Football analytics models, taken apart
Five families of models, each explained from the formula up, with the data and code to rerun them.
Elo ratings
One number per team, nudged after every game by how surprising the result was.
Read the model SHOT 02Markov chain ratings
A random walk that drifts toward whoever beats the teams that beat everyone else.
Read the model SHOT 03Win probability
Score, clock, down, distance and field position turned into one live percentage.
Read the model SHOT 04Expected points
What a first down at the opponent's 35 is worth before the snap, and what EPA measures after it.
Read the model SHOT 05Clustering tiers
k-means groups players whose projections overlap, so tiers come from the data, not from rank order.
Read the modelRatings & rankings
How to turn a season of scores into one number per team: Elo, least squares, win percentage, Markov chains, and how to tell whether any of them is right.
Elo Rating Formula for Football: K-Factor and Home Field
One update equation from chess, three football-specific adjustments, and the exact numbers a public NFL model used to make it work.
Football Rating Systems: Markov Chain and Massey Methods
A rate matrix built from wins and losses, an equilibrium that ranks every team, and the linear-algebra cousins that compete with it.
Ranking Algorithms and the NFL: Offensive Rankings Explained
A first look at how a season of results turns into a single number per team, and why offense and defense often get ranked separately.
RatingsMassey Least Squares Ratings: The Simple Version
One regression equation, solved across an entire season of games at once, turns final scores into a single rating per team.
RatingsWin Percentage Formula: Why It Falls Short as a Ranking
Win percentage counts results; it cannot tell a hard schedule from an easy one, and two fixes attempt to close that gap in different ways.
RatingsHow Consistent Are Football Rating Systems With Each Other?
Two rank-correlation statistics turn 'these rankings look similar' into a number, and the number usually tells a more specific story than a glance does.
RatingsEigenvector Power Ratings for Football Teams
A ratio of who-beat-whom, turned into a matrix, whose dominant eigenvector becomes the power rating.
RatingsCrowdsourced Football Rankings: How Polls Aggregate Votes
A points system aggregates dozens of individual ballots into one list, resting on the same statistical idea behind a century-old ox-weighing contest.
Win probability
Building, testing and reading in-game win probability, from the first logistic regression to calibration and overfitting.
Football Win Probability: Building the NFL Model (Part 1)
Every win-probability chart on a broadcast reduces a football game to a single number between 0 and 1 — this is the first of four parts on how that number gets built from raw play-by-play data.
Win probWin Probability, Uncertainty and Overfitting in NFL Models
A single win-probability number looks precise, but every one of them carries uncertainty the display doesn't show — and overfitting is the fastest way to make that uncertainty invisible.
Win probProbability vs. Outcome in a Win Probability Model
A single football game only happens once, which makes a probability estimate about it strange to verify — this piece works through what a win-probability number is actually claiming, and what would have to be true for it to be wrong.
Win probLogistic Regression vs Random Forest: Win Probability (Part 2)
Part 2 of the win-probability series compares three ways to fit the same six features to a probability, and explains why most modern public models settled on tree-based methods.
Win probA Good Win Probability Model: Brier Score & Log Loss (Part 3)
Part 3 of the win-probability series asks what 'good' even means for a probability model, and works through the two standard scoring rules used to answer it.
Win probWin Probability: Feature Engineering & Evaluation (Part 4)
Part 4 closes the win-probability series with the feature engineering that turns raw game-state columns into interaction terms, and the season-based validation that checks whether the finished model can be trusted.
Win probDo Down and Distance Matter to Win Probability?
Down and distance are genuine inputs to a win probability model, but their effect on the displayed number is frequently too small for a reader to notice next to score and clock — which raises a real usability question, not a modeling one.
Win probHow nflfastR, ESPN and the NYT Bot Talk About Win Probability
The same underlying idea — a probability estimate for who wins — gets described three different ways by nflfastR, ESPN's FPI and the New York Times 4th Down Bot, and the differences say something about what each one is actually built to do.
Win probHow to Read an NFL Win Probability Plot
A win-probability plot compresses an entire football game into one line, and reading it well means knowing which parts of that line are informative and which are just noise.
Expected points & forecasting
What a down and a yard line are worth, and what forecasts of players, coaches and seasons can and cannot tell.
Expected Points and EPA in the NFL: How the Model Works
A three-page 1971 paper and a modern gradient-boosted model answer the same question: what is a given down, distance, and field position actually worth in points?
