Loading the backtester
BETABETADescribe a bet the way you'd say it out loud. It runs against 173,183 real player-games.
A book never offered "over 24.5 points" to a bench player who scores two a game, so counting his game as a loss measures nothing. This sets how much of the line a player had to be averaging beforehand to count. It is a convention, not a measurement: there is no historical record of who was really offered what, so every result says which setting produced it.
Plays the result day by day with real stakes, so you can see the swings, not just the total. Each day's bets are sized from that morning's bankroll. "Double after a loss" doubles up to 6 times in a row (64x the first bet), never stakes more than the bankroll, and ends the run when it can't cover the next stake. It is chasing losses, included as a stress test. Nothing here changes which games count.
Picked to show the three answers you can get, not because any of them is a good bet. The verdicts are recomputed from the data every night, so they change when the games do.
Breakeven is not 50%. At -110 you risk $110 to win $100, so a strategy has to hit better than 52.4% of the time to make money; at -130 it needs 56.5%. The tool prices every bet at whatever odds you set, and the hit rate you need moves with it. A rate above that line with a real sample is an edge; below it is a losing bet no matter how good it feels.
Check the price you'd actually be offered. Props are frequently not -110. Books shade the popular side to -120 or worse. The price needed figure on every result is the number to compare against the one on your screen: if the book is offering worse than that, the strategy loses money however good the hit rate looks.
Compare against the base rate. Some bets hit often for everyone (RB "under their average" hits more than half the time because one big game inflates an average). The comparison card shows whether your condition actually changes the odds, or just rides the base rate.
Averages are computed entering each game, using only earlier games that season (minimum two), so the backtest never peeks at the result it's grading. "Their season average" as a line is a simple synthetic line, one input a book would use but not the line it posted, which also weighs role, matchup, injuries and money; a fixed number like 229.5 tests one line against every player, which can flatter or punish a strategy.
2016–2026. FBS player game logs. 2022–2025 were built from ESPN play-by-play published by the open-source sportsdataverse project; every other season comes from CollegeFootballData.com box scores. Spot checks against official season totals land within about 1–2%. QB rushing yards are shown before sack yardage, unlike official NCAA stats. CFBD box scores carry no target counts, so seasons built from them record no targets at all. A strategy that uses targets excludes those seasons and says which, and if that leaves nothing the run is refused rather than answered with a different stat. No historical betting lines are included yet, so a run prices every bet at the single number you choose rather than the price actually offered on the day. Crak'd is a research tool, not betting advice.
Finished games land the next morning (7am ET), from CollegeFootballData.com. Seasons before 2026 are fixed historical data. Last updated 2026-10-16.
College football changed a lot after 2020: the transfer portal, NIL, and a full extra playoff round all reshaped rosters. Seasons before then are real data, but treat them as a different sport rather than more of the same. The look-back slider starts at a recent window for this reason; widen it deliberately, and check whether a strategy still holds up on recent years alone before trusting it.