Inside the Lab

The live record

No settled bets yet: the first simulated decisions are made about 90 minutes before opening night's first tip, and graded after the games. Expect few bets early, by design.

Simulated bets at real prices; no real money. Updated with every publish. Wins and losses alike.

Live win probability (the game-flow line on finished games, Prediction tools): Scored on 41 games this season, at six fixed points of each game, 18.1% closer to the result than the score and clock alone. The curve values are the research, not shown; what is shown is the method and this record. Model v2, parameter set #fa3972d7.

How it works

• About 90 minutes before each tip, the Lab prices every market it can (game lines and player props) and records all of them, not just the ones it bets: it learns from every price, so it is not fooled by only looking at its favourites. It re-checks after lineups and after any injury news, and records those too.
• It places a simulated bet only where its edge clears the bar for that kind of bet. No real money is ever bet.
• After the game, each bet is graded and compared with the price just before tip-off (the closing line). Game bets are also compared with the closing price at Pinnacle, the sharpest sportsbook. For player props we only have DraftKings' own closing line, a weaker check that we keep in mind.
• 50% is not break-even. At the usual −110 price you need 52.4% to break even; winning half your bets loses about 4.5% of everything staked.
• Closing-line value is the early proof. Results are noisy for months; beating the closing price consistently shows real skill much sooner.

What to expect early on

Our model has to earn trust before it bets: it starts cautious and gets more say only as its record shows it deserves to, by a rule fixed in advance. So in the first weeks the Lab will place few or no bets and the record above may read 0 for a while. That is the system being honest, not idle: every priced market is still recorded and graded, and that is what the trust is learned from.

How ideas get tested

Many ideas (strategies) compete at once on fresh games they have never seen. Most lose. The rule for calling a winner was fixed before the first bet: it accounts for how many ideas are being tried at once, so a lucky streak cannot pass as skill and adding more ideas cannot manufacture a winner.

An AI analyst starts reviewing the record once the Lab has 20 game days of graded results (it has not run yet: a handful of nights is too little to learn from). From then on it reviews it every week, writes down new ideas, launches new experiments and pauses dead ends (a paused idea keeps its record and can come back). A person reads everything it does.

The honesty rules

• Forward only: every bet is recorded before the game, and frozen.
• Every result is kept, especially the bad ones.
• A model is promoted from Experimental to Proven only by a rule fixed in advance: a long run of settled bets over many game days, measured against closing prices, beating the closing line with room to spare, and a profit. The same rule demotes.
• Promotion is switched off until that rule has itself been tested against random results.

What we will never show, and why

• Individual bets. Publishing them would make us a tipster; the aggregate record is what matters.
• Strategy rules. They are the research itself, and they change as ideas win and lose.
• Podcast audio or transcripts. Not ours to publish; we show short credited summaries.
• Anything about you. There are no accounts, and we never track who reads what.

Nothing here is betting advice. If you bet, bet responsibly: 21+. Gambling problem? Call 1-800-GAMBLER.

Where our numbers come from →