EPAWhat Social Media Data Reveals About Football Fandom
Millions of geotagged and game-referencing posts turn out to be a genuine dataset for questions no box score can answer: how far a fan base actually extends, and how a game's outcome moves a crowd's mood in real time.
EPAForecasting QB Performance With Multilevel Regression
A rookie with four starts and a ten-year veteran should not be forecast the same way — multilevel regression is the tool built specifically for blending individual and league-wide evidence in the right proportion.
EPAHow to Grade Football Forecasts After the Season
Being right once proves nothing about a probabilistic forecast — grading it properly means checking two separate things: how close the numbers were, and whether they were honest about their own confidence.
EPAWhy Coaching Changes Are Hard to Judge with Data
Teams fire coaches after bad seasons, and bad seasons tend to be followed by better ones anyway — which makes isolating a coaching change's real effect a genuine statistical problem, not a simple before-and-after comparison.
EPAUncertainty in Expected Points Estimates: Why It Matters
An expected points model outputs one number per situation, but that number is an estimate with its own margin of error — and the error bars are wider in some game states than others.
Methods
The statistics under the models: clustering, cross-validation, outliers, classification errors and reproducible work.
K-Means Clustering for Fantasy Football Tiers
A ranked list draws a hard line between picks 12 and 13; clustering draws the line where the data actually breaks.
MethodsConfusion Matrix, Precision, and Recall for Football Models
A model that's right 90% of the time can still be useless — it depends entirely on what the other 10% cost.
MethodsStatistical Decision-Making vs. Prediction in Football
Predicting what a team will do and calculating what it should do are not the same question, and a model built for one answers the other only by accident.
MethodsReproducible Research in Sports Analytics
A chart is a claim. A chart with the code and data behind it is a claim someone else can actually check.
MethodsFeature Selection vs. Sample Selection in Football Models
Adding a better stat and adding a better game are both called 'improving the data' — they fix completely different failure modes.
MethodsOutliers and Cook's Distance in Football Models
Not every extreme game is dangerous to a model — only the ones sitting where the model has little other data to lean on.
MethodsCross-Validation for Sports Prediction Models
A model that trains on next month's games and tests on this month's isn't validated — it's cheating, just slowly enough to miss.
MethodsCounterintuitive Findings and Publication Bias in Sports
A finding that upends conventional wisdom gets shared faster than one that confirms it — which is exactly why it needs to survive more checking, not less.
Field notes
How football analytics is done, used and hired for.
Where Football Analytics Is Discussed and Published
Most of what moves football analytics forward is published in the open, by people who expect their code and their numbers to be checked.
NotesHow to Get a Job in Sports Analytics: Real Paths In
Every documented path into a football analytics job starts with published, checkable work, not a private tip about who's hiring.
NotesSituational Thinking in Football Analytics: Where Data Helps
A third-and-two and a third-and-eleven are the same down and distance category on a stat sheet and almost unrelated plays on the field.
NotesBuilding an NFL Analytics Department: Roles and Workflow
An analytics department only creates value at the moment a model's output changes a decision a coach or scout was about to make.
NotesHow to Ask for Help With Analytics Code, and Get It
The question that gets answered is almost never the one that describes the problem best — it's the one that's easiest for a stranger to reproduce.
NotesThe State of Football Analytics: Why the NFL Lagged Baseball
Baseball had Bill James by 1980 and a public play-by-play archive soon after; football's equivalent data infrastructure is barely a decade old.
NotesRacial Framing in Quarterback Commentary: The Research
Three decades of content-analysis research, and a more recent computational study, converge on the same specific pattern in how quarterbacks are described by race.
Frequently asked questions
Which rating system is the most accurate for the NFL?
No single system wins every season. Elo carries information from past games and adjusts for home field and margin; least-squares systems such as Massey use only the current season. The backtest page shows how the site’s Elo has predicted games in every season since 2000.
What is the difference between win probability and expected points?
Expected points estimates how many points the next score is worth from the current down, distance and field position. Win probability estimates the chance of winning the game, so it also depends on the score and the clock.
Do I need to code to follow the model guides?
No. Each guide explains the idea in words and formulas first. Short Python snippets are included for readers who want to rerun the numbers on the open nflverse data.
How are fantasy football tiers built with k-means clustering?
Players are described by a few numbers, such as points per game and week-to-week volatility, scaled to the same range, and k-means groups players whose numbers sit close together. The clustering guide walks through the whole process.