In-App Purchase Prediction
“Which players will make an in-app purchase?”
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 make an in-app purchase?
Only 2-5% of free-to-play players ever make a purchase. Showing the wrong offer at the wrong time trains players to ignore your store entirely. A game generating $30M in annual IAP revenue that improves conversion from 3% to 4% adds $10M. The signal is not in demographics alone; it is in the sequence of gameplay behaviors, social influences, and store browsing patterns that precede a first purchase.
Quick answer
The most effective way to predict IAP conversion is to connect store browsing behavior, level progression, session patterns, and social influence data in a graph-based ML model. Players who browse specific item categories after failing hard levels, while their guild mates have recently purchased, convert at 8x the base rate. SAP SALT benchmark shows relational graph ML achieves 91% accuracy vs 75% for XGBoost on flat tables.
Approaches compared
4 ways to solve this problem
1. Heuristic segmentation
Bucket players into segments (whale, dolphin, minnow, free) based on install cohort and early session behavior, then target each segment with pre-built offers.
Best for
Fast to implement. Works for broad-stroke monetization when you do not have a data science team.
Watch out for
Segments are static and miss individual timing. A player ready to buy right now gets the same treatment as one who will never spend.
2. Logistic regression on player features
Train a purchase propensity model on aggregated features like days since install, total sessions, and level reached.
Best for
Interpretable model that gives a directional read on which player attributes correlate with spending.
Watch out for
Cannot capture the temporal sequence of events leading to a purchase. The order of browsing, failing, and browsing again matters more than the totals.
3. Collaborative filtering
Recommend items based on what similar players purchased, using player-item interaction matrices.
Best for
Decent for item recommendation once a player has some purchase history to anchor similarity.
Watch out for
Cold-start problem is severe for first-purchase prediction. Most free players have zero purchase history, so there is nothing to collaborate on.
4. KumoRFM (relational graph ML)
Point Kumo at your players, sessions, store views, purchases, and level progress tables. The GNN learns the temporal, cross-table journey from install to first purchase.
Best for
Best accuracy for first-purchase prediction. Captures store browsing intent, progression frustration, and guild spending norms in a single model.
Watch out for
Requires normalized relational tables with clear foreign keys. Will not help if your monetization data is a single flattened event stream.
Key metric: SAP SALT benchmark: relational graph ML achieves 91% accuracy vs 75% for XGBoost on flat tables in purchase propensity tasks.
Why relational data changes the answer
The journey from free player to first purchase is not captured in any single table. The signal lives across store views (browsing frequency and categories), level progress (frustration from repeated failures), sessions (increasing engagement depth), and social connections (guild mates who recently purchased). A flat feature table reduces this rich behavioral sequence to static aggregates like 'total store views = 8,' destroying the temporal patterns that distinguish genuine purchase intent from idle curiosity.
Relational models connect these tables and learn sequences like 'player who failed level 28 seven times, then browsed the gem pack category three times in two days, while four of six guild members recently bought.' SAP SALT benchmark shows relational graph ML achieves 91% accuracy vs 75% for XGBoost on flat tables in propensity tasks. For a game with 2M DAU, that accuracy gap is the difference between a well-timed offer that converts and a generic pop-up that trains players to dismiss your store.
Predicting a first purchase from aggregated stats is like predicting a car sale from the number of times someone visited a dealership lot. You miss that they test-drove the specific model twice, asked about financing, and their neighbor just bought the same car. The purchase decision is a sequence of connected events across multiple touchpoints. Graph ML reads that entire sequence instead of counting lot visits.
How KumoRFM solves this
Graph-learned player intelligence across your entire game ecosystem
Kumo models the journey from install to first purchase as a relational graph connecting sessions, store views, level progress, and social connections. It learns that players who view a specific item category 3+ times after failing a hard level, while their guild mates have recently purchased, convert at 8x the base rate. The model distinguishes between curiosity browsing and purchase intent by analyzing temporal patterns across the player network.
