7.9 DATA‐DRIVEN ERROR COSTS
In this era of big data, businesses should leverage the information in their existing databases in order to help uncover the optimal predictive models. In other words, as an alternative to assigning error costs because “these cost values seem right to our consultant” or “that's how we have always modeled them,” we would instead be well advised to listen to the data and learn from the data itself what the error costs should be. Let us illustrate the power of data‐driven error costs by continuing our example.
Recall that our only nonzero costs were CostFP = $10 and CostTP = − $40. Fortunately, however, we have access to data that would give us a better idea of CostTP, namely the predictor Sales per Visit. This predictor provides the average amount of money spent per visit for each customer. So, if we calculate the mean Sales per Visit across all customers, we could use this as a better estimate...