How Likely Are You to Draft a Starter?

luckiestman · May 3, 2023

Cool data. Makes me wonder how much you could narrow it down without much loss. Like how many guys started 48 games but didn’t get to 80 (16 game seasons your set ).

you could also profile positions and questions like didn’t start year one but then was an X year starter.

or you could know when to cut bait. If didnt start X games within first X years the likelihood of ever starting is ?.

it’s a good data set

joe dokes · May 3, 2023

That's great stuff.
I wonder if its possible to get more granular by some sort of points system assigned to fewer-than-5-year starters. My sense is that 5-year starters are stars or near-stars. Or at a minimum, there just aren't that many of them. And while teams should get credit in this context for drafting them, I wonder if only counting *them* is the best way to assess drafting. Believe me, this is not a criticism of your heavy lifting here. (lifting I am not capable of).

Morgan's Magic Snowplow · May 3, 2023

Cool data. How are you defining a starter? For example, if I guy starts 9 games out of 16 in a given year, is he classified as a starter? What if a guy plays only 8 games due to a big injury but starts them all?

pappymojo · May 3, 2023

This is very interesting. I don't want to ask you to do work, but I have always wondered about the ability to draft a good player based on draft slot. Basically, for so many years we have made excuses for the Patriots because they have been picking so late in the first round, is there any truth to that, and if so would it make more sense to consider picks 1-20 as the true first round, and to consider 21-40 as the true second round, etc.?

SMU_Sox · May 3, 2023

I would check out this study - do you use Python at all?

You can build and refine tiers and definitions too.

SMU_Sox · May 3, 2023

Oh and one more to look at. Chase Stuart took a look at AV by draft slot too:

Ferm Sheller · May 3, 2023

pappymojo said:
This is very interesting. I don't want to ask you to do work, but I have always wondered about the ability to draft a good player based on draft slot. Basically, for so many years we have made excuses for the Patriots because they have been picking so late in the first round, is there any truth to that, and if so would it make more sense to consider picks 1-20 as the true first round, and to consider 21-40 as the true second round, etc.?

I think that this data gives some indication. If you look at just round one, it appears that HoFers skew heavily toward being top 10-12 draft picks (but you also have to factor in every picked in round two and later -- I haven't looked at it closely yet).

Dotrat · May 3, 2023

This is great stuff. I wonder if it's also possible to weight teams' records. I may be wrong, but it seems to me that more players are more likely to start for losing teams, and especially for teams that consistently finish at or below .500.

Devizier · May 3, 2023

SMU_Sox said:
Oh and one more to look at. Chase Stuart took a look at AV by draft slot too:

View attachment 64334

The high variance on the back end of that decay is as good evidence for the "lottery tickets" argument as any.

InstaFace · May 3, 2023

Devizier said:
The high variance on the back end of that decay is as good evidence for the "lottery tickets" argument as any.

yeah, it took me a second, but those diamonds on the back end of the decay chart are not individual player outcomes plotted as a scatter, they're average-AV (so average-of-an-approximation

) for all players drafted at that pick #, meaning it gives no real sense of the probability of drafting someone useful, because 10 guys who are zeroes are averaged with one guy who's a 30 AV and one guy who's a 12 AV and so the average is a 4. Or rather, it hides that information behind the averaging.

BJ, this is cool info, but I would probably do this from a # of starts perspective (to capture useful players whose career was cut short by injury, or other quirks of a career that often has ups and downs). Some threshold like 30 career starts (or 20, or 50, I dunno). But I think "5-year starter" has a lot of noise associated with the definition that obscures what we're really trying to get at, which is a draft success that is materially useful to the drafting team.

For that matter, the 5-year threshold is odd too, since players who aren't stars (but might be useful) tend to change teams after 3 or 4 years. Only first-rounders come with a 5th year team option, right? So I would at least dial that back to 4-year starters.

Then for me, the next step I'd take with your data would be calculating some value-above-replacement-draft kind of metric, where you have some sort of set expectation based on the draft slots that a team is drafting with, and they have some outcomes (in AV, or just the binary "was he a starter long enough to clear this bar?" flag), and we can calculate the extent to which every team's drafts each year were successes or failures. And add up which team has done better or worse or by how much. And assess whether success or failure is something that can be consistently replicated by an organization or is just kinda random. All of which would be very cool insights.

