Eight Furlong · Melbourne

One place for equine data, analysis, welfare and buying.

We film horses walking and running, measure them with pose models we built and own, and join every measurement to a racing database of 137,879 horses and 2.7 million runs. Buyers use it before the hammer. Trainers use it in the barn every morning. Racecourses can use it on the day.

One walk, three views. Side-on with our 29-point model, then away and back with our 26-point front/rear model.Every landmark drawn is the model's own output
137,879
horses in our racing database, with career records back to 1986
25,010
yearlings measured on parade video, 2019 to 2026
3.7M
frames of walking read in one training barn this year
29
landmarks per horse from a model we trained from scratch
What we do

Measure the horse. Keep the record. Test the claim.

Four things, and every one of them feeds the others. The video gives the measurement, the database gives the outcome, and the test decides what we are allowed to say.

Data

The racing database

Australian results back to 1986, ten years of stewards' notes, in-race stride for 368,000 runs, sale catalogues and prices, and our own measurements, joined by horse and refreshed nightly.

Analysis

Models we own

A 29-point side-on pose model, a 26-point front and rear model and a trainer's Balance method, all trained on frames we labelled. No licensed weights, so the measurements are ours to sell.

Welfare

The same horse, against itself

Walks measured day after day in the barn, and, as a concept, on raceday, so a change in how a horse moves is seen before it is felt.

Buying

Documents before the hammer

Bid lists, price bands and a read on every lot, locked and hashed before the sale, then graded against what the horses made and what they went on to do.

Solutions

Six products. One method behind all of them.

Each one is built on the same pose models and the same database, and each one carries the numbers that held when we tested them.

01 · Yearlings

25,010 yearlings. What the walk tells you, and what it can't.

Every yearling at the Inglis, Magic Millions and NZB sales since 2019, measured whole-video with our own model. We have followed 7,365 of them through their racing careers and tested every claim blind on the crop the model never saw.

Side-on walk with 29 landmarks
Side-on. 29 landmarks, hooves to poll, near and off legs told apart on 97 to 100% of frames.
Walking toward the camera with the front model's landmarks
Front. Knee, fetlock and hoof spans.
Standing horse with construction lines
Build. Trunk, leg and height ratios inside the sale.

What the walk reads is distance and build, and the parts of a yearling's page that hold up blind are the trainer, the stable, the dam and the broodmare sire. So that is what the card shows: a distance lean, the build inside its sale, comparables, and the points that stack at the hammer.

2.5×
more likely to be a stayer in the most-likely fifth by build: 24.7% against 9.7%, held blind on the 2023 crop
20% → 53%
metropolitan placers across the 2023 crop as the points that hold stack up at the hammer
+0.14
walk reach against the horse's own breeze stride 18 months on; out of sample +0.18. The ring pays for neither.
1.7×
more repeatable than our previous model when the same horse is measured at two sales
02 · Ready2Race and Horses in Training

The clock is the market. The stride is the edge.

Every breeze-up catalogue prints the time. We read the stride the horse used to run that time, from the breeze video at the sensor's own resolution, and checked against the on-horse sensor where the sale publishes one. A horse that is slow on the clock but long in the stride for that clock is the one the ring leaves behind.

The pass at the camera. Inglis Ready2Race 2026, Wangaratta. Stride length and cadence are read on this stretch only, never on the warm-up canter or the replay.
3×
metropolitan winners by the end of the three-year-old season: 17 of 269 flagged slow-clock horses against 21 of 968, p = 0.001
−23%
bought below the catalogue: a median of $46,000 against $60,000 for the rest of the measured lots
+45%
prize money per sale dollar: $0.71 against $0.49 for the rest of the catalogue
3×
rated 80 or better: 6.0% of flagged horses against 1.8% of the slow-clock horses priced alongside them

2026, locked before the sale. Fifteen horses on the Magic Millions Gold Coast list, each with a bargain price, a ceiling and a walk-away price fixed before lot one. Four were bought at or under the ceiling for $117,500 combined, 27% under what the ring normally pays for the profile, and the rule walked away from every lot that went past its ceiling. Inglis Ready2Race: 193 parade walks and 66 breeze passes read, ten horses on the bid list with three prices each, hashed on 13 October before the first lot sold.

Flagged by the stride, before they racedSalePaidPrize moneyReturnRecord
FeroceNZB 2023$160,000$1,611,87510×Won in town at 3 · rated 108
WarmongerNZB 2022$165,000$1,133,4256.9×Won in town at 3 · rated 115
Tavi TimeInglis 2021$130,000$1,128,3758.7×Won in town at 3 · rated 105
King of ThunderNZB 2023$50,000$478,4259.6×Won in town at 3 · rated 90
Tycoon HarryMM 2022$70,000$365,9905.2×Won in town at 3 · rated 87
BarrengarryInglis 2024$10,000$197,40020×Won in town at 3 · 2 wins from 5 starts

Prize money is the Australian racing record to October 2026; the 2024 horses are early in their careers. Across the three 2025 sales the rule flagged 59 horses, bought at a median of $60,000 against $90,000 for the rest of the catalogue; they are starting their racing now.

03 · The database

Results back to 1986. Every note. Every stride.

