The Draft Pick Value Curve

What's the expected career value of the 42nd pick in the NFL draft? Depending on the model, somewhere between 9% and 47% of the first pick's value. The range is wide because the question contains several smaller questions inside it: what counts as value, whether to adjust for what the player costs, and whether to optimize for average value or high upside.

People have been going back and forth on draft valuations for over 30 years.

Prior Work

It starts with the Dallas Cowboys. In the early 1990s, Cowboys VP Mike McCoy built a point-based trade chart at Jimmy Johnson's request, assigning 3,000 points to the first pick and declining steeply from there. McCoy derived the values from four years of actual draft-day trades. Importantly, it didn't look at how the players turned out, just from what teams had historically swapped for each other. Coaches carried the chart from job to job, and as Belichick later noted, "everybody probably uses about the same value chart." For twenty years, the Johnson chart was the agreed-upon exchange rate for NFL draft capital.

Nobody really tested it against reality until 2005, when Cade Massey (then at Yale) and Richard Thaler (Chicago, later a Nobel laureate) circulated a working paper that eventually became The Loser's Curse. They tracked every player drafted from 1991 to 2002 across five full NFL seasons and categorized each player-season into one of five tiers: not in the league, on the roster but not starting, backup (8 or fewer starts), starter, or Pro Bowl. A simple performance-based framework.

According to Massey and Thaler, the Johnson chart overvalued top picks by roughly 2x. The on-field production drop from pick 1 to pick 32 was much shallower than the chart suggested. First-round picks spent about as many player-seasons out of the league entirely (8%) as they did making Pro Bowls (9%). Whether a player had a better career than the next player drafted at his position was barely better than a coin flip: about 52%.

Then they factored in compensation. Top picks were getting $50-70 million contracts, and Massey-Thaler defined surplus value as performance value minus salary cost. Performance declined with draft position, but compensation declined faster. Surplus value actually increased through the first round, peaking in the early second. The first overall pick had the lowest surplus value of any pick in round one.

As Thaler does so well, he attributed the overvaluation to four psychological biases working in concert: non-regressive predictions (scouts fail to regress toward base rates), overconfidence (confidence grows with information even when accuracy doesn't), the winner's curse (the highest bidder in an auction with uncertain values systematically overpays), and the false consensus effect (teams overestimate how much other teams covet the same player, creating urgency to trade up).

In the years since, there have been more models and more value curves. Kevin Meers at the Harvard Sports Analysis Collective ran the numbers using Pro Football Reference's Career Approximate Value and found that the first pick produced about 5x the value of pick 94, while the Johnson chart implied a ratio closer to 24x. Chase Stuart at Football Perspective built his own chart from 38 years of data and got a much flatter curve than Johnson's. Everyone who looked at actual performance found the same thing: teams were systematically overpaying to move up.

The 2011 CBA changed the math. The new rookie wage scale capped first-round salaries. Sam Bradford signed for 6 years, $78 million ($50M guaranteed) as the first pick in 2010. Cam Newton, first pick in 2011, got 4 years, $22 million. That's a $56 million reduction in one year. Meers went back to his model and found the efficiency of the first overall pick jumped 3.4x. Slash the cost of top picks by two-thirds while production stays the same, and the surplus value math inverts completely.

The rookie wage salary caps had an especially big impact on the draft math for quarterbacks. Research from analysts at major football analytics firms has shown that a franchise QB on a rookie deal is probably the single most valuable asset in professional sports. Four years of top-10 quarterback play at a fraction of market rate puts almost any team in a Super Bowl window. For non-QBs, surplus value still tends to peak somewhere in the late first to early second round. And a 2024 paper from Wharton by Brill and Wyner makes an interesting counter to the Massey-Thaler framework: maybe teams aren't trying to maximize expected value at all. Maybe they're paying for the probability of landing a transformational player, the right tail of the distribution, and that's a rational thing to optimize for if championships are won by elite talent rather than roster-wide efficiency.

So where does this leave PerThirtySix. In short, we take a similar approach as analysts before us, but with a new data source. We measure on-field outcomes using pVAR, which combines player statistics, season awards, and per-snap grades.

How We Build the Curve

With 20 years of player-level pVAR scores, we can calculate what each pick slot historically produces. The challenge is noise, especially in the later rounds where any single slot has a small sample and high variance.

Variable-Bandwidth Gaussian Smoothing

Our approach: smooth the raw averages using a Gaussian kernel whose bandwidth widens as pick numbers increase.

For any two picks $i$ and $j$, the weight $w_{i,j}$ that pick $j$ receives when estimating the expected value of pick $i$:

The bandwidth grows with draft position:

Three parameters: $C = 2.5$ (overall spread), $L = 5$ (baseline offset so the top of the draft isn't under-smoothed), $S = 2.5$ (growth rate). The bandwidth roughly follows the square root of the pick number: tight at the top, wide at the back.

The expected pVAR for pick $i$ is a normalized weighted average across all pick slots:

The Shape of the Curve

Steep at the top, flattening through day two, approaching zero through day three. Pick 1 expects about 47.4 pVAR. Pick 32 expects 31.7. Pick 100: 14.9. Pick 200: 6.4.

The steepest part of the curve is between picks 1 and 15, a drop of about 8 points across 14 picks. After that it flattens out. The gap between pick 32 and pick 64 (about 10 points) is comparable to the gap between pick 1 and pick 15. For trade purposes, moving from 32 to 15 buys roughly the same expected pVAR gain as moving from 15 to 1, at a fraction of the cost in draft capital.

