The Form Book

New here? Start with this.

The Form Book does two honest things, kept firmly apart:

🩺

Welfare Watch — which jumps horses are most at risk of not completing a race (pulling up or falling). Built for connections, vets and welfare officers — not a betting tool.

💷

Betting Watch — mid-priced each-way place runners in big fields, where the bookmaker's terms misprice. Not favourites, not tips — a small, honest, structural edge.

The rule for both: nothing ships unless it still works on races the model never saw — and we show you every idea we rejected, not just the wins. Tap any i for a plain-English explanation of a term.

📅

Today

Every race at every meeting, in time order — the whole day at a glance, next race highlighted. Each race is read through two separate lenses, kept in their own columns so they never blur together:

🩺 Health — the runner most at risk of not completing (a welfare/injury signal, jumps only). 💷 Bet — whether the race is each-way territory where our edge lives. A race can flag one, both, or neither — blank is honest, not a gap.

The card, live

Time

Race

🩺 Health · injury risk

💷 Bet · each-way

🩺 Health names the jumps runner most likely not to complete (pull up / fall) — a welfare flag, not a tip. 💷 Bet marks a big-field handicap where the each-way place edge applies (live odds name the actual picks). Flat races get no Health flag here — a fall/pull-up signal is jumps-only; the full Race Card still scores every Flat runner's relative risk.

🩺

Welfare Watch

In plain English: this is an early-warning list of the jumps horses most likely not to get round today — to pull up, fall, or be pulled up. It is not saying a horse is injured, and it can't tell you one will be — nobody can know that in advance. It's a "keep an eye on these" list for owners, trainers and vets, scored from the ground, the horse's age, its own record of falls, and how sound its yard and its sire tend to be. After they run, we check which flagged horses actually failed to finish — and keep an honest score of how often the warning was right.

2.5×higher risk

The horses this model flags highest-risk fail to complete 26.4% of the time, versus 10.5% for the lowest — validated on 149,000 held-out jumps runs it never saw (2023–26), ranked cleanly across all ten risk bands.

Today's watch-list — highest risk first

jumps runners

Runner	Risk	Score	Why flagged

What drives the score

six validated factors

Ground

Heavy and soft ground tires horses and brings on falls.

Good 14% → Heavy 20.5% DNF

Age

Older horses are more fragile and slower to recover.

3–5yo 12.7% → 9yo+ 20.8%

Own history

A horse that has pulled up or fallen tends to again.

never 13.5% → 3+ 21.7%

The yard

Some trainers' horses fail far more often — and it persists.

13.4% ↔ 22.6% · corr 0.64

Sire

Soundness is heritable — some sire lines throw more non-finishers.

persists · corr 0.48

Damsire

The broodmare-sire line adds a further soundness signal.

persists · corr 0.38

What this is. A pattern-based risk indicator, not a diagnosis. It says a horse is statistically more likely not to complete — not why, nor that anything is wrong. For vets, welfare officers and connections to watch flagged runners, and — with veterinary input — to understand and reduce risk. Public UK & Irish jumps results, 2015–2026.

Track record — does the flag actually come true?

out-of-sample

Flag	Runs	Failed to complete	vs field

Jumps vs Flat — the flag across both codes

out-of-sample

Day by day — every flag graded on the result

Date	Flags	Graded	HIGH DNF	Field DNF

The flag's actual calls — named, right and wrong

📈

On the Up Live from the form book

Not who won last time — who's trending. Yards whose strike-rate is climbing well above their own season, and young horses whose form figures are improving run on run. A real snapshot from the form book — the deployed site refreshes it nightly on the rig.

Yards on the up

last 30 days vs season

Trainer	Runs (30d)	Strike	Season	Trend

Horses on the up

young, improving figures

Horse	Age	Last 4	Figure trend	Note

How it's built. A yard is "on the up" when its recent strike-rate sits meaningfully above its own season baseline — measured, not eyeballed. A horse is "on the up" when its speed figures rise run-on-run while it's still young enough to keep progressing. When the rig posts the first nightly feed, these fill with real names and the Sample layout tag disappears.

🏇

Betting Watch

Most tipping sites sell you the favourite dressed up as a system. We tested whether any public-data model can beat the win market for profit — properly, out of sample, eight different ways including a full gradient-boosting model. It can't. So we don't pretend. What's below is the one edge that survived the tests.

The honest basis. No public-data model beats the win market for profit — we tested it eight ways, it can't. The one edge that survives is structural, not predictive: in big-field handicaps the fixed each-way place terms can misprice a horse's true place chance, especially on extra-place offers. So we back mid-priced (8–19) runners to place, five to a race at most, logged before the off and graded in the open. Prices are a guide only, not betting advice.

