Skip to content

Weight of Evidence (WOE) — In-Depth Guide

What is WOE?

Weight of Evidence measures how different a bin's event rate is from the population average:

Positive WOE = more good events than expected = lower risk
Negative WOE = more bad events than expected = higher risk
WOE of 0 = the bin's risk matches the population average

In scorecard modeling, WOE transforms categorical or binned numerical features into a continuous monotonic scale that has a linear relationship with the log-odds of the target event.

Why Use WOE in Scorecards?

  • Linearizes log-odds — enables standard logistic regression
  • Handles missing values — missing can be treated as its own bin
  • Bin-level interpretability — each bin gets a transparent risk score
  • Monotonicity — easier to validate and explain to regulators
  • Handles outliers — binning + WOE naturally caps extreme values

WOE vs Other Encoding Methods

Method Pros Cons
WOE Linearizes log-odds, handles outliers, interpretable Requires target, loses absolute feature scale
One-Hot No target leakage, preserves categories High dimensionality, no ordinal relationship
Target Encoding Simple, captures target correlation Risk of overfitting, no explicit monotonicity

WOE Calculation Methods

All methods take arrays of good and bad counts per bin and return WOE values.

Method Formula When to Use
Standard ln(dist_good / dist_bad) Default when both classes have sufficient counts in every bin
Adjusted (Laplace) ln((good + s) / (good_total + 2s) / ((bad + s) / (bad_total + 2s))) Handles zero-count bins via smoothing. Good default for most use cases.
Empirical Logit ln((good + 0.5) / (bad + 0.5)) Agresti correction. Standard in SAS-based scorecard development. Good for small bin counts.
Signed sgn(good - bad) * ln(max / min) Preserves directional information explicitly. Useful when sign matters for regulatory reporting.
Weighted ln(dist_good / dist_bad) * (bin_total / n_total) Downweights small bins to prevent outsized influence from tiny populations.

Usage:

from ScoreCardModel import WOETransformer

WOETransformer(method='adjusted', laplace_smoothing=1e-6)
WOETransformer(method='empirical_logit')
WOETransformer(method='signed')
WOETransformer(method='weighted')

Choosing a Method

  1. Large dataset, balanced classes → Standard WOE
  2. Zero-count bins possible → Adjusted (Laplace) WOE
  3. Small dataset or SAS migration → Empirical Logit
  4. Regulatory focus on direction → Signed WOE
  5. Unbalanced bin sizes → Weighted WOE

WOE Transformer Features

Rare Level Lumping

Automatically merge categorical levels with population below min_bin_pct into a single RARE bin:

WOETransformer(rare_lumping=True, min_bin_pct=0.05)

Unseen Category Handling

Controls how unseen categories in test/validation data are handled:

Strategy Behavior
'zero' WOE = 0 (no evidence, no adjustment)
'mean' WOE = mean of all known WOE values
'min' WOE = min of all known WOE values (conservative)
'raise' Raise error 
WOETransformer(unseen_strategy='zero')

Missing Value Treatment

Strategy Behavior
'separate' Treat NaN as its own "MISSING" bin
'mode' Impute with most common bin
'raise' Fail on missing values 
WOETransformer(missing_strategy='separate')

WOE Diagnostics

Every fitted WOETransformer exposes diagnostics for regulatory review:

Diagnostic Function Description
Monotonicity check_monotonicity(woe_map, ordered_bins) Returns increasing / decreasing / non-monotonic. Uses Spearman rank correlation.
IV Contribution iv_by_bin(good, bad, good_total, bad_total) Per-bin IV breakdown. Identifies which bins carry predictive power.
Bin Statistics bin_statistics(bin_series, target) Per-bin counts, event rates, WOE values, and population %.
Chi-Square Test woe_chi_square(bin_series, target) Test of independence between bins and target.
Midpoint Correlation midpoint_correlation(bin_edges, woe_values) For numeric features: correlation between bin midpoints and WOE. High = good linearity. 
from ScoreCardModel.weight_of_evidence.diagnostics import check_monotonicity

wt = WOETransformer().fit(X, y)
for col, woe_map in wt.woe_maps_.items():
    direction, strength = check_monotonicity(woe_map, ordered_bins)
    print(f"{col}: {direction} (strength={strength:.2f})")

Information Value (IV)

IV quantifies a feature's predictive power:

IV Range Label Recommendation
< 0.02 Useless Reject
0.02 – 0.1 Weak Accept only with business justification
0.1 – 0.3 Medium Good candidate
0.3 – 0.5 Strong Excellent candidate
> 0.5 Suspicious Investigate for data leakage

IV is accessed directly on a fitted WOETransformer:

print(wt.iv)  # dict of {feature_name: iv_value}

Interpreting WOE Patterns

  • Monotonic increasing — risk decreases as the feature value increases
  • Monotonic decreasing — risk increases as the feature value increases
  • U-shaped — both extremes are high-risk (common for age, LTV ratios)
  • Erratic / non-monotonic — likely overfitting or poor binning. Consider merging adjacent bins.

Common Pitfalls

  1. Zero-count bins — Laplace smoothing mitigates this (use adjusted method)
  2. Perfect separation — IV > 0.5 is suspicious; investigate for data leakage
  3. Non-monotonic WOE in production — monitor for pattern reversal over time
  4. Too many bins — each bin should have at least 5% of population
  5. Extreme WOE values — cap at ±3.0 to prevent single bins from dominating scores

Regulatory Context

WOE-based scorecards are well-suited for regulatory compliance:

  • Basel II/III — scorecards must be interpretable, stable, and well-documented
  • CCAR — model use must be clearly defined with performance monitoring
  • FCRA / Reg B — adverse action reason codes can be derived from per-feature point contributions

The transparency of WOE — each bin maps to a specific point contribution — is a key advantage over black-box models in regulated environments.