Move the scorecard's splits and watch the weight of evidence change.
A weight-of-evidence binning tool I built for the SecureBank fraud case in EN.705.603 at Johns Hopkins. The counts are aggregated fraud and legitimate transactions per micro-bin on the training months; no transaction rows are in this page. optbinning chose the dashed splits for my scorecard model and they load as the default; drag the solid ones to see what a different cut does to WOE and information value.
Training rows, January 2019 to October 2020, labelled only. Drag a split line to move it; double-click the chart to add a split; click a split's handle at the top to delete it. WOE = ln(legitimate share / fraud share) per bin, so a negative WOE means the bin is riskier than average. IV is the sum over bins. The dashed splits are what optbinning chose for the scorecard; the monotonic button loads its best solution with the fraud rate forced to rise or fall across the bins.
| Bin | Rows | Share | Fraud | Fraud rate | WOE | IV |
|---|