InstaFace · May 3, 2023

From the explanation that PFR's founder, Doug Drinen, gave about the Approximate Value metric:

"Essentially, AV is a substitute for --- and a significant improvement upon, in my opinion --- metrics like 'number of seasons as a starter' or 'number of times making the pro bowl' or the like. You should think of it as being essentially like those two metrics, but with interpolation in between. That is, 'number of seasons as a starter' is a reasonable starting point if you're trying to measure, say, how good a particular draft class is, or what kind of player you can expect to get with the #13 pick in the draft. But obviously some starters are better than others. Starters on good teams are, as a group, better than starters on bad teams. Starting WRs who had lots of receiving yards are, as a group, better than starting WRs who did not have many receiving yards. Starters who made the pro bowl are, as a group, better than starters who didn't, and so on. And non-starters aren't worthless, so they get some points too."

I submit that their "Drafting-team AV" number, which should be in the same dataset you're using, is probably a better metric than either of the ones you're looking at. It better-captures the full range of player quality, makes allowances for players who are just-under or massively-above your thresholds, and allows us to more precisely answer the questions I posed in my last reply. It's also usable only 4 seasons of play removed from the draft year, so although the story isn't fully closed on the careers of players in the 2018 draft, it pretty much is for the purpose of assessing what the drafting team got from them, for all the players who weren't so good that the drafting team decided to sign them to an extension or keep them in free agency (which is a small minority). You definitely wouldn't have to stop at the 2013 Draft.

More significantly, it gets at the problem you're talking about in post #16, which is that the draftees of crap teams don't have as high a bar to clear to be a starter as the draftees on good teams (in general), and answers it through the AV metric. Teams and positional units that are collectively crap will generate less AV to be distributed among their various players, even if any team's stars will still come off well by the metric.

InstaFace · May 5, 2023

They have, in fact, shared how they arrive at the AV numbers.

https://www.pro-football-reference.com/about/approximate_value.htm

Doug Drinen developed it in a series of forum threads which are also linked at the bottom of that page, if you want to read the back-and-forth between him and other nutcases.

InstaFace · May 5, 2023

So I didn't want to be left out of the fun, so I went through the minor bother of downloading all the draft histories and evaluations from PFR, from 1995 (when they went from 28 to 30 teams) to 2018 (the most recent year we could reasonably start to evaluate the quality of a pick retrospectively). And I've put my analysis up at this google sheet which is publicly accessible. This yielded a little over 6000 picks worth of data.

Here are some probabilities of a pick at a given pick # resulting in a thing, plus some regression curves which I am using to give a fairer estimate against which to evaluate a team's pick quality.

Those measures are, respectively, "Starter 3+ years", "Starter 5+ years", "Pro Bowl 1+ time", "All-Pro 1st Team 1+ time", "Played in at least 30 games", and "Accumulated at least 10 AV". I have a separate one for the average expected EV by draft pick, for the drafting team (DrAV).

The regressions are either Exponential or Logarithmic, whatever resulted in higher R^2. Interestingly, the regressions for the DrAV one were about the same for Exponential and Logarithmic and so I got to decide whether I'd prefer slightly greater accuracy at the top end (early 1st round) of the draft, or better fit for the vast majority of the back end of the drafts. I chose the latter.

Next up I'm going to:
- Assess variance of pick #s, using these value measures
- use those regression values as a baseline upon which to evaluate each team's picks over this window
- Calculate a team's Value Over Replacement Pick that they've created, positive or negative, using some weighted average of some of these measures. I'll probably give people a settings panel where they can adjust weights to their liking and decide how teams have done.

Notes:
- I'm going to focus on picks 1-256, even though a few drafts had more (and frankly I might reduce to 1-240 except that it doesn't really change anything).
- My input to the regressions started with a nearest-neighbor smoothing using the adjacent two values, even though I probably shouldn't have done that before the regression.
- AP1 is super random, Pro Bowl not much less so. AV, Games Played and Years Started are definitely better measures.
- There was such a big jump in average AV by pick right at Pick #199, even averaged over 24 years, that it was actually the impetus for me to do regression smoothing rather than use the real-world (noisier) values.

