Predicting Components of a Target Value Versus Predicting the Target Value Directly

In many Regression problems one can predict components of a target value and then combine those components to determine the target value prediction. The alternative is to predict the target value directly. A simple example is automobile insurance claims. The traditional approach is to compute Severity (the average value of claims made) and Frequency (the number of claims made per year). The product of these will then provide the average money paid annually to the customer (henceforth called claim rate). On the other hand, one can derive the claim rate for each customer and use this as the target value. Intuitively one would think that the latter approach (predicting the target value directly) should provide better results but, in fact, the former approach is better. We investigate the difference in performance of these two approaches (called component and composite predictors respectively) and illustrate the difference. We demonstrate this difference using ten Machine Learning algorithms.