Sharpen: recent winners beaten 20+ lengths next time over jumps

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-332b49e9 (Rival: the lift is a one-bad-run rebound after close recent form, need) passed discovery: A/E 1.23 on 1600 bets (p 0.004), same-race control 0.98. O-10a4d6e5
2. AskedAmong jumps handicappers beaten 20+ lengths, do recent winners (best_last3_beaten = 0) beat their price more than horses with a close placed finish, and why does this reverse the parent niche? answered
3. MechanismA recent win is the clearest proof of current ability. The heavy beating that followed often came after a rise in the weights or class and a hard race, which is a temporary setback rather than a loss of form, so the horse recovers to near its winning level.
4. Why the market misses itPunters read a heavy defeat straight after a win as being found out by the handicapper, and let the horse drift, even though the winning form is only weeks old.
5. Predictedpositive, A/E 1.25 to 1.45 on 500+ bets, in both GB and IRE
6. Falsified ifA/E at or below 1.15, or GB below 1.05, or no higher than the placed-but-not-won group (1.152)
7. Rule sketchis_handicap = 1 AND race_type IN ('Hurdle', 'Chase') AND last_beaten_lengths >= 20 AND best_last3_beaten = 0

Mechanism and rationale

A recent win is the clearest proof of current ability. The heavy beating that followed often came after a rise in the weights or class and a hard race, which is a temporary setback rather than a loss of form, so the horse recovers to near its winning level.

The rule

is_handicap = 1 AND race_type IN ('Hurdle', 'Chase') AND last_beaten_lengths >= 20 AND best_last3_beaten = 0

Reads: is_handicap, race_type, last_beaten_lengths, best_last3_beaten

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

e9feb21517271af3dca9790539c4ba82b33c7847c28e477552513c52acd4151f

Frozen

17 Oct, 04:22

Nothing in the definition can change after this.

Confounders

• Partly post hoc: the winner subset was found by splitting the close-finish rule (1.232 on 1600 vs 1.152 on 931 without winners). The discovery result overlaps the data that suggested it.
• Opposite pattern inside the never-involved big-field niche, where excluding recent winners raised A/E. The effect may be confined to horses that showed something in the heavy defeat.
• Rating rises after the win can be large. A heavy defeat after a big rise may mean the horse is genuinely out of its depth, not having a temporary setback.
• Longshot-heavy price profile: a few big-priced winners can inflate A/E on a short window.
• Short data window (both key columns start 2025-11-01), so the sample is mostly one jumps season and could reflect seasonal or going effects.
• GB/IRE split: the earlier non-winner lift was mostly IRE (GB 1.031), so a pooled result could hide a one-region effect.

Control

{
  "description": "Same-race rivals should score about 1.0, which shows the lift belongs to the runner and not to the kind of race. The question asks whether winners carry the close-finish rebound beyond horses that were only placed, so the main comparison is the placed-but-not-won group. These are jumps handicappers beaten 20+ lengths last time whose best run in the last three was beaten more than 0 and up to 6 lengths, previously about 1.152 on 931. Recent winners must clearly beat that group. A/E no higher than about 1.15 means a win adds nothing beyond a close finish.",
  "rule": "is_handicap = 1 AND race_type IN ('Hurdle', 'Chase') AND last_beaten_lengths >= 20 AND best_last3_beaten > 0 AND best_last3_beaten <= 6"
}

Discovery result

pass broad rule · 17 Oct, 04:14

Bets

669

84 winners

A/E

1.34

95%: 1.07 to 1.65

SP return

+23.3%

level stakes

Betfair SP return

+72.6%

667 bets with a BSP

Strike rate 12.6%. p = 0.0063. Same-race control: the other runners in these races had A/E 0.97, 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 5 survived.

• perturbation: Perturb heavy-beating cut down to 16 lengthspass

  879 bets · A/E 1.23 (1.01 to 1.48) · control 0.98 · 17 Oct, 04:22

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-274911.30

  after 2026-04-273881.14

• perturbation: Perturb heavy-beating cut up to 25 lengthspass

  496 bets · A/E 1.33 (1.01 to 1.72) · control 0.98 · 17 Oct, 04:22

Sealed holdout (vault)

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

Forward record

• forward vs bar {'min_bets': 100, 'min_ae': 1.05, 'max_p': 0.2}info

  0 bets · · 17 Oct, 04:24

  Breakdown

  betsA/E0.71.01.3

  forward0–

Other experiments

• neighbour: >= 20 -> 16 (-20%)pass

  879 bets · A/E 1.23 (1.01 to 1.48) · control 0.98 · 17 Oct, 04:14

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-274911.30

  after 2026-04-273881.14

• neighbour: >= 20 -> 24 (+20%)pass

  528 bets · A/E 1.36 (1.04 to 1.74) · control 0.97 · 17 Oct, 04:14

  Breakdown

Lineage

Agent runs

• designer v1ok · 32s · 17 Oct, 04:13

  Output (schema-checked)

