Generalise: last-time winners of any race type running in a handicap at 9/1+ are over-bet (winner-tag money)

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-220ad315 (Raised handicap winners sent off at 9/1+ are over-bet, at any gap sinc) pointed the predicted way but failed the bar: A/E 0.84 on 1561 bets (p 0.933), control 1.02. O-76793565
   • H-cae0acb5 (Rival: unraised last-time handicap winners at 9/1+ are over-bet too (w) pointed the predicted way but failed the bar: A/E 0.82 on 144 bets (p 0.741), control 1.02. O-feebd3f3
2. AskedDo all last-time winners running in a handicap at 9/1+ (raised, unraised, or coming from a maiden/novice win) share the same ~0.83 deficit, so the effect belongs to the winner tag rather than the mark? answered
3. MechanismA last-time win draws money from punters who back recent winners and from form-figure systems, even when the horse has drifted to a long price. The drift reflects real negatives: the rise, a class step, conditions, or a lucky last win. The winner's '1' in the form keeps a floor under its price, so it ends up slightly too short.
4. Why the market misses itThe '1' in the form figures is the most visible thing on the card. Recency and tag money shorten these horses relative to what the drift is signalling. Raised and unraised pools both came out at about 0.82–0.84, which fits a cause shared by every last-time winner.
5. Predictednegative, A/E 0.82 to 0.90 on 1,800+ bets
6. Falsified ifA/E 0.95 or higher on 1,000+ bets, or the deficit vanishes once the raised-handicap-winner pool is removed (the non-raised remainder at 0.97+ on 300+ bets).
7. Rule sketch

The question's answer so far: No. The pooled winner-tag population scored 0.88 on 2,634 (p 0.084), weaker than the raised-winner pool's 0.84. The non-handicap and unraised additions dilute the deficit, so the tag alone does not carry the full ~0.83 effect.

Mechanism and rationale

A last-time win draws money from punters who back recent winners and from form-figure systems, even when the horse has drifted to a long price. The drift reflects real negatives: the rise, a class step, conditions, or a lucky last win. The winner's '1' in the form keeps a floor under its price, so it ends up slightly too short.

The rule

last_position = 1 AND is_handicap = 1 AND sp_decimal > 9

Reads: last_position, is_handicap, sp_decimal

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

508341be38201dfcb65792948b1bc4f967918daacf09a7017b55c0e749cb0052

Frozen

not frozen

The definition can still change.

Confounders

• Overlaps heavily with the killed H-220ad315 raised-winner pool (1,561 of the expected ~1,800 bets), so the broad result may simply reproduce its p≈0.08. The non-raised refinement is the real test of the generalisation.
• Favourite-longshot bias at 9/1+: long-priced handicap runners may be over-bet whatever their form. The last-time-loser control at the same prices separates this from the tag effect.
• First-time handicappers after a maiden or conditions win: punters cannot see a handicap mark, so they may be priced differently from handicap winners (checked by the last_is_handicap = 0 refinement).
• Code mix shifts between seasons (the flat-heavy later period scored worse), so a pooled result may reflect when the bets fell rather than the tag.
• Field size: long prices are more common in big fields, where every runner's win chance is low; the same-race control absorbs this.
• Country split: the Irish unraised pool went 0 from 25, so any Irish-driven deficit rests on a tiny sample. Check GB and IRE separately.
• Price-band instability: earlier pools put the deficit in different bands (9/1–14/1 for raised, 16/1+ for unraised), so no band refinement is pre-registered.

Discovery result

fail broad rule · 16 Oct, 16:56

Bets

2,634

118 winners

A/E

0.88

95%: 0.73 to 1.05

SP return

-36.7%

level stakes

Betfair SP return

-13.9%

2,540 bets with a BSP

Strike rate 4.5%. p = 0.93. Same-race control: the other runners in these races had A/E 1.02, 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 · 35s · 16 Oct, 16:53

  Output (schema-checked)

