How it works, in plain words.

This site helps you judge a bet before you place it. It gives a chance for every game and every player pick. And it keeps a public record of how its picks did.

The numbers are early. They are still proving themselves. So Our record shows every settled pick with its result and date. Look there before you trust any number here.

Two models make the numbers. Model A (ALL_IN) is built by Grok. Model B (FULL_SLATE) is built by Muse. Where both exist, the headline number is the simple average of the two. Neither model is a proven winner. Both are forecasters. The record is the proof, and it is still being built.

Game forecasts (Who wins?)

Each game gets one headline number: the team we think wins, and how often we think it wins — for example, “Knicks 87%”. Under it, Model A (ALL_IN) and Model B (FULL_SLATE) show the two numbers behind it, where their numbers are in this snapshot. A game close to 50-50 is marked as a close call — barely better than a coin flip, not a pick to trust.

A chance comes from past games. How well teams score and stop the other team. Recent form. Rest. Playing at home. Recent games count most. News that comes after the numbers are made — injuries, late lineup changes — isn’t in them. So check the news before you use a pick. Every number carries its date.

One guardrail applies to Model B. If Model B’s chance is more than 15 points away from the book’s chance, we pull it back to within 15 points. See Enhancements below for why.

Player picks

Player picks give a chance a player goes over or under a line. Points, shots on goal, catches and yards, and more. The chance comes from recent games — the most recent ones count most. It also looks at how good the other team’s defense is against that kind of player. Warnings flag injuries and other reasons to be careful.

Every player-pick chance goes through calibration before it is shown. That means we checked it against past seasons and adjusted it so the quoted chance matches how often it actually hit. See Calibration below.

Touchdown picks work the same way. How often a player gets the ball near the goal line. Plus the weather and how the game should play out. The chance shown is our guess, not a book’s price. Hold it next to a real payout and see if the payout is worth it. Touchdown picks only cover quarterbacks, running backs, receivers, and tight ends. And if the gap between our chance and the book’s price is more than double, the pick stays on the board with a “Longshot” tag next to it. That kind of gap usually means our number is wrong, not that we found a bargain — the tag is just a label, and it does not change the chance shown.

Picks in the same game can affect each other. If teammates are on the same ticket, one scoring can make the other less likely. So the real chance of the whole ticket is lower than just combining the single chances.

Simulations

We play each game 10,000 times on a computer before we show you a number. The computer uses the same inputs as the model — team strength, home field, rest — and adds the normal randomness of a real game. Then it counts how often each outcome happened.

For games, the simulation gives a second opinion on who wins, and a range for the total points and the margin. The range is shown as a band, for example “total 41 to 58”. That band is the middle 90% of the simulated games. 5% of the simulated games were below it. 5% were above it. The band is wide because real games are noisy. A narrow band would be a lie.

For player props, the simulation plays the player’s game 2,000 times and counts how often he goes over the line. Then we compare that number with the model’s calibrated chance. If the two disagree by more than about 10 points, the row says so in plain words: “Our model and the simulation disagree by N%”. A disagreement does not mean one of them is wrong. It means the pick is uncertain, and you should know that.

The simulation is also how we price totals and spreads. The chance a total goes over, or a team covers the spread, is the share of the 10,000 simulated games where it happened.

What the simulation cannot do: it cannot see injuries, weather changes, or a coach’s game plan. It only knows what past games looked like.

Calibration

A model that says “70%” should be right about 70 times in 100. Most models are not. They are too sure of themselves. Calibration fixes that.

Before any player-pick chance is shown on this site, it passes through a walk-forward calibration check. Walk-forward means we tested the model on past seasons it had never seen, in order, the way it would have run live. Then we built a table that adjusts the model’s chances so they match how often they actually hit. Only the adjusted chance is shown. This rule has applied to every prop since November 3, 2026.

The calibration tables are not rebuilt every day. They are rebuilt only when 100 or more new graded picks have piled up for that market. That is because a small sample makes the table worse, not better. Until the table is rebuilt, the site keeps quoting the old table.

