Methodology

How our two models — one for MLB, one for college football — turn information into published picks, and how every one of those picks is graded. It is written in plain language and is deliberately light on internals: we describe what goes into the models and how we hold them to account, not the recipe.

The short version

Project every game. Each model builds its own view of every game on the day’s slate.
Compare it with the price. That view becomes a chance of winning for each bet, which is set against the chance implied by the sportsbook’s price.
Publish only real edges. A pick publishes only when the model beats the price by enough to matter, and only after automated guardrails have checked it.
Same stake, every time. Every published pick risks one unit — no bigger bets on “stronger” plays.
Grade everything. Every pick is logged before the game with its exact line and price, then graded automatically from the final score.

What goes into a projection

MLB

Each game is projected from the matchup itself.

How each lineup has been scoring, over the full season and lately
The starting pitchers’ underlying skill, not just the runs they have allowed
Each bullpen, and how hard it has been worked in recent days
The ballpark
The temperature and wind at first pitch

The moneyline, run line and total all come from that one projection, so they never contradict each other.

College football

Each game gets a projected margin and total, built from:

Team talent — recruiting, plus the players each program gains and loses through the transfer portal
Established power ratings
Play-by-play efficiency
Scoring
Returning production — how much of last season’s team is back

Why we trust the process

Every input is well-established baseball or college football information with a long record of relating to results. No hunches, and no following the crowd.
The model’s opinion is recorded on every game it looks at, not just the ones it picks, so it is judged on the whole slate rather than a flattering handful.
Changes are batched and re-tested against the full history before they ship — never a same-day reaction to one bad beat.
The record can’t be curated: the Track Record rebuilds itself from the same files the daily brief writes.

What an edge is

Every price a sportsbook posts implies a chance of winning: a heavy favourite’s price implies a high one, a long shot’s a low one. The model makes its own estimate of that same chance. The edge is the gap between the two — how much more likely the model thinks a bet is to win than its price suggests.

A pick publishes only when that gap is big enough to be worth a bet. The bar is fixed — it does not move after a good or a bad run — and we don’t publish where it sits. No edge, no pick: a day with nothing on the card is a valid day.

Picks are priced, and graded, at one sportsbook (Bovada), so the record reflects prices you could actually have taken rather than the best price anywhere on the market.

Guardrails

A model will sometimes produce an answer the real world doesn’t support — most often because something it was fed is wrong. So automated guardrails check the data and every projection before anything can publish, and a pick that doesn’t pass stays off the card.

We keep the details of those checks private. Picks they hold back are still recorded and graded on our side, so we can tell whether a guardrail is helping or costing us.

How picks are graded and logged

Flat one-unit stakes. Every published pick risks one unit. A win pays at the price we logged; a loss costs the unit.
Logged before the game. Each pick is written to the ledger before first pitch or kickoff, with the exact line and price. Once a game starts, its picks are locked.
Settled from the final score. Grading is automatic, from the official final score. A push — a result landing exactly on the line — returns the stake and counts as neither a win nor a loss. A game that isn’t played doesn’t count.
Parlays. Each parlay card risks one unit. If a leg pushes or its game isn’t played, that leg drops out and the card is paid on the legs that remain; a card left with a single leg is a push.
How the record is figured. Win rate and return on investment leave pushes out; return on investment is units won per unit risked.
Nothing is edited after the fact. Results live in the same ledger files the daily brief writes, and the Track Record rebuilds itself from them every day, losses included.

Closing line value

Between the moment a pick is published and the start of the game, its price usually moves. If it moves our way — the same bet costs more at game time than when we published it — we beat the close. Over enough picks, beating the close is the earliest honest sign that a model is finding real value, because it doesn’t depend on how a few games happen to finish.

We measure it like for like: the price we logged against the same sportsbook’s own last price before the game. It is still accumulating — only picks logged since we began capturing closing prices carry it — so treat today’s figure as early.

Read this before you bet

The sample is small. A few weeks of results can look brilliant or awful by chance alone. The win-loss and return figures will swing for a long time before they settle into anything meaningful.
Closing line value is still building up. It is the better early signal, but it needs many more picks before it proves anything.
Past results don’t guarantee future ones.
Nothing here is advice. Houdini Wagers is an independent analytics project, published for information and entertainment. It is not betting or financial advice. Only bet what you can afford to lose, and only where it is legal for you.