APBRmetrics

Hello everyone!

awhile back I released that Mamba metric, I was kind of using this years win projections as a test run to see how it does in real world testing, and thought it did pretty decent so far all things considering. I basically got minutes from I think it was basketball monster, slightly adjusted some minutes based upon what I thought (IIRC, all i did was lower a rookie or two, I think I only did it on the lakers or maybe the nuggets but it wasnt anything too crazy) and just applied it flat, without any regression to 41 wins or age adjustment or anything (that 15 wins for the wizards unfortunately stands out here haha).

I was wondering though, when it comes to building all in one metrics, what considertions should I make when it comes to like, the quality of the RAPM? Originally since it was mostly a proof of concept type thing, and since it worked out pretty good overall, I just left it at that, but after I release it hopefully sometime within the next week or two, Im thinking of actually honing in on the actual "RAPM" side of things, rubber band adjustment, custom weighing for blowouts, home court, all those things that currently I haven't really accounted for. was wondering if there were any guidelines there or if anyone with experience building things out here think thats shown a substantial boost, id think it would help a lot but maybe not.

as a side note

RAPM seems a bit inaccessible for alot of people to run, I remember blessed J.E put his code out there but I was thinking of making some sort of rshiny dashboard sometime this month that would allow people to run it themselves and decide what adjustments they want to make, and then export it as a csv file or something, just since coding itself or getting it can be a bit intimidating for sum, but im not sure maybe that sounds a bit intimidating too IDK. not sure if thats something that sounds interesting

Rn my idea is something like

You input a date range
toggle an on and off button for HC adjustments, rubber band adjustments, maybe different weight for "garbage time" minutes (wont be neaarly as strict as PBP here and i wont have it so once it hits garbage time you need it to be super close to be non garbage time), how much luck you want on offense and defense, 3pt and fts (offense and defense separately), 0 meaning no luck whatsoever and 100 being a 100% luck adjustemnt on selected category.

and itll spit out a downloadable csv file or something along those lines, feel like it could be a fun little project but wanted to see if it sounded interesting to people, i think itd be fun to try out making something like that.

I've been trying to build out some sort of visualization and or data pulling tool that allows you to get regular or tracking data or visualize it from various date ranges and filters rather than only whole seasons, and I thought doing something similar, what ive been doing for a few days since the New years but pulling daily box scores this way has been a huge pain even with the API

So yeah a TLDR::
asking for Advice on how much doing things that make RAPM test better would help an all in one metric (kinda a dumb question thinking about it lol)
and gauging interest on making a tool allowing people to get customized RAPM spit out without needing to code

been working on this data viz thing which will also have a tab where you can (if I can get it to work) be able to get datatables with for date ranges cross season or with things you cant normally see together in the nba website (so like, youd be able to select Open3s, Open 3%, Rim shots contested, and Rim % lessened than the shooters normal avg if you wanted to ), but tbh I just needed a project where I used SQL to make/pull from a database (even if maybe a bit uneccessarily tbh lol)

My experience is that corrupting RAPM with these luck adjustments, blowout weightings, rubber band adjustments etc makes for a metric which is worse for predicting the future. Such methods, at least for me, are enticing but have proven to be worse than useless. I bet using my models, in fact an algorithm does the betting... anyway, point is it's my opinion (and what I have found when building models attempting to predict future outcomes) that those adjustments corrupt the dataset more than they lift it from the grips of multicollinear uncertainty.

However, give it a go. I would never encourage closing your mind just because I have said it wasn't helpful for beating vegas.

What I am certain of, is that a good publicly available source of RAPM will be welcomed by people, so your idea of creating something like that, with some bells and whistles, is indeed a nice idea.

v-zero wrote: ↑Fri Jan 03, 2025 9:57 am My experience is that corrupting RAPM with these luck adjustments, blowout weightings, rubber band adjustments etc makes for a metric which is worse for predicting the future. Such methods, at least for me, are enticing but have proven to be worse than useless. I bet using my models, in fact an algorithm does the betting... anyway, point is it's my opinion (and what I have found when building models attempting to predict future outcomes) that those adjustments corrupt the dataset more than they lift it from the grips of multicollinear uncertainty.

However, give it a go. I would never encourage closing your mind just because I have said it wasn't helpful for beating vegas.

What I am certain of, is that a good publicly available source of RAPM will be welcomed by people, so your idea of creating something like that, with some bells and whistles, is indeed a nice idea.

Yeah a very smart man once told me alot of people and teams got helped out by J.Es dump so I was thinking a tool like that might be practically useful, so someone can like get their own RAPM data set with the adjustments they want I guess since I feel different people be doing different things or wanting different things, still gotta figure out how practical it is though. Im not sure how often current season RAPM in general get updated, but Id aim to update the PBP weekly or something along those lines, and would be shooting for end of January - all star break depending on how long this viz tool takes to make in the first place, although I feel it doesnt sound like it would be too difficult to make?

I think right now from where it stands the MAMBA thing generally has done well, I think just based on how I think the season is going to go I think it might go something around 18-19 to 11 in O/U wins and losses vs vegas, and some of the misses like the spurs are pretty clearly just since I threw it on without any sort of like age adjustment or anything like that, or just injuries with situations like the 76ers.

Right now theres like a very conservative luck adjustment thats somewhat stronger on defense, I think it was a bit helpful but I doubt it was much of a game changer at all, I think the blowout weightings and rubber band stuff I heard helped from somewhere but I haven't heard all too much about them in general, so great to hear your insight there.

When you say a metric, do you mean the RAPM itself or as a component in an all in one or a component within an overall predictive model? Just curious.

If you built a custom RAPM site, I'd be most interested in the ability to produce splits on the database. RAPM against top 10 opponents, good defenses or offenses, last 20 games, home / road, possibly by main position, sublineup groups, by half, playoffs, etc.

I know estimated errors become a bigger issue with smaller samples but I am looking for shadows that might be clues to consider not ultimate truth. Sample size can be pushed back up by allowing multi-year runs, mostly 1+ or 2+ years.

TeemoTeejay wrote: ↑Fri Jan 03, 2025 10:30 am When you say a metric, do you mean the RAPM itself or as a component in an all in one or a component within an overall predictive model? Just curious.

My models are numerous and come together under numerous algorithms. I have two versions of what you would call RAPM, they are significantly different from each other but are built on the same dataset. They approach the data from different perspectives, I can't go into detail.

I have four different models tracking box score data and extended box score data built out from the PBP. Again whilst these models look at the same data, they approach it from different angles and provide different information.

Every model is put through its paces both in isolation, and when combined with the others.

So, when I say I have found those things worse than useless, I mean I have tried them across a range of models with very varying philosophies, using a range of statistical and machine learning methods, and I have found them to make my models worse at what I want them for: prediction.

Crow wrote: ↑Fri Jan 03, 2025 3:35 pm If you built a custom RAPM site, I'd be most interested in the ability to produce splits on the database. RAPM against top 10 opponents, good defenses or offenses, last 20 games, home / road, possibly by main position, sublineup groups, by half, playoffs, etc.

I know estimated errors become a bigger issue with smaller samples but I am looking for shadows that might be clues to consider not ultimate truth. Sample size can be pushed back up by allowing multi-year runs, mostly 1+ or 2+ years.

Might try to add that later on, but could def do like last 20 games and stuff like that

right now im thinking of having the date be adjustable, so lets say, you pick a start and end date, among witht he other adjustments