What 207 Races Taught Us About Projecting the Lead

The test

The base is every race we settled from October 8–19, 2026: 207 races, 1,501 starters, each one graded against its official chart. For the leader question, 148 races had a gradable first-call leader and a complete projection. Four candidate projections were scored on the same matched races:

• The blend — our production sketch at the time: early-pace figures blended with early-speed points.
• The blend, E1 everywhere — the same blend with the first-call figure used in routes too, instead of the second-call figure.
• Early-speed points only — the simple count of demonstrated early speed, no figures at all.
• Our own points — early points we constructed ourselves from running lines, recency-weighted by lengths behind at the first call over each horse's last three starts.

Two scores for each: leader-hit (how often the projection's named leader actually led at the first call) and the Spearman correlation between the projected early order and the real one.

1. The simple count beat every blend

Early-speed points alone named the actual first-call leader in 41.2% of gradable races — 8.8 points clear of the blended sketch we were publishing, on the same races. Adding figure information to the points made the projection worse, not better. That was not the result we expected when we built the blend, and it was not close enough to argue with. The sketch on the card now projects from the points alone (this is RACE SHAPE v1.2, the version of our pace-shape annotation that followed the study).

One race shows what the table means in practice. At Keeneland on October 17, the blend projected Obliteration as the leader by the third-widest margin in the study; the points named Roscoe's Brother. The chart says Roscoe's Brother led and Obliteration was 8th at the first call. We walk through that race in full — pre-race card and chart side by side — in How to Read a Pace Card. But one race is an illustration. The 148-race table above is the evidence.

2. Our own home-built points lost

We wanted to own the input, so we built early points from the running lines themselves — no vendor numbers. They scored 31.8%: behind the product's points, and behind even the blend we were trying to replace. The construction has a flaw we can name: lengths-behind at the first call mixes a horse's own speed with the pace it happened to face, and the data we captured included no past field sizes to correct for that. The honest outcome is that our points, as constructed, are not a substitute today. That line of work is parked until a field-relative version — lengths measured against each past field's own first-call median — earns another run.

3. Printed running styles are tendencies, not roles

Every horse carries a printed style label (E, E/P, P, S). We computed styles a second way — from what each horse demonstrably did in its recent races — and the two disagreed on 61.8% of the 1,474 starters who had a computed style. The chart record shows why the label can't be taken literally: horses labeled E led at the first call only 23.2% of the time, and were 5th or worse at the first call in 31.7% of their starts. Horses labeled S — closers — still led at the first call 5.0% of the time. A style label is a leaning, not an assignment, and any method that treats it as a role is building on sand.

4. The named pace shapes did not survive

RACE SHAPE v1.0 sorted races into named types — Solo Lead, Soft Front, Pocket Trip, Meltdown, Dual Early — each with a story about how the race would run and who it favored. The stories read well. In the combined 207-race base, graded on paper at each type's own terms, every named type lost: Meltdown −27.9% ROI across 33 races, Solo Lead −63.2% across 5, Pocket Trip −70.0% across 8, and the rest of the family with them. Part of the failure was structural — the early version of the rules called 74 of the first 92 graded fields "hot," so the soft-front types the system was proudest of barely existed. All of the types were retired. A good story about a race shape is not the same thing as a measured effect, and we now treat any named pattern as a hypothesis until a settled base says otherwise.

5. A big projected gap made things worse, not better

In the feature study that preceded this backtest, projected leaders with a wide projected gap over the field performed worse, returning −77.7% ROI at gaps of 1.5 or more — the opposite of the intuition that a standout projection is a safer one. In the v1.2 run we could not re-test that cell: the harness's first-call lengths capture came back empty on the chart set, and we report that as not run, not as evidence either way. The original result stands as a single unreplicated finding, labeled as one.

6. Routes resist projection

In the 39 route races with an identifiable chart leader, the blend named the leader 10.3% of the time (4 of 39); using the first-call figure everywhere improved it only to 15.4% (6 of 39). The cell is small, but the direction was the same under every candidate: projecting the early lead in routes is weak, full stop. The card still draws the sketch for routes — as annotation, with this record behind it.

7. One hypothesis survived — carefully

Top Closer: horses with a printed S or P/S style and a raw late-pace percentile of 90 or higher. In this base they returned +2.4% ROI across 60 starters (+5.9% in the earlier feature study) — winning right at what their prices implied, with final-odds checks saying the market prices these horses fully. One detail decides whether the idea lives at all: run the same gate with behavior-computed styles instead of the printed ones and the cell nearly triples to 146 starters and loses −18.9%. The gate only holds under its printed definition. Top Closer runs on the card as a flagged hypothesis, graded nightly against settled results. It does not make picks, and one $47.20 winner accounted for 38% of its feature-study returns — leave that horse out and the cell swings from −34.8% to +2.9%. We say so because it is true.

8. The solo-early mirage

The strangest cell in the study. Defined by printed labels, the projected "solo early" leader was the worst cell we measured: −55.8% ROI across 20 races. Defined by demonstrated behavior instead — what the horses had actually done early in recent starts — the same idea flipped to +32.7% across 32 races. A flip that clean, in one cell among the many we examined, is exactly how false discoveries happen in handicapping data. It is filed as a frozen hypothesis for a future forward test. It is not on the card, and it is not in the picks.

What actually changed

• The card's pace sketch now projects the early order from early-speed points only (RACE SHAPE v1.2).
• The named pace-shape types were retired; the sketch prints race facts and the projection, not a verdict.
• Our own constructed points are parked, pending a field-relative rebuild.
• Top Closer survives as a flagged, nightly-graded hypothesis under its printed definition only.
• None of this touches the selections. Rankings follow the morning line; the sketch annotates the race and never reorders a pick. Why the morning line holds that job is its own guide: The Morning Line Is Not a Prediction.

And the grading does not stop at publication. Every settled race since has kept scoring the projection the same way, and the record — picks, sketches, payouts — settles in public each night in the results archive.