Built over two years from official results, stewards' reports, in-race sectionals, sale catalogues and our own measurements, keyed by horse and refreshed every night. It is where every number on this page is checked, and it is a product on its own.

A yearling Points card in the racing database: a +4 score, distance lean, the points known before the sale, at the hammer and in work, and value for money
A Points card, on the horse's page. Peace Treaty, Inglis Premier 2022: a +4 card, bought for $50,000 in the cheapest fifth of her sale, against a page-implied $103,000. Five wins from 11 starts and $308,925 since.
The racing database dashboard
The database itself. Every horse, result, sire and track, searchable, with last night's results already in.
137,879
horses, with colour, sex, date of birth, sire, dam and every start
2.72M
runs across 494,920 races at 919 tracks, updated nightly
843,311
per-horse stewards' notes since 2016, from every state
368,438
runs with in-race stride and sectional data, 200 m by 200 m
What is in itResults, margins, prize money, ratings and gear for every run. Stewards' notes per horse, including 153,734 that mention a vet examination or lameness. Stride length and cadence for every 200 m where the state publishes it.
What we add25,010 yearling parade measurements, 13,821 parade videos through the pose models, breeze strides for every Ready2Race lot since 2018, sale prices and reserves, and the Balance lines for 13,446 horses.
How it is usedSire and broodmare-sire leaderboards as of any date, trainer and stable strike rates before a sale, a horse's full record behind any raceday sighting, and the blind back-tests behind every claim we publish.
04 · Balance

A trainer's eye, measured on 13,000 horses.

One of Australia's most experienced trainers drew the method he reads a horse by: sixteen lines on a standing horse, in a fixed order, with the angles to look for. We turned the drawing into a measurement, ran it on every sale video we hold, and handed back a trainer's method at a scale no eye can reach.

The trainer's lines drawn by hand (solid) and the same lines built by our model (dashed), on the trainer's own photograph
Solid, the trainer's hand. Dashed, our model. The trainer's photograph and lines, our construction laid over the top.

Four of the method's anchors are not among the 29 points our model finds, so the trainer placed them by hand on 300 frames and we built each one from the points we do have, to within about a centimetre on the horse. We read the lines at the stance moment of the walk, reached a hundred times in a parade clip, and take the median. The same horse filmed at two sales now agrees with itself at up to 0.77 on the key lines.

13,446
horses on a trainer's lines: every yearling and Ready2Race parade we hold, 2019 to 2026
1 cm
the trainer's hand-placed anchors rebuilt from the 29-point model, leave-one-out
184
horses measured at one and at two: the trainer's growth claim holds, +4.8° to +3.5°, p < 0.0001
3 weeks
from a trainer's drawing to the whole population measured on it
05 · Barn monitoring

Every walk over the mat, every morning, against the horse's own last.

Two fixed cameras face each other across the walkway of a Victorian training barn. Every horse that crosses the mat is filmed from both sides, identified from the brand on its shoulder, measured on every frame, and compared with its own previous walks. A change in the walk is flagged to the trainer before anyone has to notice it.

Both sides of the horse. East over west, the same stride on both, aligned on the hoof signal rather than on timestamps.

The system has run since January. Only real walks get through: a clip is kept when a horse is on the mat, centred, the right distance from the lens and actually moving for at least two seconds, and the pair of cameras is kept or discarded together. The pose model then reads every frame, both sides, and every frame is labelled from its own points before anything is measured.

16,832
walks kept and measured, 250 days, 93 horses, 3.7 million frames
0.98
agreement when the same walk is measured twice; 0.55 before the every-frame rule
0.68%
landmark error as a share of body length on unseen video; the best generic animal model scores 1.71%
6
measures per stride, both sides: length, symmetry, head-nod, speed, frequency and fore-hind timing
06 · Track monitoring · Concept

Raceday. A concept, tested on broadcast footage.

The barn system, proposed for the racecourse. Nothing is installed yet. We tested the idea on broadcast footage: reading a horse's walk in the mounting yard and naming it from its saddlecloth. A working system would need fixed cameras in the yard and behind the barriers, so each runner could be compared with its own record before the race.

The mounting yard, filmed across the lawn. Skeleton, tracking box and the saddlecloth number are all model output, on a broadcast feed, in the rain.
Behind the gates. Horses tracked in a broadcast clip; a number is shown only after three agreeing reads from the same horse.

Tested on broadcast vision from three meetings in two states in October 2026, with no raceday training. On the best clips, where the camera looked across the yard at the far side, the model placed the full skeleton on nearly every walking frame and read the saddlecloths correctly. That shaped the proposed design: cameras filming the far straight across the open yard, and a screen for the raceday vet. It has not yet run live at a meeting, and it has not been tested on every runner.

99%
of walking frames carrying all 29 landmarks on the best yard read, across the lawn in the rain
28 of 28
saddlecloth reads correct on that runner, named within half a second of entering the straight
3
meetings tested off the broadcast, country and metropolitan, black cloths and red
Concept
no cameras installed; a pilot at a working track is the next step
Talk to us

Bring the data to your next sale, your barn, or your racecourse.

Briefings for bloodstock managers, syndicators, trainers and racing bodies. Eight Furlong is private and bespoke. We do not publish picks.

[email protected]