Sullybomb
Gelding
Most racing models are built to answer one question: which horse is most likely to win? There's a different question, and it belongs to one of the people I've recently started working with: which horses is the market most likely to shorten, and which are most likely to drift?
That sounds like the same problem. It isn't. A horse can be a perfectly reasonable favourite and still be a poor price, and a horse doesn't have to win for you to have been right that the market was underestimating it at six o'clock the evening before. The market itself — how it moves, when it moves, what it hasn't absorbed yet — is something you can model. That's T tony spencer territory, and I'll come back to it, because putting his work next to mine is the reason I'm writing this at all.
I should say who "I" is, because this is the first time I've written about any of it publicly. I run FormLab. My question is the older one, asked as thoroughly as I know how: what actually happened in a race, and what does it mean for the next one? Speed figures built from the clock, sectional and pace analysis, draw and race-shape work, trainer and market studies — and over it all, one discipline: every idea is written down and sealed before it's tested, tested once on data with no re-tries, and the result filed whether it passes or fails. Most things fail. Knowing precisely which fifty ideas don't beat the market is how you recognise the two that might.
Tony, who many of you will know, has spent years building a complete race model of his own — calibrated probabilities, model prices, pace maps, finishing-strength and trainer analysis — and, on top of it, the market-behaviour work the opening question describes. There's a third collaborator who, for professional reasons, can't be named. The point isn't the CVs. The point is that these approaches were built entirely independently — different philosophies, different data, no shared assumptions — and it turns out that's exactly where the value is. When methods like that agree, the agreement means something. When they disagree, one of them knows something the others don't, and finding out which, and when, and why, is information none of the models contains on its own. That's what we're building together.
One early result gives the flavour. Tony's trainer model ranking a horse top tells you surprisingly little by itself about what the early market will do with it. But when that top ranking comes with a clear probability gap over the second-ranked trainer — past a threshold we fixed before we started watching — those horses have been noticeably more likely to shorten from the early price into the morning market. We didn't decide what the rule ought to be and then collect examples that fitted. We tested the population, fixed the rule, and now we watch what the next qualifiers do. His finding, my discipline wrapped around it — that's the collaboration in one sentence, and everything we test together works exactly that way.
Has any of it been right yet? Here are the last two cards we ran combined shortlists on — every pick, in the order they were ranked before racing, settled at Betfair SP. Nothing left out.
Saturday 8th August: Crimson Spirit (Haydock 3.34) — 2nd at 4.2 Arctic Dawn (Newmarket 3.40) — WON 8.2 (flagged beforehand as "best value of the lot": model price 4.0 against 7.6 on offer) Democracy Dilemma (Redcar 3.47) — 5th at 2.5 Archers Bay (Ascot 3.55) — WON 4.1 Merrimack (Newmarket 4.50) — 4th at 5.7 Prince Quattro (Newmarket 5.23) — WON 4.8 French Affair (Redcar 5.31) — WON 2.9 Seven bets, four winners, +12.7 points level stakes.
Monday 10th August: Mythological Star (Windsor 5.45) — WON 3.5 (ranked the strongest profile on the card) Keep An Eye On It (Ayr 5.15) — WON 4.1 Velvet Skies (Ayr 3.45) — 10th at 11.0 Catton Lady (Ayr 2.45) — WON 18.35 (the value flag: two independent ability reads had her first, model fair price 5.5 against 12s available) Believitanducan (Kempton 5.00) — 3rd at 4.8 Harry Knows (Windsor 6.15) — 3rd at 1.9 Spirit Of Acklam (Ayr 3.15) — 4th at 7.8 Seven bets, three winners, +18.6 points level stakes.
Fourteen bets, seven winners, +31 points. And now the sentence that has to sit right underneath, because it's true: two good days prove nothing, an 18/1 and an 8/1 winner are carrying most of that profit, and anyone can show you a good weekend. What I'd actually point you at is the smaller print: the horse flagged as the danger "if the leaders overdo it" in the Kempton race duly won it at 12/1, and a horse we recorded as watch-don't-bet — because the models disagreed in a specific, interesting way — ran second at 8s. Right reasoning shows up in the losses too, and that's the part you can't fake for long.
So rather than ask anyone to be impressed by a fortnight, here's what I'll do. From now on the combined shortlists go into a recorded ledger the day they're made, settled at BSP, published wins and losses alike, and judged when the sample is a few hundred picks rather than fourteen. And on the market question specifically, I'll post selected horses here in the evening, early price timestamped, where the data says the probability of a morning shortening is well above normal — before the move, not after. If they shorten you'll see it; if one drifts the post stays up. A model demonstrated in advance is worth a thousand results advertised in hindsight, and frankly the drifters will teach us as much as the steamers.
Where this ends up, I don't know yet. The honest version of the ambition is this: the interesting prediction may not be "Horse A will win." It may be "Horse A is 9/1 tonight, and there is a materially higher-than-normal chance it's 6/1 by morning." If that can be shown to hold out of sample, consistently, it's a different way of looking at racing markets altogether — whichever side of the book you sit on.
Last thing. If you've built something over the years — ratings, a pricing method, market-movement data, sectional work, trainer modelling, exchange analysis, anything you've actually tested on data rather than observed anecdotally — and you've ever wondered what it would look like alongside genuinely independent methods, my messages are open. Nothing proprietary needs to be shared to have the first conversation, and anything discussed stays confidential. A good idea sitting on one computer is an opinion; the same idea, holding up against models that share none of its assumptions, is knowledge.
The people I'd most like to hear from will recognise themselves in one line: you'd rather find out your edge isn't real than keep believing it is.
