Building a machine learning model that minimizes prediction error is a core skill for predictive data scientists. Picking the right cost function is usually easy enough, but understanding how well your model is predicting all the nooks and crannies of a diverse data set can help surface trends and groupings that may have gone unnoticed.

During his talk at DataScience: Elevate, Morgan Hansen discusses how visualizing your predictions — either segmented manually or by algorithmically defined categories — can help you identify segments of your prediction population that may need additional attention, as well as latent variables and potential business risks.