The author applies logical positivism incorrectly. There's nothing wrong with coming up with new models and testing them against the data, and eventually getting enough variables to be significantly more effective. But such models should be a low portion of your overall probability until actually found to be more effective.
Of course, if you start out with a contrived example where there's a large amount of unobservable simple structure, then it looks like postulating unobserved structures is good. I could give a contrary example where it's impossible to do better than random and draw the opposite lesson.
Beyond that, under logical positivism there isn't a notion of truth as the author applies it. To critique LP by saying it doesn't get at the truth is kind of missing the point.
Of course, if you start out with a contrived example where there's a large amount of unobservable simple structure, then it looks like postulating unobserved structures is good. I could give a contrary example where it's impossible to do better than random and draw the opposite lesson.
Beyond that, under logical positivism there isn't a notion of truth as the author applies it. To critique LP by saying it doesn't get at the truth is kind of missing the point.