Generalise: short-priced planned moves from yards proven poor at that move, all four moves pooled

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-61d14772 (Negative: short-priced recruits to a strong new yard after a heavy bea) pointed the predicted way but failed the bar: A/E 0.81 on 84 bets (p 0.809), control 1.04. O-282acadf
   • H-432686d7 (Negative: short-priced planned moves from yards proven poor at that mo) pointed the predicted way but failed the bar: A/E 0.87 on 195 bets (p 0.82), control 1.05. O-bc2c08d8
2. AskedPooled across all planned moves (chase debut, handicap debut, new yard, fresh), are short-priced runners from yards with a long, poor record at that move over-bet? answered
3. MechanismA planned move (fresh return, new yard, handicap or chase debut) depends on the yard's preparation. Yards that keep getting it wrong do so for lasting reasons: method, facilities, how they place horses. That same handling is behind today's runner too.
4. Why the market misses itA short price on a debutant or a returner comes from the horse's own form and reputation. Punters rarely look at a yard's long record with that specific move, and the yard's headline strike rate hides it.
5. Predictednegative, A/E 0.85 to 0.93 on 400+ bets
6. Falsified ifA/E at or above 0.97 on 350+ bets, or no better than the same-race control
7. Rule sketch

Mechanism and rationale

A planned move (fresh return, new yard, handicap or chase debut) depends on the yard's preparation. Yards that keep getting it wrong do so for lasting reasons: method, facilities, how they place horses. That same handling is behind today's runner too.

The rule

sp_decimal <= 6 AND ((tx_chase_debut_runs >= 20 AND tx_chase_debut_wins <= 0.7 * tx_chase_debut_exp) OR (tx_hcap_debut_runs >= 25 AND tx_hcap_debut_wins <= 0.7 * tx_hcap_debut_exp) OR (tx_new_yard_runs >= 25 AND tx_new_yard_wins <= 0.7 * tx_new_yard_exp) OR (tx_fresh_runs >= 30 AND tx_fresh_wins <= 0.7 * tx_fresh_exp))

Reads: sp_decimal, tx_chase_debut_runs, tx_chase_debut_wins, tx_chase_debut_exp, tx_hcap_debut_runs, tx_hcap_debut_wins, tx_hcap_debut_exp, tx_new_yard_runs, tx_new_yard_wins, tx_new_yard_exp, tx_fresh_runs, tx_fresh_wins, tx_fresh_exp

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

b4401aa909f5282c4b424b5584466b204375b8ffef7974b77ac69b7ddc789816

Frozen

17 Oct, 05:13

Nothing in the definition can change after this.

Confounders

• Moves overlap: one runner can be fresh and on its handicap debut, or in a new yard and returning, so it can qualify under several clauses at once. It counts once as a bet, but the clauses are not independent evidence.
• Short-priced debutants and returners may be over-bet whatever the yard. The good-record control rule tests this.
• A few large yards with many runs may dominate the poor-record group, so the result may reflect a handful of trainers. Check the yard concentration.
• Fresh-run records only start in 2021-03 and the wins columns start a little later, so the fresh clause adds runners only later in the window.
• GB and IRE may behave differently (the new-yard split was 0.81 vs 1.35). The GB-only refinement deals with this. The column is region; country does not exist.
• The 0.7 x expected cut on a sample of 20-30 runs partly captures noise. Some poor records will regress, which pushes the result towards 1.0.

