Subscribe to DSC Newsletter

What are the advantages/disadvantages of using WOE approach viz-a-viz. using continuous variables in their original form..For e.g. Making bins of Age and then using its WOE's in the model v/s using the values of the variable Age as such in the model..

Countering the arguments in favor of WOE :-

1. WOE gives us the 'riskiness' measure of an attribute - BUT, so does age as the values of age will be 'corresponding' to the dependent variable only, as in, using values of age as given in the data can also 'measure' the 'riskiness' aspect because it will be exactly what data values are showing..For e.g. if data shows that highest % of defaulters belong to a particular age group then that the way it is..So, using WOE just to capture the riskiness is not sufficient as the data per se is capturing it

2. WOE differentiates amongst the bins in proportion to the %good / % bad ratio BUT, the collapsing of the bins also leads to loss of IV and even if it doesn't it may give us the similar predictive power which the original variable does

3. WOE can help us in monitoring the trend of the characteristic BUT, that can be done with the original variable also

4. WOE can make the characteristic linear BUT, that can be done by characteristic transformation

So why should (not) WOE be prefered over using a continuous variable per se..?

Views: 4460

Reply to This

Replies to This Discussion

I assume you use logistic regression post applying WOE transformation(?)!

X: Independent Variable

Y: Binary Dependent Variable

T(.): Transformation function 

The fundamental assumption in logistic regression is 'X' and 'Log-Odds of Y' are linearly related. But, in most of the real world problems you will not see this relation restricting you to start with the model. What you need to do then? Apply a best transformation 'T(.)' to your 'X' that makes 'T(X)' and 'Log-Odds of Y' linearly related. With some mathematics done on WOE formula you will realize that 'WOE' is one of the best transformations that makes the above said assumption valid!

Another reason could be possibly, all the independent variables will be standardized by applying WOE transformation. The co-efficients (betas) you will get will be on the same scale (in fact unit free), they can then be directly comparable for relative inferences.

Hope this helps!


Thanks Sandeep, Sorry to have missed your reply..Yeah, you're right on your inputs..I'll look deep into the best transformation through WOE in detail.. :)


On Data Science Central

© 2021   TechTarget, Inc.   Powered by

Badges  |  Report an Issue  |  Privacy Policy  |  Terms of Service