How we train, test and judge the EP Model
Every player prop gets one number, the EP Score. This is how the models behind it are built, how a change earns its place on the board, and how we grade ourselves every week.
Elite Picks Research · NFL · NBA · MLB · PGA
How a model is trained
We do not run one giant model. Each kind of prop gets its own: NFL receiving yards, MLB strikeouts, NBA points, and so on. Each one learns from several seasons of play-level history, and from this season's games as they are played.
Three kinds of information go into every read:
The player. His role, his usage and his recent form, measured over several stretches of games rather than one hot week.
The matchup and the game. The defense he is facing, the kind of game it is likely to be, and conditions like the weather and the park.
The number. The posted line and the best price across sportsbooks. A read only matters against the number you can actually bet.
Two rules hold for every input. It has to be available live, before the game, or it does not go in. And public betting percentages stay out, for reasons we have written about.
Before a model is chosen, we try thousands of combinations of inputs and weights against past seasons and keep the ones that hold up.
How a change gets tested
An idea does not reach the board because it sounds right. It goes through four steps.
Past seasons, walked forward. The model trains on what came before and is scored on weeks it has never seen, one after another, the way a real season arrives.
Every slate counts. A test's headline number includes the cold weeks. Nothing is left out because it looked bad.
Side by side with the live model. A candidate often runs in the background on real boards, scoring the same props as the live model, before anyone sees it.
Reviewed before release. A change ships only after a review of what changed and how it tested.
How we judge a week
We grade the board you saw. When a game locks, the line, the side, the EP Score and the best price are frozen, and that is what gets graded. Not a friendlier line from earlier in the week, and not a score that moved afterwards.
Then we look at the week five ways:
The top of the board. The top five scores on each slate. That is what most people act on, so it is measured first.
The 90+ band. The record and the return for every score of 90 and above.
Every category. Receptions, rushing yards, strikeouts and the rest are graded apart, so a hot category cannot hide a cold one.
Overs and unders. Each side is graded on its own.
Return, not hit rate alone. Results are counted at the price that was available. The same hit rate can win or lose depending on the odds.
Through October 7, 2026, across NBA, MLB and NFL. Hit rate is not profit, because prices vary by prop.
The weekly loop
During the NFL season the same four things happen every week, in the same order.
Each pick is graded against the box score minutes after its game ends.
The whole week is graded again against a second source of stats, and the season's record is updated category by category.
The scoring is tuned on the season so far. Reads that keep delivering hold the top tiers. Reads that have not are scaled back.
We send what held, what didn't and what changes. Then next week's board posts.
Baseball and basketball run the same loop on a daily clock: picks are graded the morning after, every day.
What the EP Score is, and isn't
- One number per prop. The strongest reads score 90 to 99.99. Everything else scores 75 to 90.
- A ranking, not a probability. It says how strong the read is at the price. It is not the chance the pick wins.
- It moves with the line. When the number or the price changes, the score changes with it.
- It shows its work. EP Model Insights list what moved a score up or down that day, in plain words.
- It is not a guarantee. A 95 can lose. Over a season, the top of the board is what we expect to hold up, and it is what we measure.
To see where the model is running hottest right now, open Model Trends in the app. For the math behind value at a price, start with expected value.
Elite Picks is a research and analytics platform. We do not accept wagers, hold gambling licenses, or facilitate any real-money transactions. All content is for informational and research purposes.
What data does Elite Picks use?
Several seasons of play-level history for each sport we cover, this season's games as they are played, and lines and odds from a wide set of US sportsbooks and exchanges, captured every 15 minutes. Public betting percentages are not used.
How often do the models change?
In the NFL, the scoring is tuned every week on the season's graded results. In daily sports, picks are graded every morning. Larger model changes ship when they have passed testing on past seasons and on live boards, not on a fixed calendar.
Why does the EP Score sometimes disagree with the price?
The model reads the player, the matchup and the game, then compares that with the line and the price on offer. When the two are far apart, the score is high. When they agree, it is not.
Where can I see how the model is doing?
Model Trends in the app shows the categories where the model is running hottest. During the NFL season we also send a weekly recap of what held, what didn't and what changes for the next week.
Is Elite Picks free?
The charts and basic stat displays are free, the same category of hit-rate splits some tools charge a subscription for. Premium unlocks the full EP Score across the board, daily pick newsletters with research attached, advanced tools, and the AI Agent.
Do you guarantee winners?
No. Sports betting is probabilistic. The EP Score tells you where the model sees the most value at the price, not which tickets will cash. Long-term results also depend on bankroll management, lower risk exposure (more straights, fewer parlays), and a consistent approach.