May four-year-old returners: the late end of the window carries no edge

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-2d115a58 (Early-spring four-year-old returners: the lift sits in March and April) passed discovery as a niche: A/E 1.29 on 322 bets (p 0.068), same-race control 0.95. O-ed858a4b
2. AskedDo four-year-old returners whose first run back falls in May score at or below their prices, as a setback-selection reading of the early-spring concentration predicts? answered
3. MechanismA lightly raced four-year-old that reappears in March or April has been brought back on schedule after a winter of growing. One that first appears in May has more often had a setback. By May the market has also seen how that crop's four-year-olds are running, so the pre-season doubt has been priced away.
4. Why the market misses itIn March and April, the form book shows only three-year-old runs that are five or more months old, so the market prices the horse on a stale profile. By May there are trends across the crop and stable form to price, and the late reappearance itself is a warning sign that bettors underweight less than the early unknown.
5. Predictednegative, A/E 0.85 to 1.02 on 400+ bets
6. Falsified ifMay returners score A/E 1.10 or more on 400+ bets, meaning the March–April concentration was noise and the full March–May window should stand
7. Rule sketch

Mechanism and rationale

A lightly raced four-year-old that reappears in March or April has been brought back on schedule after a winter of growing. One that first appears in May has more often had a setback. By May the market has also seen how that crop's four-year-olds are running, so the pre-season doubt has been priced away.

The rule

race_type = 'Flat' AND is_handicap = 1 AND month = 5 AND age = 4 AND prior_runs <= 10 AND days_since_last_run >= 150

Reads: race_type, is_handicap, month, age, prior_runs, days_since_last_run

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

1a80ab9fbe35f519634f8a47c04044d3d2b6aecbcee32a90616ff9c4e64d028f

Frozen

not frozen

The definition can still change.

Confounders

• May has more turf meetings and bigger fields than March, so field size and course mix change with the month
• March has more all-weather racing, which changes the surface mix (the turf-only refinement checks this)
• The GB and Irish season calendars differ, and earlier work shows Irish slices are too small to read, so region is not split
• A May returner's 150+ day layoff can have started in late autumn (an all-weather or late-turf run) rather than in summer, which mixes horses on schedule with horses that had a setback
• Overlap with the March–May parent rule (H126): May is the remainder of a window already seen, so the early-spring concentration may partly be selection on noise
• Longshot instability: the 16/1+ band has swung between cuts, so a few big-priced winners can dominate the overall A/E

Control

{
  "description": "Same-race other runners should score about 1.0 (0.95 to 1.05), so any shortfall in May returners is specific to them and not a property of May handicaps. The key comparison is the same profile in March and April (H-2d115a58: 1.29 on 322). The hypothesis needs May clearly below that, at or under 1.02. If May matches March and April, the effect is not early-season.",
  "rule": "race_type = 'Flat' AND is_handicap = 1 AND month IN (3, 4) AND age = 4 AND prior_runs <= 10 AND days_since_last_run >= 150"
}

Discovery result

fail broad rule · 16 Oct, 13:15

Bets

197

22 winners

A/E

1.06

95%: 0.66 to 1.60

SP return

-26.0%

level stakes

Betfair SP return

-10.4%

197 bets with a BSP

Strike rate 11.2%. p = 0.43. Same-race control: the other runners in these races had A/E 0.99, so the selection 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 · 30s · 16 Oct, 13:09

  Output (schema-checked)

