Model transparency
Pre-game model performance
Compare two closely related versions of the NHL pre-game win probability model using overall accuracy, confidence ranges, probability calibration, rolling trends, and team-level results.
About model versions
V1 is the original model retrained for the new season. V2 uses the same core approach with a small number of additional features, so the two versions are expected to produce similar results. Filters apply to every performance section below.
Overall results
Model summary
Completed games for Model V2 matching the selected season and game type.
XGBoost · Model V2 · Updated nightly · Completed games through yesterday
Who picks more winners?
Accuracy on 106 matching games with usable pregame odds and predictions. This is a smaller sample than the overall summary.
Higher accuracy does not necessarily mean higher profit. The price paid on each winner matters too.
Model picks
Net units+8.63
ROI+8.14%
Market favourites
Net units+2.84
ROI+2.67%
Simulated one-unit bets on the same 106 matching games, using the existing betting service's best captured prices. The full betting page may include additional games.
How this comparison works
Uses the existing betting evaluation service's pregame odds and prediction cutoffs. The market favourite comes from average margin-adjusted home probabilities across paired sportsbooks. Exact market ties are excluded. Both approaches are scored on the same games. Bars show accuracy; cards show simulated net profit and return on the total amount staked.
Prediction quality
Accuracy by confidence range
This shows how often Model V2’s predicted winner was correct within each predicted home-team win-probability range.
View detailed results
| Predicted home-win range | Games | Correct predictions | Accuracy |
|---|---|---|---|
| 30–40% |
3
Low sample
|
2 | 66.7% |
| 40–50% |
37
|
22 | 59.5% |
| 50–60% |
45
|
27 | 60.0% |
| 60–70% |
21
Low sample
|
13 | 61.9% |
Games
3
Correct
2
Games
37
Correct
22
Games
45
Correct
27
Games
21
Correct
13
Bins use 10% increments. Low sample means fewer than 30 completed games.
Probability quality
Probability calibration
Calibration compares average predicted home-team win probability with the actual home-team win rate in each range.
Predicted home-win chanceActual home-win rate
Similar bars mean better agreement.
View detailed results
| Predicted home-win range | Games | Average predicted | Actual home-win rate | Difference |
|---|---|---|---|---|
| 30–40% |
3
Low sample
|
39.0% | 33.3% | -5.7% |
| 40–50% |
37
|
46.2% | 40.5% | -5.7% |
| 50–60% |
45
|
54.7% | 60.0% | +5.3% |
| 60–70% |
21
Low sample
|
63.3% | 61.9% | -1.4% |
Games
3
Predicted
39.0%
Actual
33.3%
Games
37
Predicted
46.2%
Actual
40.5%
Games
45
Predicted
54.7%
Actual
60.0%
Games
21
Predicted
63.3%
Actual
61.9%
Difference equals actual home-win percentage minus average predicted percentage.
Positive values mean home teams won more often than predicted.
Trend
Overall model accuracy over time
Rolling Model V2 accuracy over the latest 25 completed games, shown at the end of each game day.
Each point shows the rolling accuracy at the end of that day. Until 25 games are available, all completed games are used.
Team analysis
Team-level model view
Select a team to compare average Model V2 probability, actual win rate, favorite and underdog records, and a rolling 10-game trend.
ANA
Rolling 10-game model probability versus actual win rate.
Updating team view…
One line shows rolling average model probability. The other shows actual
rolling win rate over the same 10-game span.
Winner accuracy
How often the predicted winner actually won. A 60% chance still allows plenty of losses.
Log loss
Measures how much chance the model gave the actual result. Being very confident and wrong is penalized strongly. Lower is better.
Brier score
Measures the gap between the predicted home-win probability and the result. Lower is better; zero would mean perfect predictions. This is not a percentage.
Confidence ranges
Games are grouped by the predicted home-team chance. Winner accuracy counts correct picks whether the model favoured home or away.
Confidence versus results
Home teams assigned about a 70% chance should win about 70% of the time across many games. Compare the predicted and actual bars.
Trends
Overall accuracy uses the latest 25 completed games. The team chart compares predicted chances with actual wins over 10 games. Short runs can fluctuate.
Small samples
A handful of games cannot establish reliability. Buckets with fewer than 30 games are flagged so a high percentage does not hide limited evidence.
Which games count?
Existing model results include completed games through yesterday for the selected version, season and game type. The market comparison requires extra odds coverage and uses matching games only.