"Machine learning works best in relatively stable environments, while dynamic changes and interventions make prediction alone less reliable."
Hardt and Recht explain that machine learning performs best in a stable world where historical observations remain informative about future events. Their discussion of the Fragile Families Challenge illustrates how social dynamics and uncertain future conditions can defeat extensive predictive modelling. Their distinction between prediction and reasoning about interventions directly supports the book's argument about the limits of statistical approximation in changing environments.
View Source