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 | total_spend |
|---|---|---|---|
| PLR101 | 2025-02-01 | iOS | $0.00 |
| PLR102 | 2025-01-15 | Android | $0.00 |
| PLR103 | 2025-02-20 | iOS | $4.99 |
SESSIONS
| session_id | player_id | timestamp | duration_min | store_visits |
|---|---|---|---|---|
| S101 | PLR101 | 2025-03-02 | 38 | 3 |
| S102 | PLR102 | 2025-03-01 | 12 | 0 |
| S103 | PLR103 | 2025-03-02 | 55 | 1 |
STORE_VIEWS
| view_id | player_id | item_id | category | timestamp |
|---|---|---|---|---|
| SV01 | PLR101 | ITM_GEM500 | Currency | 2025-03-02 |
| SV02 | PLR101 | ITM_SKIN_DRAGON | Cosmetic | 2025-03-02 |
| SV03 | PLR102 | ITM_GEM100 | Currency | 2025-02-28 |
PURCHASES
| purchase_id | player_id | item_id | amount_usd | timestamp |
|---|---|---|---|---|
| PUR101 | PLR103 | ITM_BATTLEPASS | 9.99 | 2025-02-25 |
LEVEL_PROGRESS
| progress_id | player_id | level | attempts | completed |
|---|---|---|---|---|
| LP01 | PLR101 | 28 | 7 | N |
| LP02 | PLR102 | 12 | 2 | Y |
| LP03 | PLR103 | 35 | 1 | Y |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT BOOL(PURCHASES.*, 0, 7, days) FOR EACH PLAYERS.PLAYER_ID WHERE PLAYERS.TOTAL_SPEND = 0
Prediction output
Every entity gets a score, updated continuously
| PLAYER_ID | PLATFORM | DAYS_SINCE_INSTALL | IAP_PROB_7D |
|---|---|---|---|
| PLR101 | iOS | 29 | 0.68 |
| PLR102 | Android | 45 | 0.09 |
Understand why
Every prediction includes feature attributions — no black boxes
Player PLR101 -- iOS, Day 29, $0 spend
Predicted: 68% IAP probability within 7 days
Top contributing features
Store views (last 3d)
8 views, 3 categories
30% attribution
Level fail-retry pattern
7 attempts on L28
25% attribution
Guild member purchase rate
4 of 6 purchased
19% attribution
Session duration trend
+15% last 7d
14% attribution
Cosmetic store dwell time
4.2 min avg
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 in-app purchase prediction
How do you predict in-app purchase conversion in free-to-play games?
Connect store browsing data, level progression events, session patterns, and social graph data into a relational model. The strongest signals are temporal sequences: a player who fails a level repeatedly, then browses the store multiple times, while their friends have recently purchased, is showing genuine intent. Flat propensity models miss these cross-table sequences.
What data drives IAP prediction accuracy?
Four tables matter most: session logs (engagement depth), store interaction data (browsing frequency and categories), level progression (frustration signals), and social connections (spending norms within friend groups). Adding guild-level spending data alone can improve prediction accuracy by 10-15%.
How do you identify first-time buyers in mobile games?
First-time buyer signals are different from repeat-buyer signals. Look for convergence of store browsing acceleration, progression frustration, and social proof (guild mates purchasing). Relational models are especially strong here because collaborative filtering has no purchase history to work with for never-purchased players.
What conversion rate improvement can IAP prediction deliver?
Games that target offers based on graph ML purchase intent typically lift IAP conversion from 2-3% to 5-8% among targeted segments. For a game with 2M DAU generating $30M in annual IAP, improving overall conversion by 1 percentage point adds $10M in revenue.
Should you use collaborative filtering or ML for game monetization?
Collaborative filtering works for item recommendation among existing spenders, but fails for first-purchase prediction where the player has no purchase history. Graph ML on relational data solves both: it predicts who will buy and what they will buy by connecting behavioral sequences across sessions, progression, and social data.
Bottom line: A free-to-play game with 2M DAU that converts 1% more free players to payers adds $10M in annual IAP revenue. Kumo detects purchase intent signals across store behavior, progression frustration, and social influence that propensity models on flat feature tables cannot learn.
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.
