Fewest prior runs: the mark has the least information

Reasoning

Why this hypothesis exists: what the lab saw, the question it asked, the mechanism it proposed, what it predicted, and the result that would prove it wrong. Written before the test was run.

1. Observed
   • H126-lightly-raced-four-year-old-returns (H126: lightly raced four year old returns) was GREEN in the earlier research: A/E 1.12 on 2958 bets (periods 1.21 / 1.01 / 1.20); now PROMISING. O-7b05ae3b
2. AskedIs the H126 lift caused by four-year-old physical maturation that the handicap mark hasn't caught up with, or just by lightly raced horses returning fresh at any age? answered
3. MechanismA horse with five or fewer runs was handicapped on three runs or so, often as a backward juvenile or three-year-old. It has the most room left to improve, and its mark is the most likely to understate its ability at four. This is stated before seeing any split.
4. Why the market misses itThe form book shows only a few modest runs, so punters price it off a thin record and an unimpressive profile.
5. Predictedpositive, A/E 1.15 to 1.25 on 700+ bets
6. Falsified ifA/E at or below the full H126 figure (1.12), or any period below 1.0 on 700+ bets
7. Rule sketchrace_type = 'Flat' AND is_handicap = 1 AND month IN (3, 4, 5) AND age = 4 AND prior_runs <= 5 AND days_since_last_run >= 150
8. question idQ-dc3ec26a
9. operator

The question's answer so far: Not maturation. In the vault the four-year-old March-April rule scored 1.02 on 1548 and all spring returners 1.03 on 6934. Non-four-year-olds scored 1.048 on 965 at discovery. Age four adds nothing measurable once tested out of sample.

Mechanism and rationale

A horse with five or fewer runs was handicapped on three runs or so, often as a backward juvenile or three-year-old. It has the most room left to improve, and its mark is the most likely to understate its ability at four. This is stated before seeing any split.

The rule

race_type = 'Flat' AND is_handicap = 1 AND month IN (3, 4, 5) AND age = 4 AND prior_runs <= 5 AND days_since_last_run >= 150

Reads: race_type, is_handicap, month, age, prior_runs, days_since_last_run

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

c44c605af41faf8efe3e1d77de91768042e7112f8502d34ec574c7228ac7d8c2

Frozen

not frozen

The definition can still change.

Confounders

• Data window: prior_runs counts only from July 2020, so in spring 2021 (and partly 2022) a four-year-old's two-year-old runs from 2019 are missing. Some 'lightly raced' horses are really exposed, and prior_runs of 0 to 2 in a handicap is often this artefact. The 3-to-5 refinement and the period split test for it.
• Lightly raced horses from top yards draw strong market support, which can push the selection edge into prices too short to pay at SP. H126 already needed BSP (SP ROI -19.6%).
• Unraced at two because of physical problems: some of these are fragile, not backward, and that could dilute or hide the effect
• Overlap with handicap-debut effects: many horses with 5 or fewer runs are running in their first or second handicap, so the effect may be about the debut and not about the size of the record
• Longshot artefact: lifts at 16/1 and bigger are common across rules and are not evidence for this mechanism, so the 14/1-or-shorter refinement must also hold
• Period instability: H126's middle period scored 1.01, so a pooled A/E can hide one period that fails

Discovery result

pass broad rule (niche) · 16 Oct, 12:12

Bets

218

25 winners

A/E

1.18

95%: 0.76 to 1.74

SP return

-15.7%

level stakes

Betfair SP return

+3.7%

218 bets with a BSP

Strike rate 11.5%. p = 0.24. Same-race control: the other runners in these races had A/E 0.98, so the selection beat its own races. Zone: discovery. Evaluator eval-1.0, data 6a4e98be18e83933

Falsification

Each attack is a deliberate attempt to kill the idea. 5 of 6 survived.

• control: Control: H126 four-year-olds with 6-10 prior runspass

  301 bets · A/E 1.21 (0.86 to 1.66) · control 0.97 · 16 Oct, 12:20

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-271731.30

  after 2026-04-271281.09

• partition: Partition: 3-5 prior runs only (data-window artefact check)pass

  193 bets · A/E 1.20 (0.76 to 1.80) · control 0.98 · 16 Oct, 12:20

Sealed holdout (vault)

Not queried. Only a frozen idea can spend its one query against the 2020-25 holdout.

Forward record

No forward record. Shadow tracking starts after the vault and counts only races run after the freeze.

Other experiments

• neighbour: <= 5 -> 4 (-20%)pass

  139 bets · A/E 1.28 (0.73 to 2.08) · control 0.97 · 16 Oct, 12:12

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-27861.57

  after 2026-04-27530.98

• neighbour: <= 5 -> 6 (+20%)pass

  291 bets · A/E 1.15 (0.80 to 1.60) · control 0.98 · 16 Oct, 12:12

  Breakdown

Lineage

Agent runs

• designer v1ok · 40s · 16 Oct, 12:10

  Output (schema-checked)

