APBRmetrics

Opening up a thread for the contest and any preliminary discussion related to it.



Most metric based entrants in the contest project individual players and aggregate. Has anyone made a systematic effort to lol at team level variables and add them on top of or into the player projections? We acknowledge context and the sum being greater than the parts so where are the steps for attempting to incorporate these considerations, beyond the subjective and little discussed ways that they may be included after the aggregation of individual inputs? Shouldn't we be trying to take awareness of team level stats, "system", coaching strategy & tactics and find ways to quantitatively capture and include them? Initial and partial like with the player tracking elements in PT-PM but then more comprehensively like with Dredge? I haven't thought a lot about what to try to include beyond Dredge but imagine there are some things worth considering. Recent historical home court advantage, performance on road, in tight games, when playing more back to back than normal, consistency (coincidence of above and below average performance or not), physical & mental management of players, performance out of timeouts, clutch / crunch performance (already included in the overall stats but maybe not weighted heavily enough), etc. Team effects are in previous year data but players, coaches and other management come & go, have different levels of awareness / use of team knowledge / inputs.

ESPN's future power rankings recognizes big picture elements. Can we take that and quantify short term effects? At least one or maybe a few projectors simulate the season to capture the detail of very different schedules. For projectors that don't would it be better if they did? Conceivably some super smart person (model builder / coder) could build a tool to allow other smart people not able to handle this step on their own to plug in their team summary strengths. Or maybe there could be simpler but still number driven manual adjustments.

Lineup efficiency and four factor adequacy / synergy might be another part of team impacts. You can measure lineup actual wins, maybe can estimate expected wins of lineups used and compare. Perhaps could adjust future win projections if there is a pattern of over or under performance. Part wisdom, part random results but may be worth to adjust expectations versus not addressing.

ESPN Summer Forecast
http://www.espn.com/nba/story/_/id/1713 ... -standings
http://www.espn.com/nba/story/_/id/1713 ... -standings

Doing some early projections the biggest one that stands out to me is Detroit projects terribly. Most of their roster is a regression candidate.

ESPN projects a more than 3 win change in 13 cases, less in 17. I am not saying it is too few, at this point of my analysis. Just an observation.

I could see some regression in Detroit's top 3 but after that, amongst who is still there, I dunno if there is much room to regress directly with those guys. I guess some of the absences may cause regression when replaced with somebody else.

Has there ever been a player who is not a "regression candidate" ?

Candidate maybe always but you could be less likely than normal. Perhaps three or more years of steady or gain and not old.

A stronger than average regression candidate would be a better way to phrase it. Most of the Pistons feature players had the best season of their career last year for example.

Pistons avg'd 25.0 years old, 7th youngest and 1.8 yrs less than league avg. If younger players are improving, shouldn't we expect them to keep improving? Or at least not suddenly reverse their growth?

By some measures, 2 of their top 3 players -- Morris and Drummond -- have already leveled off.
Kentavious and Reggie have improved each year. Isn't that a trend?

It may be legit to invoke a poor culture in Detroit. But they also have a tradition of winning.

Mike G wrote:Has there ever been a player who is not a "regression candidate" ?

Well, depending on the methodology everyone will have regression applied, but the question is the direction and magnitude. Many players are regressed "upwards."

Regression does not actually imply getting worse. It's about regressing to some mean.


I'd like to see this for a win contest: weigh by the number of new players/minutes given to new players. You shouldn't get a ton of credit for predicting a team that didn't have any significant changes.

Are we establishing a due date for win projections being posted here? Working on mine now.

It is tip-off of first game, or a day or two after if you ask nicely and no one objects.

Been gone for quite awhile. Count me in.

I have a conceptual question that I'm hoping someone can address.

How do we incorporate uncertainty of our model, into our model.

For example, if we flip a coin - our model forecasts that it will come up heads 50% of the time. I believe in my models accuracy with a 99.99999999999999999% confidence interval, so I really am certain that heads will come up 50% of the time.

However, if my model has two equally rated teams playing a game, on a neutral court, with the same amount of rest, I have less confidence in the accuracy of my model because my model is less certain than the chances of a coin flip.

We can look at the RMSE from previous year's predictions to get a sense at how accurate our model is. I'm left wondering then, how do we incorporate the previous year's RMSE into a model for a new season? Do we regress our forecasts, and if so, what do we regress to?

Using the previous example of two equally ranked teams playing a game on a neutral court - do we now say - I believe with a 70% confidence interval that team A has a 50% chance of winning the game?

I apolagize if this is elementary, but I'm having trouble grasping this.

Thanks

> ... How do we incorporate uncertainty of our model, into our model. ...

I think you're talking about how to incorporate that into the prediction.

The hope is that a team will win about half it's 50% games and loose about half of its 50% games so net error will be relatively small. Win totals is a linear scoring method, so you should be OK with just adding 0.5 wins to each team's expected total if you think the game is that close. (If you think it's an almost-certain blow out, you would add 1 to the winning team's expected total, and 0 to the other team's.)

Other scoring methods (like logistic scoring or something using the spread) may allow you to express confidence in any particular event in a more precise manner.

Mostly, confidence comes into play when you're doing the regression to rate the teams in the first place. In a suitably structured model, you can use confidence as a way to weigh the importance of old data and new data against each other when evaluating or updating your assessments of team strength.