Swap set-piece + shot-xG signals to Supabase sofascore data
Round-3 audit left 5 signals below ceiling because their source data (understat-shots) was Big-5 only and frozen at 2024. The inventory review turned up sofascore_shots in Supabase with 368k shots across
Round-3 audit left 5 signals below ceiling because their source data (understat-shots) was Big-5 only and frozen at 2024. The inventory review turned up sofascore_shots in Supabase with 368k shots across 24 leagues × 2025-26 fully populated — every league we bet in, and the situation field includes the set-piece categories we need.
Changes:
- New loadSofascoreShots() reads sofascore_shots + sofascore_xg for a
(league, season) pair, runs the v1 pre-shot xG model per shot, and returns an array of shots with computed xG, team, date, is-goal.
- preloadSofascoreData() runs once at main() startup — one query per
unique (league, season) in the ledger, all in parallel. Everything downstream reads the in-memory cache so the bet loop stays sync.
- loadSetPieceData() (SP-xGA) now tries sofascore first, understat
second. Big-5 gate on the signal dropped: now runs league-wide where data is available.
- loadSetPieceMismatchData() (SP-Mismatch) same treatment — Big-5 gate
dropped, sofascore preferred, understat + fotmob as fallback.
- loadXGDataFromShotFiles() (ShotXG) now tries sofascore first. Covers
arg/brazil/austria/mls/jLeague/etc. which fotmob-shots never had.
- get{Sp,SetPiece}* lookups route team-name resolution through
findTeamInMap so sofascore's "Leicester City" matches ledger's "Leicester" etc.
- .dockerignore: narrowed data/xg-model/ exclusion to keep just the
xg_pre_shot_v1_compact.json model file (696KB) needed at runtime.
Expected coverage lift:
- SP-xGA: 0/181 → ~150/181 (all AH bets in sofascore-covered leagues)
- SP-Mismatch: 56/181 (Big-5 only) → ~170/181
- ShotXG: 141/181 → ~175/181 (gains on SA + minor European leagues)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>