Rival: market over-penalises any last-time chase error, not only a one-off slip

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
   • H-3a4c45cd (Rival: H24 chase edge is only the faller/unseater story; finishers bea) passed discovery as a niche: A/E 1.18 on 389 bets (p 0.128), same-race control 0.98. O-07de040e
   • H-5bdc794e (Test-mechanism: one-off chase error that the sectionals show cost grou) passed discovery as a niche: A/E 1.29 on 243 bets (p 0.093), same-race control 0.97. O-9d7a47ef
2. AskedIs the 'one-off' condition needed? Do well-beaten chasers that made a jumping error last time also beat their price when they had errors in the runs before too? answered
3. MechanismPunters and layers treat any visible jumping mistake in the last run as a reason to oppose. But many chasers that err still jump well enough to win next time, especially once the error is behind them, so the error discount is too large whatever the history.
4. Why the market misses itThe remark 'mistake' plus a big beaten distance or a non-finish looks like a bad run in the form book. The market reads it as a lack of ability or of safe jumping and prices both, double-counting the cost of the error.
5. Predictedpositive, A/E 1.10 to 1.20 on 300+ bets
6. Falsified ifA/E of 1.05 or lower. That would mean the 'clean in the two runs before' condition is doing the work, and the one-off story stands.
7. Rule sketchrace_type = 'Chase' AND jumping_errors_last3 >= 2 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10)

The question's answer so far: Yes, the one-off condition is needed. Well-beaten chasers whose last-time error came on top of earlier errors scored only A/E 1.043 on 670 (p 0.375, control 1.0), against 1.224 for the one-off pool.

Mechanism and rationale

Punters and layers treat any visible jumping mistake in the last run as a reason to oppose. But many chasers that err still jump well enough to win next time, especially once the error is behind them, so the error discount is too large whatever the history.

The rule

race_type = 'Chase' AND jumping_errors_last3 >= 2 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10)

Reads: race_type, jumping_errors_last3, last_jumping_error, last_position, last_beaten_lengths

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

9b42cbbaaa309173859756ee3bdc97f38310802f2f62df79b404264e6693fab5

Frozen

not frozen

The definition can still change.

Confounders

• Repeat-error chasers skew to novices and lower classes, so class mix differs from the one-off pool; check is_novice and race_class splits
• The share of non-finishers (fell/unseated/pulled up) may differ from the parent pool; check completers and non-finishers separately
• More long shots among habitual bad jumpers; the lab-wide 16/1+ failure could drag the broad rule down, so look at the 14/1-or-shorter band
• Overlap with the existing fell/unseated forgive systems and the H24 chase 10L+ edge; the lift may be inherited from those
• jumping_errors_last3 and last_jumping_error only start at 2025-11-01, so the window is short and seasonal (one winter of chasing); the sample may be thin
• Edge in this family has been BSP-dependent (SP ROI negative); judge on A/E against the market, not SP profit

Control

{
  "description": "Same-race other runners should score about A/E 1.0 (0.95 to 1.02), showing the lift is not a property of these chases. The key comparison is the one-off error pool (exactly one error in the last three, that one last time) under the same well-beaten/non-finish condition: the rival story predicts the repeat-error pool scores about as well as the one-off pool (1.17 to 1.22 region); if the repeat pool scores at or below 1.05 while the one-off pool holds, the 'clean before' condition is doing the work.",
  "rule": "race_type = 'Chase' AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10)"
}

Discovery result

fail broad rule · 17 Oct, 01:08

Bets

670

73 winners

A/E

1.04

95%: 0.82 to 1.31

SP return

-14.3%

level stakes

Betfair SP return

+12.8%

665 bets with a BSP

Strike rate 10.9%. p = 0.37. Same-race control: the other runners in these races had A/E 1.00, so the selection beat its own races. Zone: discovery. Evaluator eval-1.0, data 6a4e98be18e83933

Falsification

No attacks yet. Only promising ideas are attacked.

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.

