READING THE BOARD
Hockey expected goals: why xG is not a win probability
An expected-goals total describes modeled scoring opportunities; it is not a percentage chance of winning. A hockey team credited with 0.8 xG does not therefore have an 80% win probability. Read the quantity, event sample and model definition before using the number to judge a pick.
What the model is estimating
MoneyPuck's model documentation describes its expected-goals model as assigning a scoring probability to a shot. It lists inputs including shot location, shot type, the preceding event, manpower and empty-net status. It separately describes a pregame model for predicting the winning team. Those are different prediction targets.
This guide uses invented probabilities, not MoneyPuck data or current NHL results. It does not reproduce or validate that provider's model. The distinction matters even when a provider displays both statistics on the same game page.
Add expected counts, not winning chances
Imagine a fixed list of three fictional shot opportunities. For this arithmetic exercise, their supplied scoring probabilities are 0.10, 0.25 and 0.45. These are separate opportunities, not a rebound sequence in which a goal would cancel later shots.
| Opportunity | Supplied probability of a goal | Expected-goal contribution |
|---|---|---|
| A | 10% | 0.10 |
| B | 25% | 0.25 |
| C | 45% | 0.45 |
The expected count is 0.10 + 0.25 + 0.45 = 0.80 goals. A realized count in this toy exercise can be zero, one, two or three; it cannot be 0.80. The expectation is an average across repetitions under the supplied assumptions, not a promised score. Adding expected contributions does not require independence, provided the supplied probabilities refer to the same defined experiment.
Even the chance of scoring once is a different question
Add the extra assumption that these three outcomes are mutually independent. Then the chance of no goals is (1 − 0.10) × (1 − 0.25) × (1 − 0.45) = 0.37125. The chance of at least one goal is therefore 1 − 0.37125 = 62.875%, not 80%.
Independence is stipulated only for this calculation. It is not a claim about actual hockey shots. If the outcomes are dependent, the three marginal probabilities alone do not determine the chance of at least one goal. Neither calculation supplies the opponent's scoring distribution, so neither gives a winning probability.
Keep an observed game separate from the next game
Summarizing chances already created asks a retrospective question. Forecasting the next matchup requires assumptions about opportunities that have not happened yet. Reusing last game's shot list silently assumes the future will supply that same list. It also leaves opponent scoring, game state and the relevant outcome definition unresolved.
A reader checking an xG claim can retain four things: the provider and model version; the included games and observation time; the situations included, such as manpower or empty-net filters; and the exact question being answered. Compare like-for-like samples. If someone converts an xG total directly into a winning percentage, ask for the intervening model and its retained evaluation record.
Apply the distinction to our board
Monitor Sports Picks uses the experimental method documented here; this article does not add an expected-goals model or a new data feed. Our forecast-evaluation guide explains how to assess probability claims. When comparing hockey outcomes, also retain the market's treatment of regulation ties and later play instead of assuming every winner label refers to the same event.
MoneyPuck model documentation checked October 22, 2026. All shot values and calculations above are original hypothetical examples. No actual game results, predictive advantage, provider endorsement or independent validation is claimed.