Introduction
I scroll through the media section every day or so to see what new, interesting things have been written. The other day Duilio’s article on what a typical Hall of Fame member looks like and how some current and recently retired players stack up to those was super interesting. Every sport has a group of dedicated nerds who have developed a system for determining the chance someone has to make the Hall of Fame. Its all basically voodoo science - more goes into being a Hall of Famer than just how your stats compare to others (although that is a major part of it). Football actually has its own metric that is based off approximate value. AV is cool and something I’ve dabbled in during my time on this site, however its a pain to calculate and although I’ve jumped down that rabbit hole as well and am in the middle of exploring how it works I decided to try to formulate my own system thats a cross between what they use in basketball and the one they use in football. Its pretty simple - the entire thing is based off a logarithmic regression that I then take the data from and plug into another series of equations to normalize the data and stretch it out. I’ll get into that in the technical details section of this article but I know that not everyone cares about the math but still wants a basic explanation. Basically I take a variety of inputs and compare the values from players who aren’t in the Hall of Fame to those who are and then determine what typically constitutes a Hall of Famer in terms of “statistics” and then try to fit that data to general expectations a little with some of my own weighting systems.
The “Technical” Details
As I explained above this is all based off a logarithmic regression. I took a variety of different inputs ranging from how many Ultimus rings a player has to where they stack up on the leaderboards for career and season stats. Those became the “X” inputs in the regression. The “Y” input was a binary system simply based off whether the player was a Hall of Famer or not.
For some of these x-inputs I instituted by own weighting system to try to differentiate between a “great” player and just a “very good” or “average” one. A simplistic explanation of this would be that although I could’ve just used the pure number of awards won by a player that fails to recognize that not all awards are created equal. Winning MVP, OPOY, or DPOY is much harder than winning your positional award. So I went in and assigned values to each of those awards - essentially weighting it so that MVP was the most valuable award, OPOY and DPOY the second most, all others on a scale below those. I did the same type of custom weighting with the all-time stats lists to try to differentiate between a player who is the all-time leader in a career/season stat and a player who is 9th all time in that category. Both are impressive, however players higher on the list should receive more credit for their accomplishments. I’m happy to answer any questions about my weightings if people are confused, I just don’t want to get too bogged down in the weeds. Trying to just touch on the methodological high points.
I went and pulled the values relating to my chosen inputs for a ton of active, recently retired, and all-time great players who don’t happen to be in the Hall of Fame. I then sorted the players into positional groups. To make sure I had large enough samples I combined cornerbacks and safeties into one large defensive back grouping and defensive ends and defensive tackles into one large defensive line group. By grouping players into their position group it ensured I wasn’t comparing players who had no real chance of winning an MVP against quarterbacks like Mike Boss who won multiple of them.
For players who have played multiple positions I tried to sort them into the group I considered to be the position that was the most “impactful” on their Hall of Fame chances. For example, if someone played defensive line and cornerback and was named to five probowls while a cornerback but made none on the defensive line I sorted them into the defensive back group. This isn’t a perfect system and I’m sure I misplaced a few players but it was the easiest way to handle something that just doesn’t happen that often in traditional football.
I then ran the regression for each of the groups and took the coefficients to construct the first equation. For transparency the r-squared values for all of the regressions I ran are as follows:
QB: .6482
RB: .5794
WR: .6379
TE: .7779
DL: .8366
LB: .5807
DB: .6326
K/P: .6983
These aren’t the greatest r-squared values but this is just an initial run of this system and I’m hoping I can increase the model’s fit with more data over time.
Of note at this point for all my fair rubs fans who have read this far I excluded offensive line because I didn’t consider any of the current Hall of Famers to be “traditional” offensive linemen - i.e. that it was their most “impactful” position on their Hall of Fame candidacy. In the future we’re definitely going to see some Hall of Fame offensive linemen, but as of right now I had nothing to compare current players to so the model wouldn’t function properly.
