Pablo

NFHS is talking about eliminating the rule on locked hands next year. In the discussions I've had, the general idea is that this is going to dealt with by calling it stalling instead.

I can't speak to that, but just that I think there is enough cross-over now to get decent relative rankings of D1, D2, and D3. But without an easier way to get the data, it ain't going to happen.

I can't access the dataset right now (calculations are running) but I have thousands of matches of non-D1 wrestlers against D1 opponents. Now, the question is how many other D1 opponents do those D1 opponents have? That's a bigger question. But I think it's better than you think. For example, I mentioned Henschel who wrestled Rhett Koenig from Minnesota at Jim Koch. Koening wrestled Ben Lunn from SIUE at the Cougar Clash, and Lunn wrestled Midlands, and now we are the mix of D1s. Yeah, only one path of 3 steps is fairly week, I agree, but it's the first one I found. There could be dozens more. Or take Ashton Jackson, Purdue 125 (behind Ramos). Lost to Caleb Smith, Nebraska, but beat Ben Aranda, Clev St, so has a solid D1 record. But has matches against wrestlers from Wheeling, Grand Valley St and Cornerstone, so now is connected to non-D1. I got these connections without working very hard. Sure, you won't find the big guys with many non-D1 matches, but non-top wrestlers, including back-up wrestlers on good teams, have wrestled a LOT, and in a lot of open events. I don't have any way to affirm a connection between everyone, but I think it's fairly extensive, more than you might think.

If I had convenient access to the data, it would be easy to expand Pablo rankings to allow for including everyone in D2 and D3 (and probably even NAIA). There is enough crossover matches among the divisions to get good comparisons. It would be time consuming - probably more than 50000 matches and 10000 wrestlers, but time is something that can be waited out. I asked Henschel earlier this season if he ever thought about how he'd compare to the guys in the D1 nationals. He claimed he didn't worry about it and this was a fan thing. I don't buy it in the least. For example, he did some trials for some national team last year, so he's put himself into the mix of other top wrestlers. He has to think about where he can fit in. I had one of his matches from this season in my rankings database (he wrestled a JV guy from Minnesota and won), but just the one, so it got culled out.

But nominally he was one of those "elite" wrestlers, no? I mean, ranked #1 all season, clearly he wasn't among the rest the pack.

OK, this is probably be my last update of results on this thread. New Pablo content will be posted in a new thread when I get a chance. But here's what I have learned. Using the bigger dataset as described above, I now have updated rankings for matches before the conference tournaments weekend. These were calculated using an updated valuation model using the rematch data outcomes to assess the predictive ability of events. I won't post the full rankings, just some assessment data. As I mentioned earlier in the thread, my starting benchmark are conference tournament seeds. What we can do is to compare how well do seeds do at predicting outcomes vs Pablo rankings. I didn't give you enough time to help me out with the data, but I have gone through all the conference tournaments and calculated how often the matches went according to seeds. I didn't do anything real rigorous, but I just went through the brackets on Track, and if they didn't have the seeds there, I looked them up and had them for reference. I can't guarantee the results to be perfect, but they are probably pretty close. I have no historical context to assess this, so I will just provide the results. For the conference tournaments last weekend, the higher seeded wrestler (better seeded?) won 75.4% of the time (in almost 1300 matches). I didn't include matches that were injury defaults or medical forfeits, only if they wrestled out. I also didn't include matchups of non-seeded wrestlers. The one place where I know there could be some issues is with the handful of wrestlers who changed weight classes, wrestling in weight classes they've not been in all season. I'm pretty sure that they wouldn't get seeded in the tournament, or would be seeded low, but that doesn't reflect how good they are. But I also think Pablo is going to call their matches as losses, so I don't think it affects the comparison. OK, so the seeds, which are determined by the magic of the seeding people with their brilliant understanding of wrestling and their having watched matches and stuff got 75.4% correct. I am not able to look at just the seeded matches in Pablo, but I did look at all the matches in the conference tournaments (the difference is that Pablo considered matches between unseeded wrestlers where they exist). The number of correct matches was 77.0%. Therefore, this version of Pablo rankings did better than the seeds. Now, it's not by much. If you look at the matches by the seeded wrestlers, if you used Pablo rankings to compare them, you would get about 20 more matches predicted correctly than if you used seeds (out of almost 1300). But that's a difference. This is very promising. If nothing else, it shows that Pablo, who knows nothing about wrestling and only knows who wrestled, who won, and by what score, is at least just as smart at seeding wrestlers as the current approaches. It will be interesting to see what happens with nationals. That's the next test. There is another way to think about this. Of those matches in the conference tournaments, about 600 of them were rematches from earlier in the season, so I can use them in my match-pair dataset. I've done that, and while it has made some small tweaks to the model, the conclusions are the same as what I posted above. However, one thing I also did was to look at the new additions, the rematches that took place last weekend. What I found is that, if the two wrestlers had met up previously in the season and then met up again in the tournament, the wrestler who won in the regular season won .... 76.6% of the time in the tournament. In one respect, that's satisfying because it says Pablo is basically reflecting the case of, well, who won before? Let's expect them to win again. However, I'm not as happy, because I'd like to hope that by using scores/outcomes, I can learn more about the wrestlers than just who won. There are some things I need to do to improve the fitting algorithm and maybe that will improve things a bit. I'll keep working on it in the background, but the good news, we are at a good starting place. The TL;DR Summary: Pablo would have seeded the conference tournaments better than they were.