Control

{
  "description": "Primary: the other runners in the same races, which should come in near A/E 1.0 (previous short-priced planned-move controls were about 1.05). The rule has to sit clearly below them. Second check, for the confounder that short-priced debutants and returners are over-bet whatever the yard: short-priced runners making one of the same four moves for yards with an adequate or good record at that move (wins at least 0.9 x expected, same run minimums). These should come in near 1.0. If they are as low as the rule, the effect belongs to short-priced planned moves in general, not to the yard.",
  "rule": "sp_decimal <= 6 AND ((tx_chase_debut_runs >= 20 AND tx_chase_debut_wins >= 0.9 * tx_chase_debut_exp) OR (tx_hcap_debut_runs >= 25 AND tx_hcap_debut_wins >= 0.9 * tx_hcap_debut_exp) OR (tx_new_yard_runs >= 25 AND tx_new_yard_wins >= 0.9 * tx_new_yard_exp) OR (tx_fresh_runs >= 30 AND tx_fresh_wins >= 0.9 * tx_fresh_exp))"
}

Discovery result

pass broad rule · 17 Oct, 05:11

Bets

413

75 winners

A/E

0.80

95%: 0.63 to 1.00

SP return

-28.5%

level stakes

Betfair SP return

-16.7%

396 bets with a BSP

Strike rate 18.2%. p = 0.98. Same-race control: the other runners in these races had A/E 1.07, so the selection did not beat its own races. Zone: discovery. Evaluator eval-1.0, data 6a4e98be18e83933

Falsification

Each attack is a deliberate attempt to kill the idea. 6 of 6 survived.

• perturbation: Perturb: looser record cut (0.8x exp), shorter price cap 5.0, run minimums +20%pass

  462 bets · A/E 0.93 (0.76 to 1.11) · control 1.03 · 17 Oct, 05:13

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-273170.87

  after 2026-04-271451.05

• perturbation: Perturb: wider price cap 7.5, run minimums -20%, 0.7x cut keptpass

  630 bets · A/E 0.89 (0.73 to 1.07) · control 1.03 · 17 Oct, 05:13

Sealed holdout (vault)

Bets

1,804

17 Oct, 05:13

A/E

1.02

p = 0.66

BH threshold then

0.0018

over every vault test so far

Outcome

Failed

Forward record

No forward record. Shadow tracking starts after the vault and counts only races run after the freeze.

Lineage

Agent runs

• designer v1ok · 46s · 17 Oct, 05:05

  Output (schema-checked)

