AI Picks

A betting bot with its receipts on the table. It bets only when the model's probability beats the best available price by a margin, sizes each bet in units, and grades every pick after the fight. The page lets you replay its history with your own bankroll and staking rules.

How the bot picks

A leg is the bet type: winner, method (KO, submission or decision) or rounds. The winner leg comes from the value model; the method leg from a market-aware props model; the round leg is switched off. The bot places a winner pick only when the expected value at the best available price is at least +5% (at odds up to 5.00) and a method pick at +10% (odds up to 8.00). Expected value is the average return per unit staked over the long run, from the model's probability against the price; positive means the odds pay more than the risk implies.

The bar under the header holds the leg filter on the left and the Sizing chip on the right; the chip summarises how stakes are shown in money and opens the sizing panel. The leg filter (All legs, Winner, Method, Rounds) applies to everything on the page, including a re-run of the bankroll as if you had followed only those picks.

Stake sizing

The bot sizes in units, from 0.1 to 10, with quarter-Kelly: it bets more when the edge is bigger, at a quarter of the full Kelly amount to cut swings, capped at 10 units (2 for method picks). This panel decides what a unit is worth to you. Units × bankroll (recommended) prices each unit as a fixed share of your starting bankroll; Model units shows raw units with no money; % of bankroll stakes the same share of your current bankroll on every pick, ignoring the bot's sizing, and compounds both ways; Flat amount is a fixed stake.

Starting bankroll is the money you set aside. 1 unit = % of start sets the unit value. Rebase unit every N picks recomputes the unit from the running bankroll at checkpoints, so gains compound and stakes shrink after losses; 0 never rebases. The line under the fields shows where that sizing would have taken the starting bankroll and the ROI on turnover. Risk simulation replays your history 1,000 times with random results drawn from the model's own probabilities and adds a grey band to the bankroll chart: the range 95% of those runs stay inside. Its tiles give the chance the bankroll ever halved, the chance it was wiped out, the average end and the 95% range.

Open picks

Bets waiting for their fight, grouped by event: each event header has a date tile, the countdown, the number of picks and the money at stake. Each row: the result state (OPEN), the photo of the fighter the bot backed, the pick and the fight, a FREE badge on the main event, the stage (T-14d, T-7d or T-1d: when the pick was made), the decimal odds taken and the bookmaker, Model · market (the model's win % in cyan beside the market's vig-free %, with a short bar for each), the EV at that price (green from +20%), the stake under your sizing, and P/L once settled. Tap a row to expand its details.

The main event of every card is free. On the free plan the other picks on a card are blurred until they settle; "N more picks on this card, unlock to see them" opens the card unlock. Picks are made 14, 7 and 1 day before the event, so a card bought early fills in as they land, and the bot may pass on a card entirely if it finds no edge.

Track record

The track record sits right under the header. It is computed in your browser from the settled picks of the selected leg and the sizing you chose. Bankroll is what you would have now if you had placed every settled pick that way, with the change from the start. ROI on turnover is profit divided by everything staked, also called yield; it is not bankroll growth, which divides by the starting bankroll and is usually smaller. Record is wins–losses with the hit rate and the number of void picks (cancelled fights, draws, no contests: refunded and counted nowhere). Open is the stake on picks still waiting.

The bankroll chart plots the balance after every settled pick in fight-date order. The dashed line is what the model expected: where the bankroll would be if every pick had returned exactly its expected value. Sitting above it means running hot, below it running cold; the "vs expected" figure in the corner is the gap. With Risk simulation on, the grey band shows the 95% range. Hit rate can sit under 50% while ROI is positive, because underdog wins pay more than favourite losses cost.

History

Every graded pick, grouped by event with the record and profit for that card. The stage filter (All stages, T-14d, T-7d, T-1d) shows how picks made 14, 7 or 1 day out performed; earlier lines have historically carried a bigger edge. Rows carry the same columns as open picks plus the result: Win, Loss, Void or Pass. A REPLAY badge marks history recomputed after the fact with the current model and closing prices: a backtest, not money that was on the line. Show more events extends the list.

Pick details

Why the bot bet. The Numbers column gives the model probability, the market's vig-free probability, the edge vs market in percentage points, the price and bookmaker, the expected value, the stake in money and in the bot's units, and a 0 to 100 confidence figure that does not change the stake. The Source line names which model produced the pick. Model view is a short narrative from the model's method and round probabilities. For replayed picks, "Original live pick" shows what the first version of the bot actually did. The footer gives the stage, the revision number (how many times the pick was revised at the 14, 7 and 1 day checks) and the last update.