State of the machine
The standing picture: what we hold, what we've learned, what's betting, and where the edge hunt points. Everything links to the underlying experiment — nothing here is vibes.
In plain English: this is the machine's report card — the things it has learned that held up under testing, and how its 35 betting strategies are doing. The big lessons so far: the bookmakers' main “who wins” price is very hard to beat, so the real money is in the markets they care less about (corners, cards, player props); referees and big skill mismatches are the strongest tells; and “which player scores” is far more predictable than “who wins the match.” Every claim below links to the experiment that proved it — nothing here is a hunch.
The machine in numbers
Laws — findings that survivedeach links to its experiment
- The closing line is still unbeaten at its own game — but the gap is closing: engine v9 (standings/coach/age features) runs ROI −1.46% vs v8's −2.95% across 21 seasons — and v10/v11 proved more team features make it WORSE (−1.87%, −2.44%): v9 is frozen as champion. The money remains in markets the big line doesn't obsess over. → experiment #216
- Favorite-longshot bias is real in our own data: blind away-backers end seasons at ~$654, longshot lovers ~$703, odds-on bankers ~$982. Prices worsen as they lengthen. → experiment #216
- Referees are the strongest non-market signal: fouls AUC 0.60, cards next; card-happy refs correlate with draws (won all 4 backtest seasons — live forward trial running).
- Levels beat trends: no rolling metric's recent direction adds anything its level didn't already carry (105k team-matches).
- Mismatch controls the flow: in big skill/Elo gaps the corners go ~77/23 and goals ~83/17 to the stronger side. → experiment #200
- Player questions beat match questions: the scorer model's AUC 0.762 (up from 0.737 with the MLS/CL/South-America player data) is the most predictive object in the shop (match models cap ~0.60). → experiment #196
- Market errors are slightly predictable (R² ≈ 0.07) — and the catastrophic misses are almost always a huge favorite falling. → experiment #145
- Nulls: clutch/bottling is noise, rain is a placebo, "match stakes" predicts nothing, and stacking more team features onto 1x2 models is flat — the market already has them.
Gap intelligencewhen we disagree with the market
21,280 model-vs-market disagreements dissected · overall model advantage negative (-0.01571) · gap-survival classifier holdout AUC 0.6778 · full breakdown →
Hurricane test (do wrong models average right?): engine A Brier 0.19805, engine B 0.19611, engine mean 0.19499, market 0.19016 — over 4,560 shared predictions.
The tournamentengine × rule × staking, every season from a fresh $1000
The frontierwhere the edge hunt points next
- Soft markets: cards/corners/props odds archive started 2026-07-04 — the referee, style and mismatch signals bet here, where margins are lazy.
- Player counting stats: generalize the scorer model to shots/tackles/fouls and price the books' combo "specials".
- Reaction speed: confirmed lineups land ~1h before kickoff; slow books don't reprice instantly.
- Breadth: big-5 league granular history backfilling daily via the quota-budgeted catchup.
Honest watch-outs
- Thousands of tested combos ⇒ leaderboard tops are partly luck. Tournament winners get forward trials, not belief.
- Progression staking (martingale/fibonacci) flatters small samples right up until it ruins you — read the worst-season column.
- Poisson engines carry 4 seasons of stored predictions vs the GBM's 21 — their cells are noisier.
- All numbers are paper money at real prices; real execution would face limits and slippage the paper doesn't.