If you want to play around with this data or take advantage of some of the data-prep steps I've done, feel free to Make A Copy on google sheets and go to town on it.

InstaFace · May 5, 2023

BaseballJones said:
I'm working through the 2000-2018 drafts (the caveat being, again, that this was data culled and sorted before the 2022 season, so if you look it all up now it'll take some adjustments, but it's still plenty big enough of a sample size for the purposes of this study). Here's some more thoughts.

So firstly, yeah, my data was pulled fresh last night, and it's just a straight copy/paste from the PFR Draft pages for each year. So it should be the same underlying format as whatever you're using (I added a Year column and a sequence number in the front for easier sorting, but that's it), if you want to update yours, or just start from mine.

I looked at every single pick and slotted them over that 19-year period. In other words, looking at all 19 #1 picks, all 19 #117 picks, all 19 #232 picks, etc. From 1 through 261 (there weren't many pick number 261 but still). Over this time period, here's all the draft slots the Pats have had from 2000-2018:

...

So here's their average draft slot for these rounds:

....

Here's the average career AV for those draft slots from 2000-2018:

...

So if you take the Pats' picks from 2000-2018, here's the expected career AV from all those picks:

...

Here's what the Pats, over that time, have actually gotten in career AV.

So I think this sort of pick-position-averaging actually tends to remove a lot of information. For one thing, the decay in pick value as pick number increases is exponential, not linear, so an arithmetic averaging probably under-sells the expected value of picking towards the end of rounds as the Pats usually did. For just about any measure other than "Odds of playing 30 Games", which is, somewhat oddly, almost linear.

What I'm going to do instead is calculate the expected value of each specific pick the teams each made, whatever number it was in that year, across a range of value measures. And then look at what they got with that pick, above/below expectation. I think that'll reduce the noise and also permit easier subdividing by era, by position, by pick #, by team, and so on.

Where I need advice (or clearer thinking) is on two things:

(1) Binary or continuous value measures? Simple flags of "did they get a starter who started for X years?" ignore the value difference between getting, say, a starter for 2 years vs 0, or a starter for 5 years vs 10. It's fine to use that in some contexts, but measures that have a fuller value range are probably conveying more information. But it's using those simple thresholds, of did the guy clear this bar or not, do end up somewhat mitigating the concerns about "a bad team will play more draft picks because their existing players suck more" (because a team's context usually changes every few years, and because a simple Yes/No threshold test reduces the need to even think about the team context factors that might turn a 30 AV player into a 50 AV player, or a would-be 3-year starter into a 7-year starter (or the reverse)). It focuses the information on less noisy facts. It's been a while since information theory class, but I think that helps if and only you have enough data points to get a good estimate of central tendency even with the information loss of reducing it to a boolean Yes/No. Do we? Each pick position only has N=24 (or fewer at the tail end). I just don't know whether binary measures will end up having more or less information loss than the continuous ones.

(2) So those continuous value measures - things like AV, Games Played, or # of Years Started - might be preferable. On the third hand, per the PFR commentary above, half the reason AV was created was to be an improvement over the cruder measures like Games or Years Started. So maybe AV is our do-it-all stat for value capture? Or should we take several of them and weight them together to assess a pick's value more holistically? A high-AV player who doesn't end up starting that many years because he gets hurt (say, star RBs) is still very valuable for the years he does play. A lower-end defensive lineman who might not make the pro bowl but is very good at avoiding injury and staying on the field and accumulates lots of games played, even if he's not always the primary starter, has value too, just in a different way. AV does attempt to interpolate between these different approaches to value, but maybe we should be using all of them anyway, just so we don't over- or under-state a player's value to a team. Which ones would you focus on?

