How to Build an NHL Player Prop Odds Screen and Fair-Odds Model

An NHL prop odds screen collects shots and points prop prices from several books, maps every player and stat to one ID, and removes the margin per market. A fair-odds model then prices each prop independently, typically from projected ice time and shot rates with a Poisson or negative binomial distribution, and flags stale lines against consensus.

Two systems, not one

An NHL prop odds screen and a fair-odds model are separate builds that share a data store.

  • The screen collects prop prices from several books, normalises them into one schema, and shows where books disagree.
  • The fair-odds model produces a probability for each prop, either by removing the margin from market prices or by pricing the prop independently from player data.

The useful output comes from comparing the two: a book price, the market's devigged consensus, and the model's own number, side by side with timestamps.

Step 1: collect and normalise prop prices

Pull prices from an odds API or your own collection, and store every snapshot rather than overwriting the latest. Each row needs, at minimum: book, player, stat, line, side, price, and capture time. Record when a market disappears too, because a pulled prop is information.

Identity is the expensive part. Books spell names differently and use their own IDs, so map every player to one canonical ID and keep the mapping table under version control. Stat definitions need the same care: shots on goal are not shot attempts, and points are goals plus assists. Why prop data is harder to track covers the identity problem in depth.

Store each book's settlement rules per prop type, including whether overtime counts and what happens if a player is scratched. Two props with the same label and different rules are different markets.

Step 2: devig each market separately

Margins on props are wider and less uniform than on main lines, so remove the margin per market, not with one global factor. See how sportsbook margin works for the underlying arithmetic.

For a two-way over and under, an illustrative example: over 2.5 shots on goal at -130 and under at +100. Implied probabilities are 130 divided by 230, about 56.5 percent, and 50.0 percent, summing to about 106.5 percent. Dividing each by the total gives a fair over of about 53.1 percent, or roughly -113 in American odds.

Proportional scaling is the simplest method. Power or Shin-style methods shift more of the margin onto the longer price, which often fits props better. Pick one, apply it consistently, and test it against closing prices.

One-sided markets are harder. Anytime goal scorer and many points ladders are offered as overs only, so there is no opposite side to normalise against. Options are to estimate the book's typical margin on that market type from two-way markets, to derive the fair price from a full alternate-line ladder, or to rely on the independent model. Whatever the method, alternate lines for one player must stay monotonic: the fair probability of 3 or more shots cannot exceed that of 2 or more.

Step 3: price props independently

Market consensus tells you where books agree. An independent model gives a second opinion.

Shots on goal. Project expected shots as projected time on ice multiplied by the player's shots on goal rate per 60 minutes, split by even strength and power play, then adjust for the opponent's shot suppression, the expected game script, and home or away. Public sources such as the NHL's own stats pages, Natural Stat Trick, and MoneyPuck provide time on ice and shot rates.

Turning the mean into a probability needs a distribution. Poisson assumes variance equals the mean. Player shot counts usually vary more than that, because ice time and game state vary from night to night, so a negative binomial often fits better. Illustratively, with a projected mean of 2.9 shots:

  • Poisson gives about 55.4 percent for 3 or more shots.
  • A negative binomial with variance 3.8 gives about 52.6 percent.

Against the illustrative market fair price of 53.1 percent above, the distribution choice alone decides which side of the market the model sits on. Fit the dispersion from historical game logs rather than guessing it.

Points. A simple start treats points as Poisson with a per-game rate, so the probability of at least one point is 1 minus e raised to minus the rate. A rate of 0.75 gives about 52.8 percent. That ignores clumping: when a player's line scores, several linemates collect points on the same goal, so linemate props are correlated. Model points from projected team goals and the player's share of them for a more coherent result. Modelling NHL totals covers the team goal side.

Step 4: detect stale lines

A stale line is a book price that has not updated while the rest of the market has. Useful signals:

  • Age: time since that book's price last changed, compared with others on the same prop.
  • Consensus gap: the difference between the book's implied probability and the devigged consensus, after converting different line numbers onto one scale with the fitted distribution.
  • News trigger: consensus moved shortly after a lineup, line combination, or goalie confirmation, and this book did not.

Rule out false positives before flagging anything. A different settlement rule, a suspended market, a scratched player, or a line number mismatch can all look like a stale price. Requiring at least two other books to agree on the consensus reduces noise from a single outlier.

Step 5: validate the model

Log the model's fair probability at the time each price was observed, then compare with the devigged closing price and with results. Check calibration by bucketing predictions and comparing predicted with observed hit rates, and track a proper scoring rule such as log loss or Brier score. Prop samples per player are small, so pool across players carefully and expect wide intervals. How to validate a betting value model covers the method.

This page describes data and modelling method and is not betting advice.

Frequently asked questions

How do you devig a player prop with only one side offered?
There is no opposite price to normalise against, so estimate the book's typical margin on that prop type from comparable two-way markets and remove it, or derive fair prices from a full ladder of alternate lines, or lean on an independent model. Check that the resulting probabilities across a player's alternate lines decrease as the line rises.
Should NHL shots on goal props use Poisson or negative binomial?
Start with Poisson for simplicity, then test for overdispersion in each player's game logs. There is good reason to expect it: projecting a per-game mean from ice time and shot rate makes the realised count a mixture of Poissons, and any such mixture has variance above its mean, which is what negative binomial adds a parameter for. Treat the size of the effect as something to measure on your own logs rather than a published result. In an illustrative case with a mean of 2.9, the two disagreed by almost three percentage points on 3 or more shots.
How do you detect a stale prop line?
Compare each book's price with the devigged consensus of other books, adjusted to the same line number, and with how long ago each price last changed. Flag a book that has not moved after consensus shifted on news such as a goalie confirmation. Then rule out settlement differences, suspensions, and scratches before treating it as stale.
Do NHL shots on goal props include overtime?
Settlement rules vary by book, so record them per book rather than assuming. Many books count overtime toward player stat props. Shootout attempts are kept as a separate statistical category and are widely understood not to count toward a skater's shots on goal, but confirm it against the NHL's own statistics policy and the book's rules rather than assuming. Mixing books with different rules in one consensus price is a common source of false stale-line flags and mispriced alternate lines.