How the models are tested and graded

How a model is tested

• Walk-forward. In a test, every game is priced only from what was known before it. Ratings and stats see only earlier games, so a result never leaks into its own prediction.
• Tuned, then frozen. Settings are chosen on one span of seasons (for UFC, of fights). The model is then frozen and scored on later ones it never saw.
• Scored on the probability. A test scores the probability the model gave, not only whether its pick won, so a confident miss counts against it more than a close one.
• A change has to hold up. A new input or setting ships only if it improves on the current model where it was tuned and still does on the held-out games. Most ideas do not pass, and they stay out.

Each sport's model page shows how its model has been tested and how it has done:

How live picks are locked and graded

• Locked before the start. A pick is stored before kickoff, tip-off or the fight. It can update as lines and news change until then; the version in place at the start is the one graded, and it cannot be edited afterward.
• Graded after the final. Once the result is final, the pick is graded automatically.
• Winners are the record. Each sport's record is winners called: locked picks that named the winner, out of all graded picks. Misses stay in, and nothing is deleted.
• Backtests are kept apart. Games priced after the fact are labelled as backtests and never count toward the record.
• Set beside the betting favorite. Every record shows what the favorite (the side with the shorter no-vig moneyline when the pick was locked) did on the same games, and how our picks went when we took the other side. A record under 30 graded picks is marked too early to read. Where a pick was locked before a sport's winner call followed the market, it is marked as the model alone. Every counted pick is in the public ledger, as a table, CSV or JSON.

Calibration

Every pick comes with a probability, and calibration checks whether those probabilities mean what they say: of the picks the model gave about 70%, about 70% should have won. Each sport's Results and model pages show this bucket by bucket, and the live buckets fill in as picks grade. A 70% pick still loses about three times in ten; that is the model working as designed.

What the models look at

Categories only. The exact inputs, and how they are weighted, are not published.

Football prediction pages also show the model's own margin and total beside the market's. Those are context, not picks: the model makes no spread or over/under call, and they are not part of the record.

Not published: the exact inputs and their weights, the settings and cut-offs, the rating numbers, the size of any adjustment, and what was tried and rejected. Always public: every locked pick, its result, the record and the calibration.

Player props

Props boards and player cards are research, not picks.

• History against the line. How often a player has cleared a posted number in their own past games, overall and by split (recent games, home or away, the opponent and more), with a range beside each rate so a small sample is not mistaken for a trend.
• The books' view. Beside each hit rate, the books' own chance of the over, with their margin (the vig) removed where both sides are posted, so a player's history is compared with the price rather than with 50%.
• Projected ranges. Where a card shows a median and range for a player's next game, it is context only: never a line marker, an over/under call or a pick.
• Prop picks. The Props tab under Predictions comes from a separate prop model. Its reads appear there only once they have been checked against real closing lines; until then they are recorded, not published.

What we don't do

• No against-the-spread or over/under records, for our picks or for any team. Winners are the record.
• No parlay builders, round robins or bet planners.
• No "lock of the day" and no guarantees. Every pick is a probability, and probabilities miss.
• No bets taken and no personalized betting advice. See About and the Guide for how to use the site.