How the HoFW Model Works

What drives the score

Coefficients are fit on standardized inputs, so bar width directly compares how much each feature matters to the model's induction predictions. Bigger bar = stronger signal.

Offensive linemen (T/G/C) use a separate formula. Without individual box-score stats, OL are scored on an honors-based proxy built from four inputs, heaviest first:

1. All-Decade Teamheaviest single input in the OL formula
2. First-team All-Proeach selection counts heavily
3. Pro Bowl selectionsmoderate weight per selection
4. Career value (wAV)lightest weight, often missing for players drafted before 1980

Reading the tiers

Cutoffs are set from where real Hall of Famers actually land on the scale, so each tier's size tracks the Hall's own historical selectivity of roughly 1.1% of eligible players.

HoFW vs. Induction Forecast

Two outputs, two different questions. They're easy to mix up, so check which one you're looking at.

Since the Class of 2027 there is one pool: every retired player competes together, and the Hall expects to elect five or six people a year in total. Part of that class is spoken for before any modeled player competes, so the queue budgets it in three parts.

The model ranks players only. Coaches and contributors are not scored: coaching careers need a separate model this site hasn't built, and contributors (owners, executives, broadcasters, scouts) have no common numeric record to score at all.

The two reservations are calibrated differently, on purpose. The coach and contributor still get protected spots, one finalist each on the 15-name slate, but once there they face the same room and the same 80% bar as the thirteen players. So each is budgeted like any other slate place: an expected class of five or six spread across fifteen finalists. A finalist place is not a guaranteed induction, which is why their row is a fraction of a slot rather than two. Their old election rate isn't a guide to this. Until 2023 they mostly got their own yes-or-no vote and nearly always passed, and in the three years they had to compete for votes, 2024 through 2026, none of them got in.

Long-retired candidates are the other case. They lost their protected route entirely: their historical rate came almost entirely from the senior committee, which no longer exists. Across the last ten classes (excluding the one-off Centennial Class of 2020), sixteen long-retired players were elected and fifteen of them arrived through that committee. Exactly one won a place on the regular ballot. So their reservation is scaled from that merit path alone, which is why it is a fraction of a slot rather than more than one.

So the long-retired figure rests on a single player. It's the right thing to measure, with about the worst sample size possible. It is the first number I expect to revise once the Class of 2027 is known. What Changed in the Hall of Fame Vote explains why the merge made this necessary.

Model accuracy

Validated two ways: cross-validated AUC-ROC on held-out induction outcomes, and a per-class rank check against every real induction class since 2005, testing whether real inductees ranked near the top when their player rows were held out from fitting.

Why pre-1999 careers get excluded from training

My stats window starts in 1999. Some of the tracked real Hall of Famers have careers that fully or partly predate it, so their production would be understated if I used it as-is. I originally tried to patch that with a correction feature and it made things worse, so every truncated-career row is now excluded entirely when the model is fit. A known-incomplete data point should not shape the formula for everyone else.

Already-enshrined players in this group get no score at all. They show as Enshrined with no number, because a number built on partial data would be worse than none. Players from that era who are not yet inducted still get scored from the cleanly-fit model and carry a small asterisk. I am backfilling real pre-1999 career totals year by year, sourced by hand from Pro-Football-Reference, to shrink this group over time.

Where the data comes from

Everything from 1999 on comes from nflverse, the open NFL data project, released under a Creative Commons licence. Everything before 1999 comes from Pro-Football-Reference.com and Stathead. So does every Pro Bowl and All-Pro selection back to 1983, plus the games-started and Approximate Value figures the offensive-line formula uses.

None of it is scraped. I collected every Pro-Football-Reference and Stathead file by hand, year by year, mostly with their own export tools, and copied the honors from their published pages. What this site shows is the model's output, a score and a forecast. It doesn't republish anyone's database.

There is no free, openly licensed alternative for the pre-1999 seasons. I looked. The commercial sports-data feeds start in 2000 or 2010, and the open datasets that do cover the 1980s trace back to Pro-Football-Reference anyway. They did the work of compiling that era, so I'd rather credit them directly.

This site is free and has no ads, which is part of why I'm comfortable building on that data. If that changes, I'll have to revisit this first.

What it cannot see

HoFW scores the case a player built on the field. The real vote turns on advocacy as much as evidence, and nothing in the data captures which selector is willing to argue for whom, or how a room reacts on the day.

Kickers, punters and several offensive line positions are not modeled at all. Players whose careers ended before 1999 fall outside the stats window, and a handful of real inductees only appear here through a hardcoded override.

A score answers how a player compares to the players already enshrined at their position. That is a different question from whether the committee gets to them this year, which is why the projected class and the induction probability can disagree.