Replicate: Flat handicappers heavily beaten in a big field last time beat their price

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-91a89a7c (Rival: the big-field lift is a drop into a smaller race today, not a m): the price band 'up to 7/1' carries the effect (A/E 1.51) while the rest score up to 3/1 0.82, up to 14/1 0.98, 16/1 or bigger 0.54. O-52f2a04d
   • H-9ccf4a8d (Test-mechanism: big-field beatings where the horse still finished in t) passed discovery: A/E 1.29 on 642 bets (p 0.03), same-race control 0.98. O-6a8bcbc8
   • H-f316faf4 (Generalise: big-field heavy beatings mislead in non-handicap jumps rac) pointed the predicted way but failed the bar: A/E 1.06 on 1684 bets (p 0.322), control 0.97. O-ff334f83
2. AskedDoes the market's over-penalising of big-field heavy beatings also appear in Flat handicaps, where big fields are routine? testing
3. MechanismIn big fields, horses whose chance has gone are not ridden out, and traffic compounds the losses. The beaten distance therefore overstates the true gap, more so as the field gets bigger.
4. Why the market misses itThe form book records the lengths. Punters read a 10-length beating in a 20-runner handicap as poor form, but the same performance in an 8-runner race would have been a few lengths.
5. Predictedpositive, A/E 1.06 to 1.12 on 1500+ bets, with the lift under 16/1
6. Falsified ifA/E at or below 1.02 on 1000+ bets, or the lift only at 16/1 or bigger

Mechanism and rationale

In big fields, horses whose chance has gone are not ridden out, and traffic compounds the losses. The beaten distance therefore overstates the true gap, more so as the field gets bigger.

The rule

race_type = 'Flat' AND is_handicap = 1 AND last_race_type = 'Flat' AND last_field_size >= 14 AND last_beaten_lengths >= 8 AND sp_decimal <= 15

Reads: race_type, is_handicap, last_race_type, last_field_size, last_beaten_lengths, sp_decimal

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

2fcb79832c67da7babde3d07b441eb693deeaf5dd3cb2f7555d23a9631ab3a3a

Frozen

not frozen

The definition can still change.

Confounders

• Flat draw bias in big fields: a heavy beating from a bad draw is a known excuse, and a well-drawn horse today may be priced on that. Check that the lift is not confined to particular courses or draw positions.
• Class drops after valuable big-field handicaps (big fields cluster in heritage and high-value races). Check that the lift survives when today's race_class is not lower than last_race_class.
• The general favourite-longshot bias. The sp_decimal <= 15 cap already guards against it; the lift must not rest on the 10/1 to 14/1 end alone.
• Handicap mark drops after a heavy beating: the lift may come from a lower rating rather than misread form. Check it against rating_change.
• Country and period: the result must hold in both GB and IRE and in both halves of the discovery window, because the jumps chase-debut cuts failed on exactly this.

Control

{
  "description": "The lab measures the other runners in the same races. If the beaten distance really overstates the gap, they should come out at about par (A/E around 0.97 to 1.00 after the overround), clearly below the selections. If they show the same lift, the effect belongs to the race (for example, big Flat handicaps being mispriced in general), not to the horse. A second comparison isolates the field-size part of the mechanism: the same heavy beatings taken in last-time fields of 10 or fewer, at the same prices, should show little or no lift.",
  "rule": "race_type = 'Flat' AND is_handicap = 1 AND last_race_type = 'Flat' AND last_field_size <= 10 AND last_beaten_lengths >= 8 AND sp_decimal <= 15"
}

Discovery result

fail broad rule · 17 Oct, 03:35

Bets

1,560

160 winners

A/E

0.95

95%: 0.81 to 1.11

SP return

-24.9%

level stakes

Betfair SP return

-9.0%

1,547 bets with a BSP

Strike rate 10.3%. p = 0.74. Same-race control: the other runners in these races had A/E 1.01, so the selection did not 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 · 26s · 17 Oct, 03:31

  Output (schema-checked)

