Why NHL betting lines move and our Elo model doesn't

We compared our NHL Elo model with the closing betting line on every game so far this season. The two differ by a few points, and the difference is not random.

Research as of Nov 20, 2026.

The short answer. Betting lines move through the day as news arrives. Our Elo rating moves only after a game is played. On the 323 games of 2026-27 so far (through Nov 19), the closing line and our Elo differed by 4.1 win-percentage points on average, and in 60 games (18.6%) they picked different favourites. The closing line was more accurate than our Elo, which is normal for a model built from results alone. We keep the model that way on purpose.

1. What our Elo is, and what it never sees

Every NHL team has a rating that moves after each game: up after a win, down after a loss, by more for a bigger margin, with a fixed bonus for home ice and a partial reset toward average each summer. That is the whole model. It is built from final scores only.

It never sees a betting line, an injury report, a lineup or the starting goalie. So when a star is scratched at 6 PM, our number does not change. The betting line usually does. This article is about that difference: how large it is, who is closer to the truth, and why we leave it alone. For a concrete example of a lineup move see how much a missing NHL star is worth.

2. How big the gap is

We took every 2026-27 game with a closing line (323 games, Oct 4 to Nov 19), removed the book's margin so the two sides add to 100%, and compared the home team's win chance with our Elo's.

MeasureResult
Average gap4.1 points
Median gap3.6 points
Games with a gap of 5+ points35%
Games with a gap of 8+ points11%
Games where they pick different favourites60 of 323 (18.6%)

Most of the gap is not late news. Roughly three-quarters to 85% of it is already in the first line we record (on game day for most games, the day before for about 120), and lines then drift a little further from our Elo by the close.

On average the closing line gives the home team about 1.7 points less than our Elo does. Our home-ice bonus is a fixed number; the line's is smaller, at least this season.

3. Who is more accurate

We scored both with log loss, a standard measure that rewards confident correct calls and punishes confident misses (lower is better). The closing line beat our Elo in 2026-27 by 0.0136, a small gap but a clear one. On 260 late-2025-26 games the closing line was ahead by 0.0093, but that difference was too small to tell from chance. In that late-2025-26 sample our Elo actually picked more winners than the close (57.7% vs 54.6%).

This is the normal result. A line absorbs lineups, goalies, travel and a great deal of money from people who watch every game, and it has the final day of information. A public model built from scores should expect to trail it, and ours does. We are not hiding that, and it does not change how the model is graded: on its own record, against what we wrote down before the season.

One more detail. When we put both the line and our Elo into one model, the line's disagreement with Elo carried information about who won: statistically clear, but the size is uncertain. And much of that signal came from October. In November alone the sample was too small to say. A longer sample is the only fix, and we plan to re-read it at the trade deadline and at season end.

4. Why the line and Elo differ

We tried to name the reasons: usual regulars out, backup goalies, shot quality, rest, the summer reset and popular teams. Together they line up with about two-thirds of the gap. That is a description, not proof of cause.

  • Rest. It mattered in both samples. The line gives a team on the second night of a back-to-back roughly 3 to 5 points less than our Elo does.
  • Absences and shot quality. Both mattered, but how much credit each one gets changed a lot between our two samples, so we will not rank them.
  • Backup goalies and popular teams. Small in both samples.

Two cautions. The lineup and goalie data are who actually played, which the books may only have partly known at the close. And these factors overlap, so any split of the gap among them is a convention, not a measurement. For rest see what actually predicts NHL games.

5. Does the gap shrink as the season goes on?

We expected the gap to be biggest in October, when the Elo is still carrying last summer's reset, and to shrink. It did not: the average gap was 4.15 points in October and 4.02 in November. We found no sign that it is larger early or that it narrows with games played. An early-season explanation we looked at, that the line regresses last year's teams less than our summer reset does, did not hold up once we accounted for the Elo ratings themselves, so we are not claiming it.

6. Why we keep the model honest instead of chasing the line

If we fed the closing line into our model, it would get closer to the line and our record would stop meaning anything: it would be the market's opinion with our name on it. The point of a results-only Elo is that its graded record tells you what a plain, public, rules-fixed model does by itself. The rule was written down before the season and never retuned.

That also means we never treat a line as a reason to like a team. A game is a play only if our own Elo gives one side 62% or more. The line is something we measure our model against after the fact, not something that picks games. See NHL moneyline and puck line for the record, and the base rates behind it.

7. Caveats

  • 323 games over 47 dates is a small sample. The check on late 2025-26 (260 games) is a direction check only.
  • Lineups and goalies are hindsight. Nothing here says a live lineup feed would help.
  • No cause is proven. The factors overlap and their shares of the gap are not stable.
  • This is research on how a model compares with a line. It says nothing about profit and is not a reason to bet any game.

Frequently asked questions

Why do NHL betting lines move during the day?

Books take in news and money all day: lineups, starting goalies, rest and shot-quality trends. In our 323-game sample the closing line sat 4.1 win-percentage points from our Elo on average, Most of that gap was already there in the first line we recorded, at 3.5 points.

Is the closing line more accurate than a public NHL model?

Usually, yes. On 323 games of 2026-27 the closing line scored better than our Elo on log loss, a standard accuracy measure, by 0.0136, and the gap was clear. That is normal for a model built from results alone, and it is not a failure.

Does your Elo model use betting lines?

No. It is built from final scores, home ice and a fixed rule for the summer reset. It never sees a betting line, an injury report or a starting goalie. Keeping it separate means its record measures the model itself, not a copy of the market.

What explains the gap between the line and Elo?

About two-thirds of it lines up with things we can name: rest and back-to-backs, shot quality, and who is out. Rest mattered in both samples we checked, at roughly 3 to 5 points for the team on the second night. Backup goalies and popular teams were small.

The week ahead, once a week

Last week's graded record — wins and losses — and the plays for the week ahead. One email, no daily noise, unsubscribe in a click.

Educational content only. Research results are historical, from our own data, and describe what happened, not what will happen. Must be 21+ and in a jurisdiction where sports betting is legal. If gambling is a problem for you or someone you know, call 1-800-GAMBLER or visit ncpgambling.org.