Jumps fallers at 14/1 or shorter beat their price: the market over-penalises a visible fall

KILLEDamberFailed a test or its forward record. Kept: failures are knowledge.

falsified: driven by one price band (up to 3/1): the rest score 0.92

Part of the thread “H98: brought down or unseated” 12 hypotheses

Proposed
Designed
Discovery
Promising
Falsification
Frozen
Vault
Provisional
Shadow
Validated
Live

Source: Investigator agent, v1
Made by: rival-explanation
Created: 16 Oct, 13:36
Tests logged: 16

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.

Observed
- H-9519a139 (Contrast: fell or unseated while NOT going well, at 14/1 or shorter, o) was predicted negative but came out the other way: A/E 1.13 on 367 bets (p 0.199 for the opposite direction). O-d93ddd07
- H-9f384f7c (Fell or unseated while going well: hidden form at normal prices) passed discovery but was falsified: period after 2026-04-27 A/E 0.96; IRE A/E 0.802; driven by one price band (up to 7/1): the rest score 0.97. O-ec1b9459
AskedIs the 14/1-or-shorter lift among jumps fallers and unseaters carried by fallers (last_non_finish_code = 'F') rather than unseaters, given that capped BD/UR scored only 1.021? answered
MechanismA fancied horse that fell usually walks away sound; a single fall seldom reflects a lasting jumping flaw in a horse good enough to be near the front of the market. Trainers run it back once they are satisfied with its schooling.
Why the market misses itAn F in the form figures is the most damning letter on the card. Punters fear a loss of confidence and lay off it, so the price drifts beyond what its ability merits. The unseat letter carries less stigma, which fits capped BD/UR matching its price at 1.021.
Predictedpositive, A/E 1.10 to 1.25 on 200+ bets
Falsified ifA/E below 1.04, or no higher than the capped BD/UR rule's 1.021
Rule sketchlast_non_finish_code = 'F' AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15

The question's answer so far: Not convincingly. Fallers at 14/1 or shorter scored 1.07 on 389 (CI 0.83 to 1.36, GB 1.00) against 1.021 for capped BD/UR. That gap is well inside the noise, so neither departure type clearly carries a short-price lift.

Mechanism and rationale

A fancied horse that fell usually walks away sound; a single fall seldom reflects a lasting jumping flaw in a horse good enough to be near the front of the market. Trainers run it back once they are satisfied with its schooling.

The rule

last_non_finish_code = 'F' AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND race_date >= '2025-01-01'

Reads: last_non_finish_code, race_type, sp_decimal, race_date

Pre-registration

Spec hash (sha256 of the canonical pre-spec): e0c3beaeb8fd88e59aa22109eddab4752e9e8ab3c7bc578e060548b452820781
Frozen: not frozen; The definition can still change.

Confounders: Fallers may run in better-class races than unseaters, so a class effect could pass for a letter effect; check A/E by race_class
Window mismatch: last_non_finish_code distinguishes F only from 2025-01-01 (every non-finisher was PU before), whereas last_fell_or_unseated and last_code use other windows; the rule is pinned to 2025-01-01 onward so all groups are compared on the same window
A few short-priced winners in one period may carry the result; require both halves of the window to be above 1.0
GB-only concentration, as in the going-well results (GB 1.27 vs IRE 0.80); a lift present only in GB with no mechanism for the split counts against the hypothesis
The going-well flag overlaps the faller group; earlier findings show it adds nothing, so any lift should not depend on last_fell_going_well
Chase against hurdle mix: chase fallers may differ from hurdle fallers in price and class
Control: { "description": "Same-race other runners should score about 0.98 to 1.00, showing the lift is not a property of the races fallers return in. The decisive comparison is the rival letter group: jumps BD/UR runners at 14/1 or shorter over the same window, expected near the previously measured 1.021. The hypothesis needs fallers clearly above both.", "rule": "last_non_finish_code IN ('BD', 'UR') AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND race_date >= '2025-01-01'" }

Discovery result

pass broad rule (niche) · 16 Oct, 13:40

Bets: 389; 66 winners
A/E: 1.07; 95%: 0.83 to 1.36
SP return: -24.4%; level stakes
Betfair SP return: -10.2%; 389 bets with a BSP

0.78dashed line 1.00 · interval includes 1.00: could be luck · grey tick: the rest of the field in the same races1.41

Strike rate 17.0%. p = 0.30. 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

By period

Falsification

Each attack is a deliberate attempt to kill the idea. 5 of 6 survived.

perturbation: Price cap tightened to 11/1 (sp <= 12)pass
332 bets · A/E 1.11 (0.86 to 1.42) · control 0.98 · 16 Oct, 13:46
Breakdown
betsA/E0.71.01.3
2025-10-01 to 2026-04-272221.08
after 2026-04-271101.19
Bars grow from 1.00, the market. Faded bars have fewer than 40 bets.
price_band: Fallers outside the short band: 4.0 < sp <= 15fail
309 bets · A/E 0.95 (0.65 to 1.34) · control 1.02 · 16 Oct, 13:46

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.

