Elo Rating forklaret

The core of Elo is a single equation. Given two teams with ratings RA and RB, the expected score for team A is EA = 1 / (1 + 10^((RB − RA)/400)). A team with a rating 100 points higher than their opponent is expected to win ~64% of the time; 200 points higher implies ~76%; 400 points higher implies ~91%. These are the same odds ratios that apply to chess, checkers, tennis and football - the algorithm is sport-agnostic.

After the match, each team's rating updates by K × (S − E), where S is the actual score (1 for a win, 0.5 for a draw, 0 for a loss) and K is a constant that controls how fast the ratings move. In chess, K is typically 10–40; in football, most public Elo implementations use K = 20 for regular matches and K = 30 for high-stakes matches (finals, derbies).

The net effect is that an upset (a low-rated team beating a high-rated team) moves ratings dramatically, while a predictable result barely moves them. Over time, the ratings settle on values that accurately reflect each team's long-run strength - without anyone needing to hand-code assumptions about 'form' or 'class'.

Vanilla Elo treats every match the same, but football has features that the original algorithm doesn't capture. The first is home advantage: on average across Europe's top five leagues, home sides win ~45% of matches, draws are ~25%, away wins are ~30%. A neutral Elo would underrate home sides and overrate away sides unless you bolt on a fixed home bonus (typically +80–100 Elo points) before computing the expected score.

The second is goal difference: Elo rewards a 1-0 win the same as a 6-0 win, which misses real strength information. Most football Elo systems use a goal-difference multiplier on the K factor - a 3-goal win moves the rating roughly twice as much as a 1-goal win. The exact multiplier is tuned empirically to maximise out-of-sample prediction accuracy.

The third is competition tier: a Premier League win should be worth more rating points than a League One win, because the opposition is stronger. Cross-league Elo systems (FiveThirtyEight's SPI, the ClubElo project) solve this by having teams play each other occasionally in European competition and letting the ratings converge organically. BetsPlug extends this with a separate Champions League calibration head, because knockout football has different characteristics than domestic league play.

Elo's biggest strength is that it captures long-run team strength without needing any domain knowledge beyond match results. A well-tuned Elo model on the Premier League will pick the right favourite ~70% of the time using nothing but historical results - no xG, no tactical analysis, no injury data. That's a remarkably strong baseline for a one-parameter (K) model.

The weakness is that it reacts slowly to changes that haven't yet shown up in results. When a top team loses their star striker to injury, Elo doesn't know; it only adjusts once the team starts losing matches they would previously have won. This is why modern ensembles combine Elo with faster signals like xG (which reacts within a single match) and logistic regression on injury reports (which reacts immediately when news breaks).

Inside BetsPlug's ensemble, Elo is the slow-moving anchor. When the other models disagree - one says the home team is a huge favourite, another says it's close to 50/50 - Elo's rating gap acts as the tiebreaker, because it has the longest memory and is least influenced by recent flukes. The final ensemble output tends to match the Elo lean on long-run fixtures and drift away from it only when the faster signals are confidently saying something different.

A fun experiment: run a simple Elo model on Premier League results from 2010 onward, then compare its predictions to the closing bookmaker odds. You'll find that the Elo lean matches the book's favourite about 75% of the time, and the cases where they diverge are the most interesting. Sometimes the book knows something Elo doesn't (a key injury, a tactical shift). Sometimes Elo is right and the book is wrong because the market is overreacting to a recent hot streak.

The gap between Elo and market odds is a rudimentary value signal - if Elo thinks the home team should be 1.90 and the book is offering 2.30, that's a potential edge of ~20%. But blindly trusting Elo is dangerous: the book's line incorporates information Elo doesn't have, and betting every Elo-disagreeing fixture would expose you to the worst cases in the dataset.

BetsPlug's production use of Elo is to compare it against the market-implied probability and use that delta as one input into the meta-model. Large deltas get flagged; small deltas get trusted. This is more honest than taking Elo at face value and more data-driven than taking the market at face value.