Credit Scorecard Pipeline

From application and bureau data to a points score and validation report

Credit Scorecard Pipeline From application and bureau data to a points score and validation report 01 / Sources 02 / Encode 03 / Model 04 / Scale 05 / Consume Application Data · at origination · 01 / Sources · attributes Application Data at origination attributes Bureau Data · history + outcomes · 01 / Sources · third party Bureau Data history + outcomes third party WOE Encoding · binned features · 02 / Encode · monotonic WOE Encoding binned features monotonic Define Target · 90+ DPD window · 02 / Encode · good / bad Define Target 90+ DPD window good / bad Logistic Fit · regression · 03 / Model · coefficients Logistic Fit regression coefficients Scale to Points · PDO + base · 04 / Scale · points map Scale to Points PDO + base points map Scorecard Score · decision input · 05 / Consume · per applicant Scorecard Score decision input per applicant Validation Report · KS / Gini / calib · 05 / Consume · holdout Validation Report KS / Gini / calib holdout raw attributes features credit history features delinquency outcomes WOE vectors predictors bad flag label log-odds model output points delivered scored holdout evaluation Legend primary data policy / PII async batch data store

Assembly and Target

  • • Application and bureau attributes merge into one feature set
  • • The target is a bad flag over a fixed performance window
  • • Ninety-plus days past due is the classic bad definition

Encode and Fit

  • • Weight of Evidence bins features into monotonic risk trends
  • • Logistic regression keeps the model linear in log-odds
  • • Coefficients stay inspectable for governance and challenge

Scale and Validate

  • • Log-odds scale to points via PDO and a base score
  • • Discrimination is measured with KS and Gini on holdout
  • • Calibration checks predicted versus observed default rates