Other experiments

neighbour: <= 15 -> 12 (-20%)pass
332 bets · A/E 1.11 (0.86 to 1.42) · control 0.98 · 16 Oct, 13:46
Breakdown
betsA/E0.71.01.3
2025-10-01 to 2026-04-272221.08
after 2026-04-271101.19
Bars grow from 1.00, the market. Faded bars have fewer than 40 bets.
neighbour: <= 15 -> 18 (+20%)pass
414 bets · A/E 1.07 (0.83 to 1.36) · control 0.99 · 16 Oct, 13:46
Breakdown

Lineage

H98-brought-down-or-unseatedPROMISINGH98: brought down or unseatedmade by import

H-9f384f7cKILLEDFell or unseated while going well: hidden form at normal pricesmade by test-mechanism

H-9519a139KILLEDContrast: fell or unseated while NOT going well, at 14/1 or shorter, only matches its price

Agent runs

designer v1ok · 37s · 16 Oct, 13:36

Output (schema-checked)

{
  "rule": "last_non_finish_code = 'F' AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND race_date >= '2025-01-01'",
  "refinements": [
    {
      "label": "fancied: top three in the betting",
      "rule": "last_non_finish_code = 'F' AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND race_date >= '2025-01-01' AND sp_rank <= 3"
    },
    {
      "label": "back within 60 days (walked away sound)",
      "rule": "last_non_finish_code = 'F' AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND race_date >= '2025-01-01' AND days_since_last_run <= 60"
    }
  ],
  "control": {
    "description": "Same-race other runners should score about 0.98 to 1.00, showing the lift is not a property of the races fallers return in. The decisive comparison is the rival letter group: jumps BD/UR runners at 14/1 or shorter over the same window, expected near the previously measured 1.021. The hypothesis needs fallers clearly above both.",
    "rule": "last_non_finish_code IN ('BD', 'UR') AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND race_date >= '2025-01-01'"
  },
  "pass_criteria": {
    "min_bets": 200,
    "min_ae": 1.05,
    "max_p": 0.05
  },
  "forward_bar": {
    "min_bets": 100,
    "min_ae": 1,
    "max_p": 0.2
  },
  "confounders": [
    "Fallers may run in better-class races than unseaters, so a class effect could pass for a letter effect; check A/E by race_class",
    "Window mismatch: last_non_finish_code distinguishes F only from 2025-01-01 (every non-finisher was PU before), whereas last_fell_or_unseated and last_code use other windows; the rule is pinned to 2025-01-01 onward so all groups are compared on the same window",
    "A few short-priced winners in one period may carry the result; require both halves of the window to be above 1.0",
    "GB-only concentration, as in the going-well results (GB 1.27 vs IRE 0.80); a lift present only in GB with no mechanism for the split counts against the hypothesis",
    "The going-well flag overlaps the faller group; earlier findings show it adds nothing, so any lift should not depend on last_fell_going_well",
    "Chase against hurdle mix: chase fallers may differ from hurdle fallers in price and class"
  ],
  "mechanism": "A fancied jumps horse that fell last time usually walks away sound. In a horse good enough to be 14/1 or shorter, a single fall seldom reflects a lasting jumping flaw, and trainers run it back once satisfied with its schooling. Its underlying ability is intact and its next-run win rate should match its pre-fall standing.",
  "market_miss": "An 'F' is the most damning letter in the form figures. Punters fear a loss of confidence and lay off the horse, so its price drifts beyond what its ability merits. 'UR' and 'BD' carry less stigma (they are blamed on the rider or on bad luck), which fits capped BD/UR matching its price at 1.021. The over-penalty should therefore be specific to fallers.",
  "prediction": {
    "direction": "positive",
    "size": "A/E 1.10 to 1.25 on roughly 250-350 bets",
    "falsified_if": "A/E below 1.04, or fallers no higher than the same-window capped BD/UR control (about 1.02), or the lift is confined to one period or to GB alone"
  },
  "notes": "Kept the draft rule and added race_date >= '2025-01-01', because before 2025 last_non_finish_code records every non-finisher as PU. The date makes the window explicit and puts the BD/UR control on the same window. Raised min_bets to 200 to match the prediction. The refinements come from the mechanism's 'fancied' and 'walked away sound' claims; there is no going-well refinement, because findings show that flag adds nothing."
}