What the heck is the explanation for that bump at 27 - 30? Is that a matchup issue in that if they win the first round they have an easier path to AA? Or are there wrestlers who get thrown into those seeds for non-seeding purposes? You clearly have enough data here, that's real.

Lots of data here, but I will just note that like this methodology. I've spent a lot of time analyzing outcomes-by-seeds in other places (specifically basketball) and understand your basis. In the future it could be possible to run this with Pablo rankings, and I can easily do that for the winner's bracket but when it gets to the consolation side, it gets really complicated. I know guys that have just done it via a monte carlo simulation to get a statistical result. If it is just the championship side, I can do it analytically.

Not yet. Still working on last week, looking at the outcomes for this week. I do have all the conference tournament scores from last weekend, though. Hey, if anyone wants to help, can you dig up all the outcomes for matches with seeded wrestlers? One thing I want to do is to compare Pablo predictions with the seed predictions to see if I'm in the ballpark. If you have the ability and time and can do a tournament, drop a result here. What I'd be looking for is Conference Weight class(es): # of matches the higher seed won - # of matches the lower seed won Look at all matches, including consolations (that's where it gets trickier). If one wrestler has a seed and the other doesn't, the seeded wrestler is expected to win. If neither wrestler is seeded, ignore it. This will help me a lot in terms of knowing if this is worthwhile. I will have final rankings available by probably the middle of next week.

All matches are taken into account

Self-serving side track: Unfortunately, I don't have enough data to draw any conclusions about this aspect yet. I've got outcomes for rematches as a function of time, but there's just not enough data to know exactly how much a time lag matters.

To address this part, to be fair there are kinks in the system to be worked out, and it's through your feedback that I can improve it. Last week first draft I put out certainly had issues in terms of weighting events in the model and, more importantly, a lack of data. I had 5000 data points last week. This week I have 17000. More data helps.

It's all based on an ELO type of approach, but with the finetuning of adding score differentials (which you can't do in an ELO-Chess model). So it is based completely on head-to-head outcomes. The problem with doing something like what you are asking is the system doesn't really know who is "good" and who is "bad" independent of all the other results. If all it knows is that Gable pins Pablo and Gable pins Toothache Gable, then it won't see them as any different. But if it knows that Gable with a toothache has also beaten Kerkvliet and Luffman and Andrews and Swenski, and that Kerkvliet and Luffman and Andrews and Swenski have all beaten Pablo, then yeah, that win over Gable with a toothache means a lot more, and, in fact, that win over Pablo probably doesn't mean much. I haven't shown the information here, but basically what it says is that if wrestler A beats wrestler B by X amount, then that translates into a "rating difference" of Z (where Z is a value that can be transformed into a win probability) between them. We'll call that the "match outcome." What Pablo does is give each wrestler a rating, and then compares the difference of the two ratings for the wrestlers to the "match outcome." So Gable Stevenson has a rating value of 6900, Gable with a toothache is rated 6800, and Pablo is rated 0. Healthy Gable wins by pin, such that the match outcome is 2400, let's say (in order to get there, I'd have to run around and avoid him for more than 2 minutes, in fact). So Pablo would see the differences between the wrestlers (100 and 6800) and compare them to the match outcome. The difference between the match outcome (2400) and the difference between the wrestlers (100) is huge for for Gable vs Gable; as a result, Pablo will try to make the rating difference bigger to try to minimize that deviation. The difference between Gable and Pablo looks huge, but when you program it right, you put in protections that say, "we expect a dominating performance, we got a dominating performance, it's all good" So the match between Gable and Gable means a lot more than does Gable over Pablo. Now do this over 1700 matches for 400 wrestlers at each weight class and find the ratings that minimize the overall error. Even if Gable pins Gable, you can't just set them 2400 points apart without screwing up all the other comparisons that you have. This is why I could toss all the matches where a wrestler has 1 match in the system. The rating for that wrestler will just be X away from the rating of the guy he wrestled and is not anchored to anything else. Whew, that is long! I hope that answers your question. It's very insightftul.

