Every pace map ever drawn rests on a single claim, and it is so intuitive that hardly anyone bothers to state it. Several horses in a field want the lead, they take each other on, and the early pace goes quicker than it should; the front runners empty their tanks, the field strings out behind them, and something held up comes through late to win. Where only one horse wants the lead, the argument runs the other way: it gets an uncontested tempo of its own choosing, nothing forces it to go faster than suits, and the race is won from the front.
That is the mechanism, and it is the only reason anyone counts front runners in the first place. Because the whole apparatus of lanes and pace labels is built on it, it seemed worth measuring rather than continuing to assume, and when it was measured it did not hold up.
How to measure a strung-out field
The outcome needs a number before anything can be done with it, and "strung out" is something you see rather than something anyone records, so it has to be constructed. The measure used throughout is the gap from the winner back to the third horse, in lengths, divided by the race distance in furlongs. A race whose first three finish in a heap produces a small figure; one where the winner goes clear and the third is eight lengths adrift produces a large one. Dividing by distance matters because two lengths over five furlongs describes a different race from two lengths over three miles.
The obvious influences then have to come out of it. Sprint handicaps finish closer than staying novice hurdles, soft ground spreads a field further than quick, and none of that tells you anything about a particular contest. Taken together, course, trip, going and code account for 12% of the raw variation in the measure, with class adding nothing detectable on top, so every race is compared against the average for its own kind and the remaining 88% is what the rest of this is about.
The test the mechanism failed
A forecast can fail for two quite different reasons, and distinguishing between them is the whole design of what follows. The cause being forecast might be perfectly real and simply hard to see in advance, in which case better inputs would help; or what is being treated as the cause might not be a cause at all, in which case no amount of forecasting skill will ever recover anything. Testing a forecast cannot tell the two apart, so this does not test a forecast.
Instead it takes what actually happened in 4,953 races — how many horses genuinely made the running, read afterwards from each race's own in-running account — and sets that against how strung out the same race finished. No forecast could ever have this information, and the mechanism is therefore being given every advantage available to it, marked against the answer sheet rather than asked to predict anything.
On that basis, the number of horses that actually led carries a correlation of +0.0099 with how strung out the field finished. Counting the horses that led or raced prominently, rather than led alone, returns −0.0068, which has the wrong sign as well as no magnitude, and splitting the races by date gives +0.0083 in the earlier half against +0.0144 in the later, so the finding is at least consistent with itself. A correlation of one in a hundred, measured with hindsight, is not a weak signal that sharper measurement might eventually bring out, because knowing exactly how many horses went forward turns out to tell you almost nothing about whether the field came home together.
What actually strings a field out
Something must account for it, and sorting the same races by how well backed the winner was finds it quickly enough. In the fifth of races taken by the longest-priced winners, the residual gap to third comes in at −0.007, meaning those races finish slightly tighter than their type would predict. The middle fifth sits at +0.037. The fifth taken by the shortest-priced winners reaches +0.198, comfortably the largest figure anywhere in the study, and the gradient between the two extremes is clean and runs in one direction throughout. Across all of it the correlation is +0.307, some thirty times what the front-runner count manages, and it barely moves when the data is split, at +0.3073 in the earlier half and +0.3062 in the later.
What that describes is not a hard-run race but a race that a good horse won easily. A field spreads out when one of its members is materially better than the rest, because a better horse wins by further, and it bunches when ability is distributed evenly through the field, because evenly matched horses finish close to one another. The quantity being measured was never the pace at all; it was the distribution of ability in the field, which happens to produce the visible effect that everyone has attributed to the pace.
Three separate attempts had already been made to forecast race shape before this one, from front-runner counts, from the pace of the races the field had last contested, and from the language of the in-running comments, and all three landed somewhere between 0.10 and 0.15. The conclusion drawn at the time was that race shape could not usefully be forecast, which was wrong in an instructive way. All three attempts went looking for the cause among the runners' running styles, because running styles are what a pace map is assembled from, and a question asked from inside a tool tends to arrive with the tool's own answer already attached. Three failures against one mistaken premise were read as three pieces of evidence that the thing was unforecastable, when they amounted to one piece of evidence, repeated, that it was being looked for in the wrong place.
What changed, and what has not
Part of the engine's read was conditioned on a shape inference built from counting front runners, and at the end of October that was taken out: the shape blend came out of the pace modifier, and the modifier stopped being conditioned on a scenario that cannot be forecast from what it was being forecast from. Measuring the claim cost the engine a feature, which is the outcome measurement is for.
What the map does elsewhere is unaffected. Where a horse is likely to race, whether it will lead, press, sit midfield or be ridden from the back, is a separate question from how strung out the finish will be; it is measured separately and it holds. The lanes are not what failed here, and nor is the idea that knowing where horses will race is worth something. What failed is one inference drawn from counting them.
Shape itself does move results, shifting the share of races won from the front by somewhere between nine and fourteen points across its range depending on the period measured, so a race's shape is worth knowing if it can be known. There is also a plausible cause that ought to be visible in the morning, given that the distribution of ability in a field is precisely what a betting market spends the preceding night pricing. That work is unfinished and none of it is on a card. It is being logged forward against races as they are run before any claim is made for it, since a result measured on history and a result measured on races that had not yet happened are not the same result, and this piece is already an account of what comes of believing a plausible mechanism instead of testing it. If it earns its place, it will be written up here.
One thing can be said now, and it cuts against the commercial reading of all this. A shape read assembled from the market's own prices is the market's own information handed back in a different form. It is a better description of how a race will unfold than counting front runners, and on everything measured so far it is not an edge over the prices it was built from, so nothing above should be taken as claiming that it is.
Closing
Losing a feature is not the uncomfortable part. The uncomfortable part is that a claim this central had never been checked, by us or, as far as can be established, by anyone else selling the same thing, and the reasons it went unchecked for so long are that it is intuitive, endlessly repeated, and does a decent job of explaining what people watching a race actually see. Set the count against the outcome with the answer already in hand and it comes back at one in a hundred.
A read of a race can still be useful when the story underneath it is wrong, which is roughly what seems to have happened here. The only way to find out which of those you are holding is to measure it, and the cost of measuring is occasionally having to say in public that you had it wrong.