Today's picks — live

Every day, graded — tap a day to open it

Each day is one line — how many we suggested, how many placed, the day's return and the running total — tap to see the individual horses. ✓ = placed, ✗ = unplaced. Prices are a guide only, not advice.

🧮

Staking model

Before you back anything, see what it would actually achieve. Set your stake, place-strike and odds, and the model shows the edge, the real spread of outcomes over a run, and the bank you need to survive the losing streaks. It's the honest financial picture — small edge, high variance, grind over volume.

How £1 each-way actually works

plain English

“£1 each-way” = £2 out of your pocket — £1 on the win, £1 on the place. They're two separate bets stapled together:

The win £1 pays only if it wins. On a sharp market that part roughly breaks even over time — don't expect profit there.
The place £1 pays if it finishes placed (2nd–4th, depending on field size). This is where our edge lives.

Worked example — our 31 graded bets, at £1 e/w (£62 staked):

Place side (£31 of the stake)	+27%	≈ £39 back → about +£8 profit
Win side (£31 of the stake)	~flat / down	on a sharp market → about −£9
Net — full each-way	−2%	about −£1 overall

The lesson. The money is in the place part, in fewer chosen races — not in blanket each-way. Backing seven horses at £1 e/w in one race (£14 out) hoping to “win something” is the mirage: you almost always hand money back. Pick the places worth backing, skip the rest — that's the whole game, and it's why we tip narrowly. (Live figures on the Tips tab; they update as results are graded.)

Model the outcomes

interactive

🧮 Launch the interactive staking model →

Drag the sliders; watch the edge, the outcome spread, and the losing-run cushion update live.

🐎

Race Card

The whole day, the way you'd read a racecard — grouped by meeting, then race by race, with every runner and the information we hold on each: form, official & Racing Post ratings, weight, draw, jockey and trainer. This is the reference view — the full field, nothing hidden.

Every runner now carries our 🩺 non-completion risk score . On the jumps a flag means a real chance of pulling up or falling. On the Flat the same flag is relative — non-completion is rare there (~0.3% of runners), so a flag marks a horse elevated against other Flat runners, not a horse in danger of falling. Both are validated out of sample (jumps ×1.76, Flat ×8.25).

Today's meetings

How to read it — two columns on the right. Back? marks our each-way selections: a highlighted row with a 💷 EW @price badge means this is a horse we'd back each-way to place in this race (guide price). A “·” means it's not one of our plays. Risk is the non-completion score, and every runner carries one: on the jumps it reads a real chance of pulling up or falling; on the Flat it is relative — non-completion is rare there (~0.3%), so a Flat flag marks a horse elevated against other Flat runners, not one in danger of falling. Tap a ▸ row for spotlight/comment. Working behind the score is on Data & Method.

🔍

Explore

Ask the form book about anyone. Type a trainer, a horse or a sire — see their record, then follow the threads: a trainer's horses → click a horse → its runs and breeding → click its sire → every foal it has thrown. Millions of runs, all cross-linked, like Wikipedia for racing.

🧠

Local AI vs Cloud

Where the thinking comes from — and whether we can bring it fully in-house. The master plan below is tonight's synthesis: the cloud engine reports, Qwen gives its second opinion, and the two are reconciled into one call — flagging anything Qwen caught that we're not already acting on. Underneath are the two sides in full (cloud analysis, then Qwen's take) so you can see the working. The engine stays on the cloud analysis until something better overtakes it; nothing changes a live setting without the back-test bar and your sign-off.

💷 Betting — what we're thinking

🩺 Welfare — what we're thinking

🤖 Local AI (Qwen) — its take

not yet connected

We're monitoring, not switching — yet. Keep the cloud analysis running, wire the rig's local AI in when it's ready, and compare their calls on the same numbers. We move to local when it's proven itself, and only reach for a bigger model when the results say it earns its keep. Nothing changes the live engine without clearing the same evidence bar.

📋

Knowledge — everything on the board

Every factor we've discussed — nothing dropped, including the ones that do nothing for a bet. Because a dud for betting is often gold for welfare: age has no betting signal at all, yet it's one of the strongest injury predictors we hold. So each factor carries two honest verdicts — one for betting, one for welfare/injury — and the nulls stay on the record so we never re-chase a mirage.

validatedweakpriced-in / noneto testn/a ✓ have data · ◑ proxy · ✗ needs a source

The point of the list. Most of these will move an outcome by a whisker, and some by nothing at all — but we don't decide that by hunch. Everything gets listed, then measured, then kept or dropped on the evidence. The tiny ones matter most for the tips side, where a fraction of a percent, stacked across enough races, is the whole game.

