Got a model, or a hunch?
Find out if it wins.

A place to test sports picks against real markets, and to read the picks of models that have proven themselves. Make picks by hand or send a model’s predictions from code; Pickpockt supplies the games, records the Kalshi price at the moment each pick arrives, settles the market and grades it. Your record is private until you choose to publish it. Pickpockt does not accept wagers.

No. The picks feed, every public model’s page and full record, and the leaderboards are open to anyone. An account lets you follow models, save games and record picks of your own.

No. In Studio, on the web or in the iOS app, you pick a game and a side and that’s it — the pick is priced, settled and graded exactly like one sent from code. The API is there for when you’d rather a model or a script send picks for you.

Backtests are easy to fool without meaning to: a feature that leaks the result, a price you could never have got, the best of fifty variants remembered and the rest forgotten. A forward record cannot be fooled that way. Every prediction is timestamped before the event and measured against the price on offer when it was made.

Nothing, if you make picks by hand in Studio. For a coded model: the model, and a job that sends its output. Event discovery, odds capture, settlement, grading and the dashboard are ours. The CLI needs Node 22 or later; the REST API works from anything that speaks HTTP, including an AI assistant with shell access.

No. You send predictions: the event, the market, the side, and optionally your price or confidence. Code, features, training data and model weights stay in your environment.

Only you, until you switch a model to public — and back again whenever you like, with one switch per model in Studio. There is no application and no approval. Predictions are recorded, timestamped and graded from the first one either way, and public always means that model's whole record, including everything it made while it was private.

Yes. An account can hold many models — an NFL spreads model, an NFL moneyline model, a v2 beside a v1 — each with its own record, API keys, public page and Private / Public switch. Give a model a league and a market and anything outside them is refused, so its record is exactly what it claims to be. Publish the ones you are proud of; the rest stay private.

Predictions sent through the API are final the moment they are accepted. Picks entered by hand in Studio can be changed until the event starts. After that, nothing on the event can be added, changed or removed.

Pickpockt runs its own models for NFL, CFB, NBA, MLB, UFC, ATP and WTA, and carries fixtures for leagues it does not model — soccer first, across competitions including the Premier League, La Liga, Serie A, the Bundesliga, Ligue 1, the Champions League and MLS. You can send predictions on any league the platform carries. Where we do not run a model ourselves, fixtures and results come from Kalshi, a regulated exchange whose contracts name their settlement sources.

It depends on whether the draw is something you could have backed. In soccer the moneyline is three-way, so a draw is a real side with its own price: backing a team in a 1-1 is a loss, and backing the draw wins. In two-way markets like UFC and tennis nobody can back a draw, so a drawn result voids the prediction and grades as a push. Roughly a quarter of soccer matches end level, so which rule applies is not a detail.

It is the difference between your model’s stated probability and the probability implied by a sportsbook’s odds, expressed in percentage points. It is a disagreement with the price, not a promise of profit — the record is what tells you whether the disagreements pay.

By net units at flat 1u stakes, using the Kalshi price recorded when each prediction was published (see below). The overall board ranks every public model across every league and market; each league, each market, and each league-and-market pair has its own board, ranked on just those calls — so an NFL spreads model is compared with NFL spreads models. A model needs a minimum number of graded predictions on a board before it can hold a place there, so a short hot streak cannot take it. Win rate excludes pushes. Predictions with no recorded price are excluded from units. Closing-line value is measured on moneyline predictions only. The Leaderboard can also rank by Kelly: each prediction staked at full Kelly on a 100u bankroll, sized from the model's own probability against that same book price, and skipped where the model shows no edge — which rewards calibration as well as picking winners.

Kalshi’s. When a prediction arrives we record the latest Kalshi price for that side, and every figure — units, ROI, Kelly, closing-line value, the leaderboards — is measured against it. The price is the midpoint of Kalshi’s bid and ask, before fees, so it is slightly kinder than what you would pay to take the trade. Where we don’t carry a Kalshi market for a game, which today means most NFL and NBA lines, we use the DraftKings price instead. A prediction with no recorded price at all still counts toward the record, but not toward units.

No. Pickpockt runs its own models the same way anyone can: one per league and market (PickPockt NFL Moneyline, PickPockt UFC Moneyline, and so on), owned by an ordinary account, graded by the same rules and ranked on the same boards, where any of them can lose its place. New accounts follow them so the picks feed is not empty on day one, and you can unfollow any of them like any other model.

Nothing. Pickpockt is free to build on and free to read, with no paid tier. It is supported by tips, which go to the platform rather than to individual predictors.