Last-time winners that were slow in the race are over-bet at the front of the market

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-6192e937 (Rival: lightly raced quick-returning handicap winners beat their price) was predicted positive but came out the other way: A/E 0.96 on 334 bets (p 0.392 for the opposite direction). O-3c4b3666
   • H-528222df (Raised easy handicap winners turned out within 14 days beat their pric) passed discovery: A/E 1.33 on 231 bets (p 0.013), same-race control 0.91. O-991dc8ad
2. AskedIs a last-time win the figures call slow (top speed in the bottom part of its field) over-bet when the horse is in the top three of the betting next time?
3. MechanismA horse that won while being slower at top speed than most of its field won through the shape of the race: a soft lead, a pace collapse, or rivals who failed to stay. That doesn't happen again on purpose. In a truly run race, a horse with low top speed lacks the engine to repeat. Easy winners showed real quality and beat their prices (O-991dc8ad); this is the opposite case, a win with no physical sign of ability.
4. Why the market misses itThe form book and the result say 'won'. Commentators and the public see a winner and push it into the top three of the betting. The sectional top-speed figure is not in the racecard narrative, so the price follows the visible result and not the figure. This is the lesson 'distrust what was pretty and visible; the figures pay where they disagree with what everyone saw'.
5. Predictednegative, A/E 0.82 to 0.92 on 300+ bets
6. Falsified ifA/E 0.97 or higher on 300+ bets, or A/E no lower than the same-race control.
7. Rule sketchlast_position = 1 AND last_top_speed_pct IS NOT NULL AND last_top_speed_pct <= 0.4 AND sp_rank <= 3

Mechanism and rationale

A horse that won while being slower at top speed than most of its field won through the shape of the race: a soft lead, a pace collapse, or rivals who failed to stay. That doesn't happen again on purpose. In a truly run race, a horse with low top speed lacks the engine to repeat. Easy winners showed real quality and beat their prices (O-991dc8ad); this is the opposite case, a win with no physical sign of ability.

The rule

last_position = 1 AND last_top_speed_pct IS NOT NULL AND last_top_speed_pct <= 0.4 AND sp_rank <= 3

Reads: last_position, last_top_speed_pct, sp_rank

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

609c96a7b58f74dc832a74d7fccee50717641267b597cb6426b6996315641a62

Frozen

not frozen

The definition can still change.

Confounders

• Coverage: last_top_speed_pct exists only from 2023-04 (GB) and 2024-01 (IRE), with gaps. The sample is recent and GB-heavy, so read the result by country and by period.
• Small last-time fields make the pct coarse (in a 4-runner field it can only take a few values). Refinement 1 deals with this.
• Front-runners that won from the front in slowly run races score low on top speed by design. Their low figure may reflect how the race was run, not a lack of engine, so it may not predict a fade. last_led_in_running only exists from 2025-11, too late to use as a filter.
• Code: in staying chases and marathon races top speed matters less, which may dilute the effect. Refinement 2 deals with this.
• Price band: favourites and short prices carry less overround and a different A/E baseline than 3rd-favourites. Compare with the fast-winner control at the same sp_rank, not with the whole population.
• Rise in the weights: handicap winners go up in the weights, which affects every last-time winner. The fast-winner control has the same rise, so it absorbs this.

Discovery result

fail broad rule · 16 Oct, 16:47

Bets

1,399

327 winners

A/E

0.96

95%: 0.86 to 1.07

SP return

-18.6%

level stakes

Betfair SP return

-8.4%

1,365 bets with a BSP

Strike rate 23.4%. p = 0.77. 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 · 49s · 16 Oct, 16:46

  Output (schema-checked)

