Generalise: quick-returning last-time winners at 9/1+ in any race are over-bet

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-9e92ea71 (Quick-returning last-time winners of any race running in a handicap at) pointed the predicted way but failed the bar: A/E 0.65 on 424 bets (p 0.97), control 1.02. O-3e924322
   • H-4f18f696 (Replicate: quick-returning last-time winners at 9/1+ in non-handicaps ) pointed the predicted way but failed the bar: A/E 0.46 on 106 bets (p 0.932), control 1.03. O-83cf6feb
2. AskedDo quick-returning last-time winners at 9/1+ fall short in non-handicaps too, where the winner tag is equally visible but no mark rise applies? testing
3. MechanismA fresh win draws attention, tipster lines and stable confidence, so money comes for the horse whatever the race conditions. When the horse still drifts to 9/1+, the market's real view is weak, but the tag-driven money still holds the price below its true odds. Neither the handicap mark nor the quick return explains this on its own: unraised winners scored 0.82 and quick-returning non-winners 0.99.
4. Why the market misses itThe form book shows a '1' as the most recent figure, and casual money backs it without checking how good the win really was. The horse's genuine chance is then priced too short.
5. Predictednegative, A/E 0.65 to 0.85 on 500+ bets
6. Falsified ifA/E 0.92 or higher on 450+ bets, or the deficit present in only one of handicaps and non-handicaps with the other at par

Mechanism and rationale

A fresh win draws attention, tipster lines and stable confidence, so money comes for the horse whatever the race conditions. When the horse still drifts to 9/1+, the market's real view is weak, but the tag-driven money still holds the price below its true odds. Neither the handicap mark nor the quick return explains this on its own: unraised winners scored 0.82 and quick-returning non-winners 0.99.

The rule

last_position = 1 AND sp_decimal > 9 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 14

Reads: last_position, sp_decimal, days_since_last_run

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

e854c04b9ada8a0719489585b91775fa23d2500219d481e34f0765696ebd1d92

Frozen

not frozen

The definition can still change.

Confounders

• Overlap with the earlier discovery: about 424 of the roughly 530 bets are the handicap pool already measured at 0.651 (p 0.0511). The broad result is largely a re-read of that pool, so the non-handicap leg, the falsification work and the forward record have to carry the independent evidence.
• Small win count: with 9/1+ prices, one or two big-priced winners can swing A/E. The non-handicap pool had 2 winners and a BSP ROI of +19.2, so its result can say only which way it points, not how big the effect is.
• The code mix (Flat against jumps) and the GB/IRE mix differ between periods. The deficit has to show in both codes and both regions, not in only one.
• Handicap winners turned out quickly carry a penalty before they are reassessed. That is a weight effect inside handicaps, separate from the tag-money mechanism.
• Class rises after a win (stepping-up winners have scored 1.16) may pull non-handicap or class-rise sub-pools towards par and hide the effect.
• Ordinary longshot under-performance at 16/1+ could pass for a winner-tag effect. The control and the up-to-14/1 refinement address this.

Discovery result

pass broad rule · 16 Oct, 17:03

Bets

530

17 winners

A/E

0.62

95%: 0.36 to 0.99

SP return

-53.2%

level stakes

Betfair SP return

-32.7%

494 bets with a BSP

Strike rate 3.2%. p = 0.99. Same-race control: the other runners in these races had A/E 1.03, 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.

Other experiments

• neighbour: > 9 -> 7 (-20%)fail

  759 bets · A/E 0.68 (0.47 to 0.94) · control 1.03 · 16 Oct, 17:04

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-273020.72

  after 2026-04-274570.65

• neighbour: > 9 -> 11 (+20%)fail

  381 bets · A/E 0.56 (0.26 to 1.06) · control 1.02 · 16 Oct, 17:04

  Breakdown

Lineage

Agent runs

• designer v1ok · 37s · 16 Oct, 17:03

  Output (schema-checked)