Every morning we grade yesterday’s picks and check for drift. Drift means the quoted chances and the actual results have pulled apart. We flag it when a family has 20 or more graded picks and the gap is more than 10 points (a warning) or more than 15 points (a drift flag). Right now, for example, the NFL props are not flagged: they were quoted at 48.8% on average and actually hit 47.7% over 608 graded picks, a gap of 1.1 points — well inside the warning line. When a family is flagged, the flag is shown in our notes, and the family stays under watch until the sample grows or the table is rebuilt. We do not quietly change the chances to hide a bad run. The record shows it.

The same daily grading also feeds the silent grading log. See Our record below.

Enhancements

These are the extra rules we put on top of the models. Each one exists because the plain model made a mistake we could measure.

The 15-point market cap. Live since November 8, 2026. When Model B’s chance that the home team wins is more than 15 points away from the book’s chance, we pull Model B’s number back to within 15 points of the book. The book’s chance is the book’s odds with the book’s cut taken out (see Concepts: de-vigging). The cap applies to college football, the NBA, the NHL, baseball, and soccer. It does not apply to the NFL or college basketball yet, because we have not tested it there. Why we added it: two college football games had Model B at 66% and 73% while the book had them near 29% and 45%. Both lost badly. The model was too sure of itself when it disagreed with the market that much. The cap keeps the forecast honest without throwing the model away. The cap is applied once, in the morning run. It is never re-applied during the day.

The touchdown guard. Two rules. First, only quarterbacks, running backs, receivers, and tight ends can appear as touchdown picks. If we do not know a player’s position, he is left out. Second, a touchdown pick whose gap is more than double stays on the board, but it gets a “Longshot” tag next to it. The tag is a label only — it does not change the chance shown, the tier, or whether the pick can make the card. If a line move brings the gap back to double or under, the tag comes off. Before this rule, the top 100 board picks included 55 touchdown picks and the biggest gap was over 2,200%. Touchdown picks still run hotter than other props. We watch them.

Returning starters and the backup problem. Prop models look at a player’s last few games. They do not know who else is playing. That breaks when a backup has been filling in for a hurt starter. On November 8, Braelon Allen is the example. He had been the lead back while Breece Hall was hurt for five games, so his recent games looked like a starter’s. The model gave him 90 yards and a 77% chance to clear 34.5 rushing yards — but Hall was back that day, so that number was for a backfield that no longer existed. We tested a fix that discounts a backup’s numbers from the games the starter missed. For the NFL it failed: it made rushing predictions 5% worse and receiving predictions 1.7% worse. For college football it worked big — rushing 25% better, receiving 23% better, and passing 32% better — but we cannot use it there without a reliable college depth-chart feed to tell us who the starter actually is. So the discount is not live anywhere. What is live is the morning news check: it flags a returning starter, so a pick built on a backup’s run gets a warning before you use it.

The consensus check. Before the card is picked, we compare each candidate with the sharp book’s price (Pinnacle) and the other books. If the books agree with us, the pick is marked confirmed. If they disagree, it is marked contradicted. This check never removes a pick. It is a note, and a tiebreaker when two picks are otherwise equal. We do not have a feed of where the public’s money is going, so there is no “fade the public” signal here.

Adaptive value bars. The card only takes picks whose edge over the book is at least 10%. Edge means our chance divided by the book’s chance, minus one. But the bar moves with the evidence. If a family has 30 or more graded picks and it is cold — losing money on paper, or hitting more than 15 points below its quoted chance — the bar rises to 20%. If a family is warm — breaking even or better, and hitting within 10 points of its quoted chance — the bar drops to 5%. The bar is never below 5% or above 20%.

Shrinkage toward the market. For some props, a player with only a few games of history gets his chance pulled toward the book’s chance. A small sample is noisy. The book’s price is a second opinion. The pull is stronger when the sample is smaller.

What we tested and killed. Not every idea helped. We tested a goalie adjustment for the NHL, a Bayesian shrinkage version of the receiving model, offensive-line injuries, referee tendencies, play distribution, situational spots like lookahead and letdown games, and the returning-starter discount for the NFL. None of them beat the plain model in walk-forward testing, so none of them is live. The returning-starter discount made NFL rushing predictions 5% worse, so it stays out. The receiving model’s Bayesian version is the closest call: it was tested and it did not beat the current version, so the current version stays. We list the killed ideas here so you know what we tried and why it is not in the numbers.

Concepts and methods

These are the ideas and tools behind the numbers. Each one is named plainly, because you should be able to check our work.

