What 27 seasons of data actually buys us.

The archive does not exist to make hero picks. It exists to measure everything: it trained the model, then proved the market prices games better than the model does. That measurement is the product: honest calibration, a no-vig (margin-removed) consensus, and the one edge that survives scrutiny: never paying a worse price than you have to.

READING LIVE ENGINE OUTPUT

The pipeline this data runs

SOURCE ARCHIVE

Archive

7,548 games since '99, plus our own per-book price capture.

backfill/ · raw/
WALK-FORWARD FIT

Calibrate

Elo + independent models fit walk-forward, never on future games.

calibrate.py · features.py
OUT-OF-SAMPLE TEST

Backtest

2,671 graded ATS calls (2,608 decided), 2016–25 — model vs market, no cherry-picks.

backtest.py
MARKET NORMALIZATION

De-vig & shop

Strip the juice (margin) from every book, surface the best price per side.

lines.py · closing.py
PUBLIC RECORD

Commit & grade

Hash-committed before kickoff, graded and revealed after.

commit.py · grade.py

The honesty proof — model vs market, walk-forward

Cumulative against-the-spread rate across every graded backtest game. If a line stays under the 52.4% breakeven, betting it at −110 loses money. Ours does. So does a naive read of the market — which is the point: nobody on this chart clears breakeven, and we publish that instead of hiding it.

graded backtest games
Sooth model, ATS
market close, ATS
breakeven at −110

Where the edge actually is — live from the board

Because the archive holds every book's price on every game, the engine knows the spread between books. Taking the best number instead of the worst is pure, measurable value — no prediction required. These figures are read live from board.json, the same file the public board renders.

avg best-vs-worst reading the live board…
avg best-vs-worst reading the live board…

The edge in dollars — same bets, better price

identical bets
of stake recovered by shopping
saved on the same bets
games with outright arbitrage