Documentation Index

Fetch the complete documentation index at: https://docs.xtremepush.com/llms.txt

Use this file to discover all available pages before exploring further.

Machine Learning Models

Prev Next

Xtremepush Intelligence includes four standard iGaming ML models. Each model is built for iGaming operators and produces results you can use directly for player segmentation and campaign targeting in the Xtremepush platform.

Data integration requirements

Below, lists the standard models available in your Intelligence suite, along with what they cover. This assumes a full input of supporting data from you, as described in Data Integration Requirements. You can discuss the data required with your Account Manager and the Onboarding Team.

The four standard models

Model

Purpose

Type

RFM Segmentation

Classify players into value-based segments

Statistical

Customer Retention Analysis

Track cohort retention over time

Statistical

Player Lifecycle Analysis

Estimate individual survival probability

Statistical

Churn Classification

Score each active player's probability of churning

Predictive

  • Statistical models analyse your historical player data and produce results directly from patterns in that data.

  • Predictive models are trained on labelled historical data, then applied to your current active player base to generate individual predictions.

Each of the four models is covered in full below, including what it does, what it produces, and how to use the results.

Machine Learning model delivery

The Xtremepush data team sets up Machine Learning models as part of your Intelligence onboarding. For each model:

  1. The data team analyses your historical event data to confirm model viability and check that minimum data requirements are met.

  2. The team configures and trains the model on your data.

  3. You review the results before the model goes live.

  4. The team configures computation jobs to push results to the segmentation engine on a scheduled basis (daily or weekly, depending on the model).

For the full onboarding process, including Machine Learning model stages, see Setup Overview.

RFM Segmentation

RFM Segmentation groups your players into segments based on three dimensions of behaviour:

  • Recency: how recently the player made a deposit or placed a bet

  • Frequency: how often they deposit or bet

  • Monetary: how much they wager or deposit

Each player receives an RFM score and is assigned to one of the standard segment labels.

Standard segments

The default segment labels include Champion, Loyal, Potential loyal, New, Promising, Need-attention, About-to-sleep, Can-not-lose, At-risk, Hibernating, and Lost.

You can customise segment labels. Before the model goes live, you'll see the full segment breakdown in your ML sign-off report, including the segment assigned to each group along with player counts and percentages. You can rename segment labels directly in the platform to match your own terminology, before or after go-live.

What it produces

Results are written to a player-level output table. Key columns:

Column

Description

profile_uid

Player identifier

rfm_segment

RFM segment label (for example, Champion, At-risk)

rfm_score

Composite RFM score

RFM segment assignments are pushed to the Xtremepush segmentation engine, so they're immediately available for campaign targeting.

Example use cases:

  • Send reactivation campaigns to At-risk and Hibernating segments before they lapse further.

  • Send loyalty rewards or VIP offers to Champion and Loyal players.

  • Target Potential loyal players with first-deposit bonus offers or cross-sell campaigns.

  • Exclude Lost players from re-engagement spend where propensity is below threshold.

How to read an RFM score

Each dimension is scored from 0 (worst) to 4 (best):

Score

Recency (R)

Frequency (F)

Monetary (M)

0

Lost customer

Buys rarely

Very low spender

1

Lapsing customer

Buys occasionally

Low spender

2

Average active customer

Buys regularly

Medium spender

3

Active customer

Buys frequently

High spender

4

Very active customer

Buys very frequently

Very high spender

An RFM score is written as three digits, one per dimension, in Recency, Frequency, Monetary order. For example, a score of 444 means the player is Very active, Buys very frequently, and is a Very high spender, the best possible result on all three dimensions. A score of 000 means the player is Lost, Buys rarely, and is a Very low spender, the worst possible result on all three.

Each segment maps to a set of RFM score combinations. For example, a Champion customer typically scores 444, 443, 433, 434, 343, 344, or 334: high across all three dimensions, even if not a perfect 4 on every one.

Viewing your model results

Your RFM Segmentation model runs in InfinityAI, accessible from the main menu. Open InfinityAI > Models and select your RFM model (for example, "RFM on Sports Bets").

The model page opens with a header strip showing:

  • Dataset: the dataset the model is built on (for example, Sports Bets Placed)

  • Status: whether the model is Trained

  • Trained: the date it was last trained

A Retrain button in the top right lets you manually retrain the model on demand, outside of any recurring schedule.

Below the header, the page is split into four sections: Results, Configuration, Schedule, and History.

Results

The Results section opens with a summary strip showing your total customer count and the number of players in each segment. Below that, results are broken down across the following subsections.

RFM Scores

A treemap showing customer count by segment, sized so you can see at a glance which segments hold the most players. Below it, a reference table lists every segment: its name, the RFM score combinations that map to it, a plain-language description, and your actual player count and percentage of your total base for each.

RFM Segmentation

Four bar charts comparing every segment side by side on customer count, average monetary value, average frequency, and average recency. A summary table below the charts gives the same breakdown in figures, one row per segment.

Recency, Frequency, and Monetary breakdowns

Three tabs, one per RFM dimension, each breaking your players down by that single dimension alone (for example, Very active, Active, Average active, Lapsing, and Lost for Recency) rather than by combined segment. Each tab includes the score definitions for that dimension along with its comparison charts and summary table, shown directly in the platform.