It's like a KenPom approach, although based not so much on "advanced analytics" but certainly analytics. The relevant data that I pull out of the track result is Date, winner, winner score, loser, loser score, fall time (if it's a pin), and then just yes/no flags of TF, SV or other TB. I may have the wherewithall but certainly not the motivation to break it down by scoring events (takedowns, near falls, escape, etc), which would be closer to KenPom. This is more like Sagarin. Predictive? For sure. See my post last week talking about the concept of "prediction" and how it's inherent in the concept of "better" (which includes the wrestler Gable Stevenson with a toothache). The "better" wrestler is more likely to win. This is just this year's data. I spent probably 50 hours downloading and massaging to get it here. I could probably improve that by 20% knowing better what to do , but that's just too much effort to go back to previous years.

I've got a spreadsheet that can parse the matches (mostly, it runs into some errors here and there, but mostly fine), but I need a script to GET all the matches more easily.

You don't need admin to do that. You can get an entire round for all weights, or you can get all rounds for a single weight, but it won't let you do all rounds, all weights. Life for me would be a lot easier if it did. And then they could do all rounds, all weights for all events on that day....

For new guys, apparently.

And the one you are interested in, the pin. With this data, I have a better idea of how to interpret that. Let's start at the beginning. Now that you've seen the value of all these other events in terms of predictability, how to interpret a pin? What do you think is the probability of a wrestler who wins by pin winning again? Is it comparable to a TF? Major? Dec? In fact, the answer is 81.9%. In other words, it's basically the overall. What that means is that if you know that wrestler A beat wrestler B, it does not help you to know if it is a pin or not. A pin is the equivalent to the average match outcome. Did you see that coming? Personally, I'm not all that surprised. I've always felt that pins at this level can be fluky. You make a mistake and you are in a cradle and gone. BUT there is more to it. The data are noisy (this is getting tough), but there is a real correlation between how fast the pin is. Pins in the first oh, say, 2 minutes are more like tech falls (or better, even, >90%), while pins in the last minute of regulation come in at like 70%? I don't have enough pins in overtime to make a conclusion, but for the 5 matches, it's 3 wins in the rematch, 2 losses. This is all great stuff, and I can use it. It's telling me that the value I was using for a pin previously was much too high (and that was apparent in the results), and so I've set it up to reflect that now. I think this is getting closer to what it needs to. I reran the 125s from last week with the full data and the new model, and the result is below. Remember, this is only taking into account data from before last weekend, so are not the current rankings. My next step is to run the full set of rankings for all weights and then see how it did in the conference tournaments. I'll be back when I have that information. 1 . 125:Purdue_Matt Ramos 2 . 125:Oklahoma St_Troy Spratley 3 . 125:Arizona St_Richard Figueroa 4 . 125:Northern Colorado_Stevo Poulin 5 . 125:NC St_Vince Robinson 6 . 125:Wisconsin_Nicolar Rivera 7 . 125:West Virginia_Jett Strickenberger 8 . 125:Virginia Tech_Eddie Ventresca 9 . 125:Penn St_Luke Lilledahl 10 . 125:Ohio St_Brendan McCrone 11 . 125:Nebraska_Caleb Smith 12 . 125:Iowa St_Kysen Terukina

