Methodology
How an Ascend AI Pick is made
An AI Pick on this site is a specific, dated recommendation produced by one of Ascend's models — a sports selection, an equity idea or an income build — published with the probability the model assigned it, the price it was available at, the stake it justifies and the level that invalidates it. This page explains what a candidate has to clear before it is published, how often picks are reviewed, and what is disclosed.
What qualifies as an AI Pick
It has to be a real, scheduled event
Every pick is attached to a fixture pulled from a live schedule feed and cross-checked against a second independent source. If a game cannot be verified, no pick is published for it — the board is left short rather than filled.
It has to be priced against real sportsbook numbers
Each candidate is matched to the same market at multiple books, and only to the matching market: a total is priced off posted totals, a moneyline off posted moneylines, a handicap off the posted line. The best available price and the number of books compared are shown on the pick.
The model's probability has to disagree with the market
Book prices are de-vigged into an implied probability and compared with the model's own calibrated probability. A pick only survives when the disagreement is larger than the model's own historical error on that market type.
It has to survive sizing
The surviving edge is converted into a fractional-Kelly stake. Selections whose stake rounds to nothing at a sane bankroll fraction are dropped instead of padded onto the card.
How often picks are reviewed, and how they change
- Boards
- Rebuilt every 15 minutes while fixtures are open, so prices on the page track the books rather than a morning snapshot.
- Grading
- Settled selections are graded at the released price — the price shown when the pick was published, not a better one found later.
- Calibration
- Rolling hit rate and Brier score are compared against the model's stated probabilities over a 30-day window. Drift beyond ±4 points of hit rate or +0.02 Brier raises an internal alert and the model version is re-weighted.
- Restatement
- Published accuracy is restated weekly, including the tiers and leagues where the model is losing.
Picks are not permanent. A selection can be superseded when the price moves through the model's indifference point, when a fixture is postponed, or when new information changes an input. Superseded picks stay in the archive with their original numbers rather than being deleted.
Human review versus model output
Boards, prices and probabilities are generated by the models with no manual override — nobody edits a published probability after the fact. What is human is the layer around them: the feature set, the weights, the thresholds, the universe of markets and tickers, and the written analysis. William Merrill builds and reviews that layer, and every editorial page on the site is written by him.
Accuracy, stated honestly
The headline figure on this site is 57.35% across 100,000,000 graded selections. That number is self-published from a reproducible backtest harness in this codebase. It has not been verified by a third party, and no page on this site claims otherwise. Nothing here is a guarantee of a future result, and no pick is financial or betting advice.
Commercial disclosure
Ascend sells access to its own engines as a single upfront payment for a fixed window. Sportsbooks, brokerages and tool vendors do not pay to appear on a board, and no pick placement is for sale. Where a book is named on a pick it is because that book had the best posted price at publication. Full terms are on the disclosure page.