Timeline

Job 1660 finished
16 Oct, 13:46 · engine:compute-1
hypothesis note H-49fa55f7
16 Oct, 13:46 · engine:perturb
perturbation result for H-49fa55f7 (neighbour: drop: race_date >= '2025-01-01'): pass
16 Oct, 13:46 · engine:perturb
perturbation result for H-49fa55f7 (neighbour: drop: sp_decimal <= 15): pass
16 Oct, 13:46 · engine:perturb
perturbation result for H-49fa55f7 (neighbour: drop: race_type IN ('Hurdle', 'Chase')): pass
16 Oct, 13:46 · engine:perturb
perturbation result for H-49fa55f7 (neighbour: drop: last_non_finish_code = 'F'): pass
16 Oct, 13:46 · engine:perturb
perturbation result for H-49fa55f7 (neighbour: <= 15 -> 18 (+20%)): pass
16 Oct, 13:46 · engine:perturb
perturbation result for H-49fa55f7 (neighbour: <= 15 -> 12 (-20%)): pass
16 Oct, 13:46 · engine:perturb
compute-1 took job 1660 (mutate, attempt 1)
16 Oct, 13:46 · compute-1
Job 1668 finished
16 Oct, 13:46 · engine:compute-1
H-49fa55f7 Falsification → KILLED falsified: driven by one price band (up to 3/1): the rest score 0.92
16 Oct, 13:46 · engine:falsification
falsification result for H-49fa55f7 (partition: Fallers not flagged going well): pass
16 Oct, 13:46 · engine:evaluator
compute-1 took job 1668 (attack, attempt 1)
16 Oct, 13:46 · compute-1
Job 1667 finished
16 Oct, 13:46 · engine:compute-1
falsification result for H-49fa55f7 (partition: Handicaps only): pass
16 Oct, 13:46 · engine:evaluator
compute-1 took job 1667 (attack, attempt 1)
16 Oct, 13:46 · compute-1

Hypotheses · letters · H-49fa55f7

Jumps fallers at 14/1 or shorter beat their price: the market over-penalises a visible fall

KILLEDamberFailed a test or its forward record. Kept: failures are knowledge.

falsified: driven by one price band (up to 3/1): the rest score 0.92

Part of the thread “H98: brought down or unseated” 12 hypotheses

Proposed
Designed
Discovery
Promising
Falsification
Frozen
Vault
Provisional
Shadow
Validated
Live

Source: Investigator agent, v1
Made by: rival-explanation
Created: 16 Oct, 13:36
Tests logged: 16

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.

Observed
- H-9519a139 (Contrast: fell or unseated while NOT going well, at 14/1 or shorter, o) was predicted negative but came out the other way: A/E 1.13 on 367 bets (p 0.199 for the opposite direction). O-d93ddd07
- H-9f384f7c (Fell or unseated while going well: hidden form at normal prices) passed discovery but was falsified: period after 2026-04-27 A/E 0.96; IRE A/E 0.802; driven by one price band (up to 7/1): the rest score 0.97. O-ec1b9459
AskedIs the 14/1-or-shorter lift among jumps fallers and unseaters carried by fallers (last_non_finish_code = 'F') rather than unseaters, given that capped BD/UR scored only 1.021? answered
MechanismA fancied horse that fell usually walks away sound; a single fall seldom reflects a lasting jumping flaw in a horse good enough to be near the front of the market. Trainers run it back once they are satisfied with its schooling.
Why the market misses itAn F in the form figures is the most damning letter on the card. Punters fear a loss of confidence and lay off it, so the price drifts beyond what its ability merits. The unseat letter carries less stigma, which fits capped BD/UR matching its price at 1.021.
Predictedpositive, A/E 1.10 to 1.25 on 200+ bets
Falsified ifA/E below 1.04, or no higher than the capped BD/UR rule's 1.021
Rule sketchlast_non_finish_code = 'F' AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15

Mechanism and rationale

The rule

last_non_finish_code = 'F' AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND race_date >= '2025-01-01'

Reads: last_non_finish_code, race_type, sp_decimal, race_date

Pre-registration

Spec hash (sha256 of the canonical pre-spec): e0c3beaeb8fd88e59aa22109eddab4752e9e8ab3c7bc578e060548b452820781
Frozen: not frozen; The definition can still change.

