Cohort & time controls

Model hyper-parameters

BG/NBD population parameters (used both to generate each customer's hidden purchase-rate λ and dropout probability p, and as the model's prior): r=0.90, α=3.20 (Gamma on λ), a=1.30, b=2.60 (Beta on p). This mirrors a converged MLE fit on a real transaction panel.

Portfolio stats

Customers tracked—
Sim clock—
Total predicted value—
Mean P(alive)—

Top customers by predicted CLV

    Prediction validity check

    Matured customers—
    Mean abs. error (purchases)—
    Corr(predicted, actual)—
    A customer "matures" once a full holdout window (2×T after acquisition) has elapsed. At the T mark we snapshot the model's predicted future-purchase count; once the holdout window closes we compare it against the transactions that actually happened — a genuine out-of-sample accuracy check.

    Selected customer

    Customerclick a point
    Recency t_x—
    Frequency x—
    Obs. window T—
    P(alive)—
    E[future purchases]—
    Predicted order value—
    Predicted CLV—