Customer lifetime value prediction using a ZILN model with time dependency

Abstract

It is important for marketers to predict customer lifetime value (LTV) in customer service to make a various decision. There are far more customers who don't produce any sales more than customers who produce them. Zero-Inflated Log-Normal (ZILN) model can handle this sales price distribution properly bacause it considers the probability occuring sales event and sales price distribution. However, the model requires fixed elapsed time for the observation value of each customer in the training data, and this time length in trained model cannot be changed when prediction. Because of this constraint, we cannot gather enough amount of training data because of elapsed time contraint, and cannot change the time length in prediction. To solve these drawbacks, we propose \textit{td-ZILN} model which modifies original ZILN model to handle the time dependency of sales by using the variable of time length for sales probability. By td-ZILN model, we can train the data which has the various length of elapsed time for each customer, and change the time length in the LTV during prediction.