iGaming AI CRM Analytics: Segmentation, Churn Prediction, Bonus Optimization

A large iGaming operator with a 2 million player base was losing 30% of profitable customers within the first 90 days. Mass bonuses attracted bonus hunters, campaign ROI hovered around 80%—each bonus ruble returned 80 kopecks. After implementing our system, the operator reduced losses: bonus budget cut by $200,000 annually and LTV increased by 20%, generating an additional $1.5 million.

We developed an AI CRM analytics system based on an ensemble of GradientBoosting (Gradient boosting) and KMeans segmentation. Within 6 weeks of deployment, the operator achieved 250% ROI, reduced the bonus budget by 20%, and cut churn rate by 18%. The key change was transitioning from generic RFM to GGR segmentation, which accounts for actual player profitability.

The system includes Airflow pipelines, ClickHouse storage, and Metabase dashboards. For each segment (champion, at_risk_highvalue, loyal_lowvalue, dormant), an individual bonus mechanic with different wagering requirements is selected. Built-in RG checks block bonuses for players with high risk of problematic behavior. The system updates segments and models weekly, adapting to changes in player behavior.

Problems We Solve

• Churn prediction: Standard RFM models ignore trends. We build features: session change over 7/30 days, GGR trend, support complaints. GradientBoosting achieves AUC 0.87—12% better than logistic regression.
• Bonus abuse: Mass bonuses attract hunters. Segmentation identifies 4 groups: champion (5–10% of base, 50–70% of GGR), at_risk_highvalue, loyal_lowvalue, dormant. Each receives a specific bonus type with different wagering.
• RG compliance: Built-in checks block bonuses for high RG-risk players, reducing regulatory complaints.

How to Predict Player Churn in 30 Days

We use an ensemble of GradientBoosting with parameters: n_estimators=200, learning_rate=0.05, max_depth=4. Train on 6 months of data. Features include days_since_last_session, sessions_trend, ggr_trend, bonus_expiry_ignored. The model outputs churn probability and breaks into tiers: low, medium, high, critical. For critical tier, a reactivation campaign with cashback is automatically triggered.

Why GGR Segmentation Outperforms RFM

RFM considers only recency, frequency, and monetary value. GGR = deposits – withdrawals – bonuses. It more accurately reflects true player value. For example, a player with high RFM but frequent withdrawals and bonuses is unprofitable. GGR segmentation identifies 'champion' (high GGR + activity) and 'at_risk_highvalue' (high GGR, low activity)—these are the ones to focus resources on.

How We Do It

Stack: Python, scikit-learn, PyTorch, Hugging Face (for NLP on support chats), ClickHouse (real-time analytics), Airflow (pipelines), Metabase/Tableau (dashboards). Deployed in Docker/Kubernetes.

Example segmentation code:

import pandas as pd
import numpy as np
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.cluster import KMeans

class iGamingPlayerSegmentor:
    """Segmentation by GGR and behavior"""

    def compute_player_value(self, player_data: pd.DataFrame) -> pd.DataFrame:
        """
        GGR (Gross Gaming Revenue) and LTV calculation.
        GGR = deposits - withdrawals - bonuses_paid
        """
        df = player_data.copy()
        df['ggr'] = df['total_deposits'] - df['total_withdrawals'] - df['bonuses_paid']
        df['ggr_per_month'] = df['ggr'] / df['months_active'].clip(1)
        df['deposit_frequency'] = df['deposit_count'] / df['months_active'].clip(1)
        df['withdrawal_ratio'] = df['total_withdrawals'] / df['total_deposits'].clip(1)

        # LTV forecast: simple linear projection
        df['predicted_ltv_12m'] = df['ggr_per_month'] * 12 * (
            1 - df['churn_probability_30d'].fillna(0.3)
        )

        return df

    def segment_players(self, player_value: pd.DataFrame) -> pd.DataFrame:
        """Segmentation by value and activity"""
        df = player_value.copy()

        ggr_median = df['ggr'].median()
        activity_median = df['sessions_last_30d'].median()

        def classify(row):
            high_ggr = row['ggr'] > ggr_median
            active = row['sessions_last_30d'] > activity_median

            if high_ggr and active:
                return 'champion'
            elif high_ggr and not active:
                return 'at_risk_highvalue'
            elif not high_ggr and active:
                return 'loyal_lowvalue'
            else:
                return 'dormant'

        df['segment'] = df.apply(classify, axis=1)

        return df


class ChurnPredictoriGaming:
    """Churn prediction for iGaming"""

    def __init__(self):
        self.model = GradientBoostingClassifier(
            n_estimators=200, learning_rate=0.05, max_depth=4, random_state=42
        )

