> ## Content Index
> Fetch the complete content index at: https://horse-feathers.ghost.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# Horse-Oracle Analytics
- URL: https://horse-feathers.ghost.io/analytics-2/
- Published: 2026-09-28T21:05:17.000Z
- Updated: 2026-10-03T23:05:13.000Z
- Author: Horse-Oracle
- Tags: analytics, public-record

![Racehorse running past Santa Anita Park with the San Gabriel Mountains behind it](https://showcase.horse-oracle.com/stats-report/hero-santa-anita.jpg)

Horse-Oracle · Analytics

# How the record looks

A plain-English look at the races we tested, what held up, and what is still open.

## The short version

**For racing, business, and track readers** 
**Database snapshot:** through 2026-09-26  
**Current graded prediction window:** 2026-08-25 through 2026-09-25  
**Matched EquinEdge comparison window:** 2026-08-17 through 2026-09-13

**What you are seeing:** Horse-Oracle names horses before each race. These numbers show how often the winner was in the first one, two, or three names. They do not show profit, and they do not promise what happens next.

### Where the winner appeared

**Race window:** August 25–September 25, 2026\. Results were counted only after an official result was available.

| What we checked                                                                                           | Percentage | Race count |
| --------------------------------------------------------------------------------------------------------- | ---------- | ---------- |
| **First choice won**The top name was the winner in about 31 out of 100 graded races.                      | 31.25%     | 996        |
| **Winner in first two**The winner was one of the first two names in about 51 out of 100 graded races.     | 51.15%     | 1,630      |
| **Winner in first three**The winner was one of the first three names in about 66 out of 100 graded races. | 65.99%     | 2,103      |

**Plain-English read:** The first three names gave the winner a place to look in about two out of every three graded races. It does not mean the first three finishers were listed in order.

**Settlement cutoff:** The reporting window ends September 25, but this release evaluates only authoritative results settled through September 23, 2026\. Races still awaiting a final result remain outside the graded-performance denominator and stay visible in coverage accounting.

### How we compared with EquinEdge

**Why this matters:** This is a same-race comparison. Both services had to rate the same race before it counted.

**Matched window:** August 17–September 13, 2026 · 2,852 same-race records across 37 tracks.

| What we checked                      | Horse-Oracle | EquinEdge |
| ------------------------------------ | ------------ | --------- |
| First choice won                     | 29.10%       | 30.89%    |
| Winner appeared in first two picks   | 48.60%       | 51.44%    |
| Winner appeared in first three picks | 63.11%       | 66.90%    |

**Plain-English read:** EquinEdge had the higher number in all three rows in this matched set. The first-choice gap was small enough that this sample does not show a clear winner. That also does not prove the systems are equal.

The early public favorite won 31.10% of these same races. This is a comparison point, not a Horse-Oracle result. This comparison is a different sample from the current Horse-Oracle window above.

### Coverage and races left out

A race is counted only when the saved prediction, the race identity, and the official result can all be matched.

| Race group              | Dates               | Count                       | Left out                                                    |
| ----------------------- | ------------------- | --------------------------- | ----------------------------------------------------------- |
| Current graded results  | Aug 25–Sep 25, 2026 | 3,187                       | Races still waiting for an official result or a clean match |
| Same-race EquinEdge set | Aug 17–Sep 13, 2026 | 2,852                       | Races where one side could not be paired                    |
| September source record | Sep 1–24, 2026      | 2,619 rows / 328 track-days | 44 track-days with an explicit source gap                   |

**Why this matters:** Missing source data is shown as missing. It is not treated as a win, a loss, or complete coverage.

## The details behind the numbers

The tables below show the counts, dates, races left out, and source gaps behind the summary.

### See the comparison at a glance

### How often the first choice won

Same 2,852 paired races. Longer bars mean more first-choice wins.

Horse-Oracle29.10%EquinEdge30.89%Public favorite31.10%0%20%40%

The bars are close. This sample does not establish a clear winner.

### Where the winner appeared

Same paired races. This shows the depth of each system’s short list.

Horse-Oracle · first 248.60%EquinEdge · first 251.44%Horse-Oracle · first 363.11%EquinEdge first three: 66.90%

EquinEdge’s first-three result remains in the table above.

### When the two systems agreed

Out of 2,852 paired races, both named the same first horse in 1,314.

46.07% agreed53.93% differedSame first horse: 1,314 racesDifferent first horse: 1,538 races

Agreement is useful context: the two services often read the same race similarly.

### How complete was September’s source record

Track-days with a valid DRF/Formulator card, out of 328.

86.59%valid cards284 valid track-daysCards available for theSeptember source window.44 source gapsKept visible, not countedas complete coverage.

A source gap is not treated as a successful prediction.

**How to read these pictures:** the charts make the size and shape of the results easier to see. They do not change the underlying counts or turn a close sample into a claim of superiority.

