Player Churn Prediction
“Which players will stop playing within 7 days?”
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 players will stop playing within 7 days?
Mobile games lose 75% of players within the first 3 days. Acquiring each player costs $2-$8 via paid UA, and a game with 5M MAU losing 20% of monetizing players monthly leaves $18M in annual revenue on the table. Generic retention campaigns treat every player the same, wasting live-ops resources on players who were never going to stay while missing the ones teetering on the edge.
Quick answer
The most accurate way to predict player churn is to combine session behavior, progression data, purchase history, and social connections in a graph-based ML model. Players whose friends leave, who hit level-difficulty walls, and whose session lengths decline sharply are the strongest churn signals. On the RelBench benchmark, relational models score 76.71 vs 62.44 for single-table baselines.
Approaches compared
4 ways to solve this problem
1. Cohort-based retention curves
Track D1/D7/D30 retention rates by install cohort and flag cohorts that fall below historical baselines.
Best for
Quick directional read on overall retention health across acquisition channels.
Watch out for
Treats every player in a cohort the same. Cannot identify which specific players within a healthy cohort are about to leave.
2. XGBoost on aggregated player features
Build a wide feature table with session counts, spend totals, and level progress, then train a gradient-boosted classifier.
Best for
Solid baseline when you have a data science team that knows the game deeply and can hand-craft features.
Watch out for
Misses social contagion entirely. When a player's guild goes inactive, that signal lives in the friend graph, not in a flat row.
3. Time-series models on session data
Use ARIMA or Prophet on per-player session frequency and duration to detect declining engagement trends.
Best for
Works well for detecting gradual disengagement in solo-play games with predictable session patterns.
Watch out for
Cannot incorporate progression events, social signals, or purchase behavior. The churn trigger is often a specific frustration moment, not a gradual fade.
4. KumoRFM (relational graph ML)
Point Kumo at your players, sessions, purchases, achievements, and social connections tables. The GNN learns cross-table temporal patterns automatically.
Best for
Highest accuracy with minimal feature engineering. Captures friend-graph churn contagion, progression walls, and spending pattern shifts in one model.
Watch out for
Requires relational data in normalized tables. Not the right fit if all your telemetry lives in a single denormalized event log.
Key metric: RelBench benchmark: relational models score 76.71 vs 62.44 for single-table baselines on player churn prediction tasks.
Why relational data changes the answer
Player churn is a multi-table event. The signals live across sessions (declining duration and frequency), achievements (stalled progression at a specific level), purchases (increasing time between transactions), and social connections (guild members going inactive). A flat feature table forces you to pre-aggregate all of this into static columns, destroying the temporal sequences and social network effects that actually predict departure.
Relational models read these tables as a connected graph and learn patterns like 'player whose three closest guild members stopped playing last week, who has been stuck on level 15 for five days, and whose session duration dropped 72% in three days.' On the RelBench benchmark, this multi-table approach scores 76.71 vs 62.44 for single-table baselines. In gaming, that gap translates directly into millions of dollars: the difference between catching a player during the frustration window where an intervention works and catching them after they have already uninstalled.
Predicting churn from session count alone is like predicting whether a sports fan will cancel their season tickets by counting stadium visits. You miss that their best friend stopped coming, the team lost eight straight home games, and the parking situation got worse. The full picture requires connecting the fan to their social circle, the team's performance, and the venue experience. That connected view is exactly what graph ML provides for player retention.
How KumoRFM solves this
Graph-learned player intelligence across your entire game ecosystem
Kumo connects players, sessions, purchases, achievements, and social connections into a single relational graph. It learns that players whose guild members have gone inactive, who have hit a specific level-difficulty wall, and whose session lengths have dropped 40% over 3 days are 6x more likely to churn. The social graph signal is critical: when a player's friends leave, that player follows within days. Traditional cohort models miss these network effects entirely.
