How do I choose class weights for an imbalanced classifier?

Question

AcadiFi · Accepted Answer

Class weights scale the loss contribution of each class. Three common strategies:

1. Balanced (inverse frequency): w_class = N / (K × n_class). For 2% positive dataset, w_pos = 50/2 = 25, w_neg = 50/98 = 0.51. Ratio 49:1.

2. Cost-based: w_pos / w_neg = C_FN / C_FP. Directly encodes economic costs.

3. Tuned via CV: treat weight ratio as hyperparameter, optimize via grid search.

Example: Pemberwald Life Insurance builds a policy lapse prediction model. 8.4% positive in 140,000 policies.

Balanced weights: w_pos = 1/0.084 = 11.9, w_neg = 1/0.916 = 1.09. Yields recall 0.71, precision 0.34, F1=0.46.

Cost-based: lapse loss avg $1,200 (lost future premiums - clawback); false positive (unnecessary retention call) costs $25. Cost ratio 48:1. Weights imply aggressive recall. Recall 0.84, precision 0.18, F1=0.30 — but total expected cost minimized.

Tuned CV: grid [1:1, 5:1, 12:1, 25:1, 50:1, 100:1]. Best F1 at 15:1 with F1=0.48. Best expected cost at 40:1.

Beyond static weights:
- Focal loss (Lin et al. 2017): dynamically down-weights easy examples, useful for extreme imbalance in deep learning.
- Logit adjustment: adds log(prior_ratio) to logits, mathematically equivalent to class weights in logistic regression.
- Sample weights per instance: weight by dollar exposure, e.g., $5M loan gets 10x weight of $500K loan. Aligns model loss with portfolio P&L.

Pemberwald chose 40:1 weight with XGBoost, added per-policy-value sample weights, and retrains quarterly. Expected lapse cost dropped from $14.2M to $9.8M annually in production.

How do I choose class weights for an imbalanced classifier?

Master Level II with our CFA Course

Related Questions

Practice Questions