So if we include the Brady pick (and we SHOULD, but there will be detractors from that argument), the Pats have:

- Gotten a career AV of 2,801
- Had an expected career AV of 2,305
- Gotten a net career AV of +496 over expected (which comes to +3.0 per pick)

If you remove Brady then the Pats have:

- Gotten a career AV of 2,500
- Had an expected career AV of 2,092
- Gotten a net career AV of +408 over expected (which comes to +2.5 per pick)

I'm having a bit of trouble here, because Brady's AV was 184 for his career, 168 for the Pats, and yet the difference between your with-Brady and without-Brady expected career AVs is 213, and actual-AV numbers is 301. If we replace Brady with an average pick at that position (with AV around 4.9), the difference should be less, if all else was held the same. What happened?

As for your musings about adjusting for Brady: There's no reason to ignore the Brady pick, imo. It's the most consequential outcome-altering team decision in the history of the NFL. It, like, happened, and then he went and did all that shit. The flags will fly forever. But also, at this point, it's far from the entire story of the Patriots' drafting over the last few decades. As you point out, his 168 AV for the Patriots is amazing but it's also like ~5% of the total AV you'd have expected for the Patriots over the course of time. The data is big enough at this point that conclusions won't be significantly altered even by an outlier as big as that one. So I think we keep him in, and give the Pats full credit for that miracle (which was of course the result of careful scouting and Belichick listening to different perspectives).

InstaFace · May 5, 2023

In fairness, Wilt Chamberlain averaged over 50 points a game that season, so while the 100 point game was an outlier and a target they decided to hit, it doesn't give you a misleading impression of his season, either

MainerInExile · May 5, 2023

The round-by-round breakdown seems to correspond well to the conventional wisdom around here. They hit in the first round more often than not, and they do well to find contributors in later rounds. Round two is maddening.

ZMart100 · May 6, 2023

I did something similar to what I think Instaface is suggesting a few years ago looking at starts. Position matters. Examples: WR

The parameterization I used did lead to some funky things happening late in the draft though.

slamminsammya · May 6, 2023

ZMart100 said:
I did something similar to what I think Instaface is suggesting a few years ago looking at starts. Position matters. Examples: WR
View attachment 64434
View attachment 64435
The parameterization I used did lead to some funky things happening late in the draft though.
View attachment 64436

Seems like you wanna enforce some monotone condition - how did you parametrize these?

ZMart100 · May 7, 2023

slamminsammya said:
Seems like you wanna enforce some monotone condition - how did you parametrize these?

It was awhile ago and I don't want to dig up the code. I have a vague memory that it was actually non-parametric loess, but I'm sure I was playing with other things too.

InstaFace · May 8, 2023

BaseballJones said:
Here's the reality: you are smarter than me about how to really analyze this data. I'm a regular guy that barely got through calculus, so when I do this kind of work, I really have to keep things simple. I wish I knew this stuff as well as you do, but alas.....

Thanks for your kind words. I think I'm about at the same level as you and the other data-literate people around here when it comes to crunching sports data, not least because I have a lot less experience in that data realm, and also have less subtle context about the game itself or how to control for different variables. And there's certainly nothing at the level of calculus that we're doing here regardless. Anyway, we're on the same page and same team here, let's see what we can learn together.

I think looking at it in a variety of ways gives you the best overall picture. Just as a different example, imagine a team making 70 draft picks over a ten year period. It tells one story to look at how many starters they drafted. It tells another story to look at how many superstars they drafted. AV kind of gives you both - a big AV number is only possible if a guy is (a) really, really good, and (b) plays a long time. But it doesn't tell you everything and in many ways is team dependent, so it's not necessarily a fair reflection on any individual player taken in isolation. I don't think we need to boil it down to one metric.

I know that means more work, but it appears that you and I are enjoying it so.....

This is fair. I think I will:

- Add a few other continuous metrics, such as # Years Starter, # Games Played, to go with our binary / threshold flags (and the DrAV stat which is already continuous). More recent years will have more noisy data for those stats, because ideally you'd want them to accumulate over a few more years, but we can probably make that adjustable.
- Concoct some sort of weighted average for them, but put out multiple sets of conclusions shaped by different sets of metrics or combinations thereof
- Try to get some sort of team-performance context into there, because like the Browns have been terrible forever so a metric that heavily weights Games Played will make them look better, while a metric that weights AV will make them look worse

I just removed the pick entirely and didn't replace Brady with an average pick. I just took the entirety of Brady's career (2000-2021, because, again, I am using the data from before the 2022 season) of 301 AV. I just took all 301 AV off the board from the Pats' total.

