# Successive Predictions

I am facing a problem which you could abstractly describe the following:

I have a pool of possible customers. I want to know if a customer appears during one year. I want to predict the total revenue for each customer for this year. (0 revenue is something different than not appearing)

So I have stacked/successive/dependent predictions:

1. I predict who out of the pool is going to appear
2. For these I predict what the revenue will be

So far so good. But I'd like to have more input about this specific problem and I am having a hard time finding literature / papers about this. Is there a special term / technical term for this specific problem? (all search efforts with terms like stacked prediction, successive, ... lead to different problems, just because these words are so heavily used in ML e.g. model stacking, ... ) I am also very happy about hints to good papers concerning this topic.

I have not had to deal with this exact problem, but the stages you described sound like a variation of

1. Time until next purchase (and then before or after time window in your setting). Check this post as a starting point.
2. Customer lifetime value (except it is not lifetime, but only sales for a given period in your case). I find this post and links therein quite useful.

Hope this helps mate!

In this you are trying to solve two business problem. The first business problem would be to find out the customer who have Next Purchase Day in the one year duration or Days Until next Purchase. The second problem is for the customer who have next purchase day in one year time predict their Value ( A variation of customer lifetime value)

Problem 1 : How to approach Next Purchase Day Problem or Days Until next purchase

Approach would be simple use historical transaction data to create features for RFM, segmentation and use previous year as target variable purchased or not. Use this classification problem to predict whether the person will buy or not. It is similar to churn problem just that 0,1 mapping would be different in both cases. Once you get a good model decide on threshold and get the list of customers who will buy in next year. The problem with classification maybe you consider who buys at start of the year and end of the years as same which can lead to impact on LTV model. So you can also model these problem as days to next purchase i.e. Regression problem which would be more accurate for your use case but little complex to solve

Problem 2: Customer Life time value