Rock (Slugger), Paper (Boxer), Scissors (Swarmer), and Mathematics.

I tried to run SPSS with quantitative data recently, made me realised how I missed getting punched in the head.

 

As others have said, it depends on what goes in to the model, and while it might be easy to pigeon-hole some boxers as being predominantly one style - e.g. Hopkins - Counterpuncher, Calzaghe - Swarmer, Mayweather - Boxer, I find that a bit reductionist - I've thought more around the Top Trumps kind of score out of 100, and whilst Mayweather might get let's say 95 for the Boxer style, his counterpunching ability is ALSO very high. Similarly, Hopkins was a technically very good boxer too. Then there is reach, height weight, never mind intangibles...

And then look at Usyk in his recent win against AJ - arguably he had a combination of all 3 at a high (e.g. 85-90%) level.

Which then factors in another more meta variable - who determines what the stats actually are? my scores / weightings for these things might differ from someones elses - which then gets back to lies, damned lies & statistics.

That said (i'm a software architect as day job), I find this type of study super-fascinating and it would be really interesting to look at how accurate predictive models could be - AI is a fast moving field, and with the right training sets. and data I believe predictions could be very accurate in the main - although outliers is a big area in the field - and that's where both known and unknown unknowns can come into it and affect things enormously.

 

Swarmers and boxer-punchers are nightmares of pure boxers.

Pure boxers and boxer punchers are nightmares of sluggers.

Sluggers and boxer punchers are nightmares of swarmers.

Boxer-punchers are the most complete fighters. Especially if you are a southpaw.

 

Your son should put the gloves on and put it to the test. It takes a very fit, brave and talented slugger to beat a Boxer.

 

This, to me, is the hardest part. You would probably want to crowdsource the labeling.

Hey if he wants we could set up a place where people can go and click different style labels they think apply.

Regarding what people have said about covariates, I don't think they are as important as they're being made out to be.

The style triangle is supposed to be pretty powerful. This either leaps off the page or it doesn't.

That's funny! I have to work with SAS all day. I briefly got to try out SPSS and it looked pretty cool to me actually.

I come from a non-stats programming background so all of these tools are so crazy.

Has he considered matchmaking scores like ELO?

If he built scores (not perfect, but just something to quantify it a bit), then he could quantify the effect of the style matchups and also he could see if they played any role in the largest upsets.

The database itself may be the coolest aspect of the project. I've always wanted to create one but hesitated to do any public projects because Boxrec seems like the best source but they have restrictive policies.

 

Basically the labeling would be the trickiest part. So I agree with the crowdsourcing. However even there you’d be better off with judges or trainers with a certain level of experience to vote on labels and even then you can get bias. There’s so much bias in boxing from who knows who, nationalism and style preferences. You can even find the same problems with the judging. Do we go with boxrec scorecards or get a larger consensus from boxers, judges and trainers? There have been plenty of robberies over the years.

With regards to the analysis I think style would come up huge, but personally I think skill level (arbitrary) and experience would be the most predictive factors.

It’s interesting because I don’t think the sample would be big enough for this, but there is propensity score matching which would be interesting to make apples to apples comparisons with equal experience matchups and how did those fair.

This is actually one of the more interesting threads/topics that I have used my statIstical thinking for and it’s interesting to think about the issues that would come up with the analysis. Unfortunately no one outside of ESB would pay me to do the studies or read the papers and conclusions.

 

I think ELO (or a similar system) is preferrable to propensity score matching (which my department director has drilled into me as an underappreciated tool) because we won't get nearly enough well matched fights.

ELO allows you to get meaningful info from the tomato can beatdowns in the beginning of a career. It works pretty well in football where teams only play 17 games and don't play every other team.

It would also allow us to quantify the degree to which an upset is an upset. I kind of want to really do this, or at least begin work on a database that would allow us to do some analysis.

I actuallly did do a google scholar search for boxing stats type stuff. A few pugilistically minded graduate students have taken some stabs at it but it wasn't that interesting.

 

In rare cases you may find a fighter who is all 3 and skilled in all 3 styles. Such as Duran. But even still, he was most effective and was primarily known as a swarmer.

 

I think this is both true and exactly the sort of thing that I hope @djanders' son doesn't get bogged down in.

I'd love to see a first pass, naive analysis of this and start to get a feel of the effect size.

I gotta get a boxing dataset built, that's my New Year's resolution.

 

I was reading a little bit about the method because it's not one that I use in the work that I do, however it makes a lot of sense here for boxing. Well let me know if you end up doing it, I'll find the work interesting!

 

See an ounce of disrespect for @djanders here I want to remind you, mans watched and been part of boxing since Jack Dempsey was champion so give him his dues instead of being snarky little things.