  {
    "rule": "is_handicap = 1 AND race_type IN ('Hurdle', 'Chase') AND last_beaten_lengths >= 20 AND best_last3_beaten = 0",
    "refinements": [
      {
        "label": "Win is recent: back within 60 days",
        "rule": "is_handicap = 1 AND race_type IN ('Hurdle', 'Chase') AND last_beaten_lengths >= 20 AND best_last3_beaten = 0 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 60"
      },
      {
        "label": "Not a longshot effect: 14/1 or shorter",
        "rule": "is_handicap = 1 AND race_type IN ('Hurdle', 'Chase') AND last_beaten_lengths >= 20 AND best_last3_beaten = 0 AND sp_decimal <= 15"
      },
      {
        "label": "Outside the never-involved niche, where the winner pattern reversed",
        "rule": "is_handicap = 1 AND race_type IN ('Hurdle', 'Chase') AND last_beaten_lengths >= 20 AND best_last3_beaten = 0 AND last_never_involved = 0"
      }
    ],
    "control": {
      "description": "Same-race rivals should score about 1.0, which shows the lift belongs to the runner and not to the kind of race. The question asks whether winners carry the close-finish rebound beyond horses that were only placed, so the main comparison is the placed-but-not-won group. These are jumps handicappers beaten 20+ lengths last time whose best run in the last three was beaten more than 0 and up to 6 lengths, previously about 1.152 on 931. Recent winners must clearly beat that group. A/E no higher than about 1.15 means a win adds nothing beyond a close finish.",
      "rule": "is_handicap = 1 AND race_type IN ('Hurdle', 'Chase') AND last_beaten_lengths >= 20 AND best_last3_beaten > 0 AND best_last3_beaten <= 6"
    },
    "pass_criteria": {
      "min_bets": 200,
      "min_ae": 1.1,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 100,
      "min_ae": 1.05,
      "max_p": 0.2
    },
    "confounders": [
      "Partly post hoc: the winner subset was found by splitting the close-finish rule (1.232 on 1600 vs 1.152 on 931 without winners). The discovery result overlaps the data that suggested it.",
      "Opposite pattern inside the never-involved big-field niche, where excluding recent winners raised A/E. The effect may be confined to horses that showed something in the heavy defeat.",
      "Rating rises after the win can be large. A heavy defeat after a big rise may mean the horse is genuinely out of its depth, not having a temporary setback.",
      "Longshot-heavy price profile: a few big-priced winners can inflate A/E on a short window.",
      "Short data window (both key columns start 2025-11-01), so the sample is mostly one jumps season and could reflect seasonal or going effects.",
      "GB/IRE split: the earlier non-winner lift was mostly IRE (GB 1.031), so a pooled result could hide a one-region effect."
    ],
    "mechanism": "A win in the last three runs is direct proof of recent ability at this level of code and handicap. A defeat by 20+ lengths straight afterwards usually follows a rise in the weights or class, a hard race, or an unsuitable trip or ground. That is a temporary setback, not lost form, so the horse goes back to near its winning level once conditions or the mark settle.",
    "market_miss": "Punters and form-readers weight the latest run most. They read a heavy defeat straight after a win as the horse being found out by the handicapper and let it drift, discounting winning form that is only weeks old.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.25 to 1.45 on 500+ bets, with both GB and IRE above 1.05",
      "falsified_if": "Falsified if pooled A/E is at or below 1.15, if GB A/E is below 1.05, or if the result is no higher than the placed-but-not-won control (best_last3_beaten above 0 and up to 6, about 1.152)."
    },
    "notes": "best_last3_beaten = 0 means a win in one of the last three runs. Because the last run was beaten 20+ lengths, that win must be the second or third run back. The rule sketch is kept unchanged: it uses only pre-off columns and reads no results. The pass bar is stricter than the floor (200 bets, A/E 1.1) because the hypothesis is partly post hoc and has to beat the 1.152 control. Implied size from earlier splits: about 670 winner bets at roughly 1.33."
  }

Timeline

1. Job 2900 finished

   17 Oct, 04:24 · engine:compute-1

2. forward result for H-363890c3 (forward vs bar {'min_bets': 100, 'min_ae': 1.05, 'max_p': 0.2}): info

   17 Oct, 04:24 · engine:evaluator

3. compute-1 took job 2900 (forward_check, attempt 1)

   17 Oct, 04:24 · compute-1

4. Queued forward_check on the compute lane: daily forward check of a shadow system

   17 Oct, 04:24 · engine:scheduler

5. Job 2868 finished

   17 Oct, 04:22 · engine:compute-1

6. H-363890c3 Frozen → SHADOW vault skipped: last_beaten_lengths, best_last3_beaten do not exist before the discovery zone; track it forward instead

   17 Oct, 04:22 · engine:orchestrator

7. compute-1 took job 2868 (vault, attempt 1)

   17 Oct, 04:22 · compute-1

8. Job 2861 finished

   17 Oct, 04:22 · engine:compute-1

9. Queued vault on the compute lane: frozen: one sealed vault query

   17 Oct, 04:22 · engine:scheduler

10. H-363890c3 Falsification → FROZEN survived falsification

    17 Oct, 04:22 · engine:falsification

11. falsification result for H-363890c3 (partition: Outside never-involved niche where the winner pattern reversed): pass

    17 Oct, 04:22 · engine:evaluator

12. compute-1 took job 2861 (attack, attempt 1)

    17 Oct, 04:22 · compute-1

13. Job 2860 finished

    17 Oct, 04:22 · engine:compute-1

14. falsification result for H-363890c3 (control: Control: placed but not won in last three (beaten >0 to 6 lengths)): pass

    17 Oct, 04:22 · engine:evaluator

15. compute-1 took job 2860 (attack, attempt 1)

    17 Oct, 04:22 · compute-1

16. Job 2859 finished

    17 Oct, 04:22 · engine:compute-1

17. falsification result for H-363890c3 (price_band: Not a longshot effect: 14/1 or shorter): pass

    17 Oct, 04:22 · engine:evaluator

18. compute-1 took job 2859 (attack, attempt 1)

    17 Oct, 04:22 · compute-1

19. Job 2858 finished

    17 Oct, 04:22 · engine:compute-1

20. falsification result for H-363890c3 (perturbation: Perturb heavy-beating cut up to 25 lengths): pass

    17 Oct, 04:22 · engine:evaluator