← Blog · Methodology Series
How ELO Ratings Work in Football Prediction
Published: June 2026 · 8 min read
The ELO rating system — originally designed for chess by Arpad Elo in 1960 — has become one of the most widely used mathematical frameworks for ranking football teams. At FootballMoney, ELO serves as the backbone feature in our prediction ensemble, providing a single-number representation of team strength that updates after every match.
1. The Core ELO Formula
The ELO system works on a simple principle: after each match, the winner takes points from the loser. The number of points exchanged depends on two factors:
Where:
- RA, RB = current ELO ratings of Team A and Team B
- EA = expected score for Team A (0 to 1, where 1 = certain victory)
- SA = actual result (1 = win, 0.5 = draw, 0 = loss)
- K = K-factor controlling how much a single match changes ratings
- 400 = scaling factor: a 400-point gap means the stronger team is expected to win ~91% of the time
2. How We Translate ELO to Win Probability
The expected score formula gives us a direct win probability estimate. For example, if Team A has an ELO of 1600 and Team B has 1500:
E = 1 / (1 + 10−100/400)
= 1 / (1 + 10−0.25)
= 1 / (1 + 0.562)
= 0.640 → 64.0% chance
Quick Reference: ΔELO → Win Probability
| Δ < 30 | 50-54% | Near-even |
| Δ 30-80 | 54-61% | Slight edge |
| Δ 80-150 | 61-70% | Clear edge |
| Δ 150-250 | 70-81% | Strong favorite |
| Δ > 250 | > 81% | Dominant mismatch |
3. Home Advantage Adjustment (HFA)
Raw ELO doesn't account for home advantage — a well-documented phenomenon where the home team wins approximately 45% of matches, compared to 28% for away teams. We apply a league-specific HFA bonus to the home team's ELO before computing expected scores:
| Match Type | HFA Bonus | Rationale |
|---|---|---|
| Domestic League | +80 | Strongest HFA, consistent across season |
| European Competition | +70 | Travel distances larger, crowd impact slightly lower |
| Domestic Cup | +70 | Rotation can dilute HFA |
| International | +60 | Neutral venues in tournaments reduce HFA |
| Friendly | +60 | Low stakes, rotations common |
The adjusted ELO is then used in the probability formula. For a domestic league match with ELO 1600 vs 1500 and +80 HFA:
P(home win) = 1 / (1 + 10−180/400) = 73.8%
Without the HFA, this would be 64.0% — nearly 10% lower. This is why ignoring home advantage in ELO is a critical mistake.
4. K-Factor: How Fast Should Ratings Change?
The K-factor controls how responsive ELO is to new results. Too high and ratings oscillate wildly; too low and they lag behind real form changes. FootballMoney uses a dynamic K-factor based on match importance:
- World Cup finals: K = 60 (maximum weight)
- Continental championships: K = 50
- World Cup qualifiers: K = 40
- Domestic league: K = 30
- Domestic cup: K = 25
- Friendly: K = 15 (minimum weight)
This ensures that a World Cup knockout win moves ELO significantly more than a pre-season friendly — reflecting the different information content of each match type.
5. ELO's Role in the FootballMoney Ensemble
ELO is the single most predictive feature in our model, but it's not the whole story. In our ensemble, ELO serves as the base prior — the starting point that gets adjusted by:
- Recent form (5-10 matches) — weighted by recency and competition level
- Expected goals (xG) differential — captures underlying performance beyond results
- Injury impact score — adjusts team strength for missing key players
- Motivation score — accounts for league position, must-win scenarios
- Market odds — incorporates sharp bookmaker information as a calibration signal
The Bayesian Poisson model takes this adjusted ELO difference as a prior for the goal-scoring rate parameter, producing a full probability distribution over possible scorelines — not just a single win percentage.
6. Limitations of Pure ELO
ELO alone predicts home win rates at ~52% accuracy on a 3-way classification — above random (33%) but below what's needed for profitable betting. Key limitations:
- Draws: ELO's expected score formula doesn't naturally model draws — we use a separate ordered logistic regression for draw probability estimation
- Squad changes: ELO treats teams as static entities. Major transfers, managerial changes, or injuries that aren't captured by our injury score can create rating lag
- New teams: Teams entering the system (promoted clubs, new national teams) start at 1500 with high uncertainty — it takes 15-20 matches for their ELO to stabilize
- Market efficiency: Sharp bookmaker odds often contain information ELO hasn't yet incorporated, which is why we use both signals
This is exactly why FootballMoney uses ELO as one input among many in a Bayesian ensemble, rather than relying on it as a standalone predictor.
Continue Reading
Next in series: How Bayesian Poisson Models Predict Football Scores — learn how we turn ELO differences into full probability distributions over every possible scoreline.