+- International Simulation Football League (https://forums.sim-football.com)
+-- Forum: Community (https://forums.sim-football.com/forumdisplay.php?fid=5)
+--- Forum: Media (https://forums.sim-football.com/forumdisplay.php?fid=37)
+---- Forum: Graded Articles (https://forums.sim-football.com/forumdisplay.php?fid=38)
+---- Thread: *DSFL Prospects to Graduates Similarity (Part 1) (/showthread.php?tid=26216)
*DSFL Prospects to Graduates Similarity (Part 1) - Hordle - 09-28-2020
Intro:
Good morning ISFL and DSFL players and media fans alike. Myself, Hordle, and Art were having a discussion about combining out skills and putting out a large media piece showcasing some of the previous draft talent to this year. Including build similarities and some potential things to look forward to before the draft season starts. This will hopefully be the first of several pieces covering each class. We decided to start with the QB as outside of Sim life they are one of the biggest and most judged classes of players in the draft. In the world of the sim each team only runs one QB. Which of course would lead to smaller classes due to the need for a new QB only being required due to retirement, IA or regression.
Methodology:
To start, a big shout out to the TPE tracker and the folks who made that, since I pulled all metrics from the TPE tracker.
The objective with the similarity scores (besides something for Hordle and I to write about) was to compare the builds of S26 and S25 players and match each new DSFL prospect (S26) to one of the builds in S26 of closest likeness.
A couple assumptions to start: unfortunately we’re using DSFL TPE tracker builds without any applied updates. That means a fair number (myself included) are at the base 50 TPE and that means we’re pretty limited in differentiation. As a result, a fair number of players might have near identical builds and therefore near identical matches. Regardless, we’re taking a shot with it.
The other assumption: we’re only using builds. Sure we could try to build off some more actual ‘activity’ metrics like posts, threads, hours on forum, or natural language processing of ultimate week posts (future idea there). But we’re on a time limit and have thousands of words to write breaking this down. Lets keep it simple and go with just builds, and just base builds / builds on the last update.
Now to how it was done. First, all S26 and S25 players are pulled from the TPE tracker to Google Sheets. They’re separated into two different sheets and all of their stats are normalized. Why normalized? If we leave the values as is we’re left with a pretty uneven spread that doesn’t emphasize outliers in each unique dataset. For example, DSFL players might have a mean intelligence of 60 and a standard deviation of 6. We want to reduce the range from the mean so we get a feel of who’s unique and built differently from their peers - but we want that separately for each draft class.
So each draft class is normalized with the formula (x - μ) / σx - aka individual player stat minus the mean (average) of all player stats for that attribute divided by the standard deviation of player stats for that attribute. Thanks to that, we have most of the results will lie very close to 0 (the mean).
It’s important to note here that this is normalized across all players, not just players in a certain position. And while it's normalized across positions, we are limiting matching to the same position. Why’s this important - it means QBs are extreme outliers on stats like ‘arm’ or ‘throwing’ compared to their class, especially for the small DSFL draft class of QB’s (2 at the moment).
Now that we have the normalized data for each class (it resembles the player’s name, position, and then normalized attributes), we can have fun with math and determine the similarity between the S25 class and the S26 class.
The similarity between S26 DSFL prospects and S25 DSFL ‘graduates’ aka ISFL S25 prospects is built on cosine similarity. In simple terms, cosine similarity can be thought of as a measure of how similar two vectors are - the closer cosine similarity is to 1, the more identical the two sets of real integers, while the closer cosine similarity is to -1, the more diametrically opposed are the two sets.
In more in-depth terms, cosine similarity uses a brilliant means of comparing vectors - the angle in interdimensional space between the two vectors. Let's start simple - a straight line. We have the vector [1, 1] in 2D space. I’ve drawn it below.
![[Image: eEOni0E.png]](https://i.imgur.com/eEOni0E.png)
![[Image: JhTeTZw.png]](https://i.imgur.com/JhTeTZw.png)
Now we add the vector [1, 2] and the vector [1, -1] to the graph. How do we tell the ‘similarity’ between these lines? Easy - we draw an angle in between the lines. Since they all share the first dimension (x value), we can easily compare the angles and see that, clearly, [1, 2] is much more similar to [1, 1] than [1, -1].
