Every number on this site traces back to nflverse, an open-source project that publishes clean NFL data going back to 1999 — schedules, box scores, and complete play-by-play with Expected Points Added (EPA) already computed for every play. We don't scrape or guess at anything upstream of nflverse; if a number here is wrong, it's almost certainly our math on top of their data, not the data itself.
Live in-game scores and national broadcast network come from a second, unofficial source: ESPN's public scoreboard feed. Unlike nflverse, ESPN doesn't publish or document this feed for outside use, so we treat it as a nice-to-have, not a dependency — if it ever goes away, the rest of the site keeps working, just without the live ticker. During games we check it once a minute: the score, the clock, who has the ball, down and distance, and the last play.
Live win probability on a game in progress is our own, not ESPN's: a small model fitted on every drive since 1999 (about 170,000 of them) that weighs the score, the time left, who has the ball and where, and the pre-game point spread, which matters less as the clock runs. It's checked against what actually happened: when it says 70–80%, those teams won about 74% of the time. Comeback odds are simply history: of every team since 1999 that trailed by that much with about that much time left, how many won. Both are refitted every night.
Counting readers: an open page that's on your screen sends one tiny "still here" request every five minutes, so we can tell roughly how many people are following along (say, during a game). It carries nothing about you beyond what any page view already does, stops when the tab is in the background, and isn't shared with anyone.
We run five independent rating methods side by side on purpose: Elo (sequential, reacts fast to recent results), SRS (a linear-algebra solve on point margin, adjusted for strength of schedule), an EPA-based variant of SRS, Pythagorean win expectation, and raw net EPA/play. When they agree, that's a strong signal. When they disagree, that's usually the more interesting team to watch.
The Elo rating system itself wasn't invented for football, or even for sports media — it comes from Arpad Elo, a physics professor and chess master who developed it in the 1960s to rank chess players (the US Chess Federation adopted it in 1960, the World Chess Federation in 1970). Its core idea — every competitor has a rating, the rating predicts win probability, and it updates after each result based on how surprising the outcome was — is chess math, not sports math.
What FiveThirtyEight did, starting around 2014, was adapt Elo specifically for the NFL: adding a
margin-of-victory multiplier (chess only has win/lose/draw, no "won by how much"), a home-field adjustment,
and a rest/bye-week term. Our Elo engine follows that same football-specific shape (the site shut down in
2025 and its original methodology pages no longer reliably resolve, so we're not linking directly — but
the design is theirs): the logistic win-probability formula, the margin-of-victory multiplier, and a
home-field term. We use 55 Elo points for home field (538 used 65 historically; the league-wide
home edge has been shrinking for years) and the standard divide-by-25 conversion between an Elo gap
and a predicted point spread, also a 538 convention. Since October 10, 2026 the projection also carries a
rest term: 25 Elo points, one point of spread, to the side with three or more days' more rest
than its opponent since each one's previous game — a bye against a normal week, or a normal week against a
Thursday turnaround. The Lab's bye and short-week pages are where that number comes from: the market pays about
a point for it, and the raw Elo line was missing it.
Reading the weekly charts: each week's point is a team's rating going into that week, built from every game played before it. So once a new week begins, every team gets a point for it right away — and until a team actually plays that week, the point simply repeats where they stood after their last game. A bye week shows up the same way: a flat step, because nothing happened to change it.
We simulate the rest of the season up to two million times and count how often each team makes the playoffs, wins their division, and how far they go in the postseason itself — each simulated season plays out an actual bracket, Wild Card through the Super Bowl, with the correct re-seeding rules and a neutral-site Super Bowl. Every night the server's sixteen threads run 500,000 seasons in each mode, then keep doubling the Elo run, to 1,000,000 and then 2,000,000; on a day without games, when nothing the simulation reads has changed, the seasons already stored are kept rather than rolled again. The site watches the live games every minute, and the moment one goes final it automatically kicks off a fresh simulation of both modes, 250,000 seasons each: the Elo odds first, on the site about a minute and a half later, then the coin-flip baseline about a minute behind. Then, while no other game finishes, the Elo run doubles again (500,000, 1,000,000, 2,000,000) and the coin flip goes to 500,000, back to full size in roughly a quarter of an hour — so the odds you're looking at reflect what just happened on the field, not whatever was true when the day started. Every simulated season is kept, not just the totals, which is what lets the What if? on the Playoff Odds page answer in a blink: pick some winners and it simply counts the kept seasons in which those results happened, and tells you how many that was. (In October 2026 the simulation got nearly twice as fast without a rule changing: the new code was checked against the old over a million seasons, identical.) This is a Monte Carlo simulation — named after the casino, because the core idea (run something random a huge number of times and trust the frequency of outcomes) is exactly how a roulette wheel works. Physicist Stanislaw Ulam popularized the technique in the 1940s after thinking about the odds of winning at solitaire.
