Chapter 18: Key Takeaways - Modeling the NHL

Expected goals (xG) is the cornerstone of modern NHL analytics and the most predictive team-level metric. An xG model assigns a goal probability to each shot based on distance, angle, shot type, and context. By summing shot-level xG, we measure how many goals a team "deserved" based on chance quality, independent of whether the puck actually went in. The xG differential correlates with future points at $r \approx 0.55$--$0.65$, compared to only $r \approx 0.35$--$0.45$ for actual goal differential.
Corsi and Fenwick measure territorial control and stabilize quickly. Shot attempt differentials serve as proxies for puck possession and stabilize within 20--25 games. They are useful supplements to xG, particularly for detecting team-quality changes early in the season before xG has sufficient sample size.
PDO is the most mean-reverting metric in hockey and a direct signal for regression bets. Teams with PDO above 102% are almost certainly benefiting from unsustainable luck; teams below 98% are likely unlucky. Simply betting against extreme PDO teams has been historically profitable.
Score effects fundamentally alter team behavior and must be adjusted for. Leading teams turtle (CF% drops to 43--47%), trailing teams press (CF% rises to 53--57%). Raw shot metrics are heavily contaminated by the score states a team played in. Score-adjusted metrics produce significantly more accurate team quality assessments.
Goals Saved Above Expected (GSAx) is the proper metric for goaltender evaluation. Raw save percentage is contaminated by shot quality and small-sample noise. GSAx adjusts for shot difficulty and isolates goaltender skill. However, even GSAx requires heavy regression to the mean.
Goaltender performance requires the heaviest regression of any variable in hockey. The regression constant of approximately 3,000 shots means that even after a full season, a goaltender's observed performance carries only 33--40% weight versus the league-average prior. Small-sample goaltender results (10--15 games) are almost entirely noise.
Goaltender mispricing is the single largest source of exploitable edges in NHL betting. The market overweights recent save percentage, underweights the starter-to-backup gap, and ignores shot quality context. A properly regressed goaltender model, combined with timely monitoring of goaltender confirmations, identifies persistent value.
The puck line (+/- 1.5) is one of the most interesting NHL betting markets. The 23--26% overtime rate creates a large wedge between moneyline probability and puck line cover probability. A 60% moneyline favorite covers $-1.5$ only about 33--38% of the time. Underdog $+1.5$ cover rates of 62--72% frequently exceed market-implied probabilities.
Back-to-back game fatigue is a measurable and exploitable factor. Win rates drop by 5--7 percentage points on the second game of a back-to-back, with additional penalties for travel, overtime the previous night, and backup goaltender starts. The market adjusts but often insufficiently.
Special teams contribute meaningfully to game projections. An elite power play facing a poor penalty kill gains 0.3--0.4 expected goals per game beyond even-strength projections. Modeling PP and PK separately, then adjusting for the opponent's opposing rate, adds predictive value.
The Poisson distribution is well-suited for NHL goal modeling. Goals are low-frequency events that occur approximately independently, making the Poisson a better fit than for MLB run scoring (which exhibits overdispersion). The Poisson model produces accurate moneyline, puck line, and totals probabilities from expected goal projections.
NHL betting markets are less efficient than NFL or NBA markets, creating persistent opportunities. The market is smaller, fewer sophisticated bettors focus on hockey, and late-breaking information (goaltender confirmations, schedule effects) creates informational asymmetries that the prepared bettor can exploit.