  {
    "rule": "race_type = 'Flat' AND is_handicap = 1 AND last_race_type = 'Flat' AND last_field_size >= 14 AND last_beaten_lengths >= 8 AND sp_decimal <= 15",
    "refinements": [
      {
        "label": "Bigger last-time field (18+): the mechanism says the overstatement grows with field size",
        "rule": "race_type = 'Flat' AND is_handicap = 1 AND last_race_type = 'Flat' AND last_field_size >= 18 AND last_beaten_lengths >= 8 AND sp_decimal <= 15"
      },
      {
        "label": "Fancied band up to 7/1: in the jumps parent's cuts the lift was concentrated at shorter prices",
        "rule": "race_type = 'Flat' AND is_handicap = 1 AND last_race_type = 'Flat' AND last_field_size >= 14 AND last_beaten_lengths >= 8 AND sp_decimal <= 8"
      }
    ],
    "control": {
      "description": "The lab measures the other runners in the same races. If the beaten distance really overstates the gap, they should come out at about par (A/E around 0.97 to 1.00 after the overround), clearly below the selections. If they show the same lift, the effect belongs to the race (for example, big Flat handicaps being mispriced in general), not to the horse. A second comparison isolates the field-size part of the mechanism: the same heavy beatings taken in last-time fields of 10 or fewer, at the same prices, should show little or no lift.",
      "rule": "race_type = 'Flat' AND is_handicap = 1 AND last_race_type = 'Flat' AND last_field_size <= 10 AND last_beaten_lengths >= 8 AND sp_decimal <= 15"
    },
    "pass_criteria": {
      "min_bets": 300,
      "min_ae": 1.05,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 150,
      "min_ae": 1.02,
      "max_p": 0.2
    },
    "confounders": [
      "Flat draw bias in big fields: a heavy beating from a bad draw is a known excuse, and a well-drawn horse today may be priced on that. Check that the lift is not confined to particular courses or draw positions.",
      "Class drops after valuable big-field handicaps (big fields cluster in heritage and high-value races). Check that the lift survives when today's race_class is not lower than last_race_class.",
      "The general favourite-longshot bias. The sp_decimal <= 15 cap already guards against it; the lift must not rest on the 10/1 to 14/1 end alone.",
      "Handicap mark drops after a heavy beating: the lift may come from a lower rating rather than misread form. Check it against rating_change.",
      "Country and period: the result must hold in both GB and IRE and in both halves of the discovery window, because the jumps chase-debut cuts failed on exactly this."
    ],
    "mechanism": "In big fields, a horse whose chance has gone is eased rather than ridden out, and traffic and lost ground add to its losses. The recorded beaten distance therefore overstates how far behind the horse really was, and the overstatement grows with field size. Flat handicaps have the biggest fields, so if the jumps big-field heavy-beating lift (1.187 on 2097) is real, it should replicate here. Eight lengths is used as the Flat equivalent of the jumps 20-length threshold.",
    "market_miss": "Punters and the form book take beaten lengths at face value, with no adjustment for field size. A 10-length beating in a 20-runner handicap reads as poor form, but in an 8-runner race the same run would have been a few lengths. As a result, the horse drifts beyond its true chance next time.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.06 to 1.12 on 1500+ bets, with the lift spread across the price bands up to 14/1",
      "falsified_if": "A/E at or below 1.02 on 1000+ bets; or the lift confined to 10/1 to 14/1; or the same-race other runners showing a similar lift; or the lift failing in either GB or IRE."
    },
    "notes": "I kept the rule sketch as it was. Every column in it is a pre-off catalogue column, and no part of it reads the result. A non-finisher has a NULL last_beaten_lengths, so non-finishers are excluded automatically. The price cap of 15.0 decimal (14/1) follows the finding that big-field lifts sit under 16/1. I raised the bet minimum to 300 because this is a replication with a large expected sample."
  }

Timeline

1. Job 2704 finished

   17 Oct, 03:35 · engine:compute-1

2. H-7358e6fb Discovery → KILLED fails: A/E 0.954, needs 1.05; p 0.7359, needs 0.05 or lower; periods [1.11, 0.91], each needs 1.0; A/E 0.954 against same-race control 1.013, needs 0.03 above (not a niche pass: A/E 0.954 below 1.05; p 0.7359065308960852 above 0.5; does not beat the same-race control 1.013; a period points the other way [1.11, 0.91])

   17 Oct, 03:35 · engine:evaluator

3. discovery result for H-7358e6fb (broad rule): fail

   17 Oct, 03:35 · engine:evaluator

4. H-7358e6fb Designed → RESEARCHdiscovery measuring on the discovery zone

   17 Oct, 03:35 · engine:orchestrator

5. compute-1 took job 2704 (evaluate, attempt 1)

   17 Oct, 03:35 · compute-1

6. Job 2701 finished

   17 Oct, 03:32 · engine:ai-1

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

   17 Oct, 03:32 · engine:scheduler

8. H-7358e6fb Proposed → RESEARCHdesigned pre-spec written and hashed

   17 Oct, 03:32 · agent:designer@v1

9. ai-1 took job 2701 (design, attempt 1)

   17 Oct, 03:31 · ai-1

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

    17 Oct, 03:31 · engine:scheduler

11. New idea H-7358e6fb: Replicate: Flat handicappers heavily beaten in a big field last time beat their price (replicate)

    17 Oct, 03:31 · 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.