I must be seeing some different data here because on Tom Brady's PFR page it's got him at 326 career AV (316 excluding 2022, given your data), and when you look at... oh wait, OK, the column in the draft page is actually "weighted career AV", downweighting his lesser years, and for the handful of players who played over 20 years excluding his worst years entirely. So that's an odd stat, I wish they'd just put in career AV as a column. But then you've got the DrAV column, AV accumulated for the team that Drafted this player - it's got Brady at 168. You go and look at his page, though, and his career AV for the Patriots was 285, a number you should have the same as I see it there. That's a huge difference! What on earth is this site doing here? I'm going to file a support ticket. DrAV is broken or poorly documented, and they need to just have Career AV in there as a no-shenanigans option.

Here's the latest progress.

- Switched to Logarithmic formula for (Drafting-team) AV by pick #, as that's a lot more accurate if you don't do moving-average smoothing first
- Added continuous-value metrics for # Games Played and # Years Started, per the above
- Couldn't decide from among two different methods for scoring a pick, so the spreadsheet's control panel now offers both. For continuous (counting-stat) metrics, the choices are a 100-based index or a z-score (recentered to 100). For threshold (binary yes/no) metrics, there are both linear and exponential scoring choices.
- Unsurprisingly, the standard deviations for AV, GP and YS (Years Started) all have a huge spike at pick #199.

So I went and fit StDev to a trendline and calculated the (regression-expected) StDevs for those, as well.
- While AV had a steady slope to the standard deviation (R^2 = 0.568 on a logarithmic fit), as did YS (R^2 = 0.589 on an exponential fit), GP meanwhile had basically no pattern at all, with R^2s all right around 0.1 for every regression type. So I'm plugging in a baseline StDev of 46.9 (the average StDev for all pick #s) for Games Played, and calling it a day. Any pick is equally high-variance when it comes to games played, it seems.
- For the very tail end of the regression-driven formulas for predicting odds for Pro Bowl and All-Pro selections, the predicted odds went negative, which thwarts the scoring. I have switched the regression for Pro Bowl from Logarithmic to Exponential starting after the 4th round, which doesn't fit as well overall but does not decay to negatives. For All-Pro, I switched to a Power Series fit line partway through the 2nd round (picks 39+), which fits better overall and also does not decay to negatives.
- Added scoring evaluations (based on these metrics) for each pick in our range
- Created average scores for each team

Results are visible at the same sheet.

View: https://docs.google.com/spreadsheets/d/1C9vUQZtSFjQ-3REqvLzvNwGmyQLN3fUKrszx2TPzXu8/edit#gid=949691895

And the latest dilemma:

The "Drafting-team AV" metric is wonky because of draft-day trades. For example, Eli Manning is a DrAV of 0 for the team that drafted him (the San Diego Chargers), but it's not fair to say they "missed" on him (nor that the Giants "missed" on Rivers): the Chargers got Philip Rivers and 1st- and 3rd-rounders, pretty close to full value; the Chargers are also getting credit for the quality of those picks received in trade. My sense is, the number of Eli Manning situations (draft-and-trade; shouldn't penalize drafting team) greatly outnumber the number of Lawrence Guy situations (underutilized, released / waived from a couple of teams before later finding success; drafting team should get no credit, was a 0 for the drafting team). So for the ~10% of picks where DrAV = 0 but Career AV > 0, which includes both Eli and Guy, I'm giving the team full credit for the Career AV, because between the two imperfect options that seems like the less-bad choice. A third option would be "ignore", i.e. act as if the team never made the pick, they don't get a zero and they don't get the player's result. But I'm very unsure about this. I'm happy to provide a list of the 31 players who have > 20 AV in this situation (i.e. the ones that matter) and let people look through and decide if, in general and as a group, the drafting team should get full credit, or none, or ignore. It's a reversible decision, but not one I can make into a flexible user option all that easily.


