NHL Predictions & Picks
Model predictions for every game • Picks • 2026-28 Season
Error loading data
> Our NHL prediction model generates moneyline, puck line, and over/under outputs for every game on the board each night. V1 and V2 run in parallel -- V1 with a broader game selection, V2 applying tighter edge thresholds. The model has been running since 2020 and the ML record has improved in four of five seasons: 52.7% (2020) -> 56.5% (2021) -> 61.3% (2022) -> 59.7% (2024 and 2025). All results are tracked publicly. These outputs are a starting point for your research -- not a final answer.
NFL 2025-26 Season Complete
47-38 | +100.5u
vs last season: +55.3u improvement ↑
Season begins October 2026
CFL 2025-26 Season Complete
47-38 | +100.5u
vs last season: +55.3u improvement ↑
Season begins October 2026
Season: 12-8 | +4.50u
Today's Picks • Apr 17
No picks posted yet for today
Last night: 1-0 | +0.91u
No results yet
LAST NIGHT: 1-3 • -2.17u
BETTING RECORD
1-3 | -2.17u
TEAM PICKS RECORDS
Today's Picks • Apr 17
No picks posted yet for today
MODEL PERFORMANCE
Today's Picks • Apr 17
No picks posted yet for today
MODEL PERFORMANCE
Updated Apr 5, 7:12 am
The Puck Line Dog Signal
NHL underdogs on the puck line have been the most consistent signal in the entire dataset. V1 dog ATS by year: 66.9% (2020), 61.9% (2021), 56.7% (2022), 61.7% (2023), 62.6% (2024), 61.2% (2025). Every full season over 55%, four of six over 61%. The market undervalues NHL underdogs consistently -- not every year by the same amount, but the pattern hasn't gone away. If you're using these predictions for your own research, start with the dog ATS column.
How We Use These Predictions
Every night's NHL picks come directly from this table -- games where the model shows meaningful edge, the line is still actionable, and matchup context (goalie starts, travel, rest days) supports the signal. The predictions page shows the full model view. The picks page shows which signals cleared the full research filter. Both are published so you can follow the picks or run your own process from the raw model output.