Elo ratings. A single number for how good a team is. It goes up when a team wins and down when it loses. A win against a strong team moves it more than a win against a weak team. Recent games count more than old ones: soccer ratings fade with a 400-day half-life, so a game from a year ago counts about half as much as a game from today. The NFL rating also pulls partway back toward average between seasons. Elo is the base of every game model here. On its own it is already a decent forecaster. The other features sit on top of it.

Logistic regression. A model that turns a list of numbers into a chance between 0% and 100%. It is simple, it is easy to check, and it rarely fools itself. The NFL game model is logistic regression on 11 features: the Elo gap, home field, each team’s Elo, recent form, how well each team scores and stops the other team, and rest. The NBA game model is logistic regression on 5 features. Baseball and hockey are logistic regression on Elo alone, because adding more features did not help in testing. The touchdown model is logistic regression on 20 usage features, like how often a player gets the ball near the goal line.

Gradient-boosted models. A model built from many small yes-or-no splits, stacked together. It can find patterns a straight-line model misses. It can also overfit — learn noise and call it signal. We use it for college basketball, where it beat the simpler model in 8 out of 8 tested seasons. We tested it for the NFL too, and it lost to logistic regression there (log loss 0.6765 against 0.6429), so the NFL keeps the simpler model.

Negative-binomial distributions. A way to describe how many times something happens in a game — 3-pointers made, shots on goal, rebounds. The simple version (Poisson) assumes the count is steady. Real counts are not steady: some games a player is hot, some games he is cold. The counts spread out wider than Poisson allows. That is called overdispersion. The negative binomial allows for the extra spread. We use it for NBA 3-pointers (fitted overdispersion 0.3426), NHL shots on goal (it beat Poisson in walk-forward testing on 3,010 player-games), and college rebounds and assists. For college points we tested it and rejected it: it underestimated the chance of going over low lines, so points use a normal distribution instead.

Walk-forward testing and calibration. We never test a model on the same games it learned from. We train it on past seasons, test it on the next season it has never seen, then roll forward. That is walk-forward testing. It is the only test we trust. Calibration is the step after: we compare the model’s quoted chances with what actually happened, and adjust them so a quoted 70% hits about 70 times in 100. Every prop chance on this site passes through that adjustment before it is shown.

Log loss and Brier score. Two yardsticks for how good a chance is. Log loss asks: how surprised were you by what happened? A chance of 90% on something that did not happen is a big surprise, and log loss punishes it hard. Lower is better. Brier score asks: how far was your chance from the result, squared? A chance of 70% on something that happened scores (1 − 0.70)² = 0.09. Lower is better. We report log loss for win and prop models, and the average miss in points for totals and spreads models. Every number on this page comes from walk-forward testing, not from the games the model trained on.

Monte Carlo simulation. Playing a game thousands of times on a computer with randomness added, then counting the outcomes. Games are simulated 10,000 times. Player props are simulated 2,000 times. See Simulations above.

De-vigging the book’s odds. A book’s odds include the book’s cut. If both sides of a bet are priced at −110, the two implied chances add to about 105%, not 100%. The extra 5% is the book’s cut. De-vigging takes the cut out so the two sides add to 100%. That gives the book’s true chance, which is what we compare our number with. The value of a pick is our chance divided by the book’s de-vigged chance, minus one.

Bayesian shrinkage. A way to handle small samples. When a player has only a few games of history, his average is noisy. Shrinkage pulls his number partway toward the league average, more when the sample is small and less when it is large. We tested a Bayesian shrinkage version of the receiving-yards model. It did not beat the current version in walk-forward testing, so it is not live. The current model stays.

Recency weighting and trailing averages. Recent games count more than old ones. The receiving-yards model uses a trailing-4 average: the last four games, weighted toward the most recent. NBA 3-pointers use a trailing-10 average. College props use a trailing-8 average with at least 4 games of history. A player who has been hot lately gets a higher number than a player with the same season average who has gone cold.

Efficiency and rest features. The game models look at more than wins and losses. Efficiency means how well a team scores and stops the other team, adjusted for how good the opponent was. Rest means days off and back-to-back games. The NFL model uses both. The NBA model uses rest and back-to-back flags; form and efficiency gaps were tested and dropped, because they read as noise on 7,740 test games. College basketball builds its own opponent-adjusted efficiency, because the public version only covers the top 25 teams.

