Introducing pVAR

A single number can't capture an NFL career. It can't account for the coverage scheme that inflated a corner's stats, or the torn ACL that ended a prime two years early, or the offensive line that made a running back look like a star. What it can do, if built carefully, is rank 5,000 players in an order that makes sense more often than it doesn't.

The history of sports analytics is a history of choosing what to measure. Baseball chose wins above replacement. Basketball chose Estimated Plus-Minus and DARKO. Football never settled on anything, partly because the sport is harder to decompose into individual contributions, and partly because nobody could agree on what mattered. We took a shot at it.

pVAR (player Value Above Replacement) is our attempt: a single number measuring how valuable a draft pick turned out to be, based on what a player did on the field across his career. It's the backbone of our draft redraft rankings, team report cards, and bests and busts analysis.

The Philosophy

Building a player value metric is an exercise in deciding what matters. Career counting stats reward longevity. Single-season peaks reward flash. We wanted a metric that values sustained excellence, doesn't get fooled by empty volume, and recognizes the difference between an elite player and a replacement-level starter who hung around for 12 years.

Pro Football Reference's Approximate Value, the most widely used career value metric in football analytics, has a well-documented bias toward accumulation. As an example, Ryan Tannehill, a perfectly fine but unspectacular quarterback, rates in at a 114, while Calvin Johnson comes in at 94. That shouldn't happen.

In an attempt to add weight to peak years, Pro Football Reference also has a "Weighted Approximate Value" (WAV), which is computed as

100% of the player's best season, plus 95% of his 2nd-best season, plus 90% of his 3rd-best season, plus 85% of his 4th-best season, and so on...

According to Weighted Career AV, Ryan Tannehill scores a 91 to Calvin Johnson's 78. Closer, but something still seems off.

Our approach: lean on per-play grading data to measure quality, use AV as a complement (not a foundation), and add awards to make sure truly dominant players are credited for it.

Data Sources

We pull from three sources, each capturing something the others miss:

Per-Snap Player Grades — Every player on every snap of every NFL game is graded by analysts watching film, producing a 0-100 season grade. A 60 represents replacement level. These grades form the core of our metric. This grading data is available from the 2006 season onward, which is why our analysis starts there.

Pro Football Reference (PFR) Stats — Career statistics, games played, and Weighted Approximate Value (wAV). AV is a box-score approximation of career value developed by PFR. It's imperfect (it over-credits counting stats and under-credits positions like safety and cornerback) but it captures real things that per-snap grades don't, like playoff performance and era-adjusted statistical production.

Awards — MVP, Offensive/Defensive Player of the Year, Rookie of the Year, and All-Pro selections (1st and 2nd team). Pro Bowl selections are not included; they've become a poor proxy for season performance. Awards are scraped from publicly available records and matched to draft picks by name.

From Season Grades to Career Value

A season grade captures one year. Draft value is a career question. Compressing an entire career into a single number requires choices, and those choices determine what the metric rewards.

We look at each player's career through six overlapping windows: the best single season, the best 3, 5, 7, and 10 seasons, and the full career. For each window, we compute a snap-weighted average grade. The snap weighting matters: a 90-grade season over 1,100 snaps carries more information than a 90-grade season over 200. Each season's influence is proportional to how much the player actually played.

We use "peak" windows (a player's $N$ best seasons by grade) rather than chronological windows. A player's prime might be years 2 through 6, or 4 through 8, or scattered across a decade. We don't penalize late bloomers or players whose second season was cut short by injury.

We also only count seasons on a player's dominant side of the ball, determined by career snap totals. J.J. Watt's handful of offensive snaps don't get mixed into his defensive grade.

The Formula

For each window, we subtract 60 (the replacement level) from the grade before it enters the formula. This is the same idea as WAR in baseball: we're measuring how much better a player was than the replacement-level alternative. A grade of 75 contributes 15 points. A grade of 60 contributes nothing. Below 60, the contribution goes negative.

We also apply a reliability weight based on snap count. A player needs to have logged enough snaps for us to trust the grade. Below a threshold, the contribution scales down linearly. A player with 2,000 career snaps gets $2000/6000 = 0.33$ of their career window credited. A player with 6,000+ snaps gets the full weight. This is how the metric handles the guy who played 150 snaps and earned a 92: the grade is real, but we don't trust it as much as a decade of film.