    def build_features(self, players: pd.DataFrame) -> pd.DataFrame:
        return pd.DataFrame({
            'days_since_last_session': players['days_since_last_session'],
            'sessions_trend': (players['sessions_last_7d'] - players['sessions_prev_7d']) / (players['sessions_prev_7d'] + 1),
            'ggr_trend': (players['ggr_last_30d'] - players['ggr_prev_30d']) / (abs(players['ggr_prev_30d']) + 1),
            'deposit_count_7d': players['deposit_count_7d'],
            'withdrawal_request': players['has_pending_withdrawal'].astype(int),
            'bonus_expiry_ignored': players['bonus_expiry_ignored'].astype(int),
            'support_complaint': players['has_support_complaint'].astype(int),
            'avg_session_duration_trend': players['avg_session_duration_trend'],
        }).fillna(0)

    def predict(self, players: pd.DataFrame) -> pd.DataFrame:
        X = self.build_features(players)
        probs = self.model.predict_proba(X)[:, 1]
        result = players[['player_id']].copy() if 'player_id' in players.columns else pd.DataFrame(index=players.index)
        result['churn_prob_30d'] = probs
        result['churn_tier'] = pd.cut(probs, bins=[0, 0.2, 0.5, 0.75, 1.0],
                                       labels=['low', 'medium', 'high', 'critical'])
        return result


class BonusCampaignOptimizer:
    """Bonus campaign optimization"""

    def design_retention_campaign(self, player_segment: str,
                                   player_stats: dict,
                                   budget_per_player: float) -> dict:
        """Bonus offer tailored to segment and budget"""
        campaigns = {
            'champion': {
                'bonus_type': 'cashback_vip',
                'value': min(player_stats.get('avg_weekly_ggr', 50) * 0.15, budget_per_player),
                'wagering_req': 1,  # Minimum wagering for VIP
                'message': 'Exclusive cashback for our best players'
            },
            'at_risk_highvalue': {
                'bonus_type': 'targeted_reload',
                'value': min(player_stats.get('avg_deposit', 100) * 0.25, budget_per_player),
                'wagering_req': 3,
                'message': 'Special offer — come back to us'
            },
            'loyal_lowvalue': {
                'bonus_type': 'free_spins',
                'value': 20,  # 20 free spins ~ $5-10 value
                'wagering_req': 5,
                'message': 'Loyalty bonus'
            },
            'dormant': {
                'bonus_type': 'reactivation_bonus',
                'value': min(20, budget_per_player),
                'wagering_req': 5,
                'message': "We missed you! Here's a bonus for returning"
            }
        }

        campaign = campaigns.get(player_segment, campaigns['loyal_lowvalue'])

        # RG check: no bonuses for high RG-risk players
        if player_stats.get('rg_risk_level') == 'high':
            return {'bonus_type': 'none', 'reason': 'RG restriction'}

        return campaign

    def calculate_campaign_roi(self, campaign_results: pd.DataFrame) -> dict:
        """ROI of a bonus campaign"""
        total_bonus_cost = campaign_results['bonus_value'].sum()
        incremental_ggr = (
            campaign_results['ggr_post_campaign'] -
            campaign_results['ggr_pre_campaign']
        ).sum()

        return {
            'total_bonus_cost': round(total_bonus_cost, 2),
            'incremental_ggr': round(incremental_ggr, 2),
            'roi_pct': round((incremental_ggr - total_bonus_cost) / max(total_bonus_cost, 1) * 100, 1),
            'reactivation_rate': (campaign_results['returned_to_active'] > 0).mean(),
            'bonus_abuse_rate': (
                campaign_results['withdrawal_after_bonus'] > campaign_results['bonus_value'] * 0.9
            ).mean()
        }

Our Process

1. Analytics: audit current CRM, collect data (DWH, ClickHouse), identify segments.
2. Design: define metrics (GGR, LTV, churn rate), select models.
3. Implementation: Airflow pipelines, model training, integration with bonus system.
4. Testing: A/B test on 10% traffic, compare ROI.
5. Deploy: production rollout, dashboards, documentation.

What’s Included

• Audit of current CRM and data
• DWH setup (ClickHouse, PostgreSQL)
• ML models for segmentation, churn prediction, bonus optimization
• Integration with bonus system (REST API)
• Dashboards in Metabase/Tableau
• Team training (2 workshops)
• 1 month technical support

Estimated Timeline

Segmentation Metrics Comparison

Typical Mistakes

• Ignoring RG: without built-in RG checks, bonuses may incentivize problem players, leading to fines.
• Pure RFM: RFM ignores GGR and trends, leading to false priorities.
• Mass bonuses: without segmentation, bonus budget is wasted on hunters, ROI drops below 100%.
• Weak infrastructure: without DWH and pipelines, data is stale and models perform poorly.