  {
    "rule": "sp_decimal <= 6 AND ((race_type = 'Chase' AND prior_chase_runs = 0 AND tx_chase_debut_runs >= 20 AND tx_chase_debut_wins <= 0.7 * tx_chase_debut_exp) OR (is_handicap = 1 AND prior_handicap_runs = 0 AND tx_hcap_debut_runs >= 25 AND tx_hcap_debut_wins <= 0.7 * tx_hcap_debut_exp) OR (trainer_id <> last_trainer_id AND tx_new_yard_runs >= 25 AND tx_new_yard_wins <= 0.7 * tx_new_yard_exp) OR (days_since_last_run >= 90 AND tx_fresh_runs >= 30 AND tx_fresh_wins <= 0.7 * tx_fresh_exp))",
    "refinements": [
      {
        "label": "shortest prices only (4.0 or shorter), where over-betting on reputation is strongest",
        "rule": "sp_decimal <= 4 AND ((race_type = 'Chase' AND prior_chase_runs = 0 AND tx_chase_debut_runs >= 20 AND tx_chase_debut_wins <= 0.7 * tx_chase_debut_exp) OR (is_handicap = 1 AND prior_handicap_runs = 0 AND tx_hcap_debut_runs >= 25 AND tx_hcap_debut_wins <= 0.7 * tx_hcap_debut_exp) OR (trainer_id <> last_trainer_id AND tx_new_yard_runs >= 25 AND tx_new_yard_wins <= 0.7 * tx_new_yard_exp) OR (days_since_last_run >= 90 AND tx_fresh_runs >= 30 AND tx_fresh_wins <= 0.7 * tx_fresh_exp))"
      },
      {
        "label": "long-proven poor yards (double the minimum record), testing the 'lasting reasons' claim",
        "rule": "sp_decimal <= 6 AND ((race_type = 'Chase' AND prior_chase_runs = 0 AND tx_chase_debut_runs >= 40 AND tx_chase_debut_wins <= 0.7 * tx_chase_debut_exp) OR (is_handicap = 1 AND prior_handicap_runs = 0 AND tx_hcap_debut_runs >= 50 AND tx_hcap_debut_wins <= 0.7 * tx_hcap_debut_exp) OR (trainer_id <> last_trainer_id AND tx_new_yard_runs >= 50 AND tx_new_yard_wins <= 0.7 * tx_new_yard_exp) OR (days_since_last_run >= 90 AND tx_fresh_runs >= 60 AND tx_fresh_wins <= 0.7 * tx_fresh_exp))"
      },
      {
        "label": "GB only, since the new-yard finding showed GB and IRE behave differently",
        "rule": "country = 'GB' AND sp_decimal <= 6 AND ((race_type = 'Chase' AND prior_chase_runs = 0 AND tx_chase_debut_runs >= 20 AND tx_chase_debut_wins <= 0.7 * tx_chase_debut_exp) OR (is_handicap = 1 AND prior_handicap_runs = 0 AND tx_hcap_debut_runs >= 25 AND tx_hcap_debut_wins <= 0.7 * tx_hcap_debut_exp) OR (trainer_id <> last_trainer_id AND tx_new_yard_runs >= 25 AND tx_new_yard_wins <= 0.7 * tx_new_yard_exp) OR (days_since_last_run >= 90 AND tx_fresh_runs >= 30 AND tx_fresh_wins <= 0.7 * tx_fresh_exp))"
      }
    ],
    "control": {
      "description": "I expect the other runners in the same races to sit near A/E 1.0 to 1.05, as they did in the two earlier negative rules. That control does not answer the main confounder, which is that short-priced debutants and returners may be over-bet whatever the yard. So the second control is short-priced runners making the same four moves from yards with an adequate or better record at that move (wins >= expected on the same minimum sample). The prediction holds only if the poor-yard group falls clearly below this group as well as below the same-race runners.",
      "rule": "sp_decimal <= 6 AND ((race_type = 'Chase' AND prior_chase_runs = 0 AND tx_chase_debut_runs >= 20 AND tx_chase_debut_wins >= tx_chase_debut_exp) OR (is_handicap = 1 AND prior_handicap_runs = 0 AND tx_hcap_debut_runs >= 25 AND tx_hcap_debut_wins >= tx_hcap_debut_exp) OR (trainer_id <> last_trainer_id AND tx_new_yard_runs >= 25 AND tx_new_yard_wins >= tx_new_yard_exp) OR (days_since_last_run >= 90 AND tx_fresh_runs >= 30 AND tx_fresh_wins >= tx_fresh_exp))"
    },
    "pass_criteria": {
      "min_bets": 300,
      "min_ae": 1.05,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 150,
      "min_ae": 1,
      "max_p": 0.2
    },
    "confounders": [
      "Moves overlap: one runner can be fresh, with a new yard and on its handicap debut at once. It counts as one bet, but its yard may be poor at only one of those moves.",
      "Short-priced debutants and returners may be over-bet whatever the yard. The good-yard control rule is there to test this.",
      "A few large yards may make up most of the poor-record group. Check how concentrated the yards are before reading the result as a general effect.",
      "Fresh-run yard records start only in 2021-03, and wins in 2021-04, so the fresh branch covers a shorter window than the other three.",
      "The 0.7 x expected threshold on a 20 to 30 run record partly picks yards on noise, so some regression to the mean is built in. If it is noise, it would push A/E toward 1.0, not below it.",
      "GB and IRE behaved differently in the new-yard finding (GB 0.81, IRE 1.35), so pooling the two countries could hide or blur the effect."
    ],
    "mechanism": "A planned move (fresh return, new yard, handicap debut or chase debut) depends on how the yard prepares and places the horse. Yards that keep getting a particular move wrong do so for lasting reasons: method, facilities, placing. Today's runner gets the same handling, so its chance is lower than its form suggests. The two narrower tests both pointed down (chase/handicap debut 0.87 on 195, recruits to a strong yard after a beating 0.81 on 84) but neither had enough bets on its own. Pooling the four moves is meant to get a sample big enough to settle it.",
    "market_miss": "A short price on a debutant or returner comes from the horse's own form and reputation. Punters rarely look up a yard's long record with that specific move, and a decent overall strike rate hides it. So the price does not account for how badly this yard handles this move.",
    "prediction": {
      "direction": "negative",
      "size": "A/E 0.85 to 0.93 on 400+ bets, at least 0.07 below the same-race control and below the good-yard control",
      "falsified_if": "A/E at or above 0.97 on 350+ bets, or no lower than the same-race control, or no lower than short-priced runners making the same moves from yards with an adequate record (the good-yard control)"
    },
    "notes": "I changed the sketch in one way. Each branch now requires the runner to actually be making that move: chase debut is race_type = 'Chase' AND prior_chase_runs = 0; handicap debut is is_handicap = 1 AND prior_handicap_runs = 0; new yard is trainer_id <> last_trainer_id; fresh is days_since_last_run >= 90, the same cut-off as the family's fresh rules. Without these conditions a non-null tx_* value alone could select runners that are not making the move. If the tx_* columns are already only filled in for runners making the move, the conditions change nothing. Thresholds are as in the sketch. The direction is negative, so min_ae 1.05 means A/E at or below about 0.952 (1/1.05). I raised min_bets to 300 because the point of pooling is a sample big enough to decide the question."
  }