At that point I plugged the values each player had into the equation the coefficients gave me to get an initial value that Iabeled RV (regression value). To get a percentage from this value I used the equation:
Base % = 1/(1+EXP(-(RV)))
At that point I was left with a basic percentage, however it wasn’t as immediately “accurate” as I would’ve hoped. For example, on the first run of the quarterback regression and equation Wolfie McDummy was left with a 58% chance to make the Hall of Fame. This is very analogous to how basketball calculates their HOF chance %’s but it just felt wrong to me. Franklin Armstrong - one of the greatest players in ISFL history - was sitting at a 60% chance. I realized if I wanted these numbers to be more representative I’d have to “normalize” them somehow.
The normalization process required me to build a system to compare one player’s “Base %” to another’s - and namely to someone’s who is in the Hall of Fame. I’ll give a brief overview of the steps I went through to get to my final % values that are displayed below, but do note just as my weighting choices above weren’t fully fleshed out in all the intimate details this won’t be either. I don’t want to get too deep into the weeds unless someone is extremely curious.
I first went in and found what I labeled a “null” player for each position group. Essentially a “null” player is a player who had almost no values above 0 for the x-inputs in the regression. A player who has made no probowls, has no awards, has no top-10 statistics, etc. I then subtracted every player’s “Base %” from that “null” player’s to get what I labeled “Base Inc.,” or the difference in a player’s “Base %” from that of a literal replacement level player.
I then divided that new value by the previous null player’s “Base %” to get a value for how much bigger the initial player’s “Base %” was comparatively. This left me with a more spaced out % - essentially how much better a player is than a replacement level player at their position. I then took that position and compared it to the top value, minimum value, and average value of all players in that position group in the Hall of Fame and then averaged a bunch of the different percentages to get to what I labeled “Initial HOF %.” I then jumped through a bunch of hoops comparing that new metric to the Hall of Fame players’ metrics within that position group. This eventually spat out my “Final HOF %” metric.
A Few Other Notes
I basically automated this process by the end and tried to encode this into an actual adapting model. This required me plugging in a few more priors and defining “era” parameters. Essentially the model can decide what it prefers and what it thinks of a player. It does this at a snapshot in time which is important to note because the new values I began getting occasionally contained a negative value. This DOES NOT mean that a player in question has a lower chance to make the Hall of Fame than a replacement level player, it just means at this time the model is confused on their performance in the league. I’ll do my best to highlight this confusion when we get into the actual data and explain it as best as I can.
Meat and Potatoes
Quarterbacks
Quarterbacks up first with probably a slight surprise on the list - the model considers former Austin star Easton Cole to be a virtual “lock” to make the Hall. Although I’m sure this has Joe excited it probably is slightly shocking to some of the league. The model really values QB probowls and top-10 career stats. It puts a lot of weight on being historically good for the length of a quarterback’s career. Cole chucked the ball a lot in his time in the league and leads the league in interceptions all time which knocks him a tad but is more than made up for by his incredible amount of yards and touchdowns all time. He ended up with almost double the weighting from his all-time statistics than his contemporary Wolfie McDummy - who checked in at 84% - did.
The other one that is probably shocking is Caliban being up so high on the list. That’s due to a lot of noise in the model right now. Caliban might end up being a Hall of Fame quarterback but he’s getting a huge boost from his standing on the all-time passing rating list. If he slides down that list in the coming years his overall percentage should normalize some.
I was quite curious to see Monty Jack ahead of Colby Jack but the model weighs the back-to-back Ultimus rings that Monty won more heavily that Colby’s MVP and Ultimus all in one season.
Runningbacks
Is anyone shocked to see that Hanyadi and Gump are consider “locks” under a purely statistical model? Hanyadi is one of the only runningbacks in recent memory to win an MVP which gives him a huge boost in his chances. Gump has sustained success that resulted in a high number of probowls and strong placement on the top-10 career boards to boost himself.