2017	7	233	CAR	Harrison Butker	K
2016	2	41	BUF	Reggie Ragland	ILB
2016	6	208	NWE	Kamu Grugier-Hill	OLB
2015	7	222	WAS	Austin Reiter	C
2011	7	233	GNB	Lawrence Guy	DT
2010	4	123	NOR	Al Woods	DT
2010	6	205	NWE	Ted Larsen	C
2009	3	79	PIT	Kraig Urbik	T
2009	7	226	PIT	A.Q. Shipley	C
2008	6	167	DAL	Erik Walden	DE
2006	6	191	NWE	Jeremy Mincey	DE
2005	6	207	CAR	Joe Berger	T
2005	6	213	BAL	Derek Anderson	QB
2005	7	216	MIA	Kevin Vickerson	DT
2004	1	1	SDG	Eli Manning	QB
2004	1	4	NYG	Philip Rivers	QB
2003	5	166	GNB	Hunter Hillenmeyer	LB
2003	6	185	PHI	Jeremy Bridges	G
2003	7	255	NYG	Kevin Walter	WR
2002	7	238	PHI	Raheem Brock	DE
2000	6	168	NOR	Marc Bulger	QB
2000	7	253	DET	Alfonso Boone	DT
1999	4	131	GNB	Aaron Brooks	QB
1999	6	173	CIN	Kelly Gregg	NT
1998	6	169	SEA	Bobby Shaw	WR
1998	7	207	DET	Chris Liwienski	G
1997	4	119	PHI	Damien Robinson	DB
1997	6	169	TAM	Al Harris	DB
1995	4	122	MIA	Pete Mitchell	TE
1995	7	241	ARI	Chad Eaton	DT
1995	7	242	NOR	Travis Davis	DB

Eddie Jurak · May 9, 2023

Has anyone looked at player age when drafted vs. overrall value and 4-year (ie, first contract) value?

It has sort of struck me in the past draft or two that the Pats seem to be drafting some older players... but at least some of that seemed possibly related to Covid disruptions and it isn't clear to me whether they are doing anything different than the rest of the league in this regard.

I would also assume that the pool of players who choose to enter the draft early (and thus younger) tend to be more talented, and this would have to be filtered out to really assess whether older is better.

SMU_Sox · May 9, 2023

I can’t find it but the Patriots in the last 5 years or so and I believe overall tend to draft older players. In the last 5-6 years or whatever they have been the oldest in the NFL. I’ll try and find it. They love senior bowl guys.

SMU_Sox · May 9, 2023

Almost everyone they drafted this year who I watched was a senior bowl or shrine bowl guy including 3/3 of their day 1 and 2 picks being senior bowl guys.

luckiestman · May 9, 2023

SMU_Sox said:
Almost everyone they drafted this year who I watched was a senior bowl or shrine bowl guy including 3/3 of their day 1 and 2 picks being senior bowl guys.

Do you listen to Bucky and Jeremiah? If so, what did you think about their discussion of the Colts' trait-based approach?

SMU_Sox · May 10, 2023

luckiestman said:
Do you listen to Bucky and Jeremiah? If so, what did you think about their discussion of the Colts' trait-based approach?

I think there isn’t anything new about having traits based scouting and drafting - BB has tended towards high athleticism in early picks even more than average. Think WRs round 2 for them for example or their OL picks. Bill also likes to take athletes late. Keion Crossen comes to mind. A lot of other teams and GMs take similar approaches. Coaches have been putting athletes through development for a long time. But here Bucky believes the Colts are going “all in” on this approach. I’m skeptical that going all-in on this theory is going to work out. For one even with position coaching depth most of these guys are refining core skills in the off-season with private coaches. You also have lower success rates with guys who don’t have production in college. Taking both great athletes with production is ideal but Bucky is saying the Colts are trading the former for the latter. You also get into complications with how much guys can develop on the fly. You aren’t going to start a developmental OT because his ass is going to give up way too many sacks and pressures.
Committing to developing a high traits QB isn’t new either and teams are doing s better job integrating QBs now than they were in the mid 2000s for sure. The NFL game is closer to the college game offensively than it has been in a long time. So I do think you can develop AR in that way by letting him develop with reps vs being able to do that vs the developmental OT. Some positions you can’t or shouldn’t develop in live action.
I’m curious to see how it works out as well. I believe it will work out for AR and SS did wonders for JH in Philly. But the other positions? I don’t know. Also I think AR wasn’t necessarily raw but just inexperienced. Guys who still look raw but have experience and are good athletes imo are much less likely to succeed.

