Model Information

How ShiftEdge works.

This page explains how the separate Game Winner and Game Winners models work, what their probabilities mean, how each was historically tested, and how live 2026–27 results are tracked. The Predictions page is reserved for the actual games, probabilities, player and goalie impacts, and market comparisons.

View Game Predictions View Model History View Betting History

EV is the primary metric

Expected value (EV) is the primary metric for comparing picks on ShiftEdge. The model estimates probabilities, compares them with available market prices, and uses the resulting EV to identify value. Win probability and edge provide important context, but picks should be evaluated primarily by EV—not simply by which team is more likely to win. A lower-probability side can still represent the stronger value when the available price produces higher positive EV.

How it works

The live prediction is built from the expected players and goalie for each team plus schedule context, then translated into a win probability.

1. Historical dataPlayer game history supplies the model's rolling inputs.

2. Current lineupThe most up-to-date projected lineup is used before puck drop.

3. Team contextSkater and goalie history—including rolling skater performance and the projected starter's selected rolling goalie statistics—plus rest and schedule context are calculated.

4. ModelThe active Game Winner model combines its currently selected roster, goalie, matchup, shot/possession, and schedule/form features into a win probability.

5. Market comparisonModel probability is compared with available market prices to calculate edge and EV.

Winner-168 scores are available only for independently verified matchups. Other games remain unscored until the original advanced-stat feed is restored. Only Winner-168 is authorized for future predictions; no previous-model Game Winner picks are currently being published. The notebook evaluation and prospective Winner-168 record are reported separately.

Game Winners: 168-feature L2 Logistic Regression

Five-season historical sample · 2021–22 to 2025–26

168Features

6,559Training games

5Seasons

Five-season test results: Not verified6,559 is the training sample, not an independently tested five-year record. It is not an untouched pregame holdout.

Model information

ShiftEdge currently runs Winner-168 for NHL game winner probabilities, using only verified pregame information.

How to read a prediction

EV is the most important stat when comparing picks on ShiftEdge. Win probability, market probability and edge explain the opportunity; EV combines the model probability with the available price and is the site's primary value metric.

Example: Team AIllustrative example only

58%Model Win Probability

50%Market Probability

+8.0%Edge

+16%Expected Value

How to interpret it: the model estimates Team A wins about 58% of the time, while the available price values the team closer to 50%. That gap creates an 8-point model edge. If the offered odds produce positive EV using the model's 58% estimate, the price may represent value. This does not mean Team A is guaranteed to win the individual game.

What can change a prediction?

Pregame probabilities can move as the information available to the model changes. The site continues updating future games until the official prediction is locked at puck drop.

Lineup changes

Players entering or leaving the expected lineup can alter roster continuity and the player inputs used for the matchup.

Goalie confirmation

A change in the projected or confirmed starter can affect the goalie-related inputs—including the starter's selected rolling statistics—and move the team's projected win probability.

New player data

Completed games add new player history to the system, which can change the rolling values used in future predictions.

Rest & schedule context

Days of rest, back-to-back status and the timing of the matchup are incorporated into the model and can differ from game to game.

History & transparency

Model performance and betting performance are intentionally tracked separately.

Model History

Every completed 2026–27 regular-season game with the final locked model pick and whether that pick was correct. This measures the predictive model itself.

Open Model History

Betting History

Only the wagers that qualify for the site's official 2026–27 betting record. This measures the results of the betting rules, not every model prediction.

Open Betting History

Definitions

Reference definitions for the main terms used throughout Predictions, Matchup Details, Top Picks, History and model testing.

Brier Score

Measures the squared error of the model's predicted probabilities. It rewards probabilities that are close to the actual outcomes and penalizes probabilities that are too high or too low. Lower is better, with 0 being perfect.

Edge

Model win probability minus market probability, shown in percentage points. Example: 60% model vs 52% market = +8.0-point edge.

Expected Value (EV)

The estimated average return from a 1-unit wager using the model probability and listed odds. Positive EV means the model sees theoretical value at that price. EV is the primary metric ShiftEdge uses to compare potential picks.

Goalie Impact

The change in team win probability when the exact saved model is rerun with the projected starter replaced by the next goalie; this includes the selected rolling goalie features.

Historical evaluation

Historical tests are not the same as independent pregame predictions. Only locked picks count toward the live record.

Locked Prediction

The final pregame model prediction preserved at puck drop for official model-history grading.

Log Loss

Measures the accuracy of the model's predicted probabilities while penalizing confident incorrect predictions especially heavily. Lower is better, with 0 representing perfect predictions.

Market Probability

The probability implied by the listed market price. Sportsbook probabilities are normalized to remove bookmaker margin where applicable.

Player Impact

The change in team win probability when the exact saved model is rerun with that skater replaced by the listed eligible replacement.

ROC-AUC

Measures how well the model separates winners from losers across every possible probability cutoff. A score of 0.50 is roughly random; higher values indicate better ranking ability, with 1.00 being perfect separation.

Top Pick

A game-side opportunity that meets the site's current Top Picks EV threshold.

Win Probability

The model's estimated probability that a team wins the game.