reference_odds_book_quality

Audited every book on a 454-prop WNBA board (odds-service /api/v1/odds/props/bulk?league=wnba) on 2026-07-31 at ~23:10 UTC:

BetRivers posts NO under prices — ever. It is structurally impossible to de-vig, and its over prices are wide outliers (Arike reb 2.5 at −113 while the real market was −147/−148/−154; Boston reb line 18.5 vs a 16.5 consensus). Any "edge" whose best price comes from BetRivers is an artifact. Exclude it from both the de-vig set and the best-price scan.

Pinnacle posts ONE scrape (~19:41 UTC) and is stale by tip. Using it as the fair-value anchor against a market that has since drifted manufactures edge in whichever direction the market moved — this flipped Awa Fam reb OVER 5.5 from "+1.2% vs Pinnacle" to −5.4% vs the current market. Despite feedback_pinnacle_default, on a same-night screen prefer a ≥3-book fresh two-way consensus and treat Pinnacle as a cross-check only.

Two more gotchas, same audit:

• is_live is unreliable — the same scrape second returns live:true for one book and live:false for another on games whose SDS boxscore still says status: scheduled. Confirm tip via SDS boxscore, not this flag.
• When scanning for "best price," restrict to the freshest scrape batch. Scanning all pregame quotes pulled a 17-hour-old +103 on a prop that had since steamed to −147, inventing a fake +13.5%. See feedback_devig_phantom_edge.

Which markets exist, per league (verified 2026-08-24)

odds-service requests markets from LEAGUE_MARKETS in app/services/odds_fetcher.py. Each market = 1 quota unit per call.

• NBA: 11 — points, rebounds, assists, threes, steals, blocks, turnovers, pra, pr, pa, ra.
• WNBA: 8 — the same minus steals, blocks, turnovers. WNBA books have NEVER published those: measured 3,585 steals + 2,753 blocks + 4,648 turnovers + 17,424 minutes bets over 30 days with zero closing lines. WNBA used to share NBA's list, paying quota for nothing.
• So steals/blocks/turnovers CLV arrives automatically when NBA opens — the WNBA blindness is the books, not us.

⚠️ player_steals_blocks is NOT a valid Odds API market — 422 "Invalid markets", identical to a key I made up to check. sportlib's nba.odds_api_markets lists it anyway, so that registry entry is wrong. This matters more than a missing prop: one bad key 422s the WHOLE per-event request, so trusting the registry would have stored ZERO NBA props of any kind.

How to verify a market key (do this before adding one): call the per-event endpoint, /v4/sports/{sport}/events/{id}/odds?markets=KEY.

• 200 + bookmakers = valid and priced
• 200 + zero bookmakers = valid key, this league's books don't offer it
• 422 = invalid key, will take the whole request down

Do NOT test on /v4/sports/{sport}/odds — it rejects player_points too, with a different message ("Markets not supported by this endpoint"). The two failures look alike and mean opposite things.

Shipped 2026-08-24 (odds-service b1aa3e7): NBA_MARKETS 10→11 (+turnovers), new WNBA_MARKETS = 8. Verified post-deploy: WNBA feed still returns 498 rows across the same 8 markets, so the split cost nothing and saves 2 quota units per call.