We run two modes side by side, every time, so you can see how much our ratings are actually adding over pure chance:
100 × (times they made
it) / (seasons run) — an exact fraction, not a noisy estimate. At 250,000 seasons (the run right after a final) that fraction
can only ever have 4 real decimal digits, at 2,000,000 five; anything past that would be fabricated precision.
The decimal slider on the Playoff Odds page is capped accordingly, by the run behind the mode on screen (and,
in a what-if, by the seasons that matched your picks), so what you see is always genuinely all the
information the simulation produced — never padded, never rounded away.
The NFL's real tiebreaker procedure is long: head-to-head, division record, common games (minimum four), conference record, strength of victory, strength of schedule, several points-based ranking steps, and finally a coin toss — applied differently for two-team versus three-or-more-team ties, restarting from the top every time a team is eliminated from contention.
We implement the official procedure through strength of schedule, with the NFL's two separate orders: division ties go head-to-head, division record, common games, conference record, strength of victory, strength of schedule; wild-card ties go head-to-head, conference record, common games (minimum four), strength of victory, strength of schedule. Checked against real history: re-seeding every completed season from 2020 to 2025 reproduces all 14 actual playoff teams and both #1 seeds.
Beyond that we stop short of the official points-ranking steps and coin toss, and simplify two details (the three-team head-to-head "sweep" rule, and restarting the procedure each time a team drops out of a three-way tie). If a tie survives everything we implement, point differential decides it and we flag it rather than silently pretend to be the official procedure:
None of this happens in a vacuum. Real thanks to the people and projects whose work we leaned on:
The entire data backbone of this project — schedules, scores, play-by-play with EPA back to 1999, and weekly team rosters, licensed CC-BY 4.0. Maintained by Sebastian Carl, Ben Baldwin, Lee Sharpe, and contributors. If you build anything with NFL data, start here.
Every game from 1920 to 2022 with 538's Elo ratings, licensed CC BY 4.0 -- the source of pre-1999 records, all-time series, titles and Elo on this site. We modified it: team codes are mapped to today's franchises and each game is tagged NFL, AFL or AAFC. We checked it first: all 6,421 games of 1999-2022 match nflverse's scores exactly. 538 has no affiliation with Nerdfl.
The Elo rating system itself, invented for ranking chess players in the 1960s, decades before anyone applied it to football. Every "Elo" number on this site ultimately traces back to a chess mathematician, not a sports analyst.
Founded FiveThirtyEight on March 7, 2008, and served as its editor-in-chief until May 2023. The NFL Elo model this whole engine borrows its shape from was built under his direction — without him, there's no FiveThirtyEight NFL forecasting to adapt in the first place. He's since moved on to his own newsletter, Silver Bulletin.
Not the inventors of Elo, but the ones who adapted it specifically for football starting in 2014 — the margin-of-victory multiplier, home-field adjustment, and Elo-to-point-spread conversion this engine uses are their football-specific design, refined over years of public NFL forecasting. FiveThirtyEight shut down in 2025; we're crediting the work rather than linking to it, since the original site no longer reliably resolves.
The "2.5 to 3 points" home-field rule of thumb was conventional wisdom for decades. nfelo's research shows it began trending downward starting around 2016, and their live tracker currently puts it at 0.68 points over the last 16 weeks, 2.03 over the last 5 seasons, and 2.33 since 1999 (updated nightly; last checked 2026-10-28). Our coin-flip baseline's home-field tilt is fixed at nfelo's last-16-weeks figure from 2026-09-12, 2.24 points — a yardstick shouldn't wobble night to night — converted through the same points-to-Elo path Elo mode already uses.
The EPA and win-probability models underneath nflverse's play-by-play data — built by Ben Baldwin and Sebastian Carl on the shoulders of the original nflscrapR team's public research into Expected Points models.
A high-level fantasy player who looked at a site with Elo ratings, 100,000 simulated seasons and a live win-probability model, and asked the only question that matters on a Sunday afternoon: "how are my guys doing?" The live player stats and the little gold stars exist because he asked. Compensation: this card.
Jonny's number one NFL nerd since high school, and a rabid Bears fan, which explains a lot. He asked whether 4th-quarter rushing wins games, bet the yards and the attempts would answer differently, and was right. The Lab exists because he asked. Compensation: this card, and the answer.
Chemist by degree, IT consultant by trade, and a RadioShack store manager in a former life back in North Carolina — which might explain the urge to take every number on this page apart just to see how it works. Runs TechKnowledge LLC out of Pocatello, Idaho, and is the reason any of this exists in the first place: decided 6.5 minutes was an insult and 2 was the standard, then named his AI collaborator "DAK" because it sounded funnier than anything I could've come up with myself.
Most of the code, the debugging, and an unreasonable number of tangents about EPA sign conventions came from an AI named Claude — who Jonny calls DAK, the Distiller of All Knowledge. DAK did not play football, has never thrown a spiral, and once needed three tries to get a console print statement to read the right dictionary key. But it did test every formula on this page against real numbers before shipping it, which is either impressive or the bare minimum, depending who you ask.