Backtest Explanation
The men's college basketball model results are built to answer the same customer question as the NBA results page: if the Three Check Bets process had been operating with the tested rules, and a customer risked $100 on every qualified play, what would the record, return, ROI, and market-quality indicators have been?
Walk-Forward Method
A walk-forward backtest moves through the schedule in order. Ratings, form, rest, market lines, and model features are built from information available before the game. The system is not allowed to learn from the final score before deciding whether the play qualified.
This matters more in college basketball than in many sports because the schedule is uneven. A model can look much smarter after it already knows which low-major teams were real, which conferences were weak, and which teams overperformed their box-score profile. The walk-forward structure keeps the test closer to the decision a customer would have faced before tipoff.
Three Check Process
The NCAAB card is not a blind list of every model edge. Each play has to survive a model-value check, a market/price check, and a basketball-context check before it becomes a tracked play.
The model needs enough edge against the sportsbook number. For spreads, that means the fair spread differs from the available line. For totals, it means the projected scoring environment differs enough from the market total.
The bet must still be playable at the available sportsbook price. Line shopping is part of the tested process, especially for ATS sides where a half-point can decide the result.
The final governor checks whether the matchup profile supports the model edge: site splits, rest, eFG, turnovers, free throw rate, pace, and possession environment.
ATS Model
The ATS process starts with a Balanced Growth candidate generator, then applies a site-split governor. The site-split layer asks whether the selected team has a positive site-specific net-efficiency profile or a positive site free-throw-rate matchup edge. This matters because college teams can play very differently at home, on the road, and on neutral floors.
Totals Model
The totals model is built from possession basketball. It estimates scoring by combining adjusted offensive efficiency, adjusted defensive efficiency, tempo, recent form, opponent context, and Four Factors. It then compares that projection to the market total and only keeps plays that pass the Three Check rules.
Team Totals
Team totals use the same possession and team-score logic, but the market is different. A full-game total asks how many points both teams will score together. A team total asks whether one team will score more or fewer points than its own posted number. That makes team totals useful, but they require separate sportsbook markets and separate pricing.
The current Team Totals Unified category fetches actual sportsbook team-total prices, chooses the best available line by model expected value, and limits plays with max-juice and probability governors. The latest audit now includes true-close team-total CLV, so team totals can sit beside ATS and full-game totals on the promoted customer card.
CLV
Closing-line value compares the customer bet line to the final market line before tipoff. Positive CLV means the model beat the market close. It does not guarantee that any single play wins, but it is one of the best signs that a process is finding prices the market later agrees were valuable.
The NCAAB page shows CLV where the true-close audit exists. ATS, full-game totals, and Team Totals Unified now have true-close support. Moneyline Guardrail is integrated into the season records, but its CLV is still shown separately in no-vig implied-probability percentage points because that market is priced differently than spreads and totals.
Customer Framing
Public Results Policy
The customer-facing page is designed to be read like a simple betting record. Each tested season appears in chronological order, and the internal lanes are rolled into the season where they occurred: Full-Game ATS, Full-Game Totals, Team Totals Unified, Moneyline Guardrail, and Season Combined.
Internally, Three Check Bets still keeps the lane-level detail. That includes candidate generators, safety-net filters, site-split governors, Monte Carlo totals variants, team-total price audits, and CLV refresh status. Those details are important for improving the business, but they can make a public results page harder to understand if they lead the presentation.
The public tracker therefore leads with the customer record first: bets, record, win rate, profit, ROI, and CLV status. The business archive remains available in the repository and handoff notes so future model work can audit exactly which lanes created each public row.
Important Note
The NCAAB backtest is evidence of how the saved model lanes performed under the stated rules. It is not a promise of future profit. College basketball markets change, line availability matters, neutral-site schedules matter, injuries and transfers matter, and future seasons will not match the 2025 test window exactly. The point of the page is transparency: customers can see the method, the rules, the bet count, the record, the return, and the market-quality checks.