Rival: the chase unseat lift is a jockey-change effect

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-9f59a0ed (Sharpen: chase unseaters with no jumping error recorded last time) is unstable: 2 of 4 neighbouring rules lose the effect (base A/E 1.41; e.g. drop: race_type = 'Chase' -> 0.9; drop: last_non_finish_code = 'UR' -> 0.99). O-4e460c4e
   • H-a44d6bdc (Rival: the chase unseater lift is a long-price artefact; test unseater) passed discovery as a niche: A/E 1.35 on 122 bets (p 0.084), same-race control 0.94. O-c1cf83dc
2. AskedIs the chase unseat lift carried by runs where the rider has changed since the unseat, rather than by the horse's hidden ability? answered
3. MechanismTrainers often replace the rider who came off with a stronger or more experienced jockey. That upgrade, not hidden ability, could be what wins the next race.
4. Why the market misses itThe market may mark the horse down for the U without fully crediting the change of rider, especially at mid-range prices.
5. Predictedpositive, A/E 1.40 to 1.60 on 100+ bets if the rival holds
6. Falsified ifSwitched-jockey unseaters score at or below the full pool's 1.26 to 1.40, which would mean the lift does not depend on the change of rider.
7. Rule sketch

Mechanism and rationale

Trainers often replace the rider who came off with a stronger or more experienced jockey. That upgrade, not hidden ability, could be what wins the next race.

The rule

race_type = 'Chase' AND last_non_finish_code = 'UR' AND jockey_id <> last_jockey_id

Reads: race_type, last_non_finish_code, jockey_id, last_jockey_id

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

42e3272785e0fb2bc7339b4ad6f8b150720f7077b7e7061b7375949274731d50

Frozen

17 Oct, 03:10

Nothing in the definition can change after this.

Confounders

• Claimers or conditional riders being replaced by senior jockeys: their allowance is lost, so the horse carries more weight. This works against the effect and mixes up 'upgrade' with weight.
• Trainer or yard: big yards switch to their stable jockey more often and may also place horses better after a U.
• Price band: most of the unseater profit sits at 16/1 or bigger (4 wins from 91). A few long-price winners could drive either group.
• Handicap mark: a U leaves the mark unchanged, which helps both groups equally and is not specific to the rider change.
• Last run's rider being unavailable (injured in the fall, or booked elsewhere) instead of sacked. Not every switch is an upgrade.
• Data window: last_non_finish_code only exists from 2025-01-01, so the sample is short and seasonally clustered.

Control

{
  "description": "The decisive comparison is unseaters that keep the same rider. If the rival holds, same-jockey chase unseaters should sit near par (A/E about 1.0 to 1.15) while switched-jockey unseaters carry the lift (1.40+). If both groups show similar lift (both around 1.3 to 1.4), the lift does not depend on the change of rider and the rival fails. Also, the other runners in the same races should be at par (about 0.95 to 1.0), as they were for the full unseater pool (0.94 to 0.96). Rows with a null jockey_id or last_jockey_id fall out of both groups, so the two groups will not add up to the 213-bet full pool. Report the size of the gap and do not compare the groups by subtracting one from the full pool.",
  "rule": "race_type = 'Chase' AND last_non_finish_code = 'UR' AND jockey_id = last_jockey_id"
}

Discovery result

pass broad rule (niche) · 17 Oct, 03:09

Bets

116

14 winners

A/E

1.36

95%: 0.74 to 2.28

SP return

-0.6%

level stakes

Betfair SP return

+47.1%

116 bets with a BSP

Strike rate 12.1%. p = 0.16. Same-race control: the other runners in these races had A/E 0.96, so the selection beat its own races. Zone: discovery. Evaluator eval-1.0, data 6a4e98be18e83933

Falsification

Each attack is a deliberate attempt to kill the idea. 3 of 5 survived.

• control: Same-rider chase unseaters (decisive control)pass

  97 bets · A/E 1.45 (0.81 to 2.39) · control 0.95 · 17 Oct, 03:10

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-27621.67

  after 2026-04-2735–

• partition: Switched rider, under 16/1 only (strip the long-price winners)pass

  59 bets · A/E 1.35 (0.70 to 2.37) · control 0.95 · 17 Oct, 03:10

  Breakdown

Sealed holdout (vault)

Bets

0

17 Oct, 03:10

A/E

–

p = 1.0

BH threshold then

0.0020

over every vault test so far

Outcome

Failed

Forward record

• forward vs bar {'min_bets': 100, 'min_ae': 1.0, 'max_p': 0.2}info

  0 bets · · 17 Oct, 03:10

  Breakdown

  betsA/E0.71.01.3

  forward0–

Other experiments

• neighbour: drop: race_type = 'Chase'pass

  341 bets · A/E 1.14 (0.78 to 1.62) · control 0.98 · 17 Oct, 03:10

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-271631.18

  after 2026-04-271781.11

• neighbour: drop: last_non_finish_code = 'UR'fail

  6,765 bets · A/E 0.99 (0.93 to 1.07) · control 1.01 · 17 Oct, 03:10

  Breakdown

Lineage

Agent runs

• designer v1ok · 37s · 17 Oct, 03:04

  Output (schema-checked)