The visual comparison is limited to the source window and the races where both records could be paired.

[Discuss a partnership or request analysis](https://horse-feathers.ghost.io/analytics/)[Explore race guides](https://horse-feathers.ghost.io/analytics/)[Jump to the appendix](#appendix)PAPER MODEL — NOT BETTING ADVICE · $0 STAKED

### Coverage accounting and “not graded” states

Unresolved races are excluded from graded-performance denominators but remain in the expected-population and coverage ledgers. Each expected race belongs to exactly one state:

- graded — eligible evidence and authoritative result bound
- pending — settlement or source evidence not yet due or not yet available
- quarantined — identity, field, prediction, or result evidence requires review
- excluded — permanently unavailable or invalid, with a recorded reason

**Not graded:** the grade is withheld when the saved pre-race evidence, field identity, or authoritative result cannot be bound. The machine-readable reason is recorded in the `not-graded-reason-codes` evidence object.

## Appendix: full evidence and coverage

## How much race data is here

### Full U.S. record

These are records in the database. They describe coverage, not model skill.

| Coverage measure             | Count      |
| ---------------------------- | ---------- |
| USA races                    | 14,768     |
| Runner starts                | 120,074    |
| Unique horses                | 44,509     |
| Unique jockeys               | 2,671      |
| Unique trainers              | 6,343      |
| Earliest race date           | 2022-04-02 |
| Latest race date in snapshot | 2026-09-26 |

### Current race window

This is the recent group used for the current report. It is not the same as the EquinEdge comparison group.

The recent coverage snapshot covers 2026-08-25 through 2026-09-25.

| Recent coverage measure | Count  |
| ----------------------- | ------ |
| Races                   | 3,611  |
| Runner starts           | 32,064 |
| Unique horses           | 22,323 |
| Unique jockeys          | 1,416  |
| Unique trainers         | 4,400  |

### Data depth added in the September refresh

The refreshed source record also includes these data counts. They show what was available, not what the model predicted:

| Data layer                    | Count   | What it represents                                     |
| ----------------------------- | ------- | ------------------------------------------------------ |
| Odds observations             | 688,838 | Stored market observations across the live odds record |
| Win/place/show pool snapshots | 521,187 | Pool snapshots across 3,815 races                      |
| Weather-linked race rows      | 72,242  | Race rows with weather context attached                |
| Workout histories             | 56,755  | U.S. horse starts with stored workout detail           |
| BRIS Prime Power values       | 47,010  | Measured source values retained for analysis           |

### September source coverage

For 2026-09-01 through 2026-09-24, the database contains **2,619 race rows across 328 track-days**. Valid DRF/Formulator cards were available for **284 of 328 track-days (86.59%)**. The remaining vendor-served errors were retained as explicit gaps rather than silently counted as complete coverage. Late-September BRIS ERD acquisition completed at **51 of 51 files** with zero download failures.

### Matched EquinEdge window coverage

The live database contains 3,420 USA race rows in the broader comparison-period window of 2026-08-17 through 2026-09-13\. Of these, 3,300 have stored results and 3,279 have a latest prediction record. The public EquinEdge comparison uses the narrower **2,852-race paired set** where both systems could be evaluated together.

That distinction matters. The 2,852-race figure is the correct denominator for the head-to-head table. It must not be replaced with the larger database count.

## Same-race comparison: EquinEdge

### How often the first choice won

- Horse-Oracle: **830 first-choice winners out of 2,852 races**, or **29.10%**.
- EquinEdge: **881 first-choice winners out of 2,852 races**, or **30.89%**.
- Early public favorite: **887 winners out of 2,852 races**, or **31.10%**.
- Horse-Oracle versus EquinEdge difference: **−1.79 percentage points**.
- 95% range for that difference: **−3.63 to +0.05 points**.
- Plain-English conclusion: **the sample does not establish a statistically clear first-choice winner between the two systems**.

### How deep the short list reached

- Horse-Oracle placed the winner in its first two choices in **1,386 races**, or **48.60%**.
- EquinEdge placed the winner in its first two choices in **1,467 races**, or **51.44%**.
- Horse-Oracle placed the winner in its first three choices in **1,800 races**, or **63.11%**.
- EquinEdge placed the winner in its first three choices in **1,908 races**, or **66.90%**.
- EquinEdge’s first-four figure was **77.73%**. A Horse-Oracle first-four figure was not admitted in the comparison source, so no first-four comparison is shown.

### When the first names were different

The systems selected different first horses in **1,538 races**, which was **53.93%** of the paired sample.

- Horse-Oracle’s first choice won in **332** of those races, or **21.59%** of disagreements.
- EquinEdge’s first choice won in **383** of those races, or **24.90%** of disagreements.
- Neither first choice won in the remaining **823** disagreement races, or **53.51%**.

