| Track | Race | Post | # | Horse | ML | Win % | Edge | Fav |
|---|---|---|---|---|---|---|---|---|
| FE | R1 | 4:30 PM | 4 | Cantucci WATCH | 3.5/1 | 26% | +8.0% | Demote #2 |
| FE | R3 | 5:26 PM | 3 | Wolf Beach | 2.5/1 | 24% | 0.7% | Demote #2 |
| FL | R7 | 4:03 PM | 7 | Galileans Girl WATCH | 3.5/1 | 22% | +4.2% | Demote #2 |
| FP | R6 | 4:42 PM | 6 | Juilliard WATCH | 3.0/1 | 25% | +5.7% | Demote #2 |
| MNR | R5 | 8:40 PM | 4 | Fiamma | 3.5/1 | 39% | +10.0% | On Fav |
| PID | R3 | 4:24 PM | 5 | Going Up | 4.5/1 | 34% | +18.1% | Demote #2 |
| PID | R4 | 4:51 PM | 5 | Time for an Angel | 2.0/1 | 26% | 2.8% | On Fav |
| PID | R5 | 5:18 PM | 7 | Epic Summer | 2.0/1 | 29% | 1.1% | On Fav |
| PID | R7 | 6:12 PM | 7 | Naked Eye | 2.0/1 | 32% | +5.2% | On Fav |
| PID | R8 | 6:39 PM | 7 | Garnet and Gold | 2.5/1 | 25% | +4.1% | On Fav |
| Track | Race | Horse | Edge | Fav | Result |
|---|---|---|---|---|---|
| No results for yesterday. | |||||
| Track | Race | Post | # | Horse | ML | Win % | Edge | Fav |
|---|---|---|---|---|---|---|---|---|
| No picks yet — pipeline runs after BRIS PDF ingestion. | ||||||||
Bettable = bets the live strategy actually places: turf & all-weather, full-coverage races only. Dirt is watch-only under the surface gate (the model loses on dirt in every backtest fold), so dirt picks are tracked but excluded here — this is why these figures match the headline ROI above.
Stored picks from each evening's prediction run, graded against official BQ results. Tracking starts 2026-05-30 — the first day of picks generated by the current feature set. Pending = picks not yet matched to a result (race not in BQ or scoring skipped). V6 publishes a race's top-scored horse when its honest edge is positive — the calibrated, race-normalized model win prob minus the de-vigged (vig-free) market prob — and tiers it high/mid/actionable by edge size. Backtest benchmarks shown for comparison. Pending = picks not yet matched to a result (race not in BQ or scoring skipped).
| Date | Picks | Scored | Pending | Wins | Win% | ROI |
|---|
⚠️ Overall backtest ROI: -17.9% (32.7% WR over n=4157 races). The tier ROIs below are small, selective subsets (Actionable ≈ 134 of 4157) — high-variance and NOT the live strategy’s expected return. Live betting is full-coverage, turf/AW only (dirt = watch). Treat these as diagnostic, not a forecast.
| Tier | Edge Raw | Races | Win Rate | ROI |
|---|---|---|---|---|
| HIGH | > 0.10 | 86 | 37.2% | +34.4% |
| Mid | > 0.08 | 105 | 35.2% | +24.1% |
| Actionable | > 0.06 | 134 | 35.1% | +25.5% |
| No bet | ≤ 0.06 | 1,702 | 31.8% | -14.1% |
edge_raw = model softmax - ML implied prob. Pick is always the #1 model-ranked horse; edge gates the bet.
AGREE + edge > 0 = the model picks the favorite but is more confident than the public. See the row below for its current backtest vs live numbers (both bettable-gated).
| Stance | Filter | Races | Win Rate | ROI |
|---|---|---|---|---|
| AGREE + HIGH | edge > 0.08 | — | — | — |
| AGREE + any | edge > 0 | — | — | — |
| DISAGREE + HIGH | edge > 0.08 | — | — | — |
| DISAGREE + any | edge > 0.06 | — | — | — |
| All disagree | — | — | — | — |
AGREE = model's top pick matches the ML favorite. DISAGREE = model picks a different horse. Edge thresholds use raw softmax edge. The model is excellent at confirming undervalued favorites; it struggles to find profitable alternatives when fading the public's choice.
| Track Tier | Races | Win Rate | ROI |
|---|---|---|---|
| Fair | — | — | — |
| Mid | — | — | — |
| Top | — | — | — |
At edge > 0.06, fair-track picks dominate both volume and ROI. Top-tier tracks (OP, KEE, CD, SA, AQU, GP) show negative ROI — those markets are too efficient for the edge signal to work. The best betting strategy may be: edge > 0.06 at fair + mid tracks only.
Live (post-2026-05-30) picks pivoted by your chosen Row × Column. Data: Bet = bettable turf/AW (the live strategy — matches the headline ROI); Watch = dirt (gated out, tracked but not bet); All = both combined. Cells show the selected metric over non-scratched picks; n appears underneath. Color follows ROI (green = +EV, red = −EV) regardless of metric chosen, so you can read the visual gradient as profitability while inspecting any aggregate. Cells below the min-n threshold are hidden — sample size kills inference fast in 2D cuts.
Each panel slices live (post-2026-05-30) picks by a dimension. Live column shows non-scratched bets; BT is the matching backtest cut. Red ROI ≥ 20 pts below backtest = candidate to investigate first. Cells with n<5 are too thin to draw conclusions but useful for spotting tracks/buckets we've barely sampled.
| Date | Track | Race | Horse | Edge | ML | Fav | Fin | Result |
|---|---|---|---|---|---|---|---|---|
| No graded results yet. | ||||||||