We discount performance that occurs when it doesn’t matter – when the game is more or less settled. We determine how much to discount, and when, from the data. Our starting point is the “win probability” of a team with the ball, which gives the likelihood of winning for all possible time & score-difference game situations. Below is a figure showing some of these win probabilities at notable points in the game (halftime, 2-minute warning, etc.).
We then gradually discount plays that occur when the likelihood of winning is very high or very low. We let the data tell us when and how steeply to discount.