How this whole thing works

Every prediction comes out of a trained XGBoost classifier, one number between 0 and 100 percent, built from twenty features that fall into a few plain-English buckets:

• Team strength. An Elo rating for each team that updates after every game, the same idea chess uses to rank players.
• Recent form. Offensive and defensive efficiency (EPA per play) over each team's last several games, not season-long averages that miss a hot or cold streak.
• Rest. Days between games for each side, short weeks and byes included.
• Injuries. Who's banged up, weighted heavier when it's the quarterback.
• Weather. Temperature, wind, and precipitation at kickoff, dome games treated as a controlled environment.
• The market itself. The current spread and the market's own implied win probability are two of the twenty inputs. The model gets to see what Vegas thinks before it makes its own call, and it still doesn't beat the market outright. More on that below.

Each prediction ships with the top factors that moved the number, in plain language and pointed at the team they favor, not a bare percentage you have to take on faith.

Every feature for a given matchup is built strictly from data available before that game's kickoff. Elo ratings, recent form, rest, injuries, all of it is a snapshot of the week before, never the week of. A prediction never gets to peek at a score, a stat line, or anything else from the game it's predicting.

This sounds obvious until you've seen how easy it is to get wrong by accident, a model that accidentally learns from data it shouldn't have seen looks brilliant in testing and falls apart the moment it has to call a real, unplayed game. It's enforced with a dedicated automated test, not just a coding habit, so it can't quietly regress as the pipeline changes.

The XGBoost model produces the win probability. SHAP explains which factors pushed it that direction and by how much. Claude's only job is to turn that probability and those factors into the 2 to 4 sentences of broadcast-style color you read on a matchup page.

Claude never sees the raw game data and never touches the math. It gets the number and the factor list as fixed inputs and is guardrailed against changing or inventing either one. If narration ever fails or comes back looking off, the page falls back to showing the factor list plainly rather than letting a bad sentence stand in for the model's actual call.

Every model gets evaluated the same way real forecasters do: walk-forward by season, training only on seasons already played and then grading the next one, compared against the closing betting lines for those same games with the bookmaker's cut removed. That's a demanding bar. Closing lines are one of the hardest baselines to beat using only public data, and this site doesn't beat it.

College football gets its own model, its own Elo history, and its own backtest, graded against de-vigged closing lines from CFBD (FBS versus FBS games only, since the market barely prices FCS mismatches):

Read straight: on the full sample, the NFL model calls the winner about 65.0 percent of the time against the market's 68.2 percent, a gap of roughly 3 points. CFB runs 71.7 percent against 73.8 percent, about 2 points back. Brier score, which grades not just whether the pick was right but how confident the model was while making it, tells the same story: the market scores lower (better) than the model in every single season, both sports, no exceptions.

None of that means the model is guessing. It lands in the neighborhood of a real sportsbook line using only public data, which is a genuinely hard thing to do. It just doesn't mean the model has an edge on Vegas, and this page isn't going to pretend otherwise. Every prediction on the site is labeled with a model version and a timestamp, and nothing here is betting advice.