The 15-point market cap. Covered in Enhancements above. When Model B disagrees with the book by more than 15 points, we pull it back to within 15 points. It is a guardrail, not a model.

Our record

Our record lists only bets that have finished. Date, pick, odds, result, and the profit at $100 per bet. Today’s bets stay on Our Bets until their games finish. The record started fresh on Nov 6, 2026. The card only includes games dated that day.

How the card is picked. Each day the card has 1 to 4 straight bets plus 1 Parlay of the Day. Picks are ranked by likelihood — the model’s chance it hits — among picks that pass the value screen. The value screen is an edge of at least 10% over the book (see Enhancements for the adaptive bars). The card can be any sport and any bet type. Soccer is excluded from the public record. The parlay has 2 legs, both from the value-passing pool, from different games, no leg longer than +500, and the combined odds are capped at 10 times the stake. At most 1 college football prop per card. A pick is dropped if its game has already started. Every pick is graded at $100 flat. A pick that lands exactly on the line is a push. A player who does not play stays pending.

Silent grading. Every morning we also grade the whole board in the background — every prop and every game forecast, about 25 times as many picks as the card. That grading never touches the public record. It exists only to feed the calibration tables and to show us which families are running hot or cold. On November 8, for example, it graded 355 events from the November 5 to 7 boards. The public record is only the card. The silent log is the lab notebook.

Closing-line value (CLV). We also track whether our picks beat the closing price at Pinnacle, the sharpest book. For each card pick, we record the price when we picked it and the Pinnacle price near game time. If the closing price moved our way, that is positive CLV. Positive CLV over many bets is the best early sign a model has an edge, because the closing price is the market’s best guess. So far we have 53 closes recorded, averaging about +1.6 points. That is a small sample. It is suggestive, not proof. We need around 75 closes before we draw any conclusion, and we are not there yet.

The record so far. After November 8: straights 6–6, −$157.79. Parlays 0–3, −$300. Total −$457.79 on paper. That is a losing record. We show it because the record is the point. If it stays negative, we change the selection rules and say so here.

Your money over a season

Our record also shows what happens to $10,000 over a season of 250 bets, played 10,000 times on a computer. Two ways of betting: the same $100 on every bet, and a bet sized to the edge. Bigger when the edge is bigger, smaller when it’s smaller, never more than 10% of what you have. These numbers are a teaching example, not our record. One row shows a real edge (winning 55 times in 100 at −110 odds). One row shows the same bets with no edge at all. One row shows what happens when you size bets for an edge you don’t have. Our own picks will run through the same simulation once 30 of them are graded. Until then that part of the page says so plainly. No way of betting shown here is a promise it wins. Even with a real edge, the sized bets end higher in the middle but finish ahead less often and dip further along the way.

Our models

These are the 30 models behind the numbers on this site — 29 of our models (together, Model B — FULL_SLATE) plus ALL_IN (Model A) as the 30th. Each one focuses on one job.

Each number below comes from testing the model on past seasons it had never seen. For win/loss models, the score is log loss. That’s how far the model’s chances were from what actually happened. Lower is better. For totals and spreads, the score is the average miss in points. Lower is better. R² says how much of the result the model explains versus a simple average. Higher is better. Recalibration means we adjusted the model’s chances using past seasons, so the quoted chance matches how often it actually hit.

The honest headline first: every model below is a forecaster, not a proven edge. Each one beats simple baselines like home-always and Elo-only in walk-forward testing. None has beaten the book’s odds across multiple seasons. The only model to beat the market in even one season is Eredivisie soccer, in 1 of 5 seasons. Nothing here is a guaranteed winning system. Forecasts are the product. Whether there is an edge is the lab’s open question, and the record is how we answer it.

Game models — who wins (6 models)

NFL. Logistic regression on 11 features (Elo gap, home field, each team’s Elo, recent form, scoring and defensive efficiency, rest). Trained on 2,033 games. Walk-forward log loss 0.6429, against 0.6448 for Elo alone and about 0.69 for always picking the home team. So it beats Elo alone by a hair. A gradient-boosted version was tested and lost, so the simple model stays. Moneyline backtesting shows no edge.