Confounders: Fallers may run in better-class races than unseaters, so a class effect could pass for a letter effect; check A/E by race_class
Window mismatch: last_non_finish_code distinguishes F only from 2025-01-01 (every non-finisher was PU before), whereas last_fell_or_unseated and last_code use other windows; the rule is pinned to 2025-01-01 onward so all groups are compared on the same window
A few short-priced winners in one period may carry the result; require both halves of the window to be above 1.0
GB-only concentration, as in the going-well results (GB 1.27 vs IRE 0.80); a lift present only in GB with no mechanism for the split counts against the hypothesis
The going-well flag overlaps the faller group; earlier findings show it adds nothing, so any lift should not depend on last_fell_going_well
Chase against hurdle mix: chase fallers may differ from hurdle fallers in price and class
Control: { "description": "Same-race other runners should score about 0.98 to 1.00, showing the lift is not a property of the races fallers return in. The decisive comparison is the rival letter group: jumps BD/UR runners at 14/1 or shorter over the same window, expected near the previously measured 1.021. The hypothesis needs fallers clearly above both.", "rule": "last_non_finish_code IN ('BD', 'UR') AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND race_date >= '2025-01-01'" }

Discovery result

pass broad rule (niche) · 16 Oct, 13:40

Bets: 389; 66 winners
A/E: 1.07; 95%: 0.83 to 1.36
SP return: -24.4%; level stakes
Betfair SP return: -10.2%; 389 bets with a BSP

0.78dashed line 1.00 · interval includes 1.00: could be luck · grey tick: the rest of the field in the same races1.41

Strike rate 17.0%. p = 0.30. 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

By period

Falsification

Each attack is a deliberate attempt to kill the idea. 5 of 6 survived.

perturbation: Price cap tightened to 11/1 (sp <= 12)pass
332 bets · A/E 1.11 (0.86 to 1.42) · control 0.98 · 16 Oct, 13:46
Breakdown
betsA/E0.71.01.3
2025-10-01 to 2026-04-272221.08
after 2026-04-271101.19
Bars grow from 1.00, the market. Faded bars have fewer than 40 bets.
price_band: Fallers outside the short band: 4.0 < sp <= 15fail
309 bets · A/E 0.95 (0.65 to 1.34) · control 1.02 · 16 Oct, 13:46

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.

Other experiments

neighbour: <= 15 -> 12 (-20%)pass
332 bets · A/E 1.11 (0.86 to 1.42) · control 0.98 · 16 Oct, 13:46
Breakdown
betsA/E0.71.01.3
2025-10-01 to 2026-04-272221.08
after 2026-04-271101.19
Bars grow from 1.00, the market. Faded bars have fewer than 40 bets.
neighbour: <= 15 -> 18 (+20%)pass
414 bets · A/E 1.07 (0.83 to 1.36) · control 0.99 · 16 Oct, 13:46
Breakdown

Lineage

H98-brought-down-or-unseatedPROMISINGH98: brought down or unseatedmade by import

H-9f384f7cKILLEDFell or unseated while going well: hidden form at normal pricesmade by test-mechanism

H-9519a139KILLEDContrast: fell or unseated while NOT going well, at 14/1 or shorter, only matches its price

Agent runs

designer v1ok · 37s · 16 Oct, 13:36

Output (schema-checked)

