Today's picks All tracks Yesterday's results Track record Weekly recap Backtest
Account
Subscribe Disclaimer
90-day backtest · settled races, real payoffs

What works. What doesn’t.

We tested 14 distinct bet structures — WIN, PLACE, EX, TRI, SUPER, Pick 3/4/5 — against 90 days of settled races with real payoff data. One signal stood out. Here’s the honest breakdown.

The validated edge

HOT_CONNECTIONS at ML 3-1 or longer.

Bucket Plays (N) Win hit% Win ROI W+P combo
Loading latest stats…

The chalk row is the most important. When a famous trainer/jockey combo runs together at 1-3, the public bets it down past the edge. Same horse, same connections, same engine recommendation — but the price kills it. That’s why the picks card filters to ML 3+ only.

Bet structures that DON’T work

We’d be lying if we only showed the green numbers. Here’s every other bet structure we tested across 2,500 races.

Strategy N ROI
PLACE on every top pick1,267-20.4%
WIN on every top pick (baseline)1,249-39.0%
$1×2 EX box, angle anchor + chalk4,887-41.9%
$0.50 TRI box, top-3 angles2,217-60.0%
$0.10 SUPER box, top-41,236-89.0%
$1 Pick 3 single ticket844-97.1%
$1 Pick 4 single ticket222-100%

What this means for you

1. Stick to the picks card
Not every angle scanner output is a profitable bet. The picks card is filtered to the only validated subset.
2. The edge is in WIN
The validated alpha lives in the WIN pool. PLACE-only converges to chalk; exotics blow it up. We don’t prescribe ticket shapes — we publish the edge.
3. Skip exotic boxes
EX/TRI/SUPER/Pick N look attractive (big payouts) but lose 40-100% of stake on average.
4. Volume varies
Some days produce 5-10 qualified picks. Some days zero. The card never pads to fill space.

Closing-line value

Are our picks beating the market’s closing line?

Loading CLV data…

Calibration check

When the model says X% conviction, does it actually win X% of the time?

Model conviction N picks Bin midpoint Actual win rate Drift
Loading reliability bins…
A perfectly-calibrated model has bin midpoint = actual win rate. Drift > 0 means the model is over-confident; drift < 0 means under-confident.
Methodology
Backtest period: rolling 90 days ending today, refreshed every Sunday. Sample: every settled US thoroughbred race in the period with morning-line + finish data, ~2,500 races. ROI computed on a unit-stake basis from actual settled win_payoff in the race-results data. Hot-trainer/jockey thresholds: top-60 over rolling 30-day windows with min 15 starts (trainers) / 20 starts (jockeys). HOT_CONNECTIONS = horse where both connections meet thresholds. Source: Brisnet results API.

Takeout accounting: Published WIN ROI numbers are computed on actual settled tote payoffs — net of the host track’s WIN takeout (typically 15-19%). They reflect what was actually paid at the window, not theoretical morning-line implied returns. Gross-of-takeout returns would be higher; we publish net because that’s what real-money outcomes look like.

Caveats: The 12-15% number derives from a window that includes the discovery period for the HOT_CONNECTIONS filter. Closing-line value (CLV) is the more rigorous test — we’re accumulating closing-odds data and will publish CLV here once we’ve crossed N≥60 settled picks with capture coverage. Past performance is not indicative of future results.

Sample-size honesty: 785 picks across 90 days is enough to detect a real edge but not enough to rule out variance bands of ±5pp on the ROI estimate. Treat the directional finding as the signal, not the precise number.