Skip to content

PREPACK_RECON.md

Latest commit

 

History

History
47 lines (37 loc) · 2.57 KB

PREPACK_RECON.md

File metadata and controls

47 lines (37 loc) · 2.57 KB
Raw

Prepack / parlay surface recon (2026-07-17)

The Wave-10 correlation-aware parlay engine (autonomy/parlay.py) prices combos but has no live market surface wired to it. This is the recon that scopes that wire-up, taken against the public series list (/trade-api/v2/series?category=Sports, 2,954 sports series).

What exists

21 combo-family series, all football/basketball (no MLB or WNBA):

Family Series Semantics
Cross-game ML packs KXNFLPREPACK2ML KXNFLPREPACK3ML KXNBAPREPACK2ML KXNBAPREPACK3ML KXNCAAFPREPACK2ML KXNCAAFPREPACK3ML KXNCAAFPREPACK4ML KXNCAAMB2ML KXMULTIPREPACK N moneylines across games ("2 ML Basketball Combo")
Same-game parlays KXNFLPREPACKSGP KXNFLPREPACKSGPSPREAD KXNCAAFPREPACKSGP KXNCAAMBSGP correlated legs within one game
Segment combos KXNFLPREPACK1HFT KXNFLPREPACK1Q1H half/full-time and 1Q/1H stacks
Misc / futures combos KXNFLCOMBO KXNFLPREPACK KXWCPREPACK KXWBCPREPACK KXWCGOALCOMBO KXMLBAWARDCOMBO KXWCAWARDCOMBO mixed; some are futures, not per-game

What is NOT observable yet

Every prepack series carries zero markets at any status (checked unsettled/open/closed on KXNFLPREPACK2ML and KXNFLPREPACKSGP): the 2026-season shells exist but no events have been listed, and last season's markets are purged. Leg encoding (how a market names its component legs) cannot be learned until the first fall listings appear (~Aug for NCAAF/NFL).

Staged plan (fall wave)

  1. Registry: add the per-game combo families to sports_markets.py with discover=False now; flip on when the pricing path lands.
  2. Leg classifier: on first listings, parse leg encoding from ticker/title/rules; each leg should resolve through the existing registry classifier (classify) so leg fair values come from the same sources that price the standalone markets.
  3. Pricing: cross-game ML packs = independence product of leg fair values (pure vig check); SGP families = the copula engine's actual edge case — Wave-10 validation showed same-game stacks quote ~independence while true joint probability runs ~+0.12 higher on aligned 3-leg stacks.
  4. Grading: one source per family (prepack_ml, prepack_sgp, prepack_segment), challenger-only, never auto-staked (combos multiply variance; staking stays human-gated).

Why not now

The engine's edge is correlation pricing; the correlated families (SGP, segment combos) are football/basketball only, out of season until August. In-season effort stays on MLB/WNBA evidence accrual (Wave-12) instead.