Fangraphs - Remaining Standings Projections

[igor67]

2 hours ago, Thurston Fluff said:

 

If the goal is to predict an outcome shouldn't the model be secondary to accuracy? Instead of sticking to the model that doesn't work why not find one that does for this team? 

Accuracy in what sense though? If your model just has a number that says the Brewers are special to get the math right it's not accurate in a scientific sense. You want to be able to figure out what the Brewers are doing and see if other teams are starting to catch on. I wouldn't say that costs them credibility (different poster) unless you can show another model that does as good with other teams and also projects the Brewers better. My recent recollection is that most models have been struggling to accurately project the Brewers.

[ClosetBrewerFan]

7 minutes ago, igor67 said:

Accuracy in what sense though? If your model just has a number that says the Brewers are special to get the math right it's not accurate in a scientific sense. You want to be able to figure out what the Brewers are doing and see if other teams are starting to catch on. I wouldn't say that costs them credibility (different poster) unless you can show another model that does as good with other teams and also projects the Brewers better. My recent recollection is that most models have been struggling to accurately project the Brewers.

The models do not have a "devil magic" adjustment factor for the Brewers. 

[gregmag]

2 hours ago, Underachiever said:

They are probably assuming that the Brewers will have nothing to play for when they clinch everything with 20 games left. 

I know you’re kidding, but I was actually thinking about this. There is a drag factor once teams clinch. Put it this way: My wife got me tickets to come up to Milwaukee for the final home series of the season, and I’m very intrigued about whom I’ll get to see. Unless the Brewers collapse between now and then, I ain’t seeing Miz. 

[owbc]

5 minutes ago, gregmag said:

I know you’re kidding, but I was actually thinking about this. There is a drag factor once teams clinch. Put it this way: My wife got me tickets to come up to Milwaukee for the final home series of the season, and I’m very intrigued about whom I’ll get to see. Unless the Brewers collapse between now and then, I ain’t seeing Miz. 

I'm currently trying to pawn off Miz and Harrison onto someone else in my fantasy league but I probably should have shipped them off before their load management became common knowledge. 

[Thurston Fluff]

1 hour ago, igor67 said:

Accuracy in what sense though? If your model just has a number that says the Brewers are special to get the math right it's not accurate in a scientific sense. You want to be able to figure out what the Brewers are doing and see if other teams are starting to catch on. I wouldn't say that costs them credibility (different poster) unless you can show another model that does as good with other teams and also projects the Brewers better. My recent recollection is that most models have been struggling to accurately project the Brewers.

You don't have to figure out what the Brewers are doing. You just have to take an average the model misses by over a set period of time and adjust it accordingly. It's just adding another stat to account for variables that are not accounted for in the model.

[igor67]

Yes you could, but the point I'm trying to make is that this is not how people are trained to build models for other situations. AKA if you did that on a test you'd fail, because mathematically you can always just add more factors to get a better answer.

[owbc]

47 minutes ago, Thurston Fluff said:

You don't have to figure out what the Brewers are doing. You just have to take an average the model misses by over a set period of time and adjust it accordingly. It's just adding another stat to account for variables that are not accounted for in the model.

This is what political pollsters have done for certain candidates who their models struggle to predict. I find it to be a recipe for disaster. 

The correct thing to do (in lieu of figuring out why the model can't correctly predict the Brewers) is to present the results as-is. Everybody can look up the bias from past years and do their own correction. 

There is eventually going to be a year when the Brewers don't live up to expectations. It's inevitable. 

[Thurston Fluff]

32 minutes ago, owbc said:

This is what political pollsters have done for certain candidates who their models struggle to predict. I find it to be a recipe for disaster. 

The correct thing to do (in lieu of figuring out why the model can't correctly predict the Brewers) is to present the results as-is. Everybody can look up the bias from past years and do their own correction. 

There is eventually going to be a year when the Brewers don't live up to expectations. It's inevitable. 

Political predictions rely on polls. Polls are known to have accuracy issues.How they pick random people to poll, one demographic being less inclined to take the poll than another, people who lie, vague or misunderstood questions, questions that require a black or white responses that doesn't account for a more nuanced answer to name a few. Which is what pollsters try to adjust for.

Baseball uses numbers from past performance. Much less ambiguity. All I'm saying is they could add one more cold hard fact. Calculate the variance of past performance against the model used to assess future performance then use it in your next projection.

[Thurston Fluff]

1 hour ago, igor67 said:

Yes you could, but the point I'm trying to make is that this is not how people are trained to build models for other situations. AKA if you did that on a test you'd fail, because mathematically you can always just add more factors to get a better answer.

If anyone said that's good enough when the model fails with regularity they should not be taken seriously. Once a model is considered "good enough"  and no longer tries to improve it becomes obsolete.

I don't think anyone tries to build an inaccurate model. Nor do I think anyone should just ignore something their model fails at and expect to be taken seriously. You're always supposed to be trying to get more accurate. It's one thing to ignore an outlier season. Quite another when the same outlier happens regularly over a decade by the same team. That's no longer a coincidental  outlier. That's a team that's telling you you're missing something. When that happens you either adjust for it or admit you haven't got an accurate model.  

 

[igor67]

Do we know they just aren't stumped? Perhaps they find other unanswered questions more pressing at the moment? Maybe a little of both and they think other questions are more likely to yield results sooner. My point was not to suggest they shouldn't be trying to find ways of improving, just that it may be a hard problem to solve in a satisfying way.