Now it's a cosine, so it's not quite the angle - it is the cosine of the angle itself. But for a simple explanation, it gets the point across. In the example above, cosine of 45 degrees is 0.71 and cosine of 180 degrees is, of course, -1.
Now you can add dimensions to this example and the only thing that changes is your ability to perceive it easily in your head. But we can do the same example with 3 dimensions and planar graphing, with the vectors [1, 1, 1] and [1, 3, 5] and [1, 2, 10]. Now we measure the angles between the planes drawn by the vectors.
(You might have the natural next question of how do we manage this for vectors without a shared dimensional ‘anchor’. Thanks to inner product space - that’s how. It's a pretty big digression and you’ll need some math smarts - but if you’re interested here’s a link on more.)
Why does this matter for us? Well using NumPy we create vectors of the player’s attributes. Now Arthur Naught or Franklin Armstrong has a vector of [Str, Agi, Arm, Thr, …] that encompasses their entire build.
Great, we have vectors representing our players flying through interdimensional space. What next? We have to compare them. And that means we grab all the S25 and S26 players of a certain position at once (since we want to limit comparisons to positions) and we grab 1 S26 prospect and all of the S25 players and ask the math … which S25 player vector has best cosine similarity to this S26 prospect vector?
We get an array response containing all cosine similarities for each S25 player vector, saying it is 0.2 cosine similarity to this S25 player, and 0.73 to this other S25 player. In that small example, we make the player match to be the latter.
And there you have it. Some simple code you can follow along with here, but really this project is a nice mathematics exercise with Python to compare the two classes.
Enjoy the series!
![[Image: dvXgVRR.png]](https://i.imgur.com/dvXgVRR.png)
Color code: 1 - LeBron James III, 2 - Ryan Negs, 0 - Jackie Daytona (S26)
Explaining the chart:
The chart above uses what’s called a parallel coordinates visualization to represent where each of the players (i.e. vectors) intersect, avoid, and resemble each other. Each line intersects with a point on a vertical axis representing an attribute. The first vertical axis is ‘strength’. In that axis, you can see that both our DSFL prospects have identical strength (the yellow and red overlap, color codes 2 and 1 respectively). Meanwhile our S25 comparator, Jackie Daytona represented by blue and code 0, has a significantly different strength normalized.
The best way to read the graph is to move across the categories and compare where the lines meet and where the lines differ. Using that methodology, we can draw some insights into why Datona was deemed a good match and cosine similarity to both DSFL prospects - and more specifically why LeBron James III was deemed a better cosine match (they share a closer intelligence and speed normalized stat than Ryan Negs).
Something to keep in mind when reviewing the chart is that the vertical axis are adjusted for clarity to show only the difference between the minimum and maximum values for the players. Essentially that means while it may look like Negs and Daytona have little in common, the space is actually much larger than the roughly 0.3 difference shown here.
-----------------
QBs:
![[Image: 20130908_mbr_sh2_144.0.jpg]](https://cdn.vox-cdn.com/thumbor/bfZz2cNPaAc1swdpd-EIlYYPoH0=/62x0:3932x2580/1200x800/filters:focal(62x0:3932x2580)/cdn.vox-cdn.com/uploads/chorus_image/image/19486995/20130908_mbr_sh2_144.0.jpg)
User: Orhanello
Color # in chart above: 1
Draft Year: S26
Position: Quarterback
Name: LeBron James III
S25 Match: Jackie Daytona
Cosine Similarity: 0.767
Other notable matches: Trent Blackburn (0.741), Matt Peterson (0.671)
Distinguishing attributes: Higher INT, Higher THR
Why it works:
Art - It’s a good start to the similarity scores when we start at QB with only 2 prospects since Kotone Shiomi retired and both of the prospects are ranked similar to the same S25 player - Jackie Daytona. But that said, I can see why the results came out the way that they did - Daytona had a higher traditional quarterback stats than their fellow S25 quarterbacks. Areas like arm, throwing, intelligence - they are either leading the pack or tied with Matt Peterson. And compared to the average DSFL prospect, LeBron James III is also very gifted in these facets of the game. Even compared to fellow DSFL QB prospect Ryan Negs, LeBron has more of the traditional QB gifts - a slightly better throwing arm and a slightly quicker read of the field. What do you think Hordle?