🧬

Breeding the cross that beats the parents

Best-sire × best-damsire isn't always best. Some crosses — a “nick” — make horses that out-run what each parent does apart. Choose a stallion in the first box (only ones actually at stud — no dead horses) to see the mares it should cover; add your mare's sire in the second box for the verdict on that exact cross. Every row shows how the foal would theoretically rate (peak Racing Post Rating), the rough fee band, and the value — rating edge per £1,000.

Stallion the sire you'd breed to Your mare is by optional — her sire

Best value most rating per £1,000 of fee

The strongest nicks overall

Lift = how far the cross's average peak RPR beats the mean of what the two parents produce apart — read it as the breeding edge. n = progeny that have run: a big lift on 60 is firmer than one on 6. Historical form, an evidence signal for a mating — not a guarantee, not advice.

🔬

Database — the raw material & method

Nothing here is asserted — it's derived. This is the raw material we hold, every factor we've tested with its honest verdict, and the one rule that decides whether a pattern is real or just noise. If a number's on this site, you can see where it came from.

1.85M

runners on record — GB & Irish results, 1988–2026

90k+

pulled-up / fell / unseated runs — the welfare model's raw material

layers stitched together: results form · Betfair odds · daily racecards

What we hold

✓ in hand · ◑ proxy / partial · ✗ needs an outside source

Data	Have it?	What's actually there
Trainers	✓	Every runner; rolling 14-day + career strike-rate & profit pre-computed
Jockeys	✓	Every ride; form + profit + claiming allowance
Owners	✓	~100% of rows; 98k distinct owners, with change-of-owner history
Pulled up / fell	✓	PU 65,832 · Fell 15,681 · Unseated 9,527 — the injury signal
Surgery / medical history	◑	Wind ops, gelding, dated medical notes — but not clinical injury records (that's the vets')
The full pre-race card	✓	60+ fields: spotlight, comments, ratings, going, draw, headgear, last run
The money (Betfair)	✓	Pre-race drift, in-running highs/lows, BSP, volumes
Breeding (sire / damsire)	✓	Full pedigree line — predicts soundness (not winning)
Every horse's full history	✓	All prior runs + record by course / going / distance

Welfare factors — validated

jumps · base non-completion rate 16.1%

Factor	Verdict	Effect on not-completing
Ground softness	validated	Good 14.2% → Soft 18.1% → Heavy 20.5%
Age (older)	validated	young 12.7% → prime 16.4% → 9yo+ 20.8%
Its own prior falls / pull-ups	validated	never 13.5% → 1–2: 17.9% → 3+: 21.7%
The yard (trainer DNF-rate)	validated	high yards 22.6% vs low 13.4% · persists corr 0.64
Sire (progeny soundness)	validated	heritable · persists corr 0.48
Damsire	validated	further soundness signal · persists corr 0.38

Combined into one score: the horses it flags highest fail to complete 2.5× as often as the lowest — proven on 149k runs it never saw.

Betting factors — place value

each-way, out-of-sample gate

Factor	Verdict	What we found
Most-raced third (relative mileage)	validated	+11% place-ROI; fit horses place reliably; holds out of sample
Field size 16–19 runners	validated	+10% vs break-even in 20+ fields, at extra-place terms
Proven on the ground / class drop	weak	small, consistent, tiny samples — watching
Recent form · hot yard · course form · age	priced in	no edge for betting — the market already knows (age still matters for welfare!)

The testing list — worked through

every variable, tested out of sample

Variable	Have it?	Status	What the test found
Weather — temp / wind / rain	✓	tested	joined a free weather API (90 courses): temperature = noise; wind real (+6.5pt) but race-level so it diluted the flag — held out; rain ≈ going, already in
Weight carried (jockey + kit)	✓	tested	absolute lbs = noise (confounds the code); RELATIVE weight works — bottom-weight fails to complete +2pt
Horse bodyweight	✗	no data	confirmed absent from the UK jumps form — a dead end unless a source starts carrying it
Owner record	✓	tested	a real 10.5pt DNF signal — but 0.61-correlated with the yard; scored, it double-counts the trainer (lift 1.76→1.73) — held out
Draw · days-since-run · jockey place-strike · run-after-layoff	✓	tested	all tested OOS: draw n/a on jumps · freshness diluted the flag (held out) · jockey priced-in · layoff = freshness

How we judge a signal — "statistically true, not potentially true". A pattern only ships if two things hold: the sample is large, and it survives out of sample — it still works on races it was never found on. With 1.85M runs we can afford to be strict, and most exciting-looking patterns don't survive. The ones on this site did. That discipline is the whole point: it's why the welfare model is real, and why we won't sell you a winner the market has already priced.

Plain-English glossary

every term, no jargon