APBRmetrics

So I know many people (including myself) have tried to predict a game outcome (win/loss or score etc) by regressing from player stats directly (lets assume that team stats is an aggregate of the stats of the players in the roster).

But the problem is to predict a future game outcome from player stats, you need to have an estimate of the player stats for the game they haven't played yet. It is like a chicken-and-egg problem.

So my question is. Have you ever tried to predict player stats from historical data (akin to a time series). I am doing that at the moment and I would very much like to hear you suggestions/solutions if any.

This is definitely on my to-do list. I've built basic projections for players in the past using box score statistics and regression models, but nothing fancy. Those were mostly just for me to use in fantasy basketball leagues with friends over the years. I'd like to build out a really strong projection system though. Something that, in addition to the more standard box score and RAPM type metrics, also incorporates aging curves, height, weight, wing-span, vertical leap, years experience in the league, teammates, and player archetypes that make use of similarity scores through clustering models or something similar. I want to build an ensemble model with at least a few different layers to it. I've never seen a player projection system that I'd want to boast about for NBA players. Everything I've come across is pretty elementary stuff. Perhaps there is something strong out there already, but I just haven't seen it. Fivethirtyeight at least tries to make use of aging curves and player similarity types in their CARMELO system, but even that isn't nearly as comprehensive as what I'd like to build out.

Yes true. In my experience (at least the models I am using) the best prediction of an upcoming game comes from an accurate projection of the players stats. Theoretically (i.e. if player projections are perfect), my models can correctly predict 71% of the games. But since there is some inaccuracy in the player projections, my game predictions are in the range of 67%.

I like your idea and I would really like to do something along those lines as well. What I am using at the moment is a very early version of that but quite more basic, so I would really like to extend it.

I have taken the idea of FiveThirtyEight which I thought was interesting but rather simple, and I adapted it.
So for each player for a specific stat, first I do a rough estimate using regression (e.g. ARIMA) and then I refine it by matching to these players with similar stat curves from the past (i.e. finding similar time-series). This gives pretty ok results but there is a lot more improvement to be made (i.e. incorporate the physical model of the player and it's changes over time).

I think player stat projections is one of the most important elements for game prediction and I want to work on this further. I would really like to discuss some of your ideas further and maybe share some work and code later on?

Isn't this what Kostya Medvedovsky's DARKO system does? It is an extraordinary system.

https://apanalytics.shinyapps.io/DARKO/

DSMok1 wrote: ↑Tue Dec 08, 2020 1:37 pm Isn't this what Kostya Medvedovsky's DARKO system does? It is an extraordinary system.

https://apanalytics.shinyapps.io/DARKO/

Need to read the details of that but it might be the case.
From the looks of it he is regressing and doing Kalman filtering (I would do particle filtering instead) and updates as new information comes along. I have to see how he does the updates but if he is doing something similar to me (regress to get the estimates as priors) and updates as games are played (as likelihoods) in a Bayesian system. And Kalman for the cleaning up of the updates (because of the noisy nature of per-game stats vs per-season stats). He might be regressing on different things though compared to what I am


Also, as he correctly points out CARMELO (538) is a single score and it compresses everything to a single dim. Such schemes are not as accurate as estimating each stat separately and then predicting using all the dimensions. 538 for example does a pretty poor job in predicting game outcomes (and probs) as the season progresses.

Thanks for the link. I ll have a closer look at that system for ideas

vzografos wrote: ↑Tue Dec 08, 2020 2:38 pm

DSMok1 wrote: ↑Tue Dec 08, 2020 1:37 pm Isn't this what Kostya Medvedovsky's DARKO system does? It is an extraordinary system.

https://apanalytics.shinyapps.io/DARKO/

Thanks for the link. I ll have a closer look at that system for ideas

Ok I read it in more details. Seems that he is using a very reasonable approach. Does not say what his actual regression method is but apart from that it is quite clear. I would really like to know how he captures the interaction effects between the different stats.
Definitely would like to see the data that comes out of the system and compare it.

Is this guy here or just in the Twittersphere?