Elo ratings were invented by the physicist Arpad Elo in the 1960s to rank chess players. The idea is elegant: every player carries a number, the gap between two numbers predicts the result, and after the game each number moves — a little if the result was expected, a lot if it was a shock. The same machinery now ranks football teams, tennis players and video-game accounts.

Horse racing breaks the textbook version immediately: chess has two players and one result, while a race has anywhere from two to forty runners finishing in a full order. This guide explains, in plain English, how the Elo framework is adapted to racing — the approach we use for RaceMetrics Ratings, published in full in our methodology paper.

Step 1: One number per horse, anchored at 1500

Every horse starts at 1500 — the average. As results come in, good performers drift up and poor ones drift down. On the RaceMetrics scale, 1550+ marks a strong performer and 1600+ is elite. The number is only meaningful relative to other numbers: a 1580 horse against a field of 1490s is heavily favoured by the model; the same horse in a Group 1 full of 1600s is an outsider.

Step 2: Turn a race into an expectation

In chess, your rating gap versus one opponent gives a win probability. In racing, we compute the horse's expectation against the whole field: given its rating and every rival's rating, what percentage of those rivals would it be expected to beat? A short-priced favourite on ratings might carry an expectation of beating 85% of its rivals; a rank outsider maybe 20%.

Step 3: Score the actual result with PRB

The result is measured the same way, using Percentage of Rivals Beaten (PRB): finish 3rd of 12 and you beat 9 of your 11 rivals — a PRB of 81.8%. PRB is the key that makes multi-runner Elo work, because it converts any finishing position in any field size onto the same 0–100 scale.

Step 4: Update the rating on the surprise, not the result

The rating change is driven by actual PRB minus expected PRB. A hot favourite that wins gains only a little — it merely did what the numbers said. The same favourite trailing home 9th of 12 loses heavily. An outsider that runs a close 2nd gains substantially without winning, which is exactly the kind of horse traditional win-only statistics miss. This "pay for surprise" property is why Elo-style ratings pick up improvement and decline faster than win-percentage tables.

The racing twist: rate the connections, not just the horse

A horse's performance is a team output — trainer form, jockey ability and pedigree all contribute. So RaceMetrics runs seven parallel Elo pools: horse, trainer, jockey, owner, sire, dam and damsire, each updated daily from every British and Irish result. Six of them are folded into a weighted Combined Score (Owner 20%, Trainer 20%, Jockey 20%, Dam 18%, Sire 12%, Damsire 10%). Validated across 25+ years of data, the highest Combined Score in a race wins 22.97% of the time — 2.15× what random chance would deliver.

How to actually use Elo ratings for horse racing

  • Read the H rating and Combined Score together. A high H with a modest Combined Score is a good horse in ordinary hands; a modest H with a top Combined Score is an ordinary horse with an elite support team — often the profile of an improver.
  • Watch rating changes, not just levels. A horse whose rating has climbed across its last three runs is telling you something the market may not have priced yet.
  • Compare within the race, not across races. The ratings' predictive power is in the gaps between today's runners.
  • Combine with conditions. Elo measures ability; it doesn't know today's ground or trip suits. Pair it with condition-specific form (that's what our Form Expert view is for).

Where to see live horse racing Elo ratings

RaceMetrics publishes daily-updated Elo-based ratings for every horse, trainer, jockey, owner, sire, dam and damsire in British and Irish racing, on every racecard — with a free tier. The complete mathematics, including the multi-runner derivation and validation tables, is open at racemetrics.co.uk/methodology.