Value Over Expected

The expected value curve gives us a baseline:

The biggest steals since 2006, by value over expected (see the full list on our bests and busts page):

The biggest busts:

The steals are all fifth and sixth rounders who outproduced the entire top 10 of their draft class. The busts are all top-6 picks who produced nothing. The asymmetry here is structural, a team makes roughly 5 picks on day three for every 1 pick in the top 10. The maximum downside on any single bust is about -48 points. The maximum upside on a single late-round hit is +92.

Hit Rates by Draft Tier

We looked at every pick from 2006 to 2020 (3,817 players) and sorted them into tiers using the same rating scale as our draft analysis tools.

Only about one in seven top-10 picks reaches Star caliber. One in five hits Elite. The "Made 1+ AP" column tops out at 27% in the top 10: three out of four top-10 picks never make a single All-Pro team.

By round 2, Star odds are 2.5%. By day 3, it's a lottery ticket. And yet 2.9% of rounds 4-5 picks still make an All-Pro team at some point. Over 1,125 picks, that's 33 players found after the first hundred: Antonio Brown (pick 195), Jason Kelce (191), Tyreek Hill (165), Richard Sherman (154), Dak Prescott (135).

Every player in that dataset, grouped by round:

Distribution and Variance

The averages above are useful. They're also misleading. The variance changes character across the rounds, and for front-office strategy the shape of the distribution matters as much as its center.

The maximum pVAR in the top 10 is 109.4. In rounds 6-7, it's 98.4 (Antonio Brown). The ceiling barely moves across the draft. The probability of reaching it drops by a factor of 40. The coefficient of variation climbs from 69% to 250%. Late-round picks are lottery tickets with an expected return near zero, but the right tail extends almost as far as it does at the top.

The median-mean split makes this concrete:

In round 1, the median and mean are nearly identical. The distribution is roughly symmetric. By round 5, the mean (8) is 8x the median (1). In rounds 6-7, the median is literally zero. More than half of late-round picks produce no measurable career value. But the mean is still 4-5, held up entirely by the long right tail of outliers. A GM drafting in round 6 is most likely getting nothing. But the average outcome is a decent bench player, because one in fifty of those picks turns into Jason Kelce.

Brill and Wyner's critique of Massey-Thaler comes back to this point. Optimizing for expected value across a roster means trading down. The average pick at 32 isn't that much worse than the average pick at 5, and more picks means more shots. But a GM optimizing for the probability of landing a franchise-altering player faces different arithmetic: 14% at a Star in the top 10, 1% in round 3.

Positional Value

The curves below apply the same Gaussian smoothing used for the overall expected value curve, but computed separately for each position (see our position analysis tool for deeper dives by position).

The table below breaks this down further, showing average pVAR for each position at each draft slot range. The color intensity reflects the value.

The first-round Star rate varies enormously by position:

Offensive linemen have the highest average pVAR (44.3) and the lowest bust rate (25%) of any position. The standard deviation is relatively tight. A first-round OL almost always produces something.

QBs have the highest Star rate (16%) but also the highest standard deviation (35.1) and a 32% bust rate. The distribution is bimodal: the ones who hit tend to hit big, and the ones who miss produce almost nothing. Edge rushers and defensive backs bust at 39-40%.

Safeties and running backs are the low-variance positions: bust rates of 28% and 32% respectively, but both cluster toward steady production rather than extremes. The tradeoff is that neither produces many Star-level careers (0% and 6%).

Model Comparison

Five models, normalized to percentage of pick #1's value at each slot:

The numbers at a few key pick slots:

The Johnson chart says pick 32 retains 20% of the first pick's value. Every performance-based model puts it at 36-67%. At pick 100, Johnson says 3%. The data says 15-31%.

The models are built on very different foundations. A comparison, including several that aren't plotted above:

Three categories emerge. The trade-market models (Johnson, Hill) use no performance data; they measure what teams have been willing to pay. The surplus-value models (Massey-Thaler, PFF/Riske, Baldwin) factor in salary, which is why they find the sweet spot in the late 1st and early 2nd round. The pure performance models (Meers, Stuart, Over The Cap, PerThirtySix) ignore salary and ask what each slot produces on the field.

Limitations

Expected values are computed from 2006-2021 draft classes (per-snap grading data starts in 2006; the end year ensures every class has had at least five full seasons).

Fifteen years is short. Some pick slots only have 12-15 observations. The Gaussian smoothing helps but doesn't eliminate noise.

The curve assumes stationarity. The 2006 and 2020 drafts are treated as draws from the same distribution, but the NFL has changed over that period.

No salary component. We measure what a pick produced on the field, not net of cost. The pre/post-2011 CBA distinction, which fundamentally altered the economics of every first-round pick, is invisible to our metric. A hit at pick 5 under the current CBA is worth far more in surplus value than the same hit in 2008, but pVAR treats them identically.

Compensatory picks create mild distortion. Picks 33-40 often include compensatory selections awarded to teams that lost free agents, which could slightly inflate performance at those slots.

The steal and bust identifications hold up against the names, and every serious model since Massey-Thaler has arrived at roughly the same conclusion about the shape of the curve, even where they disagree on specifics.

For details on how we compute pVAR, see Introducing pVAR. To explore individual draft classes, see our season-by-season analysis. For team-level drafting grades, see draft report cards.