Eddie Jurak · May 10, 2023

BaseballJones said:
A bunch of guys in the PFR draft data don't have ages attached to them (255 of them to be exact). But of the rest drafted from 2000-2018, I grouped them by age and then by career wAV. Here's how it broke down by draft age and their average career AV for that age group:

Age 29 (1 player): 9.0 AV
Age 28 (4 players): 6.5 AV
Age 27 (2 players): 1.5 AV
Age 26 (24 players): 10.0 AV
Age 25 (70 players): 8.6 AV
Age 24 (619 players): 11.1 AV
Age 23 (2,625 players): 14.7 AV
Age 22 (2,470 players): 19.2 AV
Age 21 (773 players): 26.2 AV
Age 20 (47 players): 28.0 AV

So yeah, the younger you draft them, the more likely it is that they'll give you more career AV than if you draft them older. Which makes sense because they can play in the league longer and accumulate more AV.

But what I want to figure out is what @Eddie Jurak asked, which is, what is their AV for their first contract? I am not sure I can do THAT without literally going through every player's page and add up their AV their first four years (and for some it would be five years, and no way can I parse THAT). But what I can do is look at their DrAV, which is their value to their DRAFTED team. This is not perfect, because it says nothing about their first contracts. A player might get moved in year 3 of his 4 year rookie deal, and this doesn't help us with that. Or a player could play 8 years for their team and this reflects those 8 years but doesn't tell us about his first contract.

In any case, here's the DrAV data for those age groups:

Age 29 - 9.0 DrAV
Age 28 - 5.0 DrAV
Age 27 - 1.5 DrAV
Age 26 - 7.8 DrAV
Age 25 - 6.7 DrAV
Age 24 - 8.0 DrAV
Age 23 - 10.3 DrAV
Age 22 - 13.4 DrAV
Age 21 - 19.9 DrAV
Age 20 - 21.5 DrAV

Same pattern. Draft younger players and you get more out of them, both for the team drafting them and the NFL as a whole.

But that still doesn't tell me what I really want to know. I'll keep working on this.

Nice work, though the inability to easily do certain data cuts makes it harder. I think using total value is probably an adequate proxy for 4-year value, at least in the early going when that is what you have.

I think there might be some confounding issues. To enter the draft at age 20 means the player declared early, and early-declarers are likely to be better players than those who do their full time in college. The other potential issue that comes to mind is draft position.

Super Nomario · May 10, 2023

Do you have Stathead @BaseballJones ? You could do something like this:

https://stathead.com/football/player-season-finder.cgi?request=1&match=player_season_combined&order_by=av&season_end=4&draft_year_min=2000&draft_year_max=2013

It's in Season Finder, not Draft Finder.

How Likely Are You to Draft a Starter?

ivanvamp

ivanvamp

Son of the Harpy

ivanvamp

Member

ivanvamp

Member

Member

queer eye for the next pats guy

queer eye for the next pats guy

Member

Well-Known Member

Member

The Ultimate One

ivanvamp

ivanvamp

The Ultimate One

ivanvamp

ivanvamp

ivanvamp

The Ultimate One

ivanvamp

The Ultimate One

ivanvamp

The Ultimate One

ivanvamp

ivanvamp

The Ultimate One

Well-Known Member

Member

Member

Member

The Ultimate One

canderson-lite

ivanvamp

queer eye for the next pats guy

queer eye for the next pats guy

Son of the Harpy

queer eye for the next pats guy

ivanvamp

ivanvamp

canderson-lite

ivanvamp

Member

ivanvamp