If you ask people who work with sports data why they end up with racing, the answers are rarely about the sport. They are about the shape of the data. Horse racing data analytics has some properties that football, tennis and most other sports do not, and some weaknesses that are just as real. Both sides follow, with numbers from our own archive.
One event, one winner
A race is a closed experiment. A fixed set of runners starts together, and one finishes first. In our UK archive 99.8% of races have exactly one winner, and the rest are mostly dead heats or races with no recorded winner. That sounds trivial until you compare it with a sport where the outcome is a score, a set of sets or a draw. A single clean outcome per race makes labelling easy and evaluation unambiguous.
The weakness is the size of the field. The average UK race has 9.5 runners and Hong Kong’s has 12.1, so a coin-toss guess for the winner is right about one time in ten. Accuracy figures that sound modest can still be a real improvement over that baseline, and you have to compare against the right one.
Plenty of events
The UK runs an average of 36 races on a racing day and between 11,900 and 13,300 in a normal year. Hong Kong adds between 790 and 860 a year on about 87 days. That is a large, regular sample. Hundreds of thousands of events let you test an idea on thousands of cases instead of dozens, and split by season without running out of data.
The weakness is that races are not independent. The same horses, jockeys and trainers appear again and again, and one horse can run twenty times in a few years. Treat them as repeated measures, not as separate random draws, or your error bars will be too small.
Information declared in advance
Before each race, the runners are published: weight, draw, gear, jockey, trainer, a rating. This pre-race record and the post-race result share the same structure and the same identifiers, which is unusual. It means you can build a clean view of what was known at declaration time and test against the result, with no guessing about which team sheet applied.
The weakness is what is not declared. A horse’s fitness on the day, its mood and its owner’s intentions never reach a dataset, and they matter. No amount of data engineering recovers what was never recorded.
Long, consistent history
Hong Kong results run back to 1979 on one consistent structure, 30,112 races across 47 years. The UK archive starts in 2011 with 197,526. Long runs on the same structure let you study how racing changes over time, from the size of fields to the mix of race types.
The weakness is that coverage is uneven. UK timing data starts in 2017 and reaches about 91% of British runners from 2024, with none in Ireland. Our coverage note has the details. A study that treats the whole archive as equally rich will go wrong.
Fine-grained timing
Where it exists, timing is the best part. A Hong Kong runner has position and time at up to six points in a race, and UK runners with sectionals have up to 18 splits. That turns a finishing order into a picture of how the race was run. Our sectional times guide shows a real example.
A natural hierarchy
Horses, jockeys, trainers and courses form a structure of entities that interact: 155,221 horses and 6,591 jockeys in the UK, 16,541 horses and 606 jockeys in Hong Kong. That suits hierarchical and mixed-effects models, and makes racing a good teaching dataset for them.
The weakness is thin data for the individual. A third of UK horses have run fewer than five times in the archive, so any model of an individual horse is working with very little.
So is it a good dataset?
For analysis, yes, provided you respect what it is. The strengths make it unusually easy to start. The weaknesses are the reason that most simple models do worse than they first appear to.
So, a good dataset, with caveats that are worth knowing before you start rather than after. The datasets and the API cover the UK and Hong Kong, the data model shows how the tables fit, and the checklist for machine learning datasets gives the tests to run first.
Comments are closed