  {
    "rule": "race_type = 'Flat' AND is_handicap = 1 AND month = 5 AND age = 4 AND prior_runs <= 10 AND days_since_last_run >= 150",
    "refinements": [
      {
        "label": "May returners at 14/1 or shorter (matches the passed fancied-returner rule)",
        "rule": "race_type = 'Flat' AND is_handicap = 1 AND month = 5 AND age = 4 AND prior_runs <= 10 AND days_since_last_run >= 150 AND sp_decimal <= 15"
      },
      {
        "label": "May returners on turf only (removes the March all-weather course-mix confounder)",
        "rule": "race_type = 'Flat' AND is_handicap = 1 AND month = 5 AND age = 4 AND prior_runs <= 10 AND days_since_last_run >= 150 AND surface = 'Turf'"
      },
      {
        "label": "May returners off a long layoff (240+ days, last run by early autumn, so this is a true first run of the season that came late, which is the setback signal)",
        "rule": "race_type = 'Flat' AND is_handicap = 1 AND month = 5 AND age = 4 AND prior_runs <= 10 AND days_since_last_run >= 240"
      }
    ],
    "control": {
      "description": "Same-race other runners should score about 1.0 (0.95 to 1.05), so any shortfall in May returners is specific to them and not a property of May handicaps. The key comparison is the same profile in March and April (H-2d115a58: 1.29 on 322). The hypothesis needs May clearly below that, at or under 1.02. If May matches March and April, the effect is not early-season.",
      "rule": "race_type = 'Flat' AND is_handicap = 1 AND month IN (3, 4) AND age = 4 AND prior_runs <= 10 AND days_since_last_run >= 150"
    },
    "pass_criteria": {
      "min_bets": 150,
      "min_ae": 1.05,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 100,
      "min_ae": 1,
      "max_p": 0.2
    },
    "confounders": [
      "May has more turf meetings and bigger fields than March, so field size and course mix change with the month",
      "March has more all-weather racing, which changes the surface mix (the turf-only refinement checks this)",
      "The GB and Irish season calendars differ, and earlier work shows Irish slices are too small to read, so region is not split",
      "A May returner's 150+ day layoff can have started in late autumn (an all-weather or late-turf run) rather than in summer, which mixes horses on schedule with horses that had a setback",
      "Overlap with the March–May parent rule (H126): May is the remainder of a window already seen, so the early-spring concentration may partly be selection on noise",
      "Longshot instability: the 16/1+ band has swung between cuts, so a few big-priced winners can dominate the overall A/E"
    ],
    "mechanism": "A lightly raced four-year-old that reappears in March or April has been brought back on schedule after a winter of physical development. One that first reappears in May has more often had a setback or a slow return. By May the market has also seen how that crop of four-year-olds is running, so the pre-season doubt that made early returners underpriced has been priced away.",
    "market_miss": "In March and April the only form is three-year-old form five or more months old, so the market prices the horse on a stale, under-informed profile and underrates winter improvement. By May there is crop-wide and stable form to price, and a late reappearance is a warning sign. So the market's error is gone or reversed, and May returners should be priced at or above their true chance.",
    "prediction": {
      "direction": "negative",
      "size": "A/E about 0.85 to 1.02 on roughly 150 to 250 bets (May is about a third of the March–May window), clearly below the March–April 1.29",
      "falsified_if": "May returners score A/E 1.10 or more on 150 or more bets. That would mean the March–April concentration was noise and the full March–May window should stand."
    },
    "notes": "The rule is kept as sketched. All columns are pre-off, and month is already used across this family. The '400+ bets' in the hypothesis's prediction is unrealistic: March–April alone gave 322 of the window, so May likely yields about 150 to 250. The size is restated to match, and the falsification threshold is kept at A/E 1.10. Refinements are restricted to the 14/1 cap from the passed parent rule, turf only for the surface confounder, and a 240+ day layoff to isolate the setback signal. No region split, because Irish samples have proven unreadable in this thread."
  }

Timeline

1. Job 1547 finished

   16 Oct, 13:15 · engine:compute-1

2. H-cbe10008 Discovery → KILLED fails: 197 bets, needs 300; A/E 1.057, needs 0.95 or lower; p 0.6551 (lower tail), needs 0.05; periods [1.06], each needs 0.99 or lower; A/E 1.057 against same-race control 0.993, needs 0.03 below

   16 Oct, 13:15 · engine:evaluator

3. discovery result for H-cbe10008 (broad rule): fail

   16 Oct, 13:15 · engine:evaluator

4. H-cbe10008 Designed → RESEARCHdiscovery measuring on the discovery zone

   16 Oct, 13:15 · engine:orchestrator

5. compute-1 took job 1547 (evaluate, attempt 1)

   16 Oct, 13:15 · compute-1

6. Job 1545 finished

   16 Oct, 13:09 · engine:ai-1

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

   16 Oct, 13:09 · engine:scheduler

8. H-cbe10008 Proposed → RESEARCHdesigned pre-spec written and hashed

   16 Oct, 13:09 · agent:designer@v1

9. ai-1 took job 1545 (design, attempt 1)

   16 Oct, 13:09 · ai-1

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

    16 Oct, 13:09 · engine:scheduler

11. New idea H-cbe10008: May four-year-old returners: the late end of the window carries no edge (test-mechanism)

    16 Oct, 13:09 · 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.