  {
    "rule": "race_type = 'Flat' AND is_handicap = 1 AND month IN (3, 4, 5) AND age = 4 AND prior_runs <= 5 AND days_since_last_run >= 150",
    "refinements": [
      {
        "label": "handicapped on three to five runs",
        "rule": "race_type = 'Flat' AND is_handicap = 1 AND month IN (3, 4, 5) AND age = 4 AND prior_runs BETWEEN 3 AND 5 AND days_since_last_run >= 150"
      },
      {
        "label": "14/1 or shorter",
        "rule": "race_type = 'Flat' AND is_handicap = 1 AND month IN (3, 4, 5) AND age = 4 AND prior_runs <= 5 AND days_since_last_run >= 150 AND sp_decimal <= 15"
      }
    ],
    "control": {
      "description": "Two comparisons. (1) Same-race other runners: expected near 1.00 (about 0.97 to 1.02), showing the race itself is not over-producing winners at long prices. (2) The rest of H126, the four-year-olds with 6 to 10 prior runs under the same conditions: the mechanism says this group should score clearly below the 5-or-fewer group (expected around 1.05 to 1.10). If the 6-10 group scores as well as or better than the 5-or-fewer group, the 'least information' mechanism fails even if the broad rule passes, because the edge would be about being a four-year-old back from a break, not about a thin record.",
      "rule": "race_type = 'Flat' AND is_handicap = 1 AND month IN (3, 4, 5) AND age = 4 AND prior_runs BETWEEN 6 AND 10 AND days_since_last_run >= 150"
    },
    "pass_criteria": {
      "min_bets": 300,
      "min_ae": 1.12,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 100,
      "min_ae": 1.05,
      "max_p": 0.2
    },
    "confounders": [
      "Data window: prior_runs counts only from July 2020, so in spring 2021 (and partly 2022) a four-year-old's two-year-old runs from 2019 are missing. Some 'lightly raced' horses are really exposed, and prior_runs of 0 to 2 in a handicap is often this artefact. The 3-to-5 refinement and the period split test for it.",
      "Lightly raced horses from top yards draw strong market support, which can push the selection edge into prices too short to pay at SP. H126 already needed BSP (SP ROI -19.6%).",
      "Unraced at two because of physical problems: some of these are fragile, not backward, and that could dilute or hide the effect",
      "Overlap with handicap-debut effects: many horses with 5 or fewer runs are running in their first or second handicap, so the effect may be about the debut and not about the size of the record",
      "Longshot artefact: lifts at 16/1 and bigger are common across rules and are not evidence for this mechanism, so the 14/1-or-shorter refinement must also hold",
      "Period instability: H126's middle period scored 1.01, so a pooled A/E can hide one period that fails"
    ],
    "mechanism": "A four-year-old with five or fewer career runs was given its handicap mark on about three runs, usually as a backward two- or three-year-old. Of all the H126 horses, its mark rests on the least evidence and it has the most physical improvement still to come over the winter, so its spring mark is the most likely to understate its ability at four. If the mechanism is right, the H126 edge should get stronger as the number of prior runs falls.",
    "market_miss": "The form book shows only a few modest runs and a long absence. Punters and bookmakers price the horse off that thin, unimpressive record and do not price in the winter's improvement, which the record cannot show.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.15 to 1.25 on 700+ bets, above the 6-10-runs control group",
      "falsified_if": "A/E at or below 1.12 (the full H126 figure); or any period (2020-22, 2023 to a year ago, the last year) below 1.0; or the 6-10-runs control scoring at or above the 5-or-fewer group"
    },
    "notes": "This sharpens H126 and adds only one condition (prior_runs <= 5 in place of <= 10), so it can be judged directly against H126's 1.12. Pass min_ae is raised to 1.12 because a subset that does no better than its parent adds nothing. The 3-to-5 refinement comes from the mechanism ('handicapped on three runs or so') and also strips out the pre-July-2020 run-count artefact. Since no year column exists, period agreement has to be checked by the lab's standard split. Expect the result to need BSP, as H126 did."
  }

Timeline

1. Job 1303 finished

   16 Oct, 12:20 · engine:compute-1

2. H-625f1f2d Falsification → KILLED falsified: driven by one price band (up to 14/1): the rest score 1.00

   16 Oct, 12:20 · engine:falsification

3. falsification result for H-625f1f2d (partition: Partition: already handicapped last time (not handicap debut)): pass

   16 Oct, 12:20 · engine:evaluator

4. compute-1 took job 1303 (attack, attempt 1)

   16 Oct, 12:20 · compute-1

5. Job 1302 finished

   16 Oct, 12:20 · engine:compute-1

6. falsification result for H-625f1f2d (price_band: 14/1 or shorter, excluding the 8/1-14/1 hot band): fail

   16 Oct, 12:20 · engine:evaluator

7. compute-1 took job 1302 (attack, attempt 1)

   16 Oct, 12:20 · compute-1

8. Job 1301 finished

   16 Oct, 12:20 · engine:compute-1

9. falsification result for H-625f1f2d (perturbation: Perturbation loose: prior_runs <= 6, 125+ days off): pass

   16 Oct, 12:20 · engine:evaluator

10. compute-1 took job 1301 (attack, attempt 1)

    16 Oct, 12:20 · compute-1

11. Job 1300 finished

    16 Oct, 12:20 · engine:compute-1

12. falsification result for H-625f1f2d (perturbation: Perturbation tight: prior_runs <= 4, 180+ days off): pass

    16 Oct, 12:20 · engine:evaluator

13. compute-1 took job 1300 (attack, attempt 1)

    16 Oct, 12:20 · compute-1

14. Job 1299 finished

    16 Oct, 12:20 · engine:compute-1

15. falsification result for H-625f1f2d (partition: Partition: 3-5 prior runs only (data-window artefact check)): pass

    16 Oct, 12:20 · engine:evaluator

16. compute-1 took job 1299 (attack, attempt 1)

    16 Oct, 12:20 · compute-1

17. Job 1298 finished

    16 Oct, 12:20 · engine:compute-1

18. falsification result for H-625f1f2d (control: Control: H126 four-year-olds with 6-10 prior runs): pass

    16 Oct, 12:20 · engine:evaluator

19. compute-1 took job 1298 (attack, attempt 1)

    16 Oct, 12:20 · compute-1

20. Job 1278 finished

    16 Oct, 12:20 · engine:ai-1