Features & Models to compute the probability of certain customer accepting an offer/product from a bank?

Question

What are the features & models that can be used to compute the probability of a certain customer accepting an offer/product from a bank?

After some research, I came to know of what is called 'Propensity Scoring' and it's definition is very relevant to what I am trying to find.

But I failed to find any technical papers going into any appreciable detail. I tried Google Scholar but reached a dead end because the term 'Propensity modeling' is actually a marketing term not exactly academic.

I also know from my search that a broader term is 'Campaign Response Models' but this turns out to be too broad and I am looking for something very specific. Any references/links that contain some technical details about the modeling process, the features or the choice of models will be very helpful.

Answer 1

Narrowing down your question is a binary classification task. Neural networks and svm are equally sufficient for the job. However, only decision trees use "scores" and actually show you the reasoning behind the classification while the other two are a black box. What you are after is called uplifting modelling. Also it is worth checking feature selection and sensitivity due to the nature of the task.

Answer 2

I don't think anyone can help you with features without more information about your particular domain, but you could do worse than logistic regression for modeling:

https://en.wikipedia.org/wiki/Discrete_choice

https://en.wikipedia.org/wiki/Logistic_regression

Answer 3

How much help this community can be depends on where in the process you are. If you have a dataset and are looking on ways to determine the most predictive features, then you could include some information as to what kind of data you've collected/have access to.

If instead your question is specifically "what data should I collect to predict purchases from banks, and what models should I use," you are unlikely to get a very specific answer.

Answer 4

I'm not sure how well this approach is going to work for you, but it worked for me in my scenario of propensity scoring. So, if it makes sense, try to relate your data to mine.

I had a list of customers and their purchase behaviors. From those purchase behaviors I deduced association rules to determine which customer is likely to purchase in what product category (based on their previous purchases- associated with the previous purchases of the group).

You can also include recency and frequency of the product purchases into the model to decide whether or not to recommend a particular product/offer to the customer for a specific time.

Based on the confidence, support and lift metrics of those rules for each customer, I mapped them to highly likely to purchase in this category with confidence being their propensity score.

However, based on your data and the task you want to achieve, every model hass to customized a little. This is only an idea that worked for me.