Building Your First Model

Power ratings: the simplest model that works

The most approachable model in all of sports betting is a set of power ratings. You assign every team a single number, measured in points, that says how much better or worse than an average team they are. An average team rates 0. A good team might rate +6; a bad one −4. The gap between two teams’ ratings is your prediction of the margin on a neutral field — before you adjust for who is hosting.

That last piece matters. Home teams win a little more than the raw talent gap suggests, so you add a home-field edge — a fixed bump for whoever is hosting. The size depends on the sport, but a couple of points is a reasonable starting estimate.

A worked projection

Suppose your ratings have Team A at +6 and Team B at +1, and Team A is hosting. Plug it in:

Where the ratings come from

Your ratings are only as good as what you feed them. Start simple and resist the urge to be fancy on day one:

• Points scored vs. points allowed. A team that outscores opponents by a lot is, all else equal, better. Margin of victory is a more stable signal than win-loss record.
• Adjust for strength of schedule. Beating great teams by 3 is more impressive than beating terrible ones by 20. Weight each result by the quality of the opponent.
• Update as games happen. An Elo-style approach nudges a team’s rating up or down after each game based on how the result compared to expectation — winners take points from losers, and surprises move the numbers more than expected outcomes.

You can refine all of this over time — pace, recent form, rest, injuries. But version one should be something you can compute and explain in a single sitting. A simple model you actually trust beats a complicated one you can’t debug.

Turning a projected margin into a price

A spread lean is useful, but to compare against a moneyline — or to size a bet honestly — you want a win probability. The bridge from points to probability is the historical distribution of game outcomes. For a given sport, results scatter around the expected margin with a known spread, measured in points per standard deviation. A projected edge of a few points translates into a modest probability bump; a large projected margin translates into a near-certain favorite.

The mechanics are conceptual here, so don’t get lost in the math: a bigger projected margin, relative to how much games typically bounce around, means a higher win probability. Once you have your probability, you compare it to the market’s no-vig probability — the fair number you learned to strip out in Courses 05 and 07. If your number is meaningfully higher than the de-vigged market number, that side is a candidate bet.

Acting on it: the disagreement threshold

Here’s the discipline that separates a model from a hunch: you only bet when your number differs from the market by more than your own uncertainty, and by enough to clear the vig. Your model is not exact — it has an error bar — and the book is charging you a margin on every ticket. A tiny disagreement gets eaten by both.

A practical rule for a points-based model is a minimum threshold: only act when you and the market differ by, say, 1.5 points or more. Below that, the “edge” is almost certainly model noise, not a real mispricing. Most nights, your model will agree with the market — that’s normal and healthy. The market is sharp. You’re hunting for the handful of games where you genuinely see something it doesn’t.

The central danger: overfitting

If you remember one warning from this course, make it this one. Overfitting is when your model has so many knobs that it perfectly explains the past — including the random noise in it — and then falls apart on games it hasn’t seen. A model can look brilliant on last season and be worthless next week for exactly this reason.

• Test out-of-sample. Build the model on one set of games and check it on a different set it never saw. In-sample accuracy proves almost nothing.
• Keep it simple. Fewer parameters means less room to fit noise. Every extra knob should earn its place.
• Don’t tweak after the fact. Adding a rule to “explain” a result you already know is how you fool yourself. That’s curve-fitting the past, not predicting the future.
• Respect what the model can’t see. A power rating doesn’t know the starting quarterback is out or that it’s 20 mph wind and rain. Obvious real-world information overrides a stale number.

Validating: track the right thing

The instinct is to judge a model by its win-loss record. Don’t — at least not first. Over a few weeks, wins and losses are mostly variance, and a good model can lose money short-term while a bad one gets lucky. The cleaner signal is Closing Line Value (CLV): did the line move toward your bet after you placed it? If your model consistently beats the closing number, it is finding real edges, regardless of how this week’s results landed. You’ll dig into CLV properly in Course 12 — for now, just know that it, not your bankroll graph, is how you tell whether the model works.

Key takeaways

Check yourself