College football. Elo plus efficiency. Trained on 7,193 games. Walk-forward log loss 0.5709. Beats Elo alone. Forecaster only.

NBA. Logistic regression on 5 features (Elo gap, home field, rest, back-to-back flags). Trained on 8,882 games. Walk-forward log loss 0.6260 and Brier score 0.2181, against 0.6272 for Elo alone. Form and efficiency gaps were tested and dropped — they read as noise on 7,740 test games.

Baseball. Elo alone, as a logistic regression. Trained on 17,902 games. Walk-forward log loss 0.6795. Adding more features did not help in testing, so we ship the simple version. Forecaster only.

Hockey. Elo alone, as a logistic regression. Trained on 9,275 games. Walk-forward log loss 0.6690. Same story as baseball: the extra features did not help, so the simple version ships. Forecaster only.

College basketball. Gradient-boosted model on Elo gap plus opponent-adjusted efficiency that we compute ourselves (the public version only covers the top 25 teams). Trained on 65,041 games over 12 seasons. Walk-forward log loss 0.5081 and Brier score 0.1714, beating the previous version (0.5221 and 0.1761) in 8 out of 8 tested seasons. Forecaster only. There is no Model A for college basketball.

Touchdowns (1 model)

Anytime touchdown, NFL. Logistic regression on 20 usage features, like how often a player gets the ball near the goal line. Trained on 38,671 player-games from 2019 to 2026. Walk-forward log loss 0.4798 and Brier score 0.1538. It beats the plain position base rate in all 6 tested seasons. Probability skill only — no edge claimed, because no archive of historical touchdown odds exists to test against. Known limits: no injury data, and it looks back about 4 weeks.

Player props (9 models)

Every prop chance passes walk-forward recalibration before it is quoted. That rule has applied since November 3, 2026.

NFL receiving yards. Two-part model: the chance a player gets any yards, times the yards if he does. Trained on 34,346 player-games. Walk-forward R² 0.331 against 0.240 for a trailing-4 average, and average miss 18.85 yards against 19.67. Wins in all 6 tested seasons.

NFL receptions. Two-part model with recalibration. Walk-forward R² 0.338 against 0.280 for the baseline. Wins in all 6 tested seasons.

NFL rushing yards. Two-part model with recalibration. Walk-forward R² 0.395 against 0.333 for the baseline. Wins in all 6 tested seasons.

NFL passing yards. Single Ridge model. Average miss 63.6 yards against 69.9 for the baseline, about 9% better. Wins in all 6 tested seasons. This one uses its raw chances, because recalibration did not help. Quarterbacks with no history are left out honestly.

NBA 3-pointers. Negative binomial on a trailing-10 average, because 3-pointer counts spread wider than Poisson allows. Opponent defense was tested and did not help, so it is not used. After recalibration, the average error by line is under 0.05 at every line from 0.5 to 3.5.

NHL shots on goal. Negative binomial plus recalibration. It beat Poisson in walk-forward testing on 3,010 player-games.

College basketball points, rebounds, assists. Trailing-8 averages with at least 4 games of history. Points use a normal distribution (the negative binomial was tested and rejected — it underestimated the chance of going over low lines). Rebounds and assists use the negative binomial. All three are pulled heavily toward the market price when the sample is small. Walk-forward tested on 103,167 predictions, with recalibrated errors under 0.05 at every line. College prop lines are confirmed live before a pick is shown.

A note on NBA points, rebounds, and assists, and college football props. These are priced from trailing-mean tables with the same walk-forward recalibration as the models above, and they carry their own calibration tables. They are not standalone models in the registry. We list them here so you know they exist, and we do not dress them up as more than they are.

Totals (4 models)

Ridge models that forecast total points. All are forecasters. None claims an edge. A lean needs a gap of 2 points for the NFL and 3 points for college football and college basketball.

NFL totals. Average miss 10.749 points against 10.901 for a naive baseline. Won 11 of 12 tested seasons.

College football totals. Average miss 13.18 against 13.20 for the baseline. Won 6 of 8 seasons. The improvement is tiny, about 0.2%.

College basketball totals. Average miss 13.751 against 13.800 for the baseline. Won 7 of 8 seasons (it lost only the COVID season).

NBA totals v2. Average miss 14.834 against 14.968 for the baseline. Won 5 of 5 seasons. Now live on the board. The old version lost to the naive baseline, so it was replaced.