I was a bit shocked to see Torenson as low as he is, however the model knocks him for never winning an Ultimus, not having any truly outstanding single-season statistics, and having a below average amount of awards credited to him.
Its worth talking about the four players here who show up negative. I did say I’d try to explain this phenomenon when we encountered it. I am absolutely not saying players such as Zoe Watts or Richard Gilbert are bad. Instead the model is interpreting the fact they each have probowls but no other high values in their inputs in a weird manner. It can’t wrap its head around Watts having some great seasons but not having an Ultimus, an award, or placing on any top-10 lists. This should normalize as the players continue their careers, it is a momentary blip in the data because they’re still relatively young players.
Wide Receivers
Saba #1? That seems believable to me. He has a strong amount of probowls and a lot of credit for his standing on top-10 career lists. It was more shocking to see Nate Swift clock in much lower at 34%. I think Swift probably ends up in the Hall because of his ridiculously long career, however the model can’t weigh some of the outside factors I think will help his case. He was three Ultimus rings but almost no recognition on any top-10 statistics list which knock him down a lot.
Dexter Banks makes the list but gets a lot of credit for his time as a quarterback. He was someone I went back and forth in trying to decide where to place him position wise. I settled on wide receiver because he played a lot longer there, however his highs at quarterback were much higher.
Tight Ends
Tight End was a really weird position. Most of the players who are in the Hall of Fame as tight ends had literally ridiculous careers. Paul DiMirio has credit for 12 probowls in the model. Versus L’Alto has 10. The most a currently active or recently retired tight end has is 5. That discrepancy pushes a lot of the modern tight ends down a lot. In the long run this should normalize. A lot of these tight ends are dominating their competition currently and will inevitably make the Hall of Fame, however there is so much more competition at the position its hard for them to accumulate the bulk amount of probowls and awards needed to contend with those in the Hall already currently. This regression should balance out more as tight ends from this era make the Hall.
Until that point, however, the model really likes Austin McCormick. The highest amount of probowls for an active tight end, an Ultimus ring, and a handful of tight end of the year awards pushes him into contention for one of the best tight ends of this era.
We should talk about Heath Evans while we’re here too. Similar to the issues with the runningbacks Evans is going to be fine in the long run when thinking about his candidacy for the Hall of Fame, but in the short term view of the model - again its an immediate snapshot in time of right now - it really wants to see him get a few more probowls and an another award or two. Its all because its trying to compare him to tight ends from a less competitive era at the position.
Defensive Line
If you didn’t read my little blurb on why Heath Evans is down so far on the list for tight ends in the previous section you should go back and do that because the same logic applies to the defensive line. I think Bubba Thumper is a surefire Hall of Famer, however the model is trying to compare him to players from a much different era - back when 25+ sacks a season was normal. Once more modern defensive linemen make the Hall, Bumper’s percentage value will begin to normalize. Until that point, however, woelkers is going to have to be content with having the highest percentage of anyone on the list right now.
Its worth talking about Nero Alexander too. He gets knocked quite a bit for not having any top-10 career statistics. Again, this is purely because the model is struggling to interpret differences in eras. Alexander had an incredible career and has a serious shot at enshrinement in the Hall. The same goes for Immanuel Blackstone who is suffering from the same issues.
Edit: Left out Alexander's career TFL achievements, a pretty big oversight. When edited he shoots to 98%. Probably some noise in that statistic cause I haven't fully recalibrated the regression with it in mind but he's going to shoot up close to that amount. Goes to show how much the model wants people to have career marks and pushes people down for not having those or top season marks. Just a huge era based difference.
The young players in the negatives are being hit for the same issues - the model can’t rationalize why they have probowls but aren’t climbing the all-time stats list. In the long run this should normalize and I intend to try to fix it with era weighting in my next iteration of this project, but for now I get to leave Big Edd at -8% and I get some satisfaction from triggering him.
