What Are Elo Ratings?
The Elo rating system was originally developed by Arpad Elo in the 1960s to rank chess players. Its elegant simplicity — teams gain rating points for winning and lose them for losing, with the amount depending on the expected outcome — has made it one of the most widely adopted strength-ranking systems in sport. FIFA used an Elo-based system for their official world rankings, and it underpins many of the most successful football prediction models in existence.
The core principle is straightforward: if a highly-rated team beats a lowly-rated team, few points change hands because the result was expected. If the underdog wins, a large number of points transfer. Over time, this produces a ranking that reflects true team strength more accurately than league tables, which are distorted by fixture difficulty, goal difference tiebreakers, and the number of matches played.
How the Elo Calculation Works
Every Elo system follows the same basic formula. Before a match, each team has a rating. The expected outcome is calculated using the logistic distribution:
Expected score = 1 / (1 + 10^((Opponent Rating - Team Rating) / 400))
If Team A has a rating of 1800 and Team B has a rating of 1600, Team A's expected score is about 0.76, meaning they are expected to win roughly 76% of the time. After the match, ratings are updated:
New Rating = Old Rating + K × (Actual Score - Expected Score)
The "actual score" is typically 1 for a win, 0.5 for a draw, and 0 for a loss. The K-factor determines how responsive the system is to new results.
The K-Factor: Responsiveness vs Stability
Choosing the right K-factor is the most important calibration decision in any Elo system. A high K-factor (e.g., 40) means ratings react quickly to results — great for capturing momentum shifts, but noisy and prone to overreaction. A low K-factor (e.g., 10) produces stable, reliable rankings but adjusts slowly to genuine changes in team quality.
In football, most well-calibrated Elo systems use K-factors between 20 and 40, sometimes adjusted for match importance. A World Cup final might use a higher K-factor than a mid-season league match, reflecting that more effort and preparation go into high-stakes fixtures.
Home Advantage Adjustment
Most football Elo systems add a fixed bonus to the home team's rating before calculating expected scores. This is typically between 60 and 100 Elo points, which translates to roughly a 10-15% increase in expected win probability. As we discussed in our home advantage analysis, this effect varies by league and context, so more sophisticated implementations use variable home adjustments.
Goal Margin Multiplier
A significant enhancement to basic Elo is incorporating goal margin. A 4-0 win contains more information about relative team strength than a 1-0 win. Systems like the one used by FiveThirtyEight apply a goal difference multiplier to the K-factor, so large victories produce bigger rating changes. The multiplier is typically logarithmic to prevent blowout results from having disproportionate influence.
Elo vs Expected Goals: Different Lenses on Team Quality
A question we frequently encounter is whether Elo ratings or expected goals (xG) are better for prediction. The answer is that they measure different things and complement each other well:
- Elo measures outcome strength: How good is a team at winning matches? Elo does not care whether a team wins 1-0 from a scrappy set piece or 3-1 from open play brilliance — it just sees the result and adjusts.
- xG measures process quality: How good is a team at creating and preventing high-quality chances? xG captures performance independent of finishing luck and goalkeeping variance.
A team can have a high Elo rating despite mediocre xG numbers (because they are efficient finishers and their goalkeeper consistently outperforms), and vice versa. Models that combine both signals outperform those using either alone, because they capture both the "what happened" and the "what should have happened" dimensions of team strength.
Real Examples of Elo in Action
Consider the following scenario from a recent Premier League season. In October, a promoted team sat in the bottom three with just seven points from nine matches. Their Elo rating was low, reflecting their league position. But their xG data showed they were creating 1.4 xG per match while conceding only 1.1 xG — the profile of a mid-table team being unlucky with finishing. Over the next three months, they climbed to 12th, and their Elo rating adjusted upward to reflect the improved results.
A model using Elo alone would have been slow to recognise the team's true quality. A model using xG alone might have overrated them if their finishing remained poor. The ensemble approach — combining Elo's result-based evidence with xG's process-based evidence — identified the value earliest.
Limitations of Elo Ratings
Elo is powerful but it has genuine limitations:
- No squad-level detail: Elo treats a team as a single entity. It cannot distinguish between a full-strength squad and one missing five key players to injury.
- Slow to adapt: By design, Elo changes gradually. A mid-season manager change that transforms a team's style and results will take multiple matches to fully reflect in the rating.
- Cross-league comparisons are unreliable: Elo works best within a single league where teams play each other regularly. Comparing Elo ratings across different leagues is problematic because the rating pools are partially isolated.
- Draws are undervalued: The standard Elo treatment of draws (0.5 for each team) can undervalue teams that draw frequently, particularly in leagues with high draw rates.
How PredictPitch Uses Elo in Its Ensemble Model
In our prediction engine, Elo ratings serve as one of several input features to the machine learning ensemble. Specifically:
- Elo difference: The gap between home and away team ratings is a direct input, capturing the relative strength of the two teams.
- Elo trend: The direction and magnitude of each team's Elo change over the last five to ten matches captures momentum and form effects that the raw rating alone might miss.
- League-calibrated Elo: We maintain separate Elo systems for each league, with K-factors optimized for that league's specific characteristics (e.g., higher volatility in Ligue 1 vs lower in the Bundesliga).
The ensemble model — combining XGBoost, LightGBM, and CatBoost — learns how much weight to give Elo relative to other features like xG, form sequences, head-to-head records, and squad strength metrics. In some contexts Elo is highly predictive; in others, different signals dominate. The model adapts automatically.
Using Elo for Your Own Analysis
Even without building your own model, understanding Elo ratings can improve your match analysis:
- Check the Elo gap: A match between teams rated 1750 and 1550 is heavily one-sided. Between 1650 and 1600 it is close. The gap tells you how competitive the match should be.
- Track Elo trends: A team whose Elo has risen 80 points in the last month is on a genuine upswing. One whose Elo has dropped 60 points is in decline.
- Compare Elo to league position: Teams whose Elo suggests they are stronger than their league position often represent value, as their results are likely to improve.
- Use Elo for cross-match comparison: If you are deciding between two potential bets, the one where the Elo differential more strongly favours the selection offers better structural backing.
To see how Elo-informed predictions perform across dozens of leagues, view today's predictions and explore our monthly performance data.