Win probabilities from the Poisson matchup engine
(matchup__5184e7b0) — team xG-rate ratings → expected goals → a score grid.
Every number below was issued before puck drop and is scored honestly
against the result. How it works & why it can lose to a logistic.
Click a matchup for its page — an upcoming game shows the predicted shot
map and breakdown; a finished game opens on what happened, next to the pregame call.
Every slate the engine has ever scored is browsable from the date selector.
Or build any matchup in the lab — pick the teams, drag
the rest sliders, and score it with the same engine.
Each column is a team's expected goals and each puck one of its goals, on one
scale for the whole night, so a stack taller than its column finished above its chances.
An empty-net goal is an outlined puck (its attempts are not in xG); the shootout winner's
credited goal gets no puck. The big number is the official final.
Season accuracy tracker
Log-loss (lower is better) over 135 completed games this
season, every prediction scored as it was issued. This is the public, unfalsifiable record —
the append-only live file means we cannot re-grade a past call.
Engine
log-loss
accuracy
games
Logistic (challenger)
0.639
66.7%
135
Poisson (shipped)
0.643
65.2%
135
Elo (baseline)
0.649
60.7%
135
Home-rate (floor)
0.683
58.5%
135
The shipped engine is Poisson. Its backtest twin, a plain
Logistic regression on the same ratings, edges it on winner log-loss by ~0.003 —
but only the generative Poisson produces the expected-goals score grid needed to price totals and
puck lines, so it stays the headline. Both beat Elo and the home-rate floor. Full walk-forward
methodology in how the engine works.
