Hi all,

I am trying to build a model to predict purchase behavior. Below is a detailed description of my problem:

• For a online retailer we are trying to predict which division a prospective customer is most likely to buy. Currently they have 4 divisions: Men,Women,Home and Kids.
• Customer: A customer is defined as a user who has made atleast 1 purchase.  
• Members : Member is user who is on file but has never made a purchase.
• So we want to build a model on current customer to predict which of the 4 divisions they might belong and then score the model on members to assign them to a division.

My current thought process:

• As of now we are looking at building a logistic kind of model and assign each customer to a particular division based on certain criteria so that we get a dependent variable to model upon.
• Once we have built the model we can score the Members and assign them to a cluster.

Problems:

• In order to build a Logit type model, I have to define a Dependent variable for customers to model upon. Currently I don't have a concrete approach to assign current customers to a division.
• We can assign customer to division to divisions based on their purchase behavior,visit behavior etc. But the problem with these is it doesn't distinguish between a Impulse buyer and a Regular buyer.
• I have variables such as revenue in each division,no.of orders/division,no.of items purchased in each division for customers. These "Purchase Variables" I guess can only be used to generate a Dependent variable but cannot be used while model building as the Members i eventually want to score don't have Purchase history.

Alternative Approaches:

• I have read about NBD/Pareto models being able to predict purchasing behavior. But want some help on how to approach this.
• Instead of a logit model, what if I take total lifetime revenue in each division as dependent variable and build a OLS Regression for each division to project each users LTV?
• Latent Class Regression: I am hugely limited by variables and lot of variables might be latent in nature. So can I consider LCA kind of model? If so what will be my Dependent Variable to model upon?

Its quite a lengthy post, but desperately need some ideas on how to approach this problem.

Thanks,

Hari