So let's say I have receipts data from dozens of sellers that serves millions of customers. The main covariates of interest I have are

• ID of the customers

• ID of the sellers

• date

• amount paid (or refunded, in negative)

Among dozens of other covariates. I would like to somehow predict some score/metric of amount paid frequency of transactions (using date) as economic health. Here are some concerns:

• Different sellers have different duration of data -- when a new company enters our data as first transaction of the company, I don't want this to bias as 'healthier' economy

• Same as above with customers

As for frequency, I would simply create a column as the number of days since last transaction by seller-customer. Does this sound about right?

Either way, what would the Y variable be?