A Bayesian-calibrated Poisson model targeting the most neglected outcome in NHL betting. Bookmakers systematically underprice the regulation draw. The model prices it correctly across 20+ books every night.
In a 3-way regulation market, the draw is the only outcome nobody is actively correcting. Three reasons it stays mispriced.
In a 3-way market, P(Home) + P(Away) + P(Draw) must equal 100%. Sharp and recreational money both flow into home and away. The draw absorbs whatever mispricing remains, and there is no direct counter-trade forcing it back to fair value. The correction mechanism that works in 2-way markets does not exist here.
The NHL overtime rate is not constant. It ran near 22% for years then shifted to 25.8% in 2025-26. Bookmakers anchored to long-run averages missed 87% of that shift. The model updates its draw probability estimate dynamically using a Bayesian prior that blends historical data with current-season evidence, adapting as the regime changes.
NHL 3-way regulation markets represent roughly 0.5% of total sports betting handle. Sharp syndicates do not focus there. Books do not set careful lines there. A 1-4% mispricing on the draw has persisted across every game, on every book, every night, because nobody with enough volume has corrected it.
From play-by-play data to 3-way probabilities. Four layers, fully traceable.
The model starts at the shot level. An expected goals model trained on 2.4 million NHL play-by-play events assigns a scoring probability to every unblocked shot based on shot geometry and game context: distance from net, angle, shot type, rebound status, strength state.
No shooter or goalie identity in the xG model, it learns what makes a shot dangerous, not who took it. These shot-level probabilities aggregate to expected goals for and against per team per game, forming the foundation for everything above.
The goalie matchup is one of the largest single-game variables in hockey. A confirmed backup starter meaningfully changes a team's goals-against expectation. A regression model incorporates deployment context, consecutive starts, fatigue signals, role share, along with goalie quality metrics to output an adjusted goals-against rate for each team.
This adjustment is computed fresh each game based on who is confirmed in net. The model captures backup fatigue and starter role changes before puck drop, not after the opening period.
Rather than predicting a single number like "this team will score 2.8 goals," the model runs 13 separate quantile regressions per team from the 5th to the 95th percentile. Each model predicts a different part of the goal distribution. The full distribution is integrated to produce a rate estimate that preserves the tail behavior: the probability of a 5-goal blowout or a 0-goal shutout.
This matters because the draw probability is concentrated in the region where both teams score the same number of goals in regulation. Getting the tails of the distribution right is what makes the draw probability accurate.
Both teams' goal distributions are fed into a bivariate Poisson model that generates a joint probability table across every possible regulation score. From that table, three probabilities emerge: P(Home wins in regulation), P(Away wins in regulation), P(Draw / goes to overtime). They sum to exactly 100%.
Raw Poisson systematically underestimates draws by 5-7 percentage points. Bayesian tie inflation corrects for this by blending the long-run historical OT rate with current-season evidence. When 2025-26 started producing OT games at 25-26%, the model's calibration parameter adapted within weeks while bookmakers stayed anchored to the historical average.
Standard devigging approaches don't account for how bookmakers distribute margin across three outcomes. Power devigging finds the exponent k that makes all three fair probabilities sum to exactly 100%, the only method that correctly handles vig in a 3-way market.
The draw edge calculation runs independently against 20+ bookmakers. The same model can show +40% ROI against one book and −15% against another on identical games. Sharp books have tightened draw pricing. Soft and international books haven't. Finding the best available price per game, not just any edge, is what the pipeline optimizes.
Four seasons run as a 2-way moneyline model before switching to the regulation draw in 2025-26. The draw was consistently eating into returns — moving to 3-way fixed that. Live results are tracking well above the prior model's averages.
Most bettors ignore the regulation draw entirely. Books price it as an afterthought. That mispricing is exactly where this model operates. The season runs through the end of the playoffs — every round is still a live opportunity.
Get NHL ModelDaily regulation draw picks with full context: model probability, best available book, edge %, and regime context, delivered direct to Discord.
Draw picks posted when puck lines become available for qualifying games, with a minimum 2.5% edge threshold before any pick is released. Games where no book clears the threshold are skipped entirely.
Every pick includes the model's draw probability, the fair value price, the best available odds across 20+ books, the edge %, and a clear play/skip call. The model shows which books are lagging and which have already corrected.
Each pick includes the current-season OT rate trend and how the model's calibration parameter has adapted. When the market is anchored to a stale prior and the model has updated, that gap is documented in the pick card.
All picks land directly in the Whizard subscriber Discord with the full model output card: draw probability, fair price, best book, edge %, and regime context.
Regulation draw picks available now through the end of the season. Available individually or as part of the All Access bundle.