Linebacker
Does the top three shock anyone? I don’t think so. I’m kinda surprised Bode wasn’t closer to 100% but he suffers from having to be compared to such a broad range of linebackers - like Mo Berry almost broke the model. With that in mind I think 74% is solid and reflects his chances quite well.
Dex Kennedy is suffering from having strong placements in top-10 season statistics but not having a probowl or award. The model just can’t understand how he had a top-10 all time season in some categories and didn’t win an award. Honestly, I’m confused about that too. Suzuki has a similar issue - great placing on top-10 season lists but just lacking in other categories that normally would come along with that prior success.
Hockhertz is an error since he’s also on the defensive back list - I kinda messed that one up but its a bit late since we’re like 3,000 words into this so uh. Winchester and Joestar are too young and just need a bit more time, the model’s struggling to rationalize early success with a lack of placement on lists that take time to achieve.
Defensive Backs
My favorite part of this article - the part where I get to trigger Dermot. Lavelle Jr. still has a solid shot at the Hall but the model really hates how he cracks so many top-10 career lists but doesn’t rank up high in awards. Its all returner based stuff and is entirely dependent on longevity. I can fix this in a later iteration of the model by weighting some top-10 categories differently than others and trying to delink them from awards somewhat. Until then Dermot gets to suffer.
Snuggles, Rector, and Green up high? No one should be shocked. Philip Stein is checking in seemingly pretty high at 30% but the model loves his Ultimus rings. Thats the same reason Chester Sweets is so high up even though he’s basically been inactive since he was drafted.
Bowie is the player across all positions most in the negatives. This is very similar to others mentioned before him who are in the negatives, he just hasn’t been around long enough basically. The model sees him having a high placement on a top-10 season category and doesn’t understand how he can literally have one of the best seasons ever in a category and not have won an award. Similar to Dermot I need to refocus the weighting on some of the categories and delink some from awards to solve this. Easily doable. Sorry Tonzy, I’ll get it handled next time.
Dawkins suffers from a similar issue - great top-10 season placement with a lack of an award that the model just hates.
Kickers and Punters
The top two make a lot of sense. The bottom few are probably pretty confusing here. McDairmid gets knocked by the model for great top-10 season and career statistics but a lack of probowls. That probably normalizes as he keeps playing. Kokot suffers from the same issue. Prohaska is an outlier I can’t wrap my head around.
All in all, Kickers suffer from a similar issue to that of tight ends. There’s a difference in eras and as the league has expanded the position has become so much more competitive. Gone are the days when someone could end up with 10+ probowls at the position. That really skews these numbers but should normalize as we get more modern kickers in the Hall.
[b]Whats Next?[/b]
So I’m in the process of coding this into a database similar to the TPE tracker with explanations of where all the numbers come from and the ability of people to deep dive into their statistics and potentially project things out in the future and see how they stack up if they add a few more different things to their resumes.
In the short term I’m going to keep releasing iterations of this probably every season. Once we get new Hall of Fame data things will chance as different combinations of inputs become more normalized and more eras become represented. I also want to fix a few of the issues I’ve identified with awards linking up with top-10 placements and not all top-10 placements being given the appropriate emphasis.
I also really need to work out era weighting. I tried to do it some but it just gets skewed so much. That should get fixed with more data after each season. The model should normalize a bit more to the modern era. Its just a snapshot in time with comparison to who is already in the Hall and I don't want to push it too much in the modern direction on my own. Hard to determine how far is too far.
[b]Whelp/Grading Note[/b]
Shoutout to Duilio for loaning me some of his data on the current guys in the Hall of Fame. I still had to go out and collect a lot more on my own but it helped jump start me a lot. Toss him a bonus if y’all can, he’s been really helpful
This whole thing probably has 50 hours in it or so thus far and is only going to keep increasing. Not all the math is perfect and I have a lot to figure out with it but its getting there. Any questions feel free to ask - will explain almost anything but this thing ended up like 4,000 words and I’m not trying to get too deep into the weeds.