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
PLAYERS
| player_id | install_date | platform | country | level |
|---|---|---|---|---|
| PLR001 | 2025-01-05 | iOS | US | 42 |
| PLR002 | 2025-02-18 | Android | JP | 15 |
| PLR003 | 2025-01-22 | iOS | US | 67 |
SESSIONS
| session_id | player_id | start_time | duration_min | levels_played |
|---|---|---|---|---|
| S001 | PLR001 | 2025-03-01 18:30 | 45 | 3 |
| S002 | PLR002 | 2025-03-02 09:15 | 8 | 1 |
| S003 | PLR003 | 2025-03-01 21:00 | 62 | 5 |
PURCHASES
| purchase_id | player_id | item | amount_usd | timestamp |
|---|---|---|---|---|
| PUR01 | PLR001 | Gem Pack 500 | 4.99 | 2025-02-20 |
| PUR02 | PLR003 | Battle Pass | 9.99 | 2025-02-15 |
ACHIEVEMENTS
| achievement_id | player_id | name | unlocked_date |
|---|---|---|---|
| ACH01 | PLR001 | Boss Slayer III | 2025-02-28 |
| ACH02 | PLR003 | Guild Leader | 2025-02-10 |
SOCIAL_CONNECTIONS
| connection_id | player_id | friend_id | type |
|---|---|---|---|
| SC01 | PLR001 | PLR003 | Guild |
| SC02 | PLR002 | PLR001 | Friend |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT COUNT(SESSIONS.*, 0, 7, days) = 0 FOR EACH PLAYERS.PLAYER_ID WHERE COUNT(SESSIONS.*, -7, 0, days) > 0
Prediction output
Every entity gets a score, updated continuously
| PLAYER_ID | PLATFORM | LEVEL | CHURN_7D_PROB |
|---|---|---|---|
| PLR001 | iOS | 42 | 0.14 |
| PLR002 | Android | 15 | 0.83 |
| PLR003 | iOS | 67 | 0.06 |
Understand why
Every prediction includes feature attributions — no black boxes
Player PLR002 -- Android, Level 15, Japan
Predicted: 83% churn probability within 7 days
Top contributing features
Session duration trend (7d)
-72% decline
31% attribution
Level progression stall
Stuck at L15 for 5d
24% attribution
Friend activity (active friends)
0 of 3 active
19% attribution
Days since last purchase
Never purchased
14% attribution
Tutorial completion rate
60%
12% 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 player churn prediction
What is the best ML model for player churn prediction?
Graph neural networks that operate on relational data (sessions, purchases, social connections, achievements) outperform flat-table models. On the RelBench benchmark, relational models score 76.71 vs 62.44 for single-table approaches. The key advantage is capturing social churn contagion and progression-frustration sequences that XGBoost on aggregated features cannot learn.
How early can you predict player churn in mobile games?
With relational ML, reliable churn signals appear 5-7 days before a player leaves. The earliest indicators are social: when a player's friends and guild members go inactive, that player follows within days. Traditional cohort models detect churn only after the player has already stopped playing.
What data do you need for a gaming churn model?
At minimum: player profiles, session logs, and purchase history. For best results, add achievement data, social connections (friends, guild membership), and level progression events. More connected tables means more signal. The social graph alone can improve prediction accuracy by 15-20%.
How do you reduce false positives in player churn prediction?
Most false positives come from models reacting to a single signal like a missed day. Relational models reduce false positives by requiring convergence across multiple dimensions: a missed day alone is normal, but a missed day plus declining session length plus inactive friends plus a progression stall is a strong composite churn signal.
What is the ROI of a player churn prediction model?
A mobile game with 5M MAU and 20% monthly churn among monetizing players can save $18M annually by retaining just 10% more at-risk spenders through targeted interventions (personalized offers, difficulty adjustments, re-engagement campaigns). The model pays for itself within the first month.
Bottom line: A mobile game with 5M MAU that retains just 10% more of its at-risk monetizing players saves $18M annually. Kumo captures friend-graph churn contagion and progression-wall patterns that cohort analytics miss, letting live-ops target the players who can still be saved.
Related use cases
Explore more gaming 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.