This is a useful way to understand the comparison. The systems were not making different calls in every race. They named the same first horse in **1,314 races**, or **46.07%** of the paired sample.

### Field-size check

The source also reports a field-size slice for **1,579 races with eight runners**.

| Eight-runner race measure | Horse-Oracle | EquinEdge |
| ------------------------- | ------------ | --------- |
| First choice won          | 25.59%       | 27.87%    |
| Winner in first three     | 58.52%       | 62.25%    |

This slice is descriptive. It does not replace the full-sample comparison.

## People and horse coverage

The following numbers describe the people and horses represented in the recent database window. They are **coverage and observed-result counts**, not a claim that Horse-Oracle itself predicted a jockey or horse leaderboard.

### Jockey activity leaders in the recent window

| Jockey        | Starts | Wins | Observed win rate |
| ------------- | ------ | ---- | ----------------- |
| LOPEZ P       | 213    | 46   | 21.6%             |
| ORTIZ J L     | 205    | 28   | 13.7%             |
| PRAT F        | 193    | 44   | 22.8%             |
| GAFFALIONE T  | 188    | 12   | 6.4%              |
| HERNANDEZ J A | 184    | 25   | 13.6%             |
| FRANCO M      | 175    | 31   | 17.7%             |
| RODRIGUEZ J   | 173    | 31   | 17.9%             |
| GONZALEZ E    | 165    | 26   | 15.8%             |
| SAEZ L        | 163    | 16   | 9.8%              |
| SANTANA R JR  | 154    | 26   | 16.9%             |

These are raw starts and wins in the selected window. They are not adjusted for field size, odds, track, distance, surface, or quality of mounts.

### Horse coverage

The recent window contains **22,323 unique horses** across **32,064 runner starts**. The matched EquinEdge comparison contains multiple starts per horse and is designed to compare race-level first choices, not to produce a stable horse leaderboard.

A horse leaderboard needs a separate, identity-clean contract before it is suitable for external claims. The current report therefore shows horse coverage and race-level prediction performance, but does not publish a ranking of individual horses as a Horse-Oracle skill claim.

## What a reader can take from this

1. **Horse-Oracle is operating at meaningful scale.** The recent window contains 3,611 USA races and 32,064 runner starts.
2. **The system is easiest to understand as a ranking tool.** The useful question is where the winner appeared among the first one, two, or three choices.
3. **The current Horse-Oracle window is stronger than the older matched snapshot on first-choice percentage, but the windows are different.** It is not valid to present that as a direct EquinEdge victory.
4. **The matched comparison is close at the top.** The first-choice difference is 1.79 points and its 95% range reaches zero.
5. **The report is honest about what is not yet proven.** A complete, same-clock, same-race ROI comparison is not admitted here, and Horse-Oracle’s first-four comparison is not available in the source file.

## How we counted

- One prediction record was retained per race, using the latest prediction for that race.
- A race counted as a first-choice hit when the stored top pick matched the authoritative winner.
- A top-two hit means the winner was among the stored first or second choices.
- A top-three hit means the winner was among the stored first, second, or third choices.
- The EquinEdge table uses only races where both systems were paired on the same race record.
- Percentages use the number of eligible races for that measure as the denominator.
- Confidence ranges are reported from the comparison receipt and are not treated as proof of a lead when they include zero.

## What these numbers do not prove

**What is verified**

- USA database inventory and date coverage.
- Recent race, runner, horse, jockey, and trainer coverage.
- Recent Horse-Oracle graded prediction counts and placement results.
- Model version and latest-prediction-per-race contract.

**What the same-race comparison supports**

- 2,852 paired races.
- 37 tracks.
- Horse-Oracle, EquinEdge, and early public-favorite top-choice results.
- Top-two, top-three, same-pick, disagreement, and eight-runner slice values.

**What this report does not prove**

- A complete same-clock ROI comparison.
- A Horse-Oracle first-four result in the matched source.
- A standalone horse leaderboard as proof of model skill.
- A standalone jockey leaderboard as proof of model skill.
- Ordered probability or market instruments that have not yet passed their receipt gate.

PAPER MODEL — NOT BETTING ADVICE · $0 STAKED

## Source record

- Live database snapshot: Horse-Oracle Postgres, read-only query, snapshot through 2026-09-26.
- Current graded performance query: latest prediction per race joined to the authoritative race-results view, 2026-08-25 through 2026-09-25 race window, settled results through 2026-09-23.
- Matched comparison source: `docs/HO_VS_EQUINEDGE_H2H_20260913.json`.
- Matched comparison window: 2026-08-17 through 2026-09-13.
- Paired comparison denominator: 2,852 races.
- This report is a detailed companion to the public Horse-Oracle data brief, not a replacement for the underlying receipts.