When a logistic regression technique is used for building predictive model, independent or exploratory variables are transformed using Weight of Evidence (WOE). Weight of Evidence (WOE) is a statistical concept which allows to group independent variable values based similarity of target variable distribution.
Weight of Evidence for a group is ln (% Good /%Bad) in the group.
Once fine classes are created for a continuous variable, Weight of Evidence (WOE) is calculated for each variable. For categorical variables, WOE for each category class is calculated.
Once WOE is available for all the categorical and continuous variable fine classes, a process of combining is taken forward by an analyst. The process of combining fines classes is called coarse classing.
Main point to consider in the coarse classing fining
- Coarse classes created should have at least 8-10% of overall items or customers
- Final classes have a trend based on WOE value
You could see the final coarse classes for borrower age.
For an ordinal or continuous variable, a class is combined to next or previous class based on WOE value. Smaller is the WOE difference similar is the classes in target variable distribution. Example, borrower age fine classes – age classes 72 and73- are combined first as they are closed. Then borrower age groups 76 and 77 are combined.
- Detailed example of Coarse Classing