Understanding RAPM

RAPM stands for Regularized Adjusted Plus-Minus. It sounds technical because it is technical. But the intuition is simple: hockey is a five-man game at even strength and a deeply contextual game everywhere else. If you want to estimate individual impact, you need a method that tries to separate a player from his teammates, opponents, and deployment. That is what RAPM is trying to do. Why raw plus-minus is not enough Raw plus-minus tells you what happened while a player was on the ice. It does not tell you how much of that result belongs to the player. If someone spends all night with elite linemates against weak competition, raw plus-minus will flatter him. If someone plays brutal defensive minutes with weak support, raw plus-minus can punish him unfairly. RAPM exists because hockey is not a clean one-player sport. How RAPM works in this app The RAPM pipeline in the app: - builds shift-segment data from play-by-play and shifts - aggregates event outcomes by shift - tracks who was on the ice - models different outcomes with ridge regression - separates strength states like 5v5, power play, and penalty kill - outputs offensive, defensive, and differential views across metrics like Corsi, goals, and xG In other words, the model tries to learn how outcomes move when players are present, while controlling for the fact that players share the ice with teammates and opponents. The regularized part matters a lot. Hockey data is messy and collinear. The same players appear together repeatedly. Ridge regression helps stabilize those estimates and prevents the model from overreacting to noisy lineup combinations. What RAPM is measuring In the app, RAPM includes outputs tied to: - Corsi - goals - xG - offensive and defensive components - plus-minus style differentials - power-play and shorthanded states That means RAPM is not just one number. It is a family of adjusted impact estimates. A player may: - look strong offensively but weak defensively - grade well in xG impact but not goal results - show stronger value at 5v5 than on special teams - have more value in shot suppression than in shot creation That is the point. Good evaluation should separate those things. How to read RAPM correctly A few rules matter here too: - RAPM is an estimate, not a verdict. - It is strongest over larger samples. - It should be compared within role and context. - Offensive and defensive components should be read together. - xG-based RAPM is often more stable than goal-based RAPM. - Special teams RAPM should be treated carefully because the samples are smaller. If a player has strong xG RAPM and weaker goal RAPM, that often means the underlying process is better than the scoreboard results. If a player rates well in both, the case is stronger. If RAPM, xG, and lineup impacts all agree, confidence goes up. The app gives you multiple ways to test that agreement.