Club soccer · Methodology

Show the work. Then the limits.

Every Big-5 prop number on this site traces to a logged bake, a held-out calibration row, or a captured sportsbook line, labeled separately.

Ledger refreshed July 17, 2026

The process

01
Build the player panel
Match logs from api-football and Understat join FBref season rates. Every prop row carries minutes, opponent, venue, and rolling hit-rates. The workbench shows measured history, not a forecast.
02
Train on temporal holdout
LGBM prop models fit on Big-5 club data with a strict time split. Isotonic calibration runs before serve. ECE must stay under the publication gate. Base-rate climatology is the skill benchmark.
03
Read posted lines
API-Football player props supply multi-book prices. We headline fair (model-implied American) vs best available decimal, not raw edge %. Injury status scales or suppresses headline reads.
04
Validate before publish
Hub bakes run through validate_soccer_hubs.py fail-closed. Probabilities must be coherent, prop slugs must crosswalk to player pages, and in-season fixture counts must match props slates.

Improvement vs per-player base rates on the temporal holdout. Positive means the model beats guessing the historical rate.

Mean squared error on binary prop outcomes. Lower is better, comparable across markets.

Expected calibration error after isotonic mapping. Publication gate at 8%. Served club props typically run under 3%.

validate_soccer_hubs.py must pass before JSON reaches Cloudflare Pages. No silent WARN-and-ship.

01
Hit-rates describe past samples. Model % is a calibrated estimate. Neither is a promise.
02
Model-minus-market gaps are diagnostics. CLV vs closing line is the honest betting metric when captured.
03
Availability adjustments are heuristic. Out or suspended players may be dropped from headline picks.
04
Spatial viz (shot maps, radars) is display context, not fed back into LGBM features on publish.
05
Off-season builds may ship empty fixture hubs while props and player archives remain populated.
06
NB ladder models for some markets coexist with LGBM. The canonical board picks one family per stat.