Weight of evidence

A credit team needs to turn borrower characteristics into risk signals that officers can inspect. Weight of evidence (WoE) assigns each category or numeric bin a value based on its share of defaults versus its share of non-defaults. With “bad” defined as default, WoE(bin) = ln[(bads in bin / all bads) / (goods in bin / all goods)]. Positive values indicate greater relative default risk; negative values indicate lower relative risk; values near zero resemble the overall population. section 4.4.3

Information value (IV) measures how strongly the bins of a feature separate good and bad borrowers: IV = Σ[(bad share − good share) × WoE] across bins. It helps prioritize features, but a very large IV—such as above 0.5—warrants a check for leakage or overfitting. Review the WoE pattern across ordered bins as well as the aggregate IV; a strong total does not replace inspection of individual risk groups. section 4.4.3

import numpy as np
import pandas as pd

data = {
    'AgeGroup':   ['<20', '20-24', '25-29', '30-39', '40+'],
    'GoodCount':  [90, 120, 160, 300, 280],
    'BadCount':   [10,  30,  40,   25,   5]
}     #1

df_bins = pd.DataFrame(data)

total_good = df_bins['GoodCount'].sum()
total_bad  = df_bins['BadCount'].sum()     #2

df_bins['GoodRate'] = df_bins['GoodCount'] / total_good
df_bins['BadRate']  = df_bins['BadCount']  / total_bad    #3

df_bins['WoE'] = np.log(
    (df_bins['BadRate'] + 1e-9) / (df_bins['GoodRate'] + 1e-9)
)     #4

df_bins['IV_bin'] = (
    (df_bins['BadRate'] - df_bins['GoodRate']) * df_bins['WoE']
)
IV_age = df_bins['IV_bin'].sum()   #5
print(df_bins[['AgeGroup','GoodCount','BadCount','WoE','IV_bin']])
print(f"Information Value (Age) = {IV_age:.4f}")

The pandas and NumPy example calculates group shares, WoE, per-bin IV, and total age IV. Its small constant prevents a zero share from making the logarithm undefined. section 4.4.3

For a manual credit-modeling pipeline:

• Clean the inputs first. Treat sentinel codes, missing values, and outliers deliberately; a dedicated missing bin may preserve a risk signal that imputation would hide. section 4.4.4
• Group categorical values and bin numeric features. One implementation uses ten quantile bins where possible, adds 0.5 to stabilize event and non-event proportions, and clips WoE to ±10. section 5.3.3
• Calculate IV, inspect bin-level results, and select features above a chosen threshold, such as 0.02. Retain the bin boundaries and WoE mappings so selected raw features can be converted to WoE-coded inputs. section 5.3.3 section 5.3.4
• Train and evaluate the downstream model. In the illustrated workflow, WoE-coded features feed a baseline XGBoost model; cross-validated AUC and confusion matrices are reviewed before probabilities are mapped to credit scores. section 5.0 section 5.2.3

def transform_to_woe(
    df,
    selected_features,
    cat_features,
    bin_edges_dict,
    woe_dict,
    target="target"
):
    df_woe = df[selected_features + [target]].copy()
    for feat in selected_features:
        if feat in cat_features:
            df_woe[feat] = df_woe[feat].map(woe_dict[feat])
        else:
            bins = bin_edges_dict[feat]
            df_woe[feat] = pd.cut(
                df_woe[feat],
                bins=bins,
                include_lowest=True
            ).astype(str)
            df_woe[feat] = df_woe[feat].map(woe_dict[feat])
    return df_woe


df_woe = transform_to_woe(
    df,
    selected_feats,
    cat_features,
    bin_edges,
    woe_maps,
    target=target_col
)
print("df_woe shape:", df_woe.shape)
df_woe.head(5)

The transformation function shows how stored category mappings and numeric bin edges replace selected raw columns with their WoE values. section 5.3.4

Document binning, missing-data decisions, and outlier treatment for audit and fairness review. WoE makes risk differences inspectable, but it does not resolve poor data quality or establish that a feature is safe to use. Automated binning and advanced drift checks are separate extensions to this manual approach. section 4.4.4 section 5.2.3