Skip to content

Learnings, 2026 Kentucky Derby

The founding race. Built in five hours; picked the winner; net +$435 on $85.

For the full case study see analysis/case-studies/2026-kentucky-derby/. This file extracts the wisdom, the parts of the experience that don't fit cleanly into code or config and need to be carried forward as priors.

What the model got right (and why we trust it)

  • Picked Golden Tempo at 25-1. Sensitivity scan tagged him "ROCK SOLID" across 200 weight perturbations. The math worked.
  • All three named-trainer overbets missed. Mike Smith on So Happy (5-1 live, off the board), Cox on Commandment (5-1, 7th), Mott on Chief Wallabee (7-1, 4th). Public-overbet thesis held up.
  • Post-1 multiplier was right. Renegade ran 2nd from the rail, exactly the model's fair-prob outcome (legitimate horse, can't win from there, places).

What the model got wrong (and what to remember)

Further Ado was our $20 top win bet. Out of the money.

Sensitivity said ROCK SOLID. He never showed up. Possible explanations: didn't break clean, didn't get a closer's trip, pace meltdown didn't fully materialize. We will never know exactly why, and that's the point. Sensitivity scans are necessary, not sufficient.

Cox barn-pick rule misfired.

The hand-coded "Velazquez on Further Ado = the tell" rule contributed materially to over-staking #18. Cox went 0-for-2 on the day. The barn-pick signal is real (jockey assignment in a multi-entry stable IS information), but one race is one race. Audit the rule rather than removing it.

AE penalty bit us twice.

Ocelli (#22A) ran 3rd at 70-1 with Tyler Gaffalione picking up an AE mount after his original ride scratched. Our model penalized Ocelli with a 0.50 AE multiplier. Bug fixed: AE-activated horses (those with live odds confirmed) lose the AE penalty.

Story we missed.

Cherie DeVaux became the first female trainer to win the Derby in 152 runnings. Owner Daisy Phipps Pulito is also a woman. Both were already in field.csv. The model picked the historic horse for math reasons; the writeup didn't tell that story to the people who would have cared most. Future writeups: surface biographical signal alongside numerical signal.

The methodological lesson, three layers

The single biggest mistake of the day was constraint-first ticket construction. We started with $85 and packed bets inside it. We should have: 1. Identified every positive-EV bet (Kelly core) 2. Filled remainder with high-EV/low-prob satellites 3. Reserved budget for top-pick wheel + longshot scan heuristics

Retrospective sim with the new modules: a 3-layer $85 ticket would have caught 96% of the hand-tuned upside ($418 vs $435). Quarter-Kelly alone would have caught 3% ($11). The 3-layer ticket is the right default; pure Kelly under-deploys and pure judgment is hard to repeat.

The trifecta we missed (19-1-22, $5,625): even with full Kelly + actual probable payout, this combo had EV of only +$0.15 per $1. Kelly would have staked $0.001, below any reasonable minimum. Capturing this required the heuristic layer specifically, --top-pick-wheel 8 reserves $8 for top-overlay-horse / top-3-other-fair-probs / ALL wheels. With the new heuristic, $0.04 lands on 19-1-22 → $450 payout.

Wagering rules to carry forward

  • Always run --target-spend = bankroll. Treat bankroll as a scalar, deploy the full amount.
  • Always include --include-tri and --top-pick-wheel <amount>. Reserve $5-10 for the top-pick wheel, no matter how the math says skip.
  • Always include --longshot-scan <amount>. Reserves $5-10 for live-tote-undervalued horses as exacta placers under your top pick.
  • Sensitivity must agree with cardinal AND rank for highest confidence. If only one method flags a horse as YES, treat as a contingent bet, not a primary stake.
  • Pull live odds at least twice, early (T-3h, set initial baseline) and late (T-30min, capture last-minute movement). The drift between snapshots IS the public-overbet signal.

Things to watch for in this kind of race

  • 20-horse field, post 1: 1 winner since 1975 (Ferdinand 1986). Multiplier 0.60.
  • Multi-entry stables: jockey assignment is the tell, not the official barn statement.
  • First-time blinkers: trainer's history matters (Mott 22% in last 365). Live odds will reveal whether the public has noticed.
  • AE horses: when scratches activate them, they're live. Drop the AE penalty.
  • Recent Derby winner's offspring: bloodline-narrative bets the public makes (Litmus Test sired by Nyquist). Real but variable signal.
  • Foreign shippers: Japan + UAE + Europe runners need US-equivalent Beyer adjustment. Without that adjustment, they're effectively neutral imputed (50%).

What this race teaches about the next one

  • Cherie DeVaux is now a Derby-winning trainer. Her score in [trainer_score] should bump for Preakness/Belmont (was 55, should probably be 75+).
  • Velazquez/Cox barn-pick rule misfired here, does NOT mean retire it, but DO weight it less aggressively than originally configured.
  • The Mott family all underperformed (Wm Mott's Chief Wallabee 4th, Riley Mott's Albus and Incredibolt OOM and 6th). Watch the trainer scores, Riley Mott's 55 may be too generous given small sample.
  • Florida Derby winner (Commandment) finished 7th. The FlaDerby = 100 score in the Derby config is HISTORICAL prior, not a hard rule. Worth examining whether to soften for next year.
  • The pace meltdown didn't fully materialize. Confirmed E's were Pavlovian, Litmus Test, Six Speed, Robusta, all out of the money. The pace-fit logic correctly favored closers (Golden Tempo won), but Ocelli (3rd, 70-1) was a closer-stalker that the model also penalized via the AE bug. Pace logic was right; AE bug suppressed the signal on Ocelli.