Methodology

What gets analysed

Every completed game since 2019 is ingested with full box scores and drive-by-drive detail, alongside opening and closing point spreads, moneylines and totals from up to 30 sportsbooks, plus venue-level weather at kickoff. From that the system derives, for each team and each game:

• Efficiency — points and yards per drive, yards per play, third-down and red-zone conversion, explosive-drive rate, three-and-out rate and turnovers, each computed for offence and defence separately.
• Opponent adjustment — raw form is corrected for the quality of the opposition faced, so a defence that has played three elite offences is not judged as though it played three poor ones.
• Garbage-time filtering — scoring once a game is already decided is discounted, because blowout points predict very little.
• Recency — the last 3, 5 and 8 games alongside the full season, so recent form is weighted without letting one result dominate.
• Team strength — a margin-aware Elo and a ridge-regularised least-squares power rating solved over every prior game, with an offence/defence split and an empirically fitted home-field advantage.
• Situation — rest days, short weeks, byes, travel distance, time-zone shift, altitude, neutral sites and conference games.
• Weather — temperature, wind, gusts, precipitation and snow at kickoff for outdoor venues, from a historical archive so the effect can actually be backtested rather than merely asserted.
• Market structure — the opening number, the current number, movement between them, and how far the price sits from a key number.

How the ensemble works

No single model is trusted. Around a dozen run in parallel — gradient boosted trees (XGBoost, LightGBM, CatBoost), random and extremely randomised forests, ridge, elastic-net and Bayesian linear models, a logistic classifier on the cover outcome, a small neural network, plus the two non-machine-learning baselines (Elo and power ratings).

Every one of them predicts the same quantity: how many points better the home side performs than the posted line implies. That makes their outputs directly comparable, and it sets the right null — a model that has learned nothing predicts zero, produces no edge, and therefore no bet.

A meta-model then learns how much to trust each one. Its weights are constrained to be non-negative, capped so no single model can dominate, and shrunk hard toward zero. The prior is that the market is right; a model has to earn weight against it. Crucially, those weights are fitted only on predictions the models made for games they were not trained on.

How probabilities are produced

A projected margin is not yet a probability. Two independent methods run, and both must agree before a wager is taken seriously:

• A discrete margin model. Football margins are lumpy — real games land on 3 and 7 far more often than on 2 or 5. The probability of each whole-number margin is modelled on that lattice, with the key-number weights measured from historical results rather than assumed, and the distribution re-centred so those weights don't quietly drag every projection toward 3.
• Monte Carlo. Each game is simulated 50,000 times by simulating the scoring process — possessions, then touchdowns, field goals and empty drives — rather than by adding noise to an average. The clustering at 3 and 7 emerges from the mechanics instead of being pasted on.

The two probabilities are blended. A disagreement between them is itself a warning that the projection is unstable, and it lowers the bet score.

Uncertainty is priced, not ignored

A projection of −7 from a stable, well-observed matchup is not the same as −7 from two teams with three games played and a quarterback in doubt. Extra variance is added for thin early-season samples, severe weather and uncertain availability. Wider distributions produce probabilities closer to a coin flip, which mechanically lowers the expected value and the stake.

How picks are ranked

Each candidate is scored 0–100. Expected value after vig carries the most weight, because that is what actually compounds. Cover probability above break-even, the raw size of the disagreement, and whether the edge carries the number across a key threshold contribute the rest. That score is then scaled by how much the individual models agree with each other and by the quality of the underlying data.

Bands: PASS, LEAN, PLAY, STRONG PLAY, ELITE PLAY. Only PLAY or better is eligible for the daily card. Two picks may never involve the same team, so the card cannot stack correlated exposure onto one outcome.

At standard −110 pricing the break-even rate is 52.38%. A wager needs to clear that plus the expected-value threshold before it is published.

Why closing-line value matters more than this week's record

If you take a team at −2.5 and the line closes at −4, you got a better number than the market's final, sharpest opinion. That is positive closing-line value, and it counts whether the bet won or lost.

Win rate over a few dozen wagers is almost pure noise — a 55% record over 40 bets is entirely consistent with having no edge at all. CLV stabilises far faster and is the leading indicator that a process is sound. So CLV is recorded for every published wager and reported on the performance page whether it flatters the model or not.

Guarding against fooling ourselves

• No leakage, structurally. Features are produced in a single chronological pass: each game's feature row is emitted from the current state, and only then is that game's result folded in. A feature cannot contain information from its own game or from any later one.
• Walk-forward validation. Models are retrained from scratch on each block using only games that had already finished, and are never scored on data they were trained on.
• Provider artefacts excluded. Anything whose availability depends on which data source covered which season is kept out of the model, so it cannot learn to recognise an era instead of a matchup.
• Live in-game lines excluded. An in-game price on a finished game sits close to the final margin. Treating one as a pre-game number would leak the result directly — an early version of this system did exactly that and produced a spectacular, entirely fake 58% record.
• Intervals, not point estimates. Bootstrap confidence intervals accompany every headline number. If the interval spans zero, the result is reported as not significant.

Known limitations

• College football injuries are effectively unavailable. There is no mandatory reporting, and the public feed typically carries a handful of entries for the entire sport. No injury adjustment is applied, and picks say so explicitly rather than implying a clean bill of health. NFL injury data, by contrast, is comprehensive.
• Historical odds depth varies. Multi-book history is deep for some seasons and thin for others, which limits how far back cross-book disagreement can be studied.
• No player-level projections. There is no depth chart, snap-count or usage model, so a specific starter's absence is not priced beyond the crude positional adjustment used in the NFL.
• Public betting percentages are not used, because no free source of them is reliable enough to test properly.

Measured out-of-sample results

Walk-forward, retrained every 14 days, scored only on games each model had not been trained on. These are the numbers the verdict above is based on.

Current model

Data pipeline health

Model weights, feature definitions and source code are not published. Everything above describes the approach at a level intended to let you judge whether the process is sound — which is the part that matters.