  {
    "rule": "race_type = 'Chase' AND last_non_finish_code = 'UR' AND jockey_id <> last_jockey_id",
    "refinements": [
      {
        "label": "Mid-range prices (under 16/1), where the market_miss says the rider change is under-credited",
        "rule": "race_type = 'Chase' AND last_non_finish_code = 'UR' AND jockey_id <> last_jockey_id AND sp_decimal < 17"
      },
      {
        "label": "Upgrade to an established, winning jockey (the mechanism's 'stronger or more experienced' rider)",
        "rule": "race_type = 'Chase' AND last_non_finish_code = 'UR' AND jockey_id <> last_jockey_id AND jockey_runs >= 100 AND jockey_wins >= 0.12 * jockey_runs"
      }
    ],
    "control": {
      "description": "The decisive comparison is unseaters that keep the same rider. If the rival holds, same-jockey chase unseaters should sit near par (A/E about 1.0 to 1.15) while switched-jockey unseaters carry the lift (1.40+). If both groups show similar lift (both around 1.3 to 1.4), the lift does not depend on the change of rider and the rival fails. Also, the other runners in the same races should be at par (about 0.95 to 1.0), as they were for the full unseater pool (0.94 to 0.96). Rows with a null jockey_id or last_jockey_id fall out of both groups, so the two groups will not add up to the 213-bet full pool. Report the size of the gap and do not compare the groups by subtracting one from the full pool.",
      "rule": "race_type = 'Chase' AND last_non_finish_code = 'UR' AND jockey_id = last_jockey_id"
    },
    "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": [
      "Claimers or conditional riders being replaced by senior jockeys: their allowance is lost, so the horse carries more weight. This works against the effect and mixes up 'upgrade' with weight.",
      "Trainer or yard: big yards switch to their stable jockey more often and may also place horses better after a U.",
      "Price band: most of the unseater profit sits at 16/1 or bigger (4 wins from 91). A few long-price winners could drive either group.",
      "Handicap mark: a U leaves the mark unchanged, which helps both groups equally and is not specific to the rider change.",
      "Last run's rider being unavailable (injured in the fall, or booked elsewhere) instead of sacked. Not every switch is an upgrade.",
      "Data window: last_non_finish_code only exists from 2025-01-01, so the sample is short and seasonally clustered."
    ],
    "mechanism": "After a chase unseat, trainers often replace the rider who came off with a stronger or more experienced jockey. The rival explanation says this rider upgrade, not hidden ability shown before the unseat, is what produces the next-time lift (A/E 1.40 on 213 for all chase unseaters). If so, the lift should sit in the switched-jockey runs and be absent when the same rider keeps the ride.",
    "market_miss": "The market marks the horse down for the U in the form figures. It may not fully price in that the rider who came off has been replaced by a better one, especially at mid-range prices, where the jockey booking moves the price less than it does for favourites.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.40 to 1.60 on roughly 100 to 150 bets for switched-jockey chase unseaters, with the same-jockey control near 1.0 to 1.15.",
      "falsified_if": "Switched-jockey unseaters score at or below the full pool's 1.26 to 1.40, or the same-jockey control shows a similar lift (within about 0.1 A/E of the switched group). Either would mean the lift does not depend on the change of rider."
    },
    "notes": "I kept the rule sketch as written. jockey_id, last_jockey_id and last_non_finish_code are all pre-off. last_non_finish_code is only available from 2025-01-01, so the full pool is 213 bets and the switched subset will probably fall short of the 150-bet floor. If it points the right way it should go to falsification and forward tracking rather than be killed for size. The test that matters is the gap between the switched group and the same-jockey control. A lift in the switched group alone is not enough. The NULL-safe inequality excludes rows with a missing jockey id. Earlier findings showed that flag subsets in this family do not reconcile, so compare the two groups directly and do not derive either one by subtraction. The jockey-quality cutoff in refinement 2 (100+ runs, 12%+ strike rate) is pre-registered as a rough 'established jockey' threshold and must not be tuned."
  }

Timeline

1. Job 2605 finished

   17 Oct, 03:10 · engine:compute-1

2. hypothesis note H-e2cae105

   17 Oct, 03:10 · engine:perturb

3. perturbation result for H-e2cae105 (neighbour: drop: jockey_id <> last_jockey_id): pass

   17 Oct, 03:10 · engine:perturb

4. perturbation result for H-e2cae105 (neighbour: drop: last_non_finish_code = 'UR'): fail

   17 Oct, 03:10 · engine:perturb

5. perturbation result for H-e2cae105 (neighbour: drop: race_type = 'Chase'): pass

   17 Oct, 03:10 · engine:perturb

6. compute-1 took job 2605 (mutate, attempt 1)

   17 Oct, 03:10 · compute-1

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

   17 Oct, 03:10 · engine:scheduler

8. Job 2600 finished

   17 Oct, 03:10 · engine:compute-1

9. forward result for H-e2cae105 (forward vs bar {'min_bets': 100, 'min_ae': 1.0, 'max_p': 0.2}): info

   17 Oct, 03:10 · engine:evaluator

10. compute-1 took job 2600 (forward_check, attempt 1)

    17 Oct, 03:10 · compute-1

11. Queued forward_check on the compute lane: daily forward check of a shadow system

    17 Oct, 03:10 · engine:scheduler

12. Job 2596 finished

    17 Oct, 03:10 · engine:compute-1

13. H-e2cae105 Vault → SHADOW vault SUPPORTIVE: A/E None on 0 bets, p 1 against BH threshold 0.002 over 50 vault tests

    17 Oct, 03:10 · engine:orchestrator

14. vault result for H-e2cae105 (vault query (supportive)): pass

    17 Oct, 03:10 · engine:orchestrator

15. H-e2cae105 Frozen → FROZENvault vault query started

    17 Oct, 03:10 · engine:orchestrator

16. compute-1 took job 2596 (vault, attempt 1)

    17 Oct, 03:10 · compute-1

17. Job 2589 finished

    17 Oct, 03:10 · engine:compute-1

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

    17 Oct, 03:10 · engine:scheduler