  {
    "rule": "last_position = 1 AND last_top_speed_pct IS NOT NULL AND last_top_speed_pct <= 0.4 AND sp_rank <= 3",
    "refinements": [
      {
        "label": "real fields last time (8+ runners)",
        "rule": "last_position = 1 AND last_top_speed_pct IS NOT NULL AND last_top_speed_pct <= 0.4 AND sp_rank <= 3 AND last_field_size >= 8"
      },
      {
        "label": "not chases (top speed matters more on the Flat and over hurdles)",
        "rule": "last_position = 1 AND last_top_speed_pct IS NOT NULL AND last_top_speed_pct <= 0.4 AND sp_rank <= 3 AND race_type IN ('Flat', 'Hurdle')"
      },
      {
        "label": "clearly slow (bottom quarter of its field)",
        "rule": "last_position = 1 AND last_top_speed_pct IS NOT NULL AND last_top_speed_pct <= 0.25 AND sp_rank <= 3"
      }
    ],
    "control": {
      "description": "Two comparisons. (1) The standard same-race comparison: the other runners in the same races should score near the race average (A/E about 0.97 to 1.0). The selected horses must score clearly below them, or the effect belongs to the races and not to the horses. (2) A comparison horse with the same story that the market sees: last-time winners in the top three of the betting whose winning run was fast (top speed better than at least 60% of that field). The mechanism predicts these score near or above 1.0, because their win had a physical sign of ability. If fast and slow winners score the same, the top-speed figure adds nothing and the hypothesis fails.",
      "rule": "last_position = 1 AND last_top_speed_pct IS NOT NULL AND last_top_speed_pct >= 0.6 AND sp_rank <= 3"
    },
    "pass_criteria": {
      "min_bets": 300,
      "min_ae": 1.05,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 100,
      "min_ae": 1,
      "max_p": 0.2
    },
    "confounders": [
      "Coverage: last_top_speed_pct exists only from 2023-04 (GB) and 2024-01 (IRE), with gaps. The sample is recent and GB-heavy, so read the result by country and by period.",
      "Small last-time fields make the pct coarse (in a 4-runner field it can only take a few values). Refinement 1 deals with this.",
      "Front-runners that won from the front in slowly run races score low on top speed by design. Their low figure may reflect how the race was run, not a lack of engine, so it may not predict a fade. last_led_in_running only exists from 2025-11, too late to use as a filter.",
      "Code: in staying chases and marathon races top speed matters less, which may dilute the effect. Refinement 2 deals with this.",
      "Price band: favourites and short prices carry less overround and a different A/E baseline than 3rd-favourites. Compare with the fast-winner control at the same sp_rank, not with the whole population.",
      "Rise in the weights: handicap winners go up in the weights, which affects every last-time winner. The fast-winner control has the same rise, so it absorbs this."
    ],
    "mechanism": "A horse that won while its top speed was lower than most of its field's (last_top_speed_pct <= 0.4) won through the shape of the race: a soft lead, a pace collapse, or rivals who failed to stay. That does not repeat on purpose. Next time, in a truly run race, a horse with low top speed lacks the engine to win again. Easy winners showed real quality and beat their prices (O-991dc8ad). This is the opposite case: a win with no physical sign of ability.",
    "market_miss": "The form book and the result say 'won'. Racecard narrative, commentators and the public see a recent winner and push it into the top three of the betting. The RaceiQ sectional top-speed figure is not part of the racecard story, so the price follows the visible result and not the figure. In the lab's terms: distrust what was pretty and visible, because the figures pay where they disagree with what everyone saw.",
    "prediction": {
      "direction": "negative",
      "size": "A/E 0.82 to 0.92 on 300+ bets, and below the fast-winner control (last_top_speed_pct >= 0.6) by at least 0.08",
      "falsified_if": "A/E 0.97 or higher on 300+ bets, or A/E no lower than the same-race other runners, or no lower than the fast-winner control."
    },
    "notes": "The rule sketch is kept as written. Every column is pre-off and was checked in the catalogue: last_top_speed_pct is a 0 to 1 share of the rest of its last field; last_position = 1 is last time's result, not today's. The data window is short (from 2023 for GB and 2024 for IRE), so I raised min_bets to 300 to match the predicted sample. The negative direction is pre-registered: the vault should judge the A/E threshold the inverse way (at or below about 0.95). Read any forward or shadow result by period and by GB/IRE, because the findings show early-period lifts fading and splits between the countries."
  }

Timeline

1. Job 2053 finished

   16 Oct, 16:47 · engine:compute-1

2. H-dd2d6d12 Discovery → KILLED fails: A/E 0.961, needs 0.95 or lower; p 0.2464 (lower tail), needs 0.05

   16 Oct, 16:47 · engine:evaluator

3. discovery result for H-dd2d6d12 (broad rule): fail

   16 Oct, 16:47 · engine:evaluator

4. H-dd2d6d12 Designed → RESEARCHdiscovery measuring on the discovery zone

   16 Oct, 16:47 · engine:orchestrator

5. compute-1 took job 2053 (evaluate, attempt 1)

   16 Oct, 16:47 · compute-1

6. Job 2050 finished

   16 Oct, 16:47 · engine:ai-1

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

   16 Oct, 16:47 · engine:scheduler

8. H-dd2d6d12 Proposed → RESEARCHdesigned pre-spec written and hashed

   16 Oct, 16:47 · agent:designer@v1

9. ai-1 took job 2050 (design, attempt 1)

   16 Oct, 16:46 · ai-1

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

    16 Oct, 16:45 · engine:scheduler

11. New idea H-dd2d6d12: Last-time winners that were slow in the race are over-bet at the front of the market (scout)

    16 Oct, 16:45 · agent:scout@v1

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