🎯 Reelme IIP555 β€” Cohort Prediction Suite

Production daily Ensemble v3.2 (80% ENet + 20% Cox, age-aware) + country CPU features Β· self-eval 0.9576 Β· 5-agent benchmark Β· last predict: β€”

πŸ“Š Today's Summary β€” Ensemble v3.2 Predictions (CPU country)

All 0 active cohorts (cost β‰₯ $500). Ensemble v3.2 = age-aware blend (Cox-heavy <14d / 50-50 <60d / ENet-heavy β‰₯60d) + country CPU features (cpu, cpu_us, cost_us_share). Self-eval 0.9576. Retrained daily at 11:30 AM.

πŸ—“οΈ All Cohorts β€” Click row to expand curve

Risk:
Age:
β€” cohorts
CohortAge CostCPP ROAS D0 US%Yr% Obs TDPred D360Pred BE RiskReason

πŸ”¬ Last Week Summary

Pairs added in past 7 days. Empty = no cohorts reached a new milestone yet, or no archived predictions to compare. Backtest accumulates daily.

πŸ“‰ Drift over time β€” MAPE per milestone

Rolling MAPE for each observable milestone. Lower = model more accurate. Trend up = drift, may trigger retrain.

🎯 Per-Milestone Backtest Pairs

β€” pairs
CohortMilestonePredict Date Predict AgePred ObsAbs ErrMAPE

πŸ”„ Retrain Log

DateTrigger OLD MAPENEW MAPEΞ” DeployedReason
Weekly retrain runs Sunday 13:00. Auto-rollback if NEW MAPE worse than OLD by >5pp.

πŸ”¬ Test Version

πŸ₯‡ Overall Leaderboard

RankAgentTotal (weighted)Total (simple) E1 MatureE2 MidE3 Young E4 ColdE5 OOT
Weighted: Exam 3 Γ—1.2, Exam 4-5 Γ—1.5 (cold-start = real-world value). Composite = 0.5Β·(1βˆ’MAPE_D360) + 0.25Β·BE_acc + 0.25Β·Risk_acc.

🎯 Strengths Heatmap

Green = strong, Red = weak. Identifies where each model excels.

πŸ” Per-Cohort Drill-Down

Prediction table (5 agents vs truth)

MilestoneTruth 🌳 LGBπŸ“ˆ ENet⏳ Cox 🎲 BayesπŸ”¬ Gompertz

πŸ“Š Per-Exam Cohort Detail β€” Actual Cohort Results

πŸ”Ή Cohort Context (raw features)

CohortAgeCostInstall CPIPay RatePaid US US%EU% Wk%Yr%IAP% FeeBE?

πŸ”Ή Truth ROAS Curve (actual observed)

CohortAge D0D3D7D14 D28D60D90 D180TD (now) Cancel D28Renew D28
β€” = milestone chΖ°a Δ‘αΊ‘t (cohort cΓ²n quΓ‘ young). Cancel/Renew rate = paid users cancelled/renewed by D28.

πŸ”Ή 5-Agent D360 Predictions vs Truth

CohortTruth D360 🌳 LGBπŸ“ˆ ENet⏳ Cox 🎲 BayesπŸ”¬ Gompertz Best

πŸ“– Methodology

  • Data: 223 IIP555 cohorts (Nov 2025 β†’ Jun 2026), stratified 171 train / 50 test.
  • Features (51 total): ROAS curve (9), Cancel signal (8), Renewal (8), Refund (5), Geo US+EU (4), Package mix (5), Cost (5), Velocity (4), Meta (3).
  • 5 Exams: cut-off day determines info visible β€” E1 D180 (easy) β†’ E5 D1 (real-world).
  • Scoring: composite = 0.5Β·(1βˆ’MAPE D360 clipped) + 0.25Β·(BE day within Β±30d) + 0.25Β·(Red/Yellow/Green risk class match).