Sporttery Engine

A probability analysis and virtual betting simulation system for Sporttery football. It is based on the Dixon-Coles Poisson model and Monte Carlo simulation, uses market-dominated probability fusion, allocates virtual bets with the Kelly criterion, and runs fully automatically on GitHub Actions at zero server cost. For data analysis and algorithm learning only. It does not constitute any betting advice; all outputs are virtual bookkeeping.

Core Principles (driven by 2026-08 empirical data)

1. The market is the benchmark: Across 426 matches in the ledger, the pure market (vig-removed) Brier score of 0.583 beats every model stream (0.646+). The fusion weight is therefore market-dominated (0.75), and models and third-party sources only make minor adjustments.
2. Every switch is decided by data: each step of the fusion chain is a config switch (config/prediction.json["fusion"]["post_fusion"]). scripts/ablation_replay.py decides whether each one stays or goes, by paired replay over settled ledgers.
3. Self-evolution with guardrails: any automatic enabling of the calibration layer or LGBM must pass a time-based split and a paired significance…