Why remove outliers for CLV? A single whale customer who spends 100× the median will massively skew averages and distort cluster boundaries. Your model ends up optimizing for unicorns instead of the 99% of customers you can actually influence.

Common approaches: Percentile capping (cap at the 99th percentile), IQR method (remove values beyond Q3 + 1.5 × IQR), log transformation (log(value + 1) to compress scale), or Winsorization (replace extremes with nearest non-extreme value).

1. Calculate raw values per customer: Recency (R) = days since last purchase. Frequency (F) = total transactions in a period. Monetary (M) = average spend per transaction.

2. Score each dimension 1–5 using quintiles. For Recency, lower is better (more recent = higher score). For F and M, higher is better.

3. Combine into segments: Concatenate scores (e.g., R=5, F=4, M=3 → “543”) or sum for a composite score (12/15). Typical segments: “Champions” (555), “At Risk” (155), “Lost” (111).

K-Means groups data points into K clusters by minimizing the distance between each point and its cluster center (centroid). Algorithm: Choose K (K=3 for Low/Mid/High is a good start), randomly place K centroids, assign each customer to nearest centroid, recalculate centroids as cluster means, repeat until stable.

Tips: Standardize RFM values first (z-score normalization). Use the elbow method or silhouette score to validate K. In production, K=3 to K=5 is usually sufficient.

A Random Forest is an ensemble of decision trees that vote on the outcome. Build N trees (100–500), each on a bootstrapped sample. At each split, only consider a random subset of features. For a new customer, all trees vote; majority wins. Confidence = percentage of trees agreeing.

Why it works for CLV: It handles non-linear relationships, is insensitive to feature scaling, and gives built-in feature importance rankings.