The grading component is a sum across all six windows:

where we sum over windows $N \in \{1, 3, 5, 7, 10, \text{career}\}$, $G_N$ is the snap-weighted average of a player's best $N$ seasons, and $k_N$ is the snap threshold for that window (600 for a single season up to 6,000 for the 10-year and career windows).

A player who was elite across every timescale, like Aaron Donald, scores high on all six terms. A player who had two great years and eight mediocre ones scores well on the short windows but loses ground on the longer ones. The sum rewards sustained excellence without ignoring peak performance.

Approximate Value (Adjusted)

PFR's Weighted AV captures things per-snap grades miss, most notably longevity and the accumulated value of showing up every week for a decade. But raw AV rewards accumulation too aggressively. A replacement-level player who starts 200 games can compile 60+ AV simply by being available.

We address this by subtracting a small per-game tax before AV enters the formula:

This deducts about 1.7 points per full season (17 games $\times$ 0.1). A player with 80 wAV over 140 games gets credit for 66 adjusted AV. A player with 40 wAV over 140 games gets credit for 26. The floor of zero prevents the adjustment from penalizing short-career players who barely accumulated AV in the first place.

The adjusted AV enters the formula through two terms, one measuring AV efficiency (per game) and one measuring volume:

The first term rewards high AV-per-game rates. The second term gives credit for total career production, but with the per-game tax already applied, it no longer rewards mere survival.

Award Bonuses

Neither grades nor box scores capture the consensus recognition that comes with winning MVP or making an All-Pro team. We add flat bonuses:

These bonuses are small relative to the grading component. An MVP adds 5 points to a raw score that can exceed 300 for elite players. The intent is to break ties and fill in gaps where the grading data falls short, not to let awards drive the rankings. Adrian Peterson (MVP, OPOY, OROY, 4x 1st Team All-Pro, 3x 2nd Team) collects 21 bonus points from awards — a meaningful contribution to his 89.2 pVAR, but still less than a quarter of the total.

Putting It Together

The raw pVAR is:

Scaling

Raw values aren't intuitive. We normalize them so that the scale is roughly 0-100, with a handful of all-time players exceeding 100:

The scale factor $f$ is computed dynamically: it's the $\text{pVAR}_{\text{raw}}$ of the 5th-highest player in the mature-career window (2006-2020), divided by 105. This means the 5th-best player in our dataset scores roughly 105, and everyone else is scaled accordingly.

Because the scale factor is computed from the data, the formula is self-calibrating. If we change weights or add components, the output scale adjusts without manual recalibration.

Negative values (players who washed out) are compressed logarithmically so busts don't have outsized influence on aggregate statistics.

From Player Value to Draft Analysis

A pVAR score by itself doesn't indicate whether a pick was good or bad. A 50 pVAR from the 1st overall pick is a disappointment; a 50 pVAR from pick 154 is a franchise-altering steal. To evaluate draft picks, we need to know what a given slot is expected to produce.

We build that baseline using variable-bandwidth Gaussian smoothing across 15 years of historical data. A player's value over expected is then:

This is the number that drives our redraft rankings, steal/bust labels, and team draft grades.

Limitations

Caveats:

Per-snap grades are the foundation, and they're imperfect. The grading is standardized and comprehensive, but it's still a human judgment applied to film. Some positions (offensive line, off-ball coverage) are harder to grade than others. Per-snap grading has become an industry standard, but it is one lens, not ground truth.

The 2006 cutoff. Per-snap grading data starts in 2006, so our metric can't evaluate players drafted before then. Peyton Manning (1998) only shows up because he played into the grading era, and his pre-2006 seasons are invisible to us. This means the metric is most reliable for players drafted from 2006 onward.

Position value is not accounted for. A dominant left tackle and a dominant safety contribute to winning in very different ways, but our metric treats a grade of 90 the same at every position. pVAR is position-agnostic by design; positional context is handled downstream in our draft position analysis.

AV remains biased, even after adjustment. Our per-game tax mitigates the worst of AV's accumulation bias, but it doesn't eliminate it entirely. Quarterbacks and edge rushers still tend to accumulate more AV than defensive backs and interior offensive linemen of comparable quality.

Awards are inconsistent. MVP voting and All-Pro selection are influenced by narrative, market size, and recency bias. Our bonuses are kept small enough that an award snub doesn't dramatically change a player's score, but they do introduce a human judgment layer on top of the analytical one.

The metric produces rankings where the names at the top and bottom make sense. The edge cases it gets wrong are ones any single-number system would struggle with. We'll keep refining it.