New to Modeling NHL Totals? A Step-by-Step Starting Point

An NHL total is a line on combined goals, and books commonly settle full-game totals including overtime and the shootout, with a shootout counting as one goal for the winning side. Action Network's explainer on NHL betting rules sets this out: once a game goes to overtime, it is assured of exactly one more goal on the settled score. Some books also offer regulation-only markets, and team totals vary by book, so confirm the rules for the market you are modelling.

That one rule shapes the whole model. In the regular season, a tie after 60 minutes goes to five minutes of three-on-three overtime and then a shootout; in the playoffs, overtime continues until someone scores. Either way, the settled total is the regulation total plus one.

You need game results, shot-level or game-level expected goals split by strength state, goaltender starts, and historical closing totals with prices. MoneyPuck offers free downloads of shot data and team, skater, and goalie data by season; its data page says the files are free for non-commercial use with credit, and other uses require contacting them. NHL play-by-play is the underlying source for public models. Store everything keyed by game, team, date, and strength state, and keep the timestamp of each odds snapshot.

For each team, build an attacking rate and a defensive rate. Expected goals for and against are steadier than goals, so a common approach blends the two, weighted toward xG over short samples. Separate five-on-five from special teams: a team's power play and penalty kill rates matter, and so does how often it draws and takes penalties.

Regress every rate toward the league average, more strongly early in the season. A hierarchical or Bayesian model does this naturally. Then combine attack and opposing defence, with a home adjustment, into an expected regulation goal count for each side.

Goaltending is where the biggest single game-to-game swing usually sits. Estimate each goalie's goals saved above expected per shot, shrink it heavily toward average because it is noisy, and apply it to the opponent's expected shots. The starter is often confirmed only on game day, so either wait for confirmation or price both scenarios and weight them. Backup starts on the second night of a back-to-back are a common pattern to account for.

Shot volume varies by team style, and two fast teams produce more attempts than their individual rates suggest if your model only combines them linearly. Rest, travel, and back-to-backs affect both sides. Penalty rates by referee crew are sometimes used, though the evidence needs testing on your own data before trusting it.

Poisson. The simplest choice: model each team's regulation goals as Poisson with its expected rate. The sum of two independent Poisson counts is Poisson with the summed rate, which makes the total easy to compute. Independence is an assumption, not a fact: score effects and shared pace correlate the two teams' goals, which is why bivariate Poisson and Conway-Maxwell-Poisson variants exist, and it makes the regulation-tie probability below model-dependent.

Negative binomial. Game totals often show more variance than Poisson predicts, partly because of empty-net goals and partly because the true rate on the night is uncertain. A negative binomial adds a dispersion parameter for that. Fit it and compare against Poisson on held-out games.

Simulation. The most flexible option simulates the game in time slices, with rates that change by score and time remaining. It handles empty nets and overtime naturally, at the cost of more parameters.

Empty-net goals are not random noise; they follow game state. A team trailing by one or two late pulls its goalie, which raises the chance of a goal for either side. Analyses at FiveThirtyEight and Hockey-Graphs documented a shift toward earlier goalie pulls during the 2010s, so use recent seasons when estimating how often late empty-net situations arise and how often they produce goals. Add them in simulation, or as a separate component conditional on the regulation score margin late in the game.

Because a regulation tie always adds exactly one goal, the effect depends on the line. Take illustrative regulation rates of 3.1 and 2.8 goals with independent Poissons and no empty-net component. The chance of a regulation tie is about 0.167.

For a line of 6.5, a 3-3 regulation tie is under but settles at 7, over. That moves the over probability from about 0.378 on regulation alone to about 0.427. For a line of 5.5, nothing changes: a 2-2 tie becomes 5, still under, and a 3-3 tie was already over. Lines where the regulation tie sits just below the number are the ones overtime moves.

Convert closing over and under prices to probabilities, remove the margin, and compare them with the model's probabilities at the same line. Track log loss and calibration on a held-out season. Measure closing line value on any simulated or real positions: a model whose prices consistently move toward the close is capturing information, and one that does not is probably not, whatever the short-run results show. Totals markets in a sport this low-scoring are noisy, so judge over hundreds of games.

This page describes modelling method and is not betting advice.

Settling the model on regulation while the market settles on overtime and shootouts. Using raw recent goals instead of regressed expected goals. Ignoring the starting goaltender. Treating every half-goal line the same when overtime affects them differently. Leaking information by fitting on data that includes the games being tested.