Market:
T tony spencer Pace Modelling & Reports
FORMLAB:
Cheers
Sul
That sounds like the same problem. It isn't. A horse can be a perfectly reasonable favourite and still be a poor price, and a horse doesn't have to win for you to have been right that the market was underestimating it at six o'clock the evening before. The market itself — how it moves, when it moves, what it hasn't absorbed yet — is something you can model. That's T tony spencer territory, and I'll come back to it, because putting his work next to mine is the reason I'm writing this at all.
I should say who "I" is, because this is the first time I've written about any of it publicly. I run FormLab. My question is the older one, asked as thoroughly as I know how: what actually happened in a race, and what does it mean for the next one? Speed figures built from the clock, sectional and pace analysis, draw and race-shape work, trainer and market studies — and over it all, one discipline: every idea is written down and sealed before it's tested, tested once on data with no re-tries, and the result filed whether it passes or fails. Most things fail. Knowing precisely which fifty ideas don't beat the market is how you recognise the two that might.
Tony, who many of you will know, has spent years building a complete race model of his own — calibrated probabilities, model prices, pace maps, finishing-strength and trainer analysis — and, on top of it, the market-behaviour work the opening question describes. There's a third collaborator who, for professional reasons, can't be named. The point isn't the CVs. The point is that these approaches were built entirely independently — different philosophies, different data, no shared assumptions — and it turns out that's exactly where the value is. When methods like that agree, the agreement means something. When they disagree, one of them knows something the others don't, and finding out which, and when, and why, is information none of the models contains on its own. That's what we're building together.
One early result gives the flavour. Tony's trainer model ranking a horse top tells you surprisingly little by itself about what the early market will do with it. But when that top ranking comes with a clear probability gap over the second-ranked trainer — past a threshold we fixed before we started watching — those horses have been noticeably more likely to shorten from the early price into the morning market. We didn't decide what the rule ought to be and then collect examples that fitted. We tested the population, fixed the rule, and now we watch what the next qualifiers do. His finding, my discipline wrapped around it — that's the collaboration in one sentence, and everything we test together works exactly that way.
Has any of it been right yet? Here are the last two cards we ran combined shortlists on — every pick, in the order they were ranked before racing, settled at Betfair SP. Nothing left out.
Saturday 8th August: Crimson Spirit (Haydock 3.34) — 2nd at 4.2 Arctic Dawn (Newmarket 3.40) — WON 8.2 (flagged beforehand as "best value of the lot": model price 4.0 against 7.6 on offer) Democracy Dilemma (Redcar 3.47) — 5th at 2.5 Archers Bay (Ascot 3.55) — WON 4.1 Merrimack (Newmarket 4.50) — 4th at 5.7 Prince Quattro (Newmarket 5.23) — WON 4.8 French Affair (Redcar 5.31) — WON 2.9 Seven bets, four winners, +12.7 points level stakes.
Monday 10th August: Mythological Star (Windsor 5.45) — WON 3.5 (ranked the strongest profile on the card) Keep An Eye On It (Ayr 5.15) — WON 4.1 Velvet Skies (Ayr 3.45) — 10th at 11.0 Catton Lady (Ayr 2.45) — WON 18.35 (the value flag: two independent ability reads had her first, model fair price 5.5 against 12s available) Believitanducan (Kempton 5.00) — 3rd at 4.8 Harry Knows (Windsor 6.15) — 3rd at 1.9 Spirit Of Acklam (Ayr 3.15) — 4th at 7.8 Seven bets, three winners, +18.6 points level stakes.
Fourteen bets, seven winners, +31 points. And now the sentence that has to sit right underneath, because it's true: two good days prove nothing, an 18/1 and an 8/1 winner are carrying most of that profit, and anyone can show you a good weekend. What I'd actually point you at is the smaller print: the horse flagged as the danger "if the leaders overdo it" in the Kempton race duly won it at 12/1, and a horse we recorded as watch-don't-bet — because the models disagreed in a specific, interesting way — ran second at 8s. Right reasoning shows up in the losses too, and that's the part you can't fake for long.
So rather than ask anyone to be impressed by a fortnight, here's what I'll do. From now on the combined shortlists go into a recorded ledger the day they're made, settled at BSP, published wins and losses alike, and judged when the sample is a few hundred picks rather than fourteen. And on the market question specifically, I'll post selected horses here in the evening, early price timestamped, where the data says the probability of a morning shortening is well above normal — before the move, not after. If they shorten you'll see it; if one drifts the post stays up. A model demonstrated in advance is worth a thousand results advertised in hindsight, and frankly the drifters will teach us as much as the steamers.
Where this ends up, I don't know yet. The honest version of the ambition is this: the interesting prediction may not be "Horse A will win." It may be "Horse A is 9/1 tonight, and there is a materially higher-than-normal chance it's 6/1 by morning." If that can be shown to hold out of sample, consistently, it's a different way of looking at racing markets altogether — whichever side of the book you sit on.
Last thing. If you've built something over the years — ratings, a pricing method, market-movement data, sectional work, trainer modelling, exchange analysis, anything you've actually tested on data rather than observed anecdotally — and you've ever wondered what it would look like alongside genuinely independent methods, my messages are open. Nothing proprietary needs to be shared to have the first conversation, and anything discussed stays confidential. A good idea sitting on one computer is an opinion; the same idea, holding up against models that share none of its assumptions, is knowledge.
The people I'd most like to hear from will recognise themselves in one line: you'd rather find out your edge isn't real than keep believing it is.
Market:
T tony spencer Pace Modelling & Reports
FORMLAB:
Cheers
Sul
