Austin is a CFO and statistician by trade — building financial models, translating complex data into clear decisions, and cutting through the noise that clouds most business judgment. He brought the same discipline to college football betting and built something that holds up under audit.
It started with one question: what actually predicts winners and spreads in college football?
I spent two years working through it. I bounced through every stat you can think of — efficiency, success rate, drive quality, third-down conversion, every flavor of pace and tempo. I landed somewhere I didn't expect: time of possession.
Not time of possession itself, but how elastic each team's possession is against different opponents. Some teams hold the ball no matter who they play. Some teams' possession gets pushed around. That elasticity, weighted by opponent strength, became Pace+.
Then I did the only honest test there is. For a full year I bet a dollar on every spread, total, and moneyline I could get to. Real money, real markets, every game. Pace+ was strong. It was also incomplete — there were specific kinds of matchups where it had blind spots — so I put the public efficiency models next to it and asked whether several reads on every game could fill the holes one read couldn't.
The answer was yes. But not in the way I expected, and the rest of this page is about why.
The first full backtest looked too good.
The top-tier picks hit 75%. The moneylines were near-perfect. I'm a statistician; I knew immediately the numbers were contaminated. Anyone whose backtest looks that good has a bug, not a thesis.
The bug was look-ahead bias — ratings that already knew how the season ended were "predicting" games inside it. We rebuilt the pipeline so every Week N projection uses only data from Week N−1 and earlier. The numbers came in lower, and they were real. Then we found the same bias in another corner of the model and rebuilt again. Then duplicate rows polluting the rating refresh. Fixed, rebuilt a third time. Lower each time, and truer each time.
That is when the question changed. We stopped asking the models for a better number and started asking when a model's number is actually worth betting. Three seasons of history, every variable we could measure — kickoff window, day, conference, week, line size, how far each model sat from the market — crossed with the situational spots bettors argue about. For spreads and totals that made 31,000 candidate signals. For the moneyline, where who wins is a different question than who covers, the search ran past a million patterns. Every one was graded point-in-time, at the line as it stood, then tested against random chance, then required to hold up year over year.
Almost everything died. 253 spread-and-total signals and 151 moneyline patterns survived. Those are the voters. A grade is how many of them agree on a bet once the ones pointing the other way are subtracted out — two is a C, eight or more is an A+, and one voice alone is No Edge, because one voice alone never beat the vig. The numbers that survived are smaller than the ones we threw away, which is exactly why we trust them.
Then we built the game itself.
The voters tell you when a number sits in a spot that has paid. They still don't tell you how a game plays — how many possessions there will be, what a two-score lead does to a coach's fourth-down call, how often a three-point favorite actually wins by three. If accuracy is the whole point, the only honest way to know those things is to play the game.
So we did. The engine is a play-by-play simulation: every snap, drive, kick, penalty, and clock decision drawn from what teams like these actually do, played out thousands of times per matchup. Each layer of it was measured before it went in, and the whole thing was measured across the same three seasons before it was allowed anywhere near a grade.
It earned a vote on the spread. Where the engine lands on the same side as the signals, the bet grades A+. Where it lands on the other side, the bet comes off the board. Where it holds an opinion and no signal fires, it grades a C on its own. Every game page shows its work — the full range of outcomes, and the path the margin takes minute by minute — and the line under every vote says, in plain words, where the engine has it.
Accuracy above all else has one more consequence: the record has to be public. Every pick locks at kickoff, settles at the number we published, and lands on Results — wins, losses, and pushes, next to the games we passed on. A number earned on the seasons we searched is a claim. 2026 is the test, and you get to watch it graded live, where nobody can retouch it.