{
  "rule": "last_non_finish_code = 'F' AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND race_date >= '2025-01-01'",
  "refinements": [
    {
      "label": "fancied: top three in the betting",
      "rule": "last_non_finish_code = 'F' AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND race_date >= '2025-01-01' AND sp_rank <= 3"
    },
    {
      "label": "back within 60 days (walked away sound)",
      "rule": "last_non_finish_code = 'F' AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND race_date >= '2025-01-01' AND days_since_last_run <= 60"
    }
  ],
  "control": {
    "description": "Same-race other runners should score about 0.98 to 1.00, showing the lift is not a property of the races fallers return in. The decisive comparison is the rival letter group: jumps BD/UR runners at 14/1 or shorter over the same window, expected near the previously measured 1.021. The hypothesis needs fallers clearly above both.",
    "rule": "last_non_finish_code IN ('BD', 'UR') AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND race_date >= '2025-01-01'"
  },
  "pass_criteria": {
    "min_bets": 200,
    "min_ae": 1.05,
    "max_p": 0.05
  },
  "forward_bar": {
    "min_bets": 100,
    "min_ae": 1,
    "max_p": 0.2
  },
  "confounders": [
    "Fallers may run in better-class races than unseaters, so a class effect could pass for a letter effect; check A/E by race_class",
    "Window mismatch: last_non_finish_code distinguishes F only from 2025-01-01 (every non-finisher was PU before), whereas last_fell_or_unseated and last_code use other windows; the rule is pinned to 2025-01-01 onward so all groups are compared on the same window",
    "A few short-priced winners in one period may carry the result; require both halves of the window to be above 1.0",
    "GB-only concentration, as in the going-well results (GB 1.27 vs IRE 0.80); a lift present only in GB with no mechanism for the split counts against the hypothesis",
    "The going-well flag overlaps the faller group; earlier findings show it adds nothing, so any lift should not depend on last_fell_going_well",
    "Chase against hurdle mix: chase fallers may differ from hurdle fallers in price and class"
  ],
  "mechanism": "A fancied jumps horse that fell last time usually walks away sound. In a horse good enough to be 14/1 or shorter, a single fall seldom reflects a lasting jumping flaw, and trainers run it back once satisfied with its schooling. Its underlying ability is intact and its next-run win rate should match its pre-fall standing.",
  "market_miss": "An 'F' is the most damning letter in the form figures. Punters fear a loss of confidence and lay off the horse, so its price drifts beyond what its ability merits. 'UR' and 'BD' carry less stigma (they are blamed on the rider or on bad luck), which fits capped BD/UR matching its price at 1.021. The over-penalty should therefore be specific to fallers.",
  "prediction": {
    "direction": "positive",
    "size": "A/E 1.10 to 1.25 on roughly 250-350 bets",
    "falsified_if": "A/E below 1.04, or fallers no higher than the same-window capped BD/UR control (about 1.02), or the lift is confined to one period or to GB alone"
  },
  "notes": "Kept the draft rule and added race_date >= '2025-01-01', because before 2025 last_non_finish_code records every non-finisher as PU. The date makes the window explicit and puts the BD/UR control on the same window. Raised min_bets to 200 to match the prediction. The refinements come from the mechanism's 'fancied' and 'walked away sound' claims; there is no going-well refinement, because findings show that flag adds nothing."
}

Timeline

Job 1660 finished
16 Oct, 13:46 · engine:compute-1
hypothesis note H-49fa55f7
16 Oct, 13:46 · engine:perturb
perturbation result for H-49fa55f7 (neighbour: drop: race_date >= '2025-01-01'): pass
16 Oct, 13:46 · engine:perturb
perturbation result for H-49fa55f7 (neighbour: drop: sp_decimal <= 15): pass
16 Oct, 13:46 · engine:perturb
perturbation result for H-49fa55f7 (neighbour: drop: race_type IN ('Hurdle', 'Chase')): pass
16 Oct, 13:46 · engine:perturb
perturbation result for H-49fa55f7 (neighbour: drop: last_non_finish_code = 'F'): pass
16 Oct, 13:46 · engine:perturb
perturbation result for H-49fa55f7 (neighbour: <= 15 -> 18 (+20%)): pass
16 Oct, 13:46 · engine:perturb
perturbation result for H-49fa55f7 (neighbour: <= 15 -> 12 (-20%)): pass
16 Oct, 13:46 · engine:perturb
compute-1 took job 1660 (mutate, attempt 1)
16 Oct, 13:46 · compute-1
Job 1668 finished
16 Oct, 13:46 · engine:compute-1
H-49fa55f7 Falsification → KILLED falsified: driven by one price band (up to 3/1): the rest score 0.92
16 Oct, 13:46 · engine:falsification
falsification result for H-49fa55f7 (partition: Fallers not flagged going well): pass
16 Oct, 13:46 · engine:evaluator
compute-1 took job 1668 (attack, attempt 1)
16 Oct, 13:46 · compute-1
Job 1667 finished
16 Oct, 13:46 · engine:compute-1
falsification result for H-49fa55f7 (partition: Handicaps only): pass
16 Oct, 13:46 · engine:evaluator
compute-1 took job 1667 (attack, attempt 1)
16 Oct, 13:46 · compute-1

Refinement	Bets	A/E	Grade
refinement: fancied: top three in the betting	210	1.22	green info
refinement: back within 60 days (walked away sound)	330	1.13	amber info
control: Same-race other runners should score about 0.98 to 1.00, showing the lift is not a property of the races faller	277	1.04	red info

Jumps fallers at 14/1 or shorter beat their price: the market over-penalises a visible fall

Reasoning

Mechanism and rationale

The rule

Pre-registration

Discovery result

By period

Falsification

Sealed holdout (vault)

Forward record

Other experiments

Lineage

Agent runs

Timeline

Jumps fallers at 14/1 or shorter beat their price: the market over-penalises a visible fall

Reasoning

Mechanism and rationale

The rule

Pre-registration

Discovery result

By period

Falsification

Sealed holdout (vault)

Forward record

Other experiments

Lineage

Agent runs

Timeline

By country

By price band

Refinements named in the pre-spec