Lineage

Agent runs

• designer v1ok · 31s · 17 Oct, 01:06

  Output (schema-checked)

  {
    "rule": "race_type = 'Chase' AND jumping_errors_last3 >= 2 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10)",
    "refinements": [
      {
        "label": "14/1 or shorter",
        "rule": "race_type = 'Chase' AND jumping_errors_last3 >= 2 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10) AND sp_decimal <= 15"
      },
      {
        "label": "completed last time, beaten 10L+",
        "rule": "race_type = 'Chase' AND jumping_errors_last3 >= 2 AND last_jumping_error = 1 AND last_position IS NOT NULL AND last_beaten_lengths > 10"
      }
    ],
    "control": {
      "description": "Same-race other runners should score about A/E 1.0 (0.95 to 1.02), showing the lift is not a property of these chases. The key comparison is the one-off error pool (exactly one error in the last three, that one last time) under the same well-beaten/non-finish condition: the rival story predicts the repeat-error pool scores about as well as the one-off pool (1.17 to 1.22 region); if the repeat pool scores at or below 1.05 while the one-off pool holds, the 'clean before' condition is doing the work.",
      "rule": "race_type = 'Chase' AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10)"
    },
    "pass_criteria": {
      "min_bets": 150,
      "min_ae": 1.05,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 100,
      "min_ae": 1,
      "max_p": 0.2
    },
    "confounders": [
      "Repeat-error chasers skew to novices and lower classes, so class mix differs from the one-off pool; check is_novice and race_class splits",
      "The share of non-finishers (fell/unseated/pulled up) may differ from the parent pool; check completers and non-finishers separately",
      "More long shots among habitual bad jumpers; the lab-wide 16/1+ failure could drag the broad rule down, so look at the 14/1-or-shorter band",
      "Overlap with the existing fell/unseated forgive systems and the H24 chase 10L+ edge; the lift may be inherited from those",
      "jumping_errors_last3 and last_jumping_error only start at 2025-11-01, so the window is short and seasonal (one winter of chasing); the sample may be thin",
      "Edge in this family has been BSP-dependent (SP ROI negative); judge on A/E against the market, not SP profit"
    ],
    "mechanism": "Punters and layers treat any visible jumping mistake last time as a reason to oppose. Many chasers that err still jump well enough to run to form next time, so the error discount is too large whether or not the horse also erred in earlier runs. If true, the one-off condition from the parent (clean in the two runs before) is not needed for the lift.",
    "market_miss": "The remark 'mistake' plus a big beaten distance or a non-finish reads as a bad run in the form book. The market prices it both as lack of ability and as unsafe jumping, counting the cost of the error twice. With a pattern of errors the form-book reader is even more likely to write the horse off, so the overreaction should be at least as large.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.10 to 1.20 on 150 to 300+ bets, similar to the one-off pool (about 1.2)",
      "falsified_if": "A/E of 1.05 or lower on the broad rule (or on the 14/1-or-shorter refinement) while the one-off control holds. That would mean the 'clean in the two runs before' condition does the work and the one-off story stands."
    },
    "notes": "I kept the rule sketch as written. All the columns are pre_off. The jumping-error columns only start at 2025-11-01 for both GB and IRE, so the pool may come in under 150; if it does but points the right way, the lab's thin-result path applies. The refinements come from the family findings: edges die at 16/1+, and completers beaten 10L+ hold on their own."
  }

Timeline

1. Job 2141 finished

   17 Oct, 01:08 · engine:compute-1

2. H-4eed61bb Discovery → KILLED fails: A/E 1.043, needs 1.05; p 0.3747, needs 0.05 or lower (not a niche pass: A/E 1.043 below 1.05)

   17 Oct, 01:08 · engine:evaluator

3. discovery result for H-4eed61bb (broad rule): fail

   17 Oct, 01:08 · engine:evaluator

4. H-4eed61bb Designed → RESEARCHdiscovery measuring on the discovery zone

   17 Oct, 01:08 · engine:orchestrator

5. compute-1 took job 2141 (evaluate, attempt 1)

   17 Oct, 01:08 · compute-1

6. Job 2139 finished

   17 Oct, 01:06 · engine:ai-1

7. Queued evaluate on the compute lane: discovery run of the broad rule

   17 Oct, 01:06 · engine:scheduler

8. H-4eed61bb Proposed → RESEARCHdesigned pre-spec written and hashed

   17 Oct, 01:06 · agent:designer@v1

9. ai-1 took job 2139 (design, attempt 1)

   17 Oct, 01:06 · ai-1

10. Queued design on the ai lane: write the pre-spec

    17 Oct, 01:06 · engine:scheduler

11. New idea H-4eed61bb: Rival: market over-penalises any last-time chase error, not only a one-off slip (rival-explanation)

    17 Oct, 01:06 · agent:investigator@v1

Every change is an event with an actor: an agent and its version, an engine, a human, or the importer. Current stage: Killed.