Standard regression models fail when predicting outcomes with lots of zeros mixed with positive values. Whether it's customer spending, insurance claims, or loan prepayments, forcing one model to handle both behaviors leads to poor predictions and business confusion. The core issue: zeros and positive values often come from completely different processes.

Traditional workarounds like linear regression produce negative predictions for non-negative outcomes, while log-transforms introduce arbitrary offsets and systematic bias. These approaches can't distinguish between customers who never buy and those who occasionally purchase but didn't this period. The result is a model that conflates different behaviors and provides no insight into what drives each outcome.

The two-stage hurdle model offers a principled solution by splitting the problem into two questions: will the outcome be zero or positive, and given that it's positive, what will the value be? This approach allows using different algorithms for each stage - perhaps a gradient boosting classifier for the first hurdle and gamma regression for the second. By separating these questions, you can use the right features for each task and get interpretable probabilities that help businesses understand whether low predictions stem from unlikely purchases or small purchase amounts.