  {
    "rule": "last_position = 1 AND sp_decimal > 9 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 14",
    "refinements": [
      {
        "label": "Up to 14/1 (removes the thin longshot tail that a single big winner can swing)",
        "rule": "last_position = 1 AND sp_decimal > 9 AND sp_decimal <= 15 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 14"
      },
      {
        "label": "Non-handicaps only (the generalisation leg: does the deficit exist outside handicaps?)",
        "rule": "last_position = 1 AND is_handicap = 0 AND sp_decimal > 9 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 14"
      },
      {
        "label": "Very quick return, 7 days or fewer (the dose-response check: tag money should be strongest when the win is freshest)",
        "rule": "last_position = 1 AND sp_decimal > 9 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 7"
      }
    ],
    "control": {
      "description": "Same-race controls (the other runners in these races) should come out near par, about A/E 0.98 to 1.03, as they did for the handicap (1.02) and non-handicap (1.03) pools. That would show the shortfall belongs to the horse and not to the races it runs in. The rival explanation, that quick returners at 9/1+ under-perform whatever their last result, has already been tested: quick-returning handicap non-winners at 9/1+ scored 0.989 on 7,481. A non-winner comparison group in all race types is still useful, so it is given as an explicit control. Expect A/E around 0.97 to 1.02. If it also comes out at 0.85 or lower, the winner tag is not the mechanism.",
      "rule": "last_position IS NOT NULL AND last_position > 1 AND sp_decimal > 9 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 14"
    },
    "pass_criteria": {
      "min_bets": 400,
      "min_ae": 1.05,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 150,
      "min_ae": 1,
      "max_p": 0.2
    },
    "confounders": [
      "Overlap with the earlier discovery: about 424 of the roughly 530 bets are the handicap pool already measured at 0.651 (p 0.0511). The broad result is largely a re-read of that pool, so the non-handicap leg, the falsification work and the forward record have to carry the independent evidence.",
      "Small win count: with 9/1+ prices, one or two big-priced winners can swing A/E. The non-handicap pool had 2 winners and a BSP ROI of +19.2, so its result can say only which way it points, not how big the effect is.",
      "The code mix (Flat against jumps) and the GB/IRE mix differ between periods. The deficit has to show in both codes and both regions, not in only one.",
      "Handicap winners turned out quickly carry a penalty before they are reassessed. That is a weight effect inside handicaps, separate from the tag-money mechanism.",
      "Class rises after a win (stepping-up winners have scored 1.16) may pull non-handicap or class-rise sub-pools towards par and hide the effect.",
      "Ordinary longshot under-performance at 16/1+ could pass for a winner-tag effect. The control and the up-to-14/1 refinement address this."
    ],
    "mechanism": "A fresh win draws attention: a '1' as the latest figure, tipster lines and stable confidence. When the horse runs again within 14 days, that win is the most visible thing on its form, so tag-driven money comes in whatever the race type or whether it is a handicap. If the horse still drifts to 9/1+, the informed market rates it weakly, but the residual tag money keeps the price shorter than its true chance. The effect should need both the win and the quick return. Mark changes alone (unraised winners 0.82, not significant) and quick returns alone (non-winners 0.99) do not produce it.",
    "market_miss": "Casual money backs the latest '1' without checking how good the win was or whether a quick turnout suits the horse. Because the horse has drifted to 9/1+, it gets less scrutiny from sharper money. Sharp layers take little interest in shortening a longshot's price, so the overbet price survives to the off.",
    "prediction": {
      "direction": "negative",
      "size": "A/E 0.65 to 0.85 on roughly 500+ bets, with same-race controls near par",
      "falsified_if": "Any of: A/E 0.92 or higher on 450+ bets; the deficit in only one of handicaps and non-handicaps, with the other at par (A/E 0.95 or higher on at least 100 bets); or the non-winner control also at 0.85 or lower."
    },
    "notes": "The rule sketch is kept as it is. It uses only pre-off columns, does not read the result, and drops the is_handicap condition, which is exactly the generalisation step. min_bets is raised to 400 because the prediction and the falsification test assume about 450 to 530 bets. A smaller pool would mean the coverage is not what the reasoning expects. The forward bar is raised to 150 bets because wins at 9/1+ are rare and 100 forward bets would hold only about 6 to 8 winners. The non-handicap refinement is the main test of whether the deficit generalises, but on its own it will be thin (about 106 bets, 2 winners). Read it only for direction, alongside the handicap pool, and not as a standalone result."
  }

Timeline

1. Job 2091 finished

   16 Oct, 17:04 · engine:compute-1

2. hypothesis note H-0ef44a7d

   16 Oct, 17:04 · engine:perturb

3. perturbation result for H-0ef44a7d (neighbour: drop: sp_decimal > 9): pass

   16 Oct, 17:04 · engine:perturb

4. perturbation result for H-0ef44a7d (neighbour: drop: last_position = 1): fail

   16 Oct, 17:04 · engine:perturb

5. perturbation result for H-0ef44a7d (neighbour: <= 14 -> 17 (+20%)): fail

   16 Oct, 17:04 · engine:perturb

6. perturbation result for H-0ef44a7d (neighbour: <= 14 -> 11 (-20%)): fail

   16 Oct, 17:04 · engine:perturb

7. perturbation result for H-0ef44a7d (neighbour: > 9 -> 11 (+20%)): fail

   16 Oct, 17:04 · engine:perturb

8. perturbation result for H-0ef44a7d (neighbour: > 9 -> 7 (-20%)): fail

   16 Oct, 17:04 · engine:perturb

9. compute-1 took job 2091 (mutate, attempt 1)

   16 Oct, 17:04 · compute-1

10. Queued mutate on the compute lane: parameter stability: do neighbouring thresholds hold?

    16 Oct, 17:04 · engine:scheduler

11. Job 2089 finished

    16 Oct, 17:03 · engine:compute-1

12. Queued falsify on the ai lane: passed discovery: try to kill it

    16 Oct, 17:03 · engine:scheduler

13. H-0ef44a7d Discovery → PROMISING passes: A/E 0.62 on 530 bets, p 0.023, periods [0.36, 0.8], control 1.026

    16 Oct, 17:03 · engine:evaluator

14. refinement result for H-0ef44a7d (control: Same-race controls (the other runners in these races) should come out near par, about A/E 0.98 to 1.03, as they): info

    16 Oct, 17:03 · engine:evaluator

15. refinement result for H-0ef44a7d (refinement: Very quick return, 7 days or fewer (the dose-response check: tag money should be strongest when the win is freshest)): info

    16 Oct, 17:03 · engine:evaluator