We guarantee a minimum 30% increase in bonus campaign ROI based on A/B test results. Contact us for a consultation—we'll assess your project in 2 days. Order the system implementation and see first results in just 4 weeks.

Recommender System Development: From Collaborative Filtering to Real-Time Serving

On one e-commerce project with a catalog of 300k SKUs, we boosted CTR from 1.8% to 4.4% — a 2.4x increase. The first leap came from switching from 'popular in the last 7 days' to collaborative filtering; the second from adding content features and re-ranking. The difference between showing popular items and showing personalized recommendations is measurable and significant. Below is the engineering experience that made this possible, along with architectures that actually work in production.

Collaborative Filtering: Matrix Factorization and Neural Approaches

Matrix Factorization is the classic approach for implicit feedback (clicks, views, purchases without explicit ratings). ALS (Alternating Least Squares) from the Implicit library handles user×item matrices with hundreds of millions of non-zero values in minutes on GPU. Latent factors 64–256, regularization λ=0.01–0.1 are starting parameters. Cold start problem: no history for new users or items — pure CF fails; content features or hybrid approach needed.

Neural Collaborative Filtering (NCF) replaces the dot product with a neural network. In practice, the gain over a well-tuned ALS is modest, but NCF is easier to extend with additional features (age, category, time of day). Sequence-aware models (SASRec, BERT4Rec) account for the order of interactions — state-of-the-art for session-based recommendations.

How to Choose Recommender System Architecture?

The answer depends on data, load, and cold start requirements. Below are three main approaches with selection criteria.

Hybrid architecture outperforms pure CF by 20–40% in long-tail coverage — validated on catalogs from 100k SKU.

Content-Based Filtering: When Interaction History is Scarce

Content-based recommends based on item characteristics rather than other users' behavior — solves cold start for new items. Text embeddings via sentence-transformers (multilingual-e5-base, BGE-M3) → similarity search using FAISS IndexFlatIP — query in <5 ms for 100k items. Item2Vec (Word2Vec on view sequences) yields interpretable 'similar items' in a couple hours of training.

Structured features (category, brand, price) are fed through embedding layers or gradient boosting — CatBoost handles categories without manual encoding.

Why Hybrid Models Work Better?

Production systems are almost always two-level. Stage 1 (Retrieval) — fast selection of 100–500 candidates from 300k items using ALS or Two-Tower model with vector search (FAISS, Qdrant). Stage 2 (Ranking) — heavy ranker on LightGBM or neural network with cross-features, time, device, and session context. LightFM is a good starting point for medium scale without heavy infrastructure. Our practice shows: moving from single-stage to two-stage yields a 15–25% accuracy improvement with only 20–30 ms additional latency.

Real-Time Serving: Architecture Under Load

Latency SLA — 50–100 ms at thousands of requests per second. Base recommendations precomputed (batch job hourly) → Redis by user_id → <5 ms. Real-time re-ranking via Kafka for events (clicks, cart adds) → update of context features. Feature serving — Redis with TTL (views in 24 hours, last clicked item). At 10k req/s, we deploy Redis Cluster with replication.

A/B testing is the only reliable way to measure improvements. Offline metrics do not always correlate with online. Kohavi et al., 'Online Controlled Experiments at Large Scale' (KDD 2013) — a must-read for the team. Test on 5–10% of traffic, monitor CTR, conversion, revenue per session. One of our client systems after hybridization increased revenue by 18% over a month of A/B.

Recommender System Development Timeline

The stages and typical time frames are in the table below. Costs are calculated individually based on catalog scale and latency requirements.

What's Included in Turnkey Development

1. Data audit — user×item matrix density (typically <0.1%), activity distribution, temporal patterns, cold start statistics.
2. Baseline — 'popular' as a simple threshold that is often hard to beat.
3. Iterative improvement — ALS → content features → two-stage → sequence-aware. Each step with A/B.
4. Serving infrastructure — batch precomputation, Redis, real-time re-ranking, Grafana monitoring.
5. Documentation — model card with metrics, deployment instructions, feature descriptions.
6. Team training — session on interpreting results and model fine-tuning.
7. Support — 1 month post-launch (incident fixes, pipeline tuning).

We are a team with 7+ years of experience in recommender systems, having delivered over 30 projects for e-commerce and media. We guarantee transparent A/B testing and documented metric improvements.

Want to assess the growth potential of your catalog? Contact us for a free data audit. Order recommender system development — first prototype within two weeks.

from implicit.als import AlternatingLeastSquares

model = AlternatingLeastSquares(
    factors=64,
    regularization=0.05,
    iterations=15,
    use_gpu=True
)
model.fit(user_item_matrix)

More about the mathematics of recommender systems — in specialized literature.