Even though the simple "did the same person win twice" answer is 82.8, the dataset is big enough to look more closely. For example, what about the different outcomes For example, let's start with a TF. If a wrestler wins by Tech Fall, that wrestler wins the other match 88% of the time. That makes sense, because tech falls are more likely to happen when one wrestler is much better than the other. What it says is that a wrestler who wins by a tech basically has a 94% chance of winning against that opponent. It's interesting, though that if a wrestler beats a guy by TF, if they come back and wrestle again, there is more than a 10% chance that the other guy wins. As upset rates go, that's pretty darn high considering we consider a TF to be pretty dominating Let's go to a Major decision. One thing I am seeing in the data is that major decisions are about as good as TFs when it comes to prediction, again, 88%. Now, I do see some correlation with point difference, so this might be a case where the real line between a dec/major/TF doesn't correlate with the point cutoffs in the rule book, but I will look at that later. In other words, it may be that up until, say, 10 points, the point difference reflects the quality difference, but above 10, it's all the same. I think that is the basic idea, although I don't know where that line is yet. Before going to a regular decision, let's look at overtime. In fact, if the match goes into overtime, there is only a 54% chance that the same guy wins twice. At this point, it's not much more than a coin flip. I don't think that is unreasonable. For a regular decision, it's 77%. So if a wrestler wins by a regular decision, there is almost a 1/4 chance he will lose if they wrestle again.

The first thing I did with the dataset is to look at matches I can use for predictive value. Fortunately, I have been able to find almost 1000 cases where the same wrestlers have met up twice during the season. This is awesome, because I can use that to get some insight into how to use match outcomes (btw, although 1000 sounds like a lot, that's not nearly enough to get a really good analysis; for my volleyball rankings, my match pairs dataset consists of nearly 100K example! But to be fair, in volleyball, you have conference teams playing home/home every year (and I have almost 20 years worth of data) and I actually have all the matches for all college levels (D1, D2, D3 and NAIA); unless something changes massively in how scores are available, that ain't happening in wrestling) But I have been able to learn from what I have. For example: If two wrestlers meet up twice in the season, the probability that the same wrestler wins both times is....82.8% That's a little higher than you usually see in other places. In NFL, it's like 63%. In college volleyball, it's about 78%. If you break that down, what it tells us is that the "better" wrestler wins about 91% of the time on average. That's pretty good, because when you say, "on average" it includes those matches of Gable Stevenson vs me and those matches of Gable Stevenson vs Gable Stevenson with a toothache. There are a couple of ways to interpret this result. It could be that the average match just isn't that close, and the one guy is much better than the other. The other interpretation, which I think is probably closer to reality, is that the standard deviation of quality is really low. You can see how this comes into play if you compare something like volleyball and tennis. I mentioned that volleyball has like a 78% repeat win rate, but in women's tennis, which I've looked at in the past, it's less than 70%. The great thing about these two sports is that they both have enough points and so the results are pretty easy to model. Volleyball works, and the real outcomes very closely reflect the results of a simulated model. In fact, that was my big breakthrough with Pablo rankings, figuring out the model that relates scoring to winning. Thinking that tennis should be similar, I gave that a try, and it really didn't work. Everything was there, but the performances just didn't match up. I concluded that the problem with tennis is that the standard deviation of quality is a lot higher. For starters, you have things like different surfaces, that affects how good players are, and the other problem is that there was only one competitor. In volleyball, with 6 players, if one player has a bad night, it can be offset by another player having a good night. If a tennis player has a bad day, they are on their own. I wondered if that would be an issue in wrestling, but apparently it is not. Even though it is a single competitor, it does not appear that the quality of their wrestling ability varies as much as you might see in other sports. That's thing 1 I have learned. Hopefully, this will bode well toward predictability (although I will add, I did take a look at the Pac12 tournament results; 25 rematches in the tournament, the wrestler who won during the regular season only won 17 (68%? Ouch!) Past performance does not guarantee future success....(FWIW, my Pablo rankings last week correctly predicted 18 of those matches :)) More to come