Timeline

1. Job 3177 finished

   17 Oct, 05:13 · engine:compute-1

2. H-a6871a63 Vault → KILLEDfailed the vault vault CONTRADICTED: A/E 1.019 on 1804 bets, p 0.6626 against BH threshold 0.001754 over 57 vault tests

   17 Oct, 05:13 · engine:orchestrator

3. vault result for H-a6871a63 (vault query (contradicted)): fail

   17 Oct, 05:13 · engine:orchestrator

4. H-a6871a63 Frozen → FROZENvault vault query started

   17 Oct, 05:13 · engine:orchestrator

5. compute-1 took job 3177 (vault, attempt 1)

   17 Oct, 05:13 · compute-1

6. Job 3165 finished

   17 Oct, 05:13 · engine:compute-1

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

   17 Oct, 05:13 · engine:scheduler

8. H-a6871a63 Falsification → FROZEN survived falsification

   17 Oct, 05:13 · engine:falsification

9. falsification result for H-a6871a63 (partition: Partition: handicaps only, where form-based pricing should be most exposed): pass

   17 Oct, 05:13 · engine:evaluator

10. compute-1 took job 3165 (attack, attempt 1)

    17 Oct, 05:13 · compute-1

11. Job 3164 finished

    17 Oct, 05:13 · engine:compute-1

12. falsification result for H-a6871a63 (price_band: Price band: exclude the shortest end, SP 3.0 to 6.0 only): pass

    17 Oct, 05:13 · engine:evaluator

13. compute-1 took job 3164 (attack, attempt 1)

    17 Oct, 05:13 · compute-1

14. Job 3163 finished

    17 Oct, 05:13 · engine:compute-1

15. falsification result for H-a6871a63 (partition: Partition: fresh-return and new-yard clauses only (debut clauses removed)): pass

    17 Oct, 05:13 · engine:evaluator

16. compute-1 took job 3163 (attack, attempt 1)

    17 Oct, 05:13 · compute-1

17. Job 3162 finished

    17 Oct, 05:13 · engine:compute-1

18. falsification result for H-a6871a63 (control: Control: short-priced planned moves with yard record ignored (run minimums only)): pass

    17 Oct, 05:13 · engine:evaluator

19. compute-1 took job 3162 (attack, attempt 1)

    17 Oct, 05:13 · compute-1