Spreads (3 models)

Ridge models that forecast the home team’s margin. All are forecasters. None claims an edge. A lean needs a gap of 2 points for the NFL and 3 points for college football and college basketball.

NFL spreads. Average miss 10.272 points against 10.993 for the baseline. Won 12 of 12 tested seasons.

College football spreads. Average miss 13.99 against 14.49 for the baseline. Won 7 of 8 seasons.

College basketball spreads. Average miss 10.183 against 11.406 for the baseline, about 11% better. Won 8 of 8 seasons.

Soccer (6 models)

One model per league: Premier League, Eredivisie, La Liga, Bundesliga, Serie A, Ligue 1. Time-decayed Elo (400-day half-life) feeding a model that gives three chances: home win, draw, away win. All ship as forecasters only. In every league, the market’s number has been closer to what happened than ours: the ranked-probability score is 0.2018 against the market’s 0.1949 in the Premier League, 0.1932 against 0.1876 in the Eredivisie, 0.1991 against 0.1926 in La Liga, 0.2065 against 0.1974 in the Bundesliga, 0.1985 against 0.1916 in Serie A, and 0.2086 against 0.2017 in Ligue 1 (lower is better). The Eredivisie is the best of the six: it beat the market in 1 of 5 seasons, the only model on the site to beat the market in any season. There is no public track record for soccer, because the data license is for private use only, not commercial use. Soccer is Model B only — there is no Model A for soccer.

Model A — ALL_IN (1 model, the 30th)

A second forecasting source, built by the Grok bot and supplied as a feed. It covers the NFL, college football, the NBA, the NHL, and tennis. Its core is a ratings-only chance, with market spread and total and per-game flags for quarterback, goalie, availability, and weather. The pipeline uses the ratings-only number, not the market-blended published number, so the market is not counted twice. The feed is polled every 15 minutes. Model A does not cover college basketball. The headline number on the site is the simple average of Model A and Model B where both exist.

Per-sport notes

NFL. Game model, totals, spreads, and five prop models (touchdown, receiving yards, receptions, rushing yards, passing yards). The game model barely beats Elo alone. The props are the strongest part of the suite, and the receiving model is the most tested. The props are not flagged for drift right now: quoted 48.8%, actually hitting 47.7% over 608 graded picks. No edge is proven.

College football. Game model, totals, spreads, and prop tables. The game model is solid (log loss 0.5709). Totals barely beat the baseline. College football receiving props have run cold. The 15-point market cap applies here — it was added after two big model-versus-market misses both lost badly.

NBA. Game model, totals v2 (built, not yet wired), and props for points, rebounds, assists, and 3-pointers. The game model barely beats Elo alone (0.6260 against 0.6272). The 3-pointer model is the most carefully calibrated prop on the site. NBA props are on watch for drift.

NHL. Game model (Elo alone) and a shots-on-goal prop model. The game model is Elo alone because nothing else helped. The shots-on-goal model uses the negative binomial and is recalibrated. NHL props are on watch for drift. A goalie adjustment was tested and failed — Elo already captures it — so it is not live.

Baseball. Game model only, Elo alone. The simplest model on the site, because nothing else helped in testing. No props, no totals, no spreads model yet. Forecaster only.

College basketball. Game model, totals, spreads, and three prop models (points, rebounds, assists). The game model is the strongest in the suite by its test scores, beating its previous version in 8 of 8 seasons. There is no Model A here, so the headline is Model B alone. College prop lines are confirmed live before a pick is shown.

Soccer. Six league models, Model B only, forecasters only. The market beats us in every league. The Eredivisie is the closest. No public track record, because the data license does not allow commercial use. Soccer is excluded from the public betting record.

Tennis. Model A only — there is no Model B tennis model. The numbers come from the ALL_IN feed, which logs three times a day plus an early morning push. Tennis moves fast, so the numbers carry their own time stamp and a stale warning if they are more than 14 hours old. Over the full log, the strongest tennis signal has been the small-value picks on the women’s tour, not picking winners. That is a small sample, not a proven edge.

Every model gets tested on past games before it goes live. And we keep score in Our Record, so you can see how they’ve done.

Like the site? Support it here — tips are optional and never affect any pick.