12 tutorials

Football

nflverse opens up every NFL play. Compute EPA, compare quarterbacks, and learn the modern football-analytics toolkit.

Start with the nflverse pull tutorial — one function call gets you real NFL play data — then EPA-explained-by-building-it, the metric the whole modern-football track speaks. The quarterback-comparison, tendencies-by-down, receiver-efficiency and fourth-down tutorials all assume both. nflverse data is genuinely free and genuinely complete; this track is the fastest route from zero to real analysis on this site.

Pull Your First NFL Data with nfl_data_py

Load a full season of NFL play-by-play, the nflverse way - including the real pandas-version gotcha that breaks nfl_data_py and the one-line fix around it.

FootballBeginner~9 min

A season of play-by-play loaded into pandas, with a plays-per-team summary.

Build a QB Efficiency Comparison Chart

Aggregate play-by-play to the quarterback level and build a labeled scatter of EPA per dropback against completion percentage to compare passers fairly.

FootballIntermediate~9 min

A labeled scatter of quarterbacks by EPA per play and completion rate.

Team Tendencies: Run/Pass Rate by Down and Distance

Pivot a season of play-by-play to map a team's run-pass tendencies across every down and distance - exactly the scouting grid coordinators build before a game.

FootballIntermediate~9 min

A heatmap of how pass-happy a team is by down and distance.

Receiver Efficiency: Targets, Catches, and Yards After Catch

Aggregate play-by-play to the receiver level to compare pass-catchers on targets, catch rate, YAC, and EPA — separating volume from true efficiency.

FootballIntermediate~9 min

A leaderboard of receivers by volume, efficiency, and YAC.

Point Spread to Win Probability: Calibrate the Line on 7,261 NFL Games

Turn an NFL point spread into a win probability three ways on 7,261 games with a closing line: raw bins, the 13.2-point normal model, and logistic regression.

FootballIntermediate~16 min

A closing point spread turned into a win probability three ways and checked against 7,261 decided NFL games from 1999-2025: the raw win rate at each spread (3-point favorites win 58.1%, 7-point favorites 76.0%), the one-number normal model (margin minus spread has a standard deviation of 13.2 points, so a 3-point favorite is 59.0%), and a logistic regression fit by Newton's method - scored out of sample on 2020-2025, then applied to the 16 opening-week games of 2026 and cross-checked against the vig-free moneylines.

Survival Curves: How Long NFL Head Coaches Really Last

Kaplan-Meier and log-rank by hand on 170 NFL head-coach hires: a four-season median, a shorter-leash headline that is censoring, and one split that holds.

FootballIntermediate~14 min

Every NFL head-coaching spell in the nflverse games table turned into survival data and read with a Kaplan-Meier estimator and a log-rank test written by hand: 170 hires since 2000, 25 of them still running, a median tenure of four seasons, a 'recent coaches last 40.6% less' headline that turns out to be censoring (log-rank p = 0.46), and a first-season record that does separate the curves once the comparison starts from a season-two landmark (p = 0.033).

Rank Correlation: How Much of an NFL Season Carries Over

Kendall's tau and Spearman's rho by hand on 26 NFL season pairs: order persists at a median tau-b of 0.205, and vanishes inside the playoff field.

FootballIntermediate~18 min

Kendall's tau-a and tau-b, Spearman's rho and an exact permutation null, all written by hand and run on every consecutive pair of NFL seasons since 1999. Win totals tie in 7.81% of within-season team pairs against 0.30% for point differential, which is why the tie correction gets a step of its own. Order persists in all 26 season pairs at a median tau-b of 0.205 - 269 of the 417 orderable team pairs kept their 2023 order in 2024 - but restricted to the teams that actually reached the playoffs the median falls to 0.096, and the 2023 field lands on exactly zero: 35 concordant pairs against 35 discordant.

Bradley-Terry Ratings: Strength From Who You Beat

Fit Bradley-Terry ratings by hand with the MM algorithm: schedule-adjusted strength for every NFL team, and an honest out-of-sample test against win %.

FootballIntermediate~11 min

Bradley-Terry strengths fitted by Hunter's MM algorithm, written by hand with a home-field factor and a one-game prior. On 2025, SEA, DEN and NE all went 14-3 and finish 1st, 3rd and 7th by strength, because their opponents' mean strength was 1.144, 0.704 and 0.627. Without the prior a winless team's first update is exactly zero and an unbeaten team's likelihood keeps rising with no maximum. Out of sample on weeks 13+ of 2010-2025 it ties plain log5 from the record: mean log loss 0.6647 against 0.6651, better in 8 of 16 seasons, a per-game difference whose bootstrap interval straddles zero.

EPA Explained by Building It: Analyze Play-by-Play Data

Expected Points Added is less mysterious than it sounds. Work with nflverse EPA directly, summarize it by down and play type, and learn what it really measures.

FootballAdvanced~9 min

An EPA-per-play summary by down, computed straight from play-by-play.

Fourth-Down Decisions: Which Teams Go For It?

Filter play-by-play to fourth downs, classify each call (go, punt, kick), and rank teams by how often they go for it — the signal of analytics-era coaching.

FootballAdvanced~9 min

A ranking of how aggressive each team is on fourth down.

Quantile Regression: The Whole Margin, Not Just the Average

Fit five quantile lines to 7,276 NFL margins with no scipy: the spread shifts the whole distribution, and the 80% band stays 34 points wide.

FootballAdvanced~19 min

Five conditional quantiles of the NFL point margin, fitted by hand on 7,276 games with a closing line: the check loss, an iteratively reweighted solver, and a brute-force solver that proves the fast one right. It answers what a single fitted line cannot - whether the point spread moves the middle of the margin distribution or its width as well. The five lines come out near-parallel, the 80-percent band stays about 34 points wide from a pick'em to a two-touchdown favourite, and the median line sits close enough to the identity that the market's number reads as a median rather than a mean.

Regression Discontinuity: What a Close Win Is Worth

Regression discontinuity by hand on 13,904 NFL team-games: a close win buys nothing next week, and the balance check shows close games are not coin flips.

FootballAdvanced~23 min

A regression discontinuity built by hand on every decided regular-season NFL game since 1999, seen from both sides: the team's own final margin is the running variable, zero is the cutoff and the next game is the outcome. Triangular-weighted local lines, a bootstrap over whole games and a six-window bandwidth sweep find no jump in next-week win rate, margin, line or cover. The balance check finds one where none should be: narrow winners went in favoured by 1.21 points more than narrow losers, a gap that survives dropping overtime and beats all 21 placebo cutoffs.