RFM Matrix and Details

The RFM Matrix combines all three dimensions into a single expandable table. Start at a Recency segment and drill down through Frequency and then Monetary to see customer counts and averages at every level of the combination. The Details subsection provides further supporting information.

Configuration

Shows the underlying model configuration used to build the segmentation: the column used as the customer identifier, the column used as the monetary value, and the column used as the transaction date. This reflects the same settings used when the model was created in InfinityAI.

Schedule

Lets you turn on a recurring schedule so the model automatically retrains at a set frequency and timezone, as an alternative to using the manual Retrain button. When the model reruns, the underlying data, visualisations, and any CRM syncs refresh automatically.

History

A log of every model run, whether triggered manually or by the schedule. It shows how many runs are completed, scheduled, or failed overall, and a table of individual runs with their start time, trigger, status, progress, duration, and any error.

Customer Retention Analysis

What it does

The Cohort Analysis model groups players by the period in which they first became active, their acquisition cohort, and tracks what percentage of each cohort remains active in each subsequent period.

This gives you a structured view of how well you retain players over weeks or months after acquisition, and lets you compare retention across cohorts acquired in different periods.

What it produces

Results are written to a cohort-level output table, with rows for each cohort period and columns for each subsequent retention period. Retention rates are calculated at weekly or monthly intervals (configurable).

How to use it

Cohort analysis is primarily a diagnostic and planning tool. It helps you identify:

  • How quickly you lose new players (early-period drop-off).

  • Which acquisition cohorts perform best over the long term.

  • Whether retention is improving or declining over time as you iterate on CRM programmes.

  • The typical point at which most players become inactive, which informs re-engagement timing.

Player Lifecycle Analysis

What it does

The Survival Analysis model estimates the probability that a player remains active over time. It calculates survival curves for your player population, showing the likelihood that a given player is still active at 30, 60, 90, 180, and 365 days.

The model also identifies which player characteristics are most strongly associated with churn risk. These hazard ratios show which attributes, such as deposit frequency, preferred vertical, and acquisition source, are most predictive of early exit. This output informs the feature selection for the Churn Classification model.

What it produces

Player-level output with survival probability estimates at each standard time horizon:

Column

Description

profile_uid

Player identifier

survival_30

Probability of remaining active at 30 days

survival_60

Probability at 60 days

survival_90

Probability at 90 days

survival_180

Probability at 180 days

survival_365

Probability at 365 days

The model also produces a separate feature importance output identifying the hazard ratio (impact on churn risk) for each attribute.

How to use it

Survival analysis informs your understanding of player lifecycle length and at-risk periods. It's particularly useful for:

  • Setting campaign timing: knowing when most players exit tells you the optimal intervention window.

  • Identifying which attributes most strongly predict early churn, so you can act on them at acquisition.

  • Understanding how lifecycle length differs across player segments or acquisition channels.

  • Informing the feature selection for the Churn Classification model

Churn Classification

What it does

The Churn Classification model scores each active player with a probability that they'll churn, defined as ceasing to place bets, within a specified future window.

Unlike the Survival model, which describes population-level lifecycle patterns, Churn Classification makes an individual prediction for every active player. This enables proactive, targeted intervention before churn occurs.

How it's built

The model is trained on historical player data using features derived from your Intelligence data model, including deposit patterns, wagering behaviour, preferred vertical, recency, and others identified by the Survival Analysis. The training process finds which combinations of features best predict churn.

Once trained, the model is applied to your current active player base on a scheduled basis, daily or weekly, and scores are kept current.

What it produces

Two profile attributes are written to the Xtremepush segmentation engine for each active player:

Attribute

Description

churn_score

Continuous churn probability (0 to 1). Higher values indicate greater churn risk: 0 means very likely to remain active, 1 means very likely to churn.

churn_tier

Categorical risk band derived from the score: low, medium, high, or critical.

The tier boundaries are:

Tier

Churn risk

low

Player is likely to remain active

medium

Moderate churn risk, worth monitoring

high

Elevated churn risk, intervention recommended

critical

High churn risk, priority for retention activity

How to use it

Both churn_score and churn_tier are available immediately in the Xtremepush segmentation engine.

Example use cases:

  • Target all players in the high and critical tiers with retention bonus offers before they lapse.

  • Tier your re-engagement spend: give critical players with high GGR premium treatment, and critical players with low GGR lower-cost touchpoints.

  • Exclude low tier players from re-engagement campaigns to avoid wasted spend.

  • Monitor the proportion of your active base in the critical tier over time as a leading indicator of retention health.

On model evaluation: When reviewing your ML sign-off report, check Recall, the proportion of players who did churn that the model correctly flagged in advance. A high-recall model catches most churners early. Missing a churner costs more than over-targeting a retained player, so recall matters more than overall accuracy.

Data requirements and staged deployment: The Churn Classification model requires sufficient historical data, typically 90 or more days of event history and an adequate active player population, to produce reliable predictions. If these thresholds aren't met at the time of your Intelligence deployment, the model is given a Revisit status in your ML sign-off report rather than being deployed immediately. The data team reassesses at the agreed future date and deploys once the data conditions are met. The other three models can still go live in the interim.