College Football Model: Play-by-Play Monte Carlo

Olympus Bets models college football with a play-by-play Monte Carlo engine (PBP v3) whose every mechanic — play selection, yardage, turnovers, penalties, tempo, kicking, fourth-down decisions — is sampled from empirical distributions fit on 2021–2024 play-by-play data. The published margin comes from Power Rating v1 (opponent-adjusted efficiency, talent and roster churn) plus a quarterback-availability adjustment; the Monte Carlo owns the total and is shifted path-by-path to that margin (production tag cfb_pbp_v3.1+pr1). Game projections for 2026 are generated and shown; recommended picks, edges, and units remain gated in shadow until the engine demonstrates out-of-sample skill. That split is deliberate.

Why There Are No CFB Picks Yet

An earlier version of this model produced a backtest that looked exceptional. Before publishing anything, the result was audited — and the apparent edge traced to leaked rows: games whose outcomes had contaminated the training window. The honest read was that the old approach had no demonstrated skill. It was retired, and a rule was written into the pipeline: a college football pick may not be published until the engine beats the market in data it has never seen.

Most services would have shipped that backtest. We publish every graded pick to an append-only public ledger, so shipping a broken model is a debt we would carry publicly for years. The shadow gate on picks is cheaper than the apology. Showing projections without a stake recommendation is how we stay transparent while the engine accumulates a real graded record.

How PBP v3 Works

1. State-Keyed Empirical Kernels

The engine does not assume any play-level behavior. Rush/pass selection, yardage gained, sack, interception, and fumble rates, penalty frequency and yardage, tempo (seconds per play), field-goal make rates by distance, punt and return outcomes, and fourth-down go/kick/punt decisions are each sampled from empirical distributions keyed on game state — down, distance, field position, score situation, and period. The kernel library was fit on more than 862,000 plays from the 2021–2024 seasons.

2. Data Integrity Screening

College football data is messier than NFL data. Before fitting, the play-by-play lake is validated by reconstructing every game's final score from its play sequence and comparing against the official result — a check that caught hundreds of corrupted upstream games, including some whose home and away scores were inverted at the source. Games that fail reconstruction are blocklisted from training entirely.

3. Team Strength With Honest Uncertainty

Team effects are estimated with empirical-Bayes shrinkage — a 130-team league with wild talent gaps and 12-game seasons demands it. Preseason strength uses as-of snapshots with priors for roster continuity and portal churn. Freshmen and players without college production keep roster identity and position baselines; we do not invent box scores.

4. Power Rating Margin and Quarterback Availability

Since September 23, 2026, who should win by how many comes from Power Rating v1: an opponent-adjusted team rating that also carries recruiting talent and roster churn, updated as the season is played. A quarterback-availability channel then adjusts the margin when an established starter is listed on the official injury report before kickoff, scaled by how likely he is not to start; it never fires on stale or post-kickoff news. The play-by-play sim owns the total, and its paths are shifted to the published margin so projected scores and win probability stay consistent. Our older internal Elo anchor (cfb_pbp_v3.1-anchored) is still computed beside every game so both can be graded, and it is used only as a fallback for a game with no rating.

5. Sealed-Holdout Validation

For fitting, the 2025 season was sealed. Power Rating v1 was fit on seasons through 2023, selected on 2024, and graded on 2025 exactly once as a pre-registered final exam, where it cut margin error versus the Elo anchor it replaced. That is a projection-accuracy result, not a betting edge. Live 2026 projections may use completed 2025 results as the latest prior; that is ordinary as-of modeling, not leakage into a future game. Live shadow weeks grade projections against real markets before any publication decision on picks.

Data Sources

• CollegeFootballData play-by-play — the raw play lake, integrity-screened by score reconstruction
• CollegeFootballData schedules & rosters — final scores and team talent for the power rating; 2026 team/player identity
• Captured market odds — real per-side prices archived at capture time, used for de-vigged benchmarking
• As-of strength snapshots — preseason and in-season team priors for the Monte Carlo
• Official injury reports (Covers.com) — pre-kickoff quarterback status for the availability channel

Program Status

No pick win rate or units are shown here because CFB has never published a stake recommendation we would stand behind. When the engine graduates, its record will accrue on the public ledger from pick one, like every other league. Season futures projections and weekly game projections are available in the app under College Football.

Explore Other Models