Purpose

Model Architecture: BG/NBD captures purchase dynamics (r=1.01, alpha=2.76) while Gamma-Gamma models spending variability (p=4.73, q=1202.75)
Customer Distribution: 50% classified as \"Loyal\" (6 customers, avg_palive=0.97), with 12 total customers spanning CLV range of ~$99,718\u2013$100,759
Recency Dominance: 75% of observations fall in 0-30 day recency window; expected transactions remain stable (~328) across frequency bins, suggesting recency is primary engagement driver
Segment Concentration: Champions and Potential segments each represent 16.7% of customers; \"Lost\" segment shows critically low palive (0.09)
Validation Gap: Prediction errors range 4,605\u201349,096%, with one-time buyers showing 32,702% error, indicating model struggles with

This section validates the BG/NBD model's predictive accuracy by comparing forecasted transactions against actual holdout-period observations across customer frequency groups. Strong alignment between predicted and actual values confirms the model reliably captures customer purchase behavior, which is essential for accurate CLV estimation and segmentation.

Key Findings

Prediction Consistency: Predicted transactions cluster tightly around 327–329 across all frequency groups, showing stable model output
Error Pattern by Frequency: One-time and low-frequency customers exhibit extreme error rates (32,702% and 49,096%), while high-frequency customers (5+ repeats) show minimal error (4,605%)
Sample Size Constraint: Only 2–3 customers per frequency group limits statistical reliability of error estimates

Interpretation

The model demonstrates strong predictive power for repeat customers but struggles with sparse-data segments. High-frequency customers provide sufficient transaction history for accurate parameter estimation, whereas one-time buyers offer minimal signal, making individual-level predictions inherently unreliable. The extreme percentage errors for low-frequency groups reflect small absolute differences magnified by low baseline actuals, not fundamental model failure.

Context

Validation results align with CLV distribution and segment summary findings, where Loyal and Champions segments (higher frequency) show stable CLV estimates. The small sample sizes per group suggest