How the ELO System Works

What is ELO?

ELO is a rating system originally developed for chess that measures the relative skill level of teams. In college football, it provides a more accurate ranking than traditional polls by empirically considering strength of schedule.

How Teams Gain/Lose ELO

Starting Values

Teams start each season with ELO ratings based on their previous year's final ELO ratings, regressed 5% towards the mean of 1500. New teams or teams without history start at 1500.

Model against the market

Every spread is points the home team is favoured by. A positive difference means the model likes the home team more than the book did.

Lattice vs the Opening Line - 2025

Model
Lattice Elo
-Games
-Mean gap (pts)
-Model lean (pts)
-Off by 3+
Wk Date Matchup Lattice Pre-market BovadaDraftKingsESPN Bet Consensus Difference Result

How Lattice does against each book

Two different bars, and they usually disagree. Miss is the average distance from the final margin — who forecast better. Against the spread takes the book's opening number as the line, bets whichever side the model prefers, and counts covers; a bet at the usual −110 price needs 52.4% to break even, not 50%. Every season this data covers is shown, so a single strong cell can be read against the twelve others rather than on its own.

Season Book Games Lattice miss Book miss Difference Against the spread Cover rate
2021 Bovada 750 13.01 12.67 -0.33 360–376–14 48.9%
2021 Consensus 750 13.01 12.67 -0.33 360–376–14 48.9%
2022 Bovada 759 12.53 12.09 -0.44 366–376–17 49.3%
2022 Consensus 759 12.53 12.09 -0.44 366–376–17 49.3%
2023 Bovada 778 12.77 12.38 -0.40 374–391–13 48.9%
2023 DraftKings 717 12.54 12.18 -0.36 349–354–14 49.6%
2023 Consensus 780 12.77 12.33 -0.44 370–405–5 47.7%
2024 Bovada 785 12.40 12.19 -0.21 387–380–18 50.5%
2024 DraftKings 690 12.37 12.51 +0.13 361–322–7 52.9%
2024 ESPN Bet 750 12.34 12.17 -0.17 369–371–10 49.9%
2024 Consensus 788 12.43 12.28 -0.15 403–380–5 51.5%
2025 Bovada 808 12.20 11.88 -0.32 406–388–14 51.1%
2025 DraftKings 667 12.05 11.86 -0.19 355–308–4 53.5%
2025 ESPN Bet 756 12.17 11.88 -0.29 389–364–3 51.7%
2025 Consensus 808 12.20 11.85 -0.34 423–378–7 52.8%
All seasons Bovada 3880 12.58 12.24 -0.34 1893–1911–76 49.8%
All seasons DraftKings 2074 12.33 12.19 -0.14 1065–984–25 52.0%
All seasons ESPN Bet 1506 12.26 12.02 -0.23 758–735–13 50.8%
All seasons Consensus 3885 12.58 12.24 -0.34 1922–1915–48 50.1%

Read the whole table before any one row of it. The model's miss is larger than the book's in every season against every book, and the cover rate sits within a couple of points of a coin flip. Thirteen book-seasons are on screen; at that many, one or two landing above the break-even line is what chance alone produces, and no cell here is far enough from 50% to survive being the best of thirteen. The market correction improved every one of these figures and changed none of these conclusions — which is what MODEL-60 measured when it was fitted, and what it still says here.

Every number is points the home team is favoured by, so the book's own sign is flipped. Neutral-site games have no home team and are listed alphabetically, marked “(N)”. Opening spreads come from CFBD and exist from 2021: Bovada from 2021, DraftKings from 2023, ESPN Bet from 2024.

The Lattice column is the model as the site serves it, including the market correction as it stood going into that week — never a later one, and never this game's own line. Pre-market is the same prediction without that correction: what the rating system knows without consulting the market at all. The correction is fitted to pull predictions toward the market, so it shrinks the Difference column by construction; the pre-market number is the one to read as independent. Neither affects the scoreboard above, which is scored against what actually happened. A team's first game of a season is never corrected, in any week — the term reads games already played and there are none yet, which is why September stays the model's weakest stretch and cannot be fixed from here.