Hey gang - I exceeded my quota post last week and so got locked out of the discussion on Wednesday, so sorry for cutting out. What I did in the meantime is jumped into completing the dataset and upgrading analysis. I believe I have now gotten ALL of the matches for all the D1 wrestlers. I went through every team's schedule on Track and downloaded all the results from every event that had matchups of D1 wrestlers. I also have a ton of matches of non-D1 wrestlers, but they are not quite as complete. If I found an event that had maybe 1 D1 squad there (maybe JV) and the all the opponents were non-D1, I tended to pass by it. But if there were matchups between D1 wrestlers mixed in, I did always grab all of them, even if they weren't D1 opponents. As a result, I do have a lot of guys who are from non-D1 school or even unattached to a school who have matches in the database. One thing I did to make life more manageable was to cull all the matches where it was a non-D1 wrestler who only showed up once. These matches do not contribute to the rankings, and just lead to a lot more baggage. That was something like 1200 matches that I identified. I spent the past week downloading all the scores, formatting them into a style I can use, and then cleaning them up. You gotta get rid of all the byes and the injury defaults and the medical forfeits. Also, while it is convenient to use all the scores from trac, depending on the details they don't always format like they should. Moreover, sometimes the input data is weird, because you get guys who are sometimes listed as wrestling for, for example, Navy, but other times they are listed for the Navy Wrestling Club (depending on the event). So I had to go through all the wrestlers and find the ones who were listed on multiple "teams" and consolidate them onto a single card (my goal here is to evaluate wrestlers, so I don't worry about whether events count toward the NCAA or not; as long as they are on the mat trying, I include their results). The other issue is that they get entered in, depending on who is entering, with different names, so is it Josh or Joshua? Well, they are the same guy, so I tried to find all the places where they had different names for the same wrestler. I finally got it done yesterday afternoon. The final result is a database of nearly 17000 wrestling matches with more than 3800 wrestlers. This is massive. But I've get things set up to calculate. It will take a while, and there are some things to work on (I still don't have conference tournaments added yet, but I'll get there). But in the meantime, there are things to learn...

Bartlett also lost to Mendez in the All-star classic. You can't just ignore results you don't like.

How did you determine that? I have data that informs the value of a pin. You are making things up.

So I added in the match against Moore from November (he had another match against some non_D1 guy but I didn't bother). I'm always a fan of more data 1 . 125:Purdue_Matt Ramos 2 . 125:Ohio St_Brendan McCrone 3 . 125:Arizona St_Richard Figueroa 4 . 125:Oklahoma St_Troy Spratley 5 . 125:West Virginia_Jett Strickenberger 6 . 125:Penn St_Luke Lilledahl 7 . 125:Wisconsin_Nicolar Rivera 8 . 125:Virginia Tech_Eddie Ventresca 9 . 125:Penn_Max Gallagher 10 . 125:Lehigh_Sheldon Seymour 11 . 125:NC St_Vince Robinson 12 . 125:Princeton_Marc-Anthony McGowan 13 . 125:Nebraska_Caleb Smith 14 . 125:Northern Illinois_Blake West 15 . 125:North Carolina_Spencer Moore 16 . 125:Rutgers_Dean Peterson 17 . 125:Oregon St_Maximo Renteria 18 . 125:Missouri_Gage Walker 19 . 125:Iowa St_Kysen Terukina 20 . 125:Indiana_Jacob Moran 21 . 125:Harvard_Diego Sotelo 22 . 125:Minnesota_Cooper Flynn 23 . 125:West Virginia_Jace Schafer 24 . 125:Oklahoma_Antonio Lorenzo 25 . 125:South Dakota St_Tanner Jordan 26 . 125:Cal Poly_Koda Holeman 27 . 125:Iowa_Joey Cruz 28 . 125:Northern Iowa_Trever Anderson 29 . 125:Cornell_Marcello Milani 30 . 125:Northwestern_Dedrick Navarro 31 . 125:Utah Valley_Bridger Ricks 32 . 125:Purdue_Isaiah Quintero 33 . 125:Army_Charlie Farmer 34 . 125:Binghamton_Carson Wagner 35 . 125:Princeton_Ethan Rivera 36 . 125:Campbell_Cooper Shore 37 . 125:Iowa St_Adrian Meza 38 . 125:Northern Iowa_Kyle Gollhofer 39 . 125:Virginia_Keyveon Roller 40 . 125:Pittsburgh_Nick Babin

Over the years I've been known as Pablo, Agent Buchwald and lately I've settled on The Bofa on the Sofa in much of my life. If you were The Bofa in 3rd grade, that would be amazing.