Data Science Meets Competition Climbing - Ft. AscentStats

[@mikoajkomorowski1290]

Data science and climbing, it doesn't get better than that :D

[@ScratchRick]

I hope one day I'm good enough to have my performance tracked.

[@flyingchic3n]

you can track your own performance!

[@GeekClimber]

I hope to see that becoming a reality too!

[@ScratchRick]

@flyingchic3n I track it (and share) right now on my KZbin channel. Last year I was climbing 5.9s but this year I started flashing 5.11+'s

[@ScratchRick]

@@GeekClimber thank you! I can't wait for us to meet you are the coolest! 😎

[@bloodyhell3113]

Damn i hope you become a pro

[@randomyogi]

Finally! Have been waiting for this!

[@BetaBoiBrandon]

This was everything I was hoping for. Sounds like Alexei has thought of a ton of stuff, I think one thing that'd be interesting to hear is how the lack of rest impacts the grade since qualifiers, semis, and finals usually all happen in a single weekend

[@GeekClimber]

That's a good point! Let's wait for Alexei for the answer!

[@AscentStats]

It is hard to disentangle the lack of rest question, because the order of rounds is always the same. Maybe it would be possible to compare competitions where the length of rest between subsequent rounds has been different. But in general you can assume this effect is having a small confounding effect on the estimates, so that the grades of later rounds might be slightly inflated, compared to what they would be if the athletes were fresh.

[@hetistijmen]

I think this effect would be smaller than other noise in the outside data. In my experience projecting stuff way easier than burden of dreams conditions play a large part: I might spend a humid day working on a bunch of moves but not even attempt a send because I know every single move takes more energy with less friction. Then some sessions might be shorter or less high quality because of other plans that day, having a strong or weak day, or just having had a fun night the night before and not being on top of my game. And then with the field being as small as it is stuff like cracks being uncommon in comps or the airco in the venue not keeping up with a thousand spectators will have a bigger effect. I guess in short: compounding data gonna compound.

[@hetistijmen]

Something else I just asked myself: does a shitty approach through loads of mosquitos influence a grade? Is DEET aid?

[@MichaelPennMath]

great video!

[@GeekClimber]

Glad to know you like the video, especially from a math guy like you!

[@granulategrain]

I really enjoyed the video and ideas! I don't know the approach they did for estimating the climbers grade (either from their expressed grade or as another nuisance parameter in the max likelihood fit), but it would be cool too see how the grades of climbers change as well! As problems I see with this though, I commented on the ELO video, but the ELO inflation seems waaaay large and could probably be normalized better, though it wasn't talked about at all in the video... Also, I wonder how accurate the hardest climbs are actually... The harder grades are going to be biases by larger pools and number of tries. If there was a hard problem in the semis and finals, the max likelihood will make the one in semis harder since the more people not topping it forces the grade higher, while in finals, I don't think there is enough stats to move the grade higher in the fit. That also explains probably why the slabs are seen as harder since there are probably more attempts per climber since it's usually less tiring Mind you, this is all small armchair critiques, the overall works seems great and I can't wait to look into it more!

[@GeekClimber]

Thanks for the feedback and these are all great questions! Let's wait for Alexei to answer you!

[@AscentStats]

Thanks for the thoughts! Yes, the climber's grades are also estimated. However although Geek Climber talks about maximum likelihood as one approach to statistical inference (which we do use in some contexts), in the full model for competition climbers we actually use Bayesian inference to allow for prior distributions and assessment of uncertainty in the estimates. You can see the credible interval of each climber's seasonal Elo estimates on their individual pages on our website. We also have credible intervals for the boulder grades (a "Rating HPD" is provided). You will see that semi-final boulders typically have about a 4 grade window of uncertainty, whereas for finals boulders it is more like 5 grades. That is due to the problem you alluded to with a larger sample size of ascent data available for semi-finals boulders. However we can avoid bias by having a hierarchical model that assumes a different average grade for each type of round, so that we can group information about final's boulder difficulty across many events, thereby overcoming any bias problem. Having said that, the current model definitely has limitations, and it is being actively developed :)

[@granulategrain]