Hordle- As you mentioned Art, it is pretty clear on why the results came to be. LeBron James has been putting in work since he appeared on the DSFL radar. His activity is outstanding which of course leads to his higher stats then his fellow prospect. Allowing him to earn more TPE than his current potential rival. His high INT should lead to less mistakes made by the rookie QB. If Lebron keeps up this activity, I am sure we’ll see him at TPE cap soon enough while leading a team into a deep playoff run at the very least.
Why it doesn’t:
Art - Jackie Daytona’s user, Mooty66, has a clear precedent for creating stellar offensive players - retiring RB Mako Mendonca right before recreating as Daytona. While there’s no doubt Orhanello can achieve the same, LeBron isn’t a recreate and so doesn’t have the same ‘ISFL-chops’ - something scouts find important for QBs as the position is high pressure. Perhaps just a chip for LeBron’s shoulder to push him to greatness? With that in mind, Matt Peterson might be a better comparator, as they are also a non-recreate and first time player-QB. Where LeBron ends up on the scale of first time player-QBs will be an interesting angle to keep in mind during their career.
Hordle - An interesting point with this being LeBron’s first player and picking the QB role. It may work in his favor in the future if he stays active and does well. Having the stock of having a previous player in a role that a lot of newer players are suggested to play a different role may give LeBron a bit of an unseen advantage for future drafts. Seeing that the player is willing to stick through what is considered a rough role to play may make him a valuable piece in any locker room with LeBron or anyone else he may decide to create in the future.
What to watch for/Similarities in builds:
Hordle- However not only has LeBron been active in terms of training and gaining TPE. He has been active on twitter as well. Making sure to earn what is essentially free money to continue the grind. Seeing LeBron’s stats and activity should be a beacon for any teams looking to get a QB that can lead them to a Championship victory. While LeBorn and Jackie have similar stats. They are two different types of players. While Jackie is a Game Manager, Lebron is a Pocket Passer. As the two continue on with their careers, we will see more and more differential between the players. With LeBron having a higher ceiling in terms of his ARM. We may end up seeing LeBron having a better average deep ball as compared to Jackie. Jackie has a higher SPEED ceiling which should allow him to potentially break the game open with his own running ability.
Art - The archetype for each player is a great item to bring up Hordle and I agree that while LeBron and Jackie share similar strengths in comparison to their draft cohorts it’ll be interesting to see how they differ as time goes along and TPE spending becomes limited by the archetype maximums. Keeping it in a tighter scope, I’ll be interested to see how LeBron places his TPE in the DSFL - where you only have so much TPE under the cap to place into key attributes. Does he focus on one aspect of his game or try to spread the love amongst the traditional QB attributes, following in the footsteps of his S25 match Jackie Daytona?
-----------------
![[Image: usatsi_13546104.jpg?w=1000&h=600&crop=1]](https://giantswire.usatoday.com/wp-content/uploads/sites/67/2020/02/usatsi_13546104.jpg?w=1000&h=600&crop=1)
User: Ryan Negs
Color # in chart above: 2
Draft Year: S26
Position: Quarterback
Name: Ryan Negs
S25 Match: Jackie Daytona
Cosine Similarity: 0.731
Other notable matches: Trent Blackburn (0.709), Matt Peterson (0.663)
Distinguishing attributes: Higher ARM, Higher SPD
Why it works:
Art - Starting with the build again, and as alluded to in the write up above for LeBron, it's clear why Negs was matched with Daytona - superior arm, throwing, intelligence compared to the rest of their draft cohort. While LeBron has the edge on some of these ‘traditional’ QB metrics - specifically intelligence - Negs is right up there with his fellow QB prospect in arm and throwing, the three components of the game that Daytona absolutely dominated.
Hordle- So like you said we start by focusing on the build. Daytona is a Game Manager and Ryan is a Gunslinger, which much like LeBron leads to obvious differences in the build. But there are obvious similar points being that they’re both playing the same position. Both players Intelligence will max out at the same rate. Leading to no arguable difference in that field, unless the Sim decides otherwise. But there will be a 5 point difference between Daytona and Ryan in terms of their accuracy. With Daytona maxing out at 95 and Ryan at 90. If the two were to go head to head it might lead to a noticeable difference. However it shouldn’t have too much of a major effect on Ryan and his career. Ryan is currently sitting on 70 Accuracy and is bound to pour more points into the stat come the day after draft. With a 72 in his ARM and having some points floating around it’s safe to assume that some, if not most, of those will be put into the stat so that the man can live up to the name of his archetype and throw absolute bombs down field. Maxing out at 100, it’ll be interesting to see if Ryan gets anywhere close to max with the DSFL TPE Cap at 250.
