I'm using CamDavidsonPhillips Customer Lifetime Value library to calculate CLV, and it uses a distribution based on Peter Fader's work on the subject that fits a Gamma distribution to model customer transaction rate, and a Geometric distribution to model customer dropout rate.
Using this with real store data, I occasionally end up with extreme parameter values, such as a = 45e7.
I imagine part of the reason this is happening is that the stores data . may be too sparse to fit a gamma distribution or geometric distribution. For instance, a store with 10,000 customer orders, may only have 1,000 actual customers. And only 10 of these customers might have bought from the store many times...
How should I go about thinking about sparsity for this fit? In other worse, how many customers are enough (and how well-varied do their behavior have to be) to obtain a reasonable solution?