  {
    "rule": "last_position = 1 AND is_handicap = 1 AND sp_decimal > 9",
    "refinements": [
      {
        "label": "Non-raised remainder (falsification check: deficit must survive without the H-220ad315 raised pool)",
        "rule": "last_position = 1 AND is_handicap = 1 AND sp_decimal > 9 AND (rating_change IS NULL OR rating_change <= 0)"
      },
      {
        "label": "Last win came outside a handicap (first-time handicappers / maiden or conditions winners, where the tag is the only visible signal)",
        "rule": "last_position = 1 AND is_handicap = 1 AND sp_decimal > 9 AND last_is_handicap = 0"
      }
    ],
    "control": {
      "description": "Same-race control: the other runners in these handicaps. If the deficit comes from winner-tag money, the other runners in those races should sit at about A/E 1.00–1.03 (earlier pools showed 1.02 and 1.017), so the gap is about the winner, not about long-priced handicap runners in big fields. A second check uses a different group: non-winners last time running in a handicap at the same price. If they show the same deficit, the effect is favourite-longshot bias at 9/1+, not the '1' tag.",
      "rule": "last_position IS NOT NULL AND last_position > 1 AND is_handicap = 1 AND sp_decimal > 9"
    },
    "pass_criteria": {
      "min_bets": 500,
      "min_ae": 1.05,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 200,
      "min_ae": 1,
      "max_p": 0.2
    },
    "confounders": [
      "Overlaps heavily with the killed H-220ad315 raised-winner pool (1,561 of the expected ~1,800 bets), so the broad result may simply reproduce its p≈0.08. The non-raised refinement is the real test of the generalisation.",
      "Favourite-longshot bias at 9/1+: long-priced handicap runners may be over-bet whatever their form. The last-time-loser control at the same prices separates this from the tag effect.",
      "First-time handicappers after a maiden or conditions win: punters cannot see a handicap mark, so they may be priced differently from handicap winners (checked by the last_is_handicap = 0 refinement).",
      "Code mix shifts between seasons (the flat-heavy later period scored worse), so a pooled result may reflect when the bets fell rather than the tag.",
      "Field size: long prices are more common in big fields, where every runner's win chance is low; the same-race control absorbs this.",
      "Country split: the Irish unraised pool went 0 from 25, so any Irish-driven deficit rests on a tiny sample. Check GB and IRE separately.",
      "Price-band instability: earlier pools put the deficit in different bands (9/1–14/1 for raised, 16/1+ for unraised), so no band refinement is pre-registered."
    ],
    "mechanism": "A last-time win draws money from punters who back recent winners and from form-figure systems, even when the horse has drifted to 9/1 or longer. The drift reflects real negatives: a higher mark, a step up in class, changed conditions, or a lucky last win. The '1' in the form keeps a floor under the price, so the horse goes off slightly too short. The effect should come from the tag itself, so it should appear whatever the mark change or the type of the last race.",
    "market_miss": "The '1' in the form figures is the most visible thing on the racecard. Recency bias and system money shorten these horses relative to what the drift is signalling, and the market only partly corrects. Raised (0.84 on 1,561) and unraised (0.82 on 144) winners scored almost the same, which fits a cause shared by every last-time winner rather than one tied to the handicap mark.",
    "prediction": {
      "direction": "negative",
      "size": "A/E 0.82 to 0.90 on 1,800+ bets, with the same-race control near 1.00–1.03",
      "falsified_if": "Any of: A/E 0.95 or higher on 1,000+ bets; the non-raised remainder (rating_change NULL or <= 0) at A/E 0.97+ on 300+ bets; or last-time losers at 9/1+ in handicaps showing a deficit of the same size (which would make it longshot bias, not the winner tag)."
    },
    "notes": "The rule sketch is kept as written. It uses only pre-off columns and does not read the result. For this negative-direction hypothesis, min_ae 1.05 means the inverted bar (A/E ≤ about 0.95, i.e. the lay side beats its prices by 5% or more). min_bets is raised to 500 because the hypothesis predicts more than 1,800 bets, and a thin pass would just repeat H-220ad315. Only two refinements are pre-registered: the non-raised remainder (taken directly from the falsification clause) and the first-time-handicapper slice (from the confounders). No price-band refinement is registered, because earlier findings showed the deficit's band was unstable between pools. The forward bar is raised to 200 bets because the per-bet signal is small."
  }

Timeline

1. Job 2069 finished

   16 Oct, 16:56 · engine:compute-1

2. H-560b6713 Discovery → KILLED fails: p 0.0844 (lower tail), needs 0.05

   16 Oct, 16:56 · engine:evaluator

3. discovery result for H-560b6713 (broad rule): fail

   16 Oct, 16:56 · engine:evaluator

4. H-560b6713 Designed → RESEARCHdiscovery measuring on the discovery zone

   16 Oct, 16:56 · engine:orchestrator

5. compute-1 took job 2069 (evaluate, attempt 1)

   16 Oct, 16:56 · compute-1

6. Job 2068 finished

   16 Oct, 16:54 · engine:ai-1

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

   16 Oct, 16:54 · engine:scheduler

8. H-560b6713 Proposed → RESEARCHdesigned pre-spec written and hashed

   16 Oct, 16:54 · agent:designer@v1

9. ai-1 took job 2068 (design, attempt 1)

   16 Oct, 16:53 · ai-1

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

    16 Oct, 16:53 · engine:scheduler

11. New idea H-560b6713: Generalise: last-time winners of any race type running in a handicap at 9/1+ are over-bet (winner-tag money) (generalise)

    16 Oct, 16:53 · 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.