A horse with a stride frequency of 2.40 and a stride length of 7.60. A typical sprinter’s profile.
A single model, blind to venue, predicts one optimal distance for this horse. The same answer at every track.
But tracks aren’t the same. Calibrate for the course and the predictions diverge.
| Track | Predicted Optimal Distance |
|---|---|
| Epsom | 6.5f |
| York | 5.9f |
| Ascot | 5.2f |
| Pontefract | 3.0f |
Same horse. Same mechanics. Three and a half furlongs apart.
The stride data hasn’t changed. The track has.
On flat, neutral courses the general prediction model performs well — consistently, across hundreds of horses. It underpins much of what we publish.
But not all courses are flat or neutral. The 5f track at Epsom drops 99 feet down an incline. The 5f track at Pontefract rises over 70 feet and negotiates a turn. York’s sprint course is as near as flat and straight.
These are not equivalent venues. The stride data posted at one does not mean the same thing as the same profile collected at another.
Pontefract at 3.0f in the table above makes the point starkly. No UK racehorse is running its optimal distance at three furlongs — the shortest flat race in Britain is five. That number is the track speaking — gradients and to a lesser extent turns distorting the stride profile until the raw prediction falls below any raceable distance. The horse hasn’t changed. The data has been contaminated by the venue.
Speed figure compilers solved a version of this problem long before stride data existed. Raw times are standardised to track and distance pars because the same clock time means different things at different venues. The raw number is incomplete without the course.
Stride data works the same way. Without standardising for the course, you’re measuring geometry and the horse — not just the horse.
What Track Geometry Does To Stride Data
Three features of a racecourse alter how a horse strides, independent of distance preference.
Inclines compress stride. An uphill section forces the horse to work harder to maintain speed. Stride length shortens. Stride frequency rises. Ascot’s straight course climbs over 60 feet — enough to shift stride profiles measurably against the same horse on flat ground.
Declines extend stride. Gravity assists the horse. Stride length increases while stride frequency holds relatively stable. At Epsom, horses descend nearly 100 feet over five furlongs. A horse striding downhill through that covers more ground per stride than the same horse on the flat — not because of its distance preference, but because the gradient is distorting the profile.
Turns compress stride too. The tighter the bend and the greater the proportion of the race spent on bends, the more pronounced the effect. Windsor’s partial figure-of-eight layout is dominated by turns. Lingfield’s all-weather circuit combines tight turns with short straights — enough bend work to compress stride across the full race, not just through isolated sections.
These three interact — and every British racecourse combines them differently. Every course asks slightly different questions of a horse — but it makes the task of measuring one from stride data considerably harder without accounting for the venue.
The Outer Layer
Track geometry is not the only thing that contaminates stride data.
The noise has a name. Pace.
A horse that crawled through the early sections before the race turned into a sprint finish produces a different stride profile to the same horse in a genuinely run contest. The numbers are real. They don’t represent the horse’s natural mechanics.
We filter for it. Use only races where the horse expressed its natural stride pattern and the stride data becomes a reliable indicator of biomechanical preference. Don’t filter it, and you’re modelling tactics, not the horse.
Pace is the outer layer. Strip that first. Then standardise the track.
It Isn’t Linear
Now take a middle-distance profile. Stride frequency 2.20, stride length 7.30.
| Track | Predicted Optimal Distance |
|---|---|
| Ascot | 11.2f |
| York | 11.0f |
| Epsom | 9.9f |
| Pontefract | 8.9f |
The spread has narrowed. The sprinter varied by 3.5 furlongs across four tracks. The middle-distance horse varies by 2.3 furlongs.
Over longer distances, the geometry factors start working against each other. The 5f track at Epsom is pure downhill — one effect, one direction. At a mile and a half, a horse climbs, navigates bends, and then descends through the same section the sprinters encounter. The effects partially cancel. The net distortion is smaller, but it’s still there — and it’s different at every track.
This is why a single correction factor per track doesn’t work. Each course requires its own calibration — across the full distance range, on a sound surface.
What The Numbers Say
The strongest-performing track model in the programme returns an R² of 96%. Across all track models, the average sits in the low-to-mid 80s — a consistent improvement over a single venue-blind model applied everywhere.
Mean prediction error at track level typically falls below one furlong. At some courses it sits closer to half a furlong.
These are measured against holdout data the models have never seen. Horses excluded from training, predicted blind, then compared to their actual racing distance. Not theoretical. Tested.
What This Means
Without track standardisation, the same stride data returns different predictions at different courses. The raw data shifts with the geometry — a horse striding uphill at Ascot looks different to the same horse on the flat at York. A single model reads both and treats the difference as biomechanical. It isn’t. It’s the track.
With track-calibrated predictions, the geometry is accounted for. A horse predicted at 10 furlongs at York should come back near 10 furlongs at Epsom — provided both races were truly run, on a sound surface, with the horse performing near its best ability. The other variables matter. But the track is no longer one of them.
The market sees a horse that ran eight furlongs at one course entered over ten furlongs at another and prices it as a step up in trip. Calibrated stride data may say that horse was always a ten-furlong type. The first track’s geometry was compressing the prediction. Not the horse’s ability.
Standardise the track. Measure the horse.