Why it doesn’t:
Art - Like we discussed with LeBron and Daytona, Negs is a first time create and therefore doesn’t have Daytona’s history within the ISFL to rely on as a prospect. To Hordle’s point in the LeBron piece, however, both Orhanello and Negs are active in LRs that I’ve joined with them and seem to be around for the long-term. I’m not sure what you’d be looking for to disqualify them, but I’ve seen no indication either will be anything but studs. Getting more specifically into Ryan Negs and his build, he’s gone with Gunslinger vs Daytona’s Game Manager - a decision that’s bound to adjust their trajectories at some point in their career, eh Hordle?
Hordle- I don’t believe there is much to disqualify either of them. I mean personal bias aside with Ryan being a Giant’s fan. Both of the prospects seem like great pickups for any team looking to bring them aboard. I know from my own personal talks with Ryan that he is looking for any way to get as much TPE as he can. Another interesting fact to note, from my small amount of research, there doesn’t seem to be very many Gunslingers in the DSFL if any. The top 5 QB’s last season none of them were Slingers. It will be interesting to see if this gives Ryan an advantage come the draft with the total arm of 100. If teams are looking to adopt a deep passing approach to their offense. Ryan might be the guy for you.
What to watch for/Similarities in builds:
Hordle - Much like LeBron, Ryan’s ARM ceiling is higher than Jackies. Allowing him to open the game up with some deep bombs down the field. However we could see this backfiring on Ryan with his lower ceiling in terms of Throwing Accuracy. Ryan has also been active in several locker rooms, I’ve had the pleasure of talking to him on many occasions and he is also going to be earning at a decent rate as long as he can continue his activity on twitter. Though the gap between the two isn’t huge with LeBron at 100 TPE total and Ryan at 89 TPE total, with training camp opening up on the 24th of September. We should end up seeing LeBron with 110 and Ryan with 99. Both could be a lot closer come draft time. I believe that Ryan is currently waiting on a total of 15 TPE with the Ultimus Week PT and the S25 Mock Draft. If Lebron hasn’t put in for either of those it could be very possible that we see Ryan overtake him for total TPE earned.
Art - I’m definitely excited to see how the QB battle between the two prospects plays out - both this season in the DSFL and throughout their careers in the ISFL. The nice thing about a small draft class is you get some great friendly rivalries and you can’t help but compare yourself (and be compared to) the other draftee. And while LeBron has a slight TPE advantage here, your right Hordle that Negs is just on his heels. In terms of the build, I’m interested to know how Negs is going to view decisions on TPE placement and building attributes. Whether he and Daytona will pursue similar build decisions from here on out, or if one of them will make adjustments to stand out - whether that's going heavy in a specific attribute instead of sharing the love among traditional QB attributes, or even diversifying and putting some emphasis on other secondary attributes like speed or agility.
Quote:Word count - 3173
Please split between Hordle and wonderful_art
RE: DSFL Prospects to Graduates Similarity (Part 1) - zaynzk - 09-28-2020
Very cool methodology man, you put in a impressive amount of work in this!!! I hope u make more media like this down the line it was a really cool read
RE: DSFL Prospects to Graduates Similarity (Part 1) - GlimsTC - 09-28-2020
You guys are actually insane. Very impressive work!
RE: DSFL Prospects to Graduates Similarity (Part 1) - Hordle - 09-28-2020
Thanks guys! Full credit to Art as well for being willing to work with me.
RE: DSFL Prospects to Graduates Similarity (Part 1) - wonderful_art - 09-28-2020
(09-28-2020, 11:42 AM)Hordle Wrote: Thanks guys! Full credit to Art as well for being willing to work with me.
If the grader could give me a bonus for having to work with Hordle that'd be great. /s
RE: DSFL Prospects to Graduates Similarity (Part 1) - Jimi64 - 09-28-2020
Wow, real cool idea!
RE: DSFL Prospects to Graduates Similarity (Part 1) - Orhanello - 09-28-2020
Wow this piece is amazing! Really raising the bar with work like this great job!