 days
20% attribution
Content completion rate decline
-55%
14% attribution
Single-device household
1 device
10% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about subscriber churn prediction
How early can you predict streaming subscriber churn?
With relational ML, reliable churn signals appear 2-4 weeks before cancellation. The earliest indicators are subtle: binge completion rates declining, session durations shortening, device usage narrowing to one device, and payment method issues. Traditional usage-threshold models detect churn only 3-5 days out, when intervention success rates are below 10%.
What is the best ML model for streaming churn prediction?
Graph neural networks that connect viewing history, payment data, device engagement, and subscriber profiles outperform flat-table classifiers by detecting compound behavioral patterns. The key advantage is capturing temporal sequences and cross-entity signals that aggregated features destroy.
What data do you need for a subscriber churn model?
At minimum: subscriber profiles, watch history with timestamps and completion rates, and payment records. For best results, add device usage data, household membership, plan details, and customer service interactions. More connected tables means earlier and more accurate churn detection.
How do you reduce false positives in churn prediction?
Require convergence across multiple behavioral dimensions. A viewing dip alone is not alarming (holidays, travel). But a viewing dip combined with a payment failure, device abandonment, and declining completion rates is a strong composite signal. Relational models naturally require this convergence because they see the full behavioral context.
What is the ROI of a streaming churn prediction model?
A platform with 30M subscribers at $12/month ARPU saves $43M annually by reducing churn just 1 percentage point. The cost of the model is a fraction of the retention value. Early detection (2-4 weeks ahead) is critical because retention offers (free month, content suggestions, plan adjustments) succeed at 20-30% when delivered early but below 10% when delivered at cancellation.
Bottom line: A 30M-subscriber streaming platform saves $43M annually by reducing churn just 1 percentage point. Kumo detects early signals like social graph erosion and viewing pattern decay weeks before traditional models flag declining usage.
Related use cases
Explore more media & entertainment use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