@@AscentStats Ah, I went through your website and read more of the details and things make a lot more sense. I actually read your arxiv paper before I had seen this video and used the results from that to make some toy models with my climbing record; it pretty cool to see you pop up again with even more cool stuff :) But yah, baysian makes a ton more sense in this context, and the method for unbiasing the finals seems pretty reasonable, though I guess the data is always going to be biased some considering semis and finals have an overlap of contestants, just the semis have more climbers. Reading more also answered my concern about the ELO inflation as well. Eyeballing the data, it seems in 2008, boulders had an average grade of ~2100 and now they have a grade closer to 2800, so suggesting boulders have gone from roughly 7A to 8A, correct? Do you have an ideas why this is, be it some bias in the model, competitions have gotten more competitive, climbers have gotten better, style of problems have change, etc? I'm inclined to believe the more droppable nature of comp style and newer generations that have grown up training for these comps are probably the main reasons, but I guess its hard to say without seeing videos of those 2008 tournaments, which I couldn't find... Anyway, very cool results, I'm excited to see how things develop! As a scientist who mainly does frequentist/ML inferencing in my research, its cool to see similar/different techniques applied to climbing. Great job :)

[@AbsoluteMoose]

Now we are talking, Climbing Geek!

[@notu483]

The only thing missing is LaTeX formatting for the math equations. Very interesting!

[@BugalydoshOG]

On a side note, I think Adam Ondra was the incorrect choice for a peak example... data shows that also

[@GeekClimber]

Adam Ondra was on top for a long time in the past!

[@BugalydoshOG]

@GeekClimber I think he's always been a bit overrated lol but that's just me and MattClimber lol I still remember the IFSC world comp Tomoa was winning and only lost because Adam knew the outdoor hand jam style and it urked me to see his expressive celebration (which he was even doing in the last comp he ended up not podiuming at).

[@НиколайТобиас]

​@@BugalydoshOGis mattclimber that weird fellow with huge ego from instagram that people are making fun of lately?

[@НиколайТобиас]

​@@BugalydoshOGanyway, data doesnt overrate, data displays things that were actually happening

[@coolguyASDQWEFEWFADSFAS]

I bet those comp slabs would be 3 or 4 grades lower if they were outdoor boulders.

[@AscentStats]

Absolutely! The way to think of competition grades is “they might as well have been grade X for all the chance you’ll have of sending it in 4 minutes.” Outdoor grades have no time limits, so the fact that a slab will take longer to work out and execute has no bearing on the grade. Competition grades take into account that some boulders are harder to do in a time limit than others.

[@ArchibaldVonSkip]

This man is doing gods' work. Hope he sleeps peacefully. :-)

[@TheAbd1233]

Hi, Greek is really cool that you made this video with and in conjunction with an expert in the field. I think this video is a good idea and really unique and it gives enough of an overview of the maths as well. Just a suggestion for video editing I found it really annoying when you moves your face in the video often. So the solution might be to keep your face on the right and the interviewee on the left. When you need to show other stuff like maths or climbing stuff either transition the entire screen or just the right side where your face is. This wil mean the professional talkings face will always be seen by viewers. I wish you the best of luck I think you should have colab with pros from other fields

[@Anza2700]

unrelated but what training goals are you currently working on?

[@Jbones2000]

It would be cool to see if this model is accurate. If the probability you flash a V6 is 0.1, then if you tried 10 different V6s you should be able to flash one of them. You could go to a gym and test this model out

[@Jurakubaxd]

Where can i find your list of calisthenics Skills ranked?

[@randomyogi]

Curious how alexei would value crimps, dyno, slope etc and how much that would affect a persons grade.

[@AscentStats]

One obvious idea to incorporate different styles of climbing is to categorise the different boulder problems. Slab is the easiest category since it is usually easy to recognise unambiguously and commentators usually point it out. Then you could provide a grade for each competitor on each style, so as to identify weaknesses and better predict outcomes (since we know there is usually exactly one slab, for example). The barrier to doing this is simply that there is not a curated dataset of the climbing styles for competition boulders. That would be a lot of manual labour. If you know a very keen person then we could do it :)

[@TheFreeRunPorject]

Amazing!

[@notapplicable7292]

mans got crazy eyebrows