This algorithm is based on the excellent paper by Mironchyk and Tchistiakov (2017) named "Monotone optimal binning algorithm for credit risk modeling". Any mistakes or shortcomings of the Python code are mine alone and I'd appreciate feedback on these possible errors
The weight-of-evidence (WOE) method of evaluating strength of predictors is an understated one in the field of analytics. While it is standard fare in credit risk modelling, it is under-utilized in other settings though its formulation makes it generic enough for use in other domains too. The WOE method primarily aims to bin variables into buckets that deliver the most information to a potential classification model. Quite often, WOE binning methods measure effectiveness of such bins using Information Value or IV. For a more detailed introduction to WOE and IV, this article is a useful read.
In the world of credit risk modelling, regulatory oversight often requires that the variables that go into models are split into bins
- whose weight of evidence (WOE) values maintain a monotonic relationship with the 1/0 variable (loan default or not default for example.)
- are reasonably sized and large enough to be respresentative of population segments, and
- maximize the IV value of the given variable in the process of this binning.
To exemplify the constraints such a problem, consider a simple dataset containing age and a default indicator (1 if defaulted, 0 if not). The following is a possible scenario in which the variable is binned into three groups in such a manner that their WOE values decrease monotomically as the ages of customers increase.
The WOE is derived in such a manner that as the WOE value increases, the default rate decreases. So we can infer that younger customers are more likely to default in comparison to older customers.
Arriving at the perfect bin cutoffs to meet all three requirements discussed earlier is a non-trivial exercise. Most statistical software provide this type of optimal discretization of interval variables. R's smbinning package and SAS' proc transreg are two such examples. To my knowledge, Python's solutions to this problem are fairly sparse.
My solution here takes two columns of data: a 1/0 variable and the variable to be binned. It returns a binned variable along with respective WOE values conditioned on user-defined thresholds on minimum possible bin size, minimum number of defaults in each bin and the maximum p-value allowed for a possible t-test in means between adjacent bins.
I hope my attempt here serves as a helpful stop-gap for someone looking to perform risk modelling in Python using WOE methods.