OverviewHow Should PlayStation Decide Which Players to Invest In?
This project builds a complete analytical pipeline using simulated PlayStation data: from raw session, purchase, and trophy events, through churn prediction and purchase propensity modelling, to a Next Best Action framework that assigns a specific, budget-constrained marketing recommendation to each player.
↓
Player 360 Feature Table (47 features per player)
├──→ Churn Model — who's leaving?
├──→ Propensity Model — who's buying?
└──→ CLV Estimation — what are they worth?
↓
Next Best Action Framework
(player-level recommendations + budget ceilings)
No public dataset captures PS Plus tiers, DLC behaviour, session data, and trophies together. We designed synthetic data that mirrors PlayStation's data structures and embeds realistic behavioural signals. The methodology is fully transferable to real data — with production data, we'd gain richer features (social graph, marketing response history, device signals) but the modelling approach and decision framework remain the same.
Key FindingsWhat the Data Tells Us
Churn is gradual, not sudden. Players who cancel show measurable engagement decay 1–6 months before cancellation — giving a window for intervention.
Revenue concentration is extreme. The top 10% of spenders generate the majority of revenue, justifying the entire CLV modelling exercise.
Auto-renew is the single strongest churn indicator — 85% of churners had it turned off. But combining this structural signal with behavioural trend features yields a substantially better model than either alone.
Free DLC drives paid DLC. Players who claimed the free Valhalla DLC purchased paid DLC at 149% higher rates — validating a specific personalisation hypothesis.
Model 1Churn Prediction
A Gradient Boosting classifier predicts which paid PS Plus subscribers will cancel. The model's probability scores create four risk tiers, each triggering a different intervention strategy.
Our churn model achieves AUC 0.998 — this reflects clean simulated signals, not real-world performance. With production data, expect AUC 0.78–0.85, which is still highly actionable. The project's value is the methodology: evaluation framework, threshold tuning, risk segmentation, and business framing.
Risk tiers are well-calibrated: actual churn rates increase monotonically across tiers. In production, Critical players receive retention offers, High Risk gets re-engagement content, Medium Risk gets nudge campaigns, and Low Risk receives no action.
Model 2Purchase Propensity & the Valhalla Signal
The propensity model predicts paid DLC purchases across all 5,000 players, achieving AUC 0.86 — a realistic, production-believable number. Engagement depth (games played, trophied, total play time) is the primary driver.
Initial AUC hit 1.000. Features like total_spend encoded the target variable. After removing all leaky features, AUC dropped to the realistic 0.86. Recognising and fixing leakage is a critical production skill.
The Valhalla Effect
Free DLC functions as a purchase intent signal. Track free DLC claims in real time and route claimers into targeted paid content recommendation flows. In the global model, broad engagement outranks this single flag — but for game-specific personalisation, it's directly actionable.
Model 3Customer Lifetime Value
A hybrid CLV model combines subscription revenue (tier price × 12 months × retention probability) with purchase revenue (recent spend velocity, churn-adjusted). This determines how much PlayStation should invest in each player.
(<$10)
($10–50)
($50–150)
($150–500)
The High segment is 11.6% of players but 36.2% of value — reinforcing why personalised treatment matters. You don't spend the same to retain a $500 player and a $5 player.
The FrameworkNext Best Action
The NBA framework combines churn risk, DLC propensity, and CLV into a single recommendation per player. Each action has a budget ceiling tied to player value (10% of CLV) — the framework is economically self-correcting.
Retain Critical — 0.1%
High churn + high CLV. Personal retention offer: free month, targeted discount.
Retain Standard — 1.1%
Moderate churn risk. Re-engagement: content recommendations, "we miss you" campaigns.
Upsell — 30.1%
Low churn + high propensity. DLC offers, new releases, tier upgrade prompts.
Nurture — 37.8%
Engaged but not purchase-ready. General engagement campaigns. Low investment.
The remaining 30.9% (Monitor) receive no active intervention — re-score next quarter. Knowing when not to spend is as valuable as knowing when to spend.
Player #2847 — Premium tier, 82% churn risk, $187 CLV → Retain Critical. Budget ceiling: $18.70. Action: personal retention offer (free month or targeted discount). This is what every row in the final output table looks like — 5,000 players, each with a specific action and budget.
Estimated Business Impact
| Action | Players | Net Impact | ROI |
|---|---|---|---|
| Retain Critical | 7 | $41 | 0.4x |
| Retain Standard | 53 | $102 | 0.6x |
| Upsell | 1,506 | $753 | 0.2x |
| Nurture | 1,889 | -$94 | -0.1x |
| Total | 5,000 | $801 | 0.2x |
Nurture going slightly negative is correct — the framework identifies where even lightweight spend isn't justified. At PlayStation's scale (100M+ players), these per-5,000 numbers multiply by orders of magnitude.
Looking AheadPlausible Extensions
- Sequence modelling: LSTM or Transformer on raw session event streams to capture temporal patterns aggregated features miss; learned player embeddings powering both predictions and recommendations.
- A/B testing the Valhalla signal: The 149% lift is observational. A randomised experiment with a control group would establish causality and measure true incremental lift.
- Real-time scoring: Event-driven churn updates after every session via streaming infrastructure (Kafka/Flink) and a feature store, enabling real-time intervention triggers.
- Multi-platform identity: Unified player graphs across PS4, PS5, PC, and mobile would substantially improve engagement feature quality.
- Intervention cost calibration: Replace assumed success rates with A/B test results, updated quarterly, to make the impact model self-improving.