How Elo Ratings Turn Match Results into Win Probabilities
In 1960, physicist Arpad Elo designed a rating system for the United States Chess Federation. It was meant to rank chess players. Sixty-plus years later, the same mathematics quietly powers a large share of soccer modeling β club rankings, published projection systems, and the odds compilers' internal tools all lean on Elo-style ideas. The reason it survived is its elegance: one number per team, one formula to update it, one formula to turn it into a probability.
The one idea
Every team gets a rating β a number like 1700 or 1950. The difference between two ratings tells you how the match should go, through a single logistic formula:
EA is team A's "expected score": the share of the points at stake it should take on average β a win counts as 1, a draw as 0.5, a loss as 0. The 400 in the denominator sets the scale. Watch what rating gaps mean in practice:
| Rating gap | Stronger's expected score | In plain words |
|---|---|---|
| 50 points | 0.57 | slight edge |
| 100 points | 0.64 | clear favorite |
| 200 points | 0.76 | heavy favorite |
| 400 points | 0.91 | mismatch |
A 100-point favorite is expected to take about 64% of the available points β the rest goes to the underdog through draws and upsets. Note the symmetry: the same gap always means the same expectation, whether the teams are rated 1900 and 1800 or 1300 and 1200.
How the rating updates
After each match, the rating moves toward reality. The update rule is almost comically simple:
Say a 1900-rated side hosts a 1700-rated side (200-point gap, expected score 0.76) with K = 30. A win gives the favorite 30 Γ (1 β 0.76) = +7.2 points; it only gains a little, because it was supposed to win. An upset loss costs it 30 Γ (0 β 0.76) = β22.8 points, while the underdog gains exactly that. A draw lands in between: 30 Γ (0.5 β 0.76) = β7.8 for the favorite.
The K-factor controls how jumpy the system is. Chess typically uses 16β32; soccer models often use higher values because seasons are short and a team changes fast. The total points are conserved β what one team gains, the other loses β so the system is a closed economy of reputation.
Soccer's three adaptations
Chess Elo needed tweaks before it worked on football:
- Goal margin. Beating someone 3β0 should move the needle more than scraping through 1β0. Soccer systems scale the K-factor by the margin of victory β a common approach weights each extra goal less than the one before, so 6β0 doesn't count six times as much as 1β0.
- Home advantage. The home team is effectively stronger just by being home. Most systems add roughly 100 rating points to the host before computing the expectation β that's the classic ~0.64 expected-score bump, matching the long-run home edge in league football.
- Draws. Chess Elo's "expected score" already handles draws, but soccer modelers usually want separate 1X2 probabilities. The standard solution: take the rating difference, then split the non-draw share using the league's empirical draw rate at that matchup strength. Elo gives the balance; the draw model gives the shape.
Where Elo shines β and where it doesn't
It shines at cross-league comparison. A goals-based model like our Dixon-Coles predictor needs teams to share a league to compare them β a Champions League tie between a Portuguese and a German side has no common baseline. Elo sidesteps this: any result between any two rated teams updates both, so international competitions feed straight into the same scale.
It struggles with anything the scoreline doesn't say. A 1β0 win with a rotated squad counts the same as 1β0 at full strength. A new manager's first month is invisible until the results accumulate. Promoted teams carry their lower-league rating into a division where it means something different. And the choice of K is a judgment call with real consequences: too high and the rating chases noise; too low and it sleeps through genuine decline.
Used honestly, Elo is a disciplined summary of "what the results say so far" β the right starting point, and the wrong finish line. The numbers tell you what usually happens next; everything else on this site exists for the cases where "usually" needs a second opinion.
For informational and educational purposes only. Nothing here is betting advice.
Related reading: How the Dixon-Coles Model Predicts Soccer Scores Β· Poisson vs Dixon-Coles: Where the Difference Actually Matters