Pick a variant
Six surfaces, each tuned for a different question.
The recommended HR betting surface. Full v2 calibrated pipeline with the v3 residual-correction layer (trained on 127K settled HR-prop closing lines from 2025+2026), then every pick priced longer than +800 American is dropped before display. The cap is load-bearing — uncapped lanes bleed ~15% ROI in the longshot tail because the chain over-estimates HR probability at long odds. Capped retrospective ROI: +46.64% on 134 picks (v3-calibrated) at the +0% EV gate; +66% on 152 picks at the +10% EV gate.
Same v2 calibrated chain as v3 — Bayesian shrinkage, pitcher decomposition, humid-air carry, Shin-style devig — but without the v3 residual layer applied to the no-vig consensus. The matched-control lane: cron forks at the calibration step and tracks v2 alongside v3 in parallel, so the long-run delta in ROI / hit-rate / CLV reads out the v3 residual's contribution directly. Capped at +800 American.
Original v1 multiplicative chain (league HR rate × batter factor × pitcher factor × park factor × weather × platoon) restricted to picks priced ≤+800 American. Highest-volume production surface — 458 picks across the 35-day retrospective window — with +37.47% ROI. The classic chain math is unchanged; the cap removes the longshot tier where uncapped v1 grinds to break-even.
Builds 2–3-leg HR parlays from the priced board, scoring each combo with same-pitcher correlation (~1.12× joint-prob lift per shared opposing starter; Tango/Stoll empirical estimate). Filters to combos one book actually prices end-to-end. Same-game stacks opt-in. Stake recommendations use ⅛-Kelly to control variance.
v3 residual model — calibration
Trained on every HR-prop closing line MyOddsy captured this season, joined to the actual game-day HR outcome. Recalibrates the consensus prob v2 feeds into its blend step. Active on v2.
/api/hr/v3-train runs.How the variants differ
The same data, scored five ways.
- v1 is the chain — every multiplier visible, easy to audit. Best for understanding why a pick exists. The capped variant throws out the top-end longshots that dominate v1's EV ranking.
- v2 replaces the chain with a calibrated pipeline — air-density physics, decomposed pitcher quality, lineup-slot PA PMF, tier-aware shrinkage, and a logit-space blend with the de-vigged market consensus. The v3 residual layer corrects book-specific and park-specific biases on top.
- Capped variants (v1 / v2) filter the same model's output to picks at +800 or shorter. Lower per-pick variance at the cost of dropping the longshots. Useful for unit-stake bankroll styles.
- Parlay engine consumes the v1 board and builds same-book 2-3-leg combos. Same-pitcher correlation (~1.12× joint lift) is applied when stack mode is on; otherwise legs are required to be from different games.