Team Continuity, Part I
Posted by Neil Paine on August 31, 2009
They've been having a spirited debate about retrodiction over at APBRmetrics recently, and apparently we're all going to engage in some sort of challenge over whose pet metric could "retrodict" a team's 2009 performance the best. This, of course, should be nothing new to faithful readers of the BBR blog, because we retrodicted 2008-09 several months ago. But my question is, does accurate retrodiction across the entire league really "prove" anything? For the vast majority of teams, any cheeky wee monkey is going to be able to predict performance effectively based on data from the past few years because A) roster turnover is rarely drastic enough to the point that a team's stars are no longer with the team anymore, and B) even in the case of roster shakeups, coaches almost always employ new players in the same role they had been playing in for their previous team. Since we know a lot of basketball productivity is role-specific, there's really not a lot of point in boasting that past production in your metric predicted future production when a player plays exactly the same role in each situation.
To me, it seems like the more telling cases would be ones where there was a lot of roster shakeup, where players were asked to switch roles. The mythical Team of Fred Hoibergs (the "Ames Mayors"?) would be the prototype, right? I mean, you can easily predict the productivity of a player like Hoiberg in Year Y based on Y-1, Y-2, etc. if he's being used the same way in each of the seasons (low-usage 3-point specialist)... But good luck using Wins Produced, PER, Adjusted +/-, or the like to predict his production if in Year Y he's asked to be, say, the primary scorer on the team.
Now, this kind of dramatic change doesn't happen often. Like I said, players are almost always used in comfortable roles, ones similar to those they've been asked to fill in the past. But the real progress is to be made in situations of shakeup, of turmoil, where players are asked to do something they aren't familiar with. To that end (and this is only scratching the surface, I realize, but we'll probably touch on this often in the future), I thought that instead of retrodicting every team, maybe we should only focus on teams with obvious personnel changes, where the unpredictability is going to be highest because of changing roles and team dynamics. So here's a list of the teams (since 1965) with the lowest "Continuity Scores" (continuity being the % of team minutes being filled by players who were on the roster the season before):
| Year | Team | Continuity |
|---|---|---|
| 2005 | ORL | 10.6% |
| 1980 | NOJ | 10.7% |
| 2005 | ATL | 10.8% |
| 1998 | CLE | 19.4% |
| 1991 | SAC | 20.3% |
| 2009 | LAC | 22.5% |
| 1997 | DAL | 25.9% |
| 1979 | BUF | 26.1% |
| 1973 | PHI | 26.5% |
| 2002 | VAN | 27.0% |
| 1990 | SAS | 27.1% |
| 2005 | LAL | 27.1% |
| 2004 | MIN | 27.2% |
| 1998 | DEN | 28.1% |
| 2000 | CHI | 28.2% |
| 1978 | NYN | 29.0% |
| 1999 | SAC | 29.0% |
| 2005 | HOU | 30.6% |
| 1981 | GSW | 32.6% |
| 2000 | ORL | 32.7% |
| 1978 | SEA | 33.2% |
| 1978 | BUF | 33.7% |
| 1982 | DAL | 34.1% |
| 1993 | MIN | 34.3% |
| 1982 | CLE | 34.5% |
| 2005 | NOH | 34.8% |
| 2000 | HOU | 35.1% |
| 1996 | PHI | 35.5% |
| 1978 | LAL | 35.7% |
| 1999 | CHH | 36.0% |
| 1982 | NJN | 36.2% |
| 2008 | MIN | 36.3% |
| 1983 | NYK | 36.9% |
| 1998 | BOS | 37.2% |
| 2007 | TOR | 37.3% |
| 1978 | IND | 37.8% |
| 2006 | MIL | 37.8% |
| 2004 | DEN | 37.9% |
| 1975 | ATL | 38.1% |
| 1996 | MIA | 38.2% |
| 2003 | DEN | 38.2% |
| 2004 | TOR | 38.3% |
| 1993 | MIL | 38.6% |
| 1997 | VAN | 39.1% |
| 1998 | GSW | 39.1% |
| 1974 | BUF | 39.5% |
| 1987 | CLE | 39.6% |
It seems to me that retrodicting these types of teams would be more informative than the typical squad, because your metric is going to have to rise or fall on its ability to anticipate how a player will react to a different role on the team. We're always talking about "Holy Grails" when it comes to player ratings, and figuring out a way to effectively retrodict situations like this would go a long way toward developing the best metric for building future teams.
August 31st, 2009 at 12:40 pm
This is why I want to use more information than just minutes played. Every year we have players change teams and/or these player's roles change, so I don't think we have to get too crazy and focus on one specific set of teams.
My goal is to build models that analyze a specific set of players on the court at one time and approximate things like shot distribution, rebounding rates, etc. Only time will tell if doing this is an improvement over a simple method, but I agree with the points you make about roles and how we might best predict players put into new roles.
August 31st, 2009 at 12:53 pm
Can you explain precisely what is in the numerator and denominator? Me and Dean Oliver use two different versions of what I call "Roster Stability" (see this post and Dean's reply), but this is something that has always interested me.
August 31st, 2009 at 2:21 pm
Wow, I'm only 4 years behind you in terms of thought processes, Ed... :)
The denominator is minutes by the players on Team A in year Y-1 (essentially 48*5*82, plus OT minutes). The numerator is minutes by those same players, logged while playing for Team A in year Y.
So, for example: 2005 Orlando Magic continuity % = (minutes by members of '04 Magic playing for Orlando in 2005) / (total minutes of Magic roster in 2004).
Yeah, that was probably a horrible explanation that made it even more confusing...
August 31st, 2009 at 2:28 pm
There is no easy way to describe it. I've tried.
August 31st, 2009 at 3:41 pm
just taking a casual look at the list, in the case of a few exceptions (Min 2004 for example), most of these teams sucked.
August 31st, 2009 at 5:46 pm
Neil - Maybe I'm misunderstanding how you've broken things down with your continuity score, but I think you might be able to take this a step farther and look more closely at minute distribution.
Right now if I understand right your continuity % considers any player who was a member of the team the previous year to be in continuity - meaning his role won't change dramatically, but I'd say a player who gets a significant boost in minutes (due to trading / retiring of a star or natural progression or whatever) could also be considered as new to the team for your purposes. So I guess looking at the backups of lost players who aren't clearly replaced might be a way to gauge this when you calculate continuity rather than just who is new to the lineup and who isn't.
I'm thinking of a guy like Terrell Brandon, who didn't join the Cavs between '94 and '95 but went from starting 10 games to 41 games because of Mark Price's injuries. Even Drexler was buried behind All-Star Jim Paxson as a rookie (don't think he started even 10) and came on as a completely new player his sophomore year.
September 1st, 2009 at 10:48 am
I see what you're saying -- perhaps I should consider low-minute guys who get a bump in PT to be newcomers instead of being in the continuity, since for all intents and purposes there's not a big difference between a young player finally getting "his turn" and someone coming to the team for the first time.
September 3rd, 2009 at 7:42 pm
Great article. You guys are the best