How Careless Predict Works
Careless Predict is the model behind every win probability, projected score, spread and total on Careless Sports. It covers the NFL, the NBA and FBS college football. This page explains what it looks at, how it's trained and checked, and where it falls short.
What It Predicts
For any matchup Careless Predict gives each team's chance to win, a projected score for each side, a point spread and a total. The same numbers appear on game pages, the predictors, the Lines page and the Games pages.
It is a statistical model, not a sportsbook and not betting advice. Its line is compared with the Vegas line so you can see where the two disagree.
The Data
Every finished regular-season and postseason game from ESPN, with box scores: NFL games from 2010, NBA games from the 2011-12 season and FBS college football games from 2014. Closing betting lines are stored next to the games, from ESPN and, for the NFL, from nflverse's public schedule file.
The Ratings
Each team's ratings are rebuilt game by game in date order, so a rating only ever knows about games played before it. Recent games count most: weights halve after a set number of days and last season's games get an extra discount.
- Elo: a running strength rating that moves more after a big win than a close one and drifts back toward average between seasons.
- Points ratings: each team's offense and defense from recent scores, adjusted for the opponents it played (NBA and college football).
- Yards per play and turnovers, adjusted the same way (football).
- Starting quarterback (NFL): each passer's adjusted net yards per dropback against the league average. Before a game the model reads ESPN's depth chart for the expected starter and compares him with the quarterbacks the team's ratings were built on, so a backup starting moves the line.
- Rest (NBA): playing the second night of a back-to-back.
- Home field, set to zero at a neutral site.
The Models
The differences between the two teams' ratings go into a ridge regression for the margin, which is the spread, and a logistic regression for the win probability. In the NFL the win probability comes from the spread itself instead: the chance that a game with that projected margin ends with a win, given how far real NFL margins usually land from the model's, so the win probability and the spread always agree. The total comes from the points ratings' projected score, calibrated around the current league average. Both models are symmetric: every training game is also added with the teams swapped, so neither side is favored by the math.
On the Lines page, Analyze splits the model's spread into the part each rating contributes, so you can see why it differs from Vegas.
Weekly Retraining and Checks
The models are retrained every week on all games played so far: college football and the NBA on Sundays, the NFL on Tuesdays. A new model is only published if it passes an accuracy check on recent seasons it wasn't trained on and isn't clearly worse than the current one over the last 28 days. Settings are tuned by hand on older validation seasons, never on the most recent ones.
How Picks Are Graded
Every few hours Careless Predict logs its pick and the Vegas line for each game starting in the next 36 hours. Once a game starts, the site shows that logged pre-game pick, never a fresh prediction, because today's ratings already include the result. After the final, the pick is graded against the score and the Vegas favorite. The full running record, by sport and day, plus walk-forward backtests by season, is on Model Analytics.
What It Can't See
Vegas still beats the model on calibration in every sport, and usually on picking winners too. Treat the model as a second opinion with a public track record.
- Injuries other than the NFL starting quarterback. Injuries are listed next to predictions but don't change them.
- Weather, travel, coaching changes, motivation and late lineup news after the last update before kickoff.
- Anything about a team that hasn't shown up in its results yet, which matters most in the first weeks of a season.
Careless Predict FAQ
Is Careless Predict AI?
It is a machine learning model: logistic and ridge regressions trained on team ratings built from past games. It is not a language model. The written previews and Analyze on the Lines page use a language model to explain the numbers, but every number comes from Careless Predict or the game data.
How accurate is Careless Predict?
Its live record, every pick logged before kickoff and graded after the final, is on Model Analytics, next to how the Vegas favorite did in the same games, along with season-by-season backtests.
Which sports does it cover?
The NFL, the NBA and FBS college football.
How often is it updated?
The models are retrained every week. Team ratings in between come from the latest weekly model, and NFL predictions check ESPN's depth chart for the expected starting quarterback each time.
Does it use the betting line?
No. The Vegas line is shown next to the model's numbers for comparison, but it isn't an input to the prediction.