Trying to be right for everyone

An offer sometimes starts with a sweeping promise: the best quality, lowest price, fastest delivery, and service for everyone. It sounds generous, but customers may struggle to identify a specific reason to choose it. The team faces an even harder task delivering all those promises […]

A Performance Number Needs Context

When one model beats another by a small amount, the case count and dependence between cases matter. The difference may reflect useful information, or it may change readily in another sample. The headline number cannot settle that question. The laboratory scores describe what happened in […]

Score the Whole Distribution

A forecast mean can land near the observed outcome while the rest of the predicted probability sits in unhelpful places. Distance from one predicted price cannot fully evaluate an entire distribution. The laboratory reports CRPS, a score for predictive distributions. A lower value is better […]

Directional Accuracy Is an Incomplete Score

Two models can call the direction correctly equally often while expressing very different confidence. One may use cautious probabilities and the other near certainty. When they miss, the degree of overconfidence differs even though the accuracy count is identical. Probabilistic evaluation therefore uses measures such […]