Dear Team,
I am in the process of starting a new project which mainly focuses on building a predictive model for customer churn. As we all know the outcome variable is whether the customer has churned or not where i will convert it to a binary variable i.e 1 and 0.
For such a case i would prefer to use a stepwise logistic regression and based on the outcome i would realize which are the best predictors.
However in many cases i have been many statistician prefer to build varied models and select the best model for instance in this case we could also look at the possibility of a decision tree.
So please advice what should be the deciding or influencing factors to look at building multiple models or as a practice we should always have multiple models.
Hello Shivi, as far as I know modelling is an art and I usually prefer to use different models and test them until they meet criteria to be achieved.
In ur case designing churn/attrition may be hard bound by some values hence I recommend to use a model that allows fuzzy logic. To prove myself, let us take example of a simple regression equation x>=10 and y>=23 which hardbounds the results. In case x=9 and y=28 you would want customer to be churned but your model would fail to do that. This is where fuzzy logic comes to rescue and also helps reach such patterns.
Although, I don't know ur exact problem statement I advise to go for either fuzzy logic or some soft-margin regression methods.
Hi Shivi, I think you can use some predictive analytics's methods such as regression trees, random forest, support vector machines. I think prediction would be more suited for this particular interest of yours.
- Badges
- Science topic
- Similar topics
- Geoscience
- Atmospheric Sciences
- Atmosphere