Tangotiger Blog

Thursday, February 07, 2013

WAR Database

James gives you some step-by-step in processing the sublime BR.com WAR file.?

Wednesday, February 06, 2013

MediaWiki Developer

Since I had fantastic luck in a Straight Arrow reader volunteering to maintain the Raines30.com site (thanks Reggie!), I figured, well, why not with my Wiki.  I have a stand-alone Wiki here, and the software is provided/installed from my web host.  Users have to register, and, frankly, it's grown stale if only because I stopped registration (too many spammers).

But ExpressionEngine has their own Wiki module, and I have member registration.  So, I'd like to transfer that wiki over to the EE-controlled version.  "Theoretically", it's supposed to be possible.  I don't have the appetite to learn yet another new thing that I'm not going to have any use for after it's completed.? 

Therefore, I figure there must be one Straight Arrow reader that has familiarity enough with MediaWiki and/or the EE Wiki module that he'll want to take the bull by the horns on this one.  If I'm describing you, then email me.  This is pure volunteer.

Knuckler Idol

Great idea here.  Probably the biggest constraint on a team in trying to find the next knuckler is the cost.  If you have a 1% chance at getting the next Tim Wakefield, then you need 100 pitchers willing to try this, and it's going to go for years.  But, get someone else to pay for part of that cost?, and suddenly, it gets a bit more attractive.  Anyway, regardless of how serious this is supposed to be, it sounds like great fun.  Getting former quarterbacks to try this seems like a decent way not to worry about "undoing" mechanics that may not apply here, not to mention that former pitchers are probably in denial anyway.  Anyway, just love the whole effort behind it.

Micah Owings

I love these pitcher-to-nonpitcher (or vice-versa) stories.?

PED is a workplace safety issue

Matt captures my feelings on the matter.  PED creates an entry barrier, one that may require others to use if they want to enter.  But with the health risks, perceived or real, with PED, this now becomes a health concern, a workplace safety issue.

Here's a crazy idea: what if you allow players to use PED, but they must register themselves as PED users.  Much like players get TUE (therapeutic-use exemption), they get a similar tag for PED.  Except in this case, the registration is made public.  You can even provide some disincentive, say that the player must sit out one month a year (unpaid).  But any unregistered player who tests positive gets an immediate two-year ban.  He is then reinstated as a registered PED user, even if he promises to no longer use PED. And if one month is not enough, then make it a two-month unpaid sit-out.

Crazy?  Or crazy enough to work?

?

“Never too old to be a kid”

Now this sounds like a Bucket List item.?

Why doesn’t USPS increase its stamp to 63 cents and stop Saturday delivery?

Canada should often be considered a "sandbox" for USA, a development environment to see how things work on a small scale.  The mail system is one such comparison point.  USA's population is about ten times that of Canada.  As luck would have it, USPS has about 80 billion$ in expenses, while Canada Post has 8 billion$.  Per-capita expenses therefore as the same.

On the other hand, while Canada Post is close to breakeven (8 billion$ in revenue), USPS? only collects 60 billion$ in revenue.  What is the main difference between Canada Post and USPS?  Well, Canada does not have Saturday delivery.  And, it costs 63 cents to mail a letter in Canada, while it costs only 46 cents in USA.

Canada Post collects one dollar in revenue for every one dollar in expense.  USPS collects 75 cents in revenue for every one dollar in expense.  Therefore, USPS needs to increase their revenue by +33%.  (That is, they want to collect 80 billion$ in revenue, not 60 billion$, so an extra 20 billion on their base of 60 billion$).

And 46 cents plus 33.33% is 61 cents.  That's very comparable to Canada's 63 cents.

FedEx and UPS also operate in Canada, so the competition is also there.

So, what is stopping Congress from authorizing the 63 cent stamp?  (In which case, there's going to be a MASSIVE run on the Forever Stamp, making it one of the best investments ever, and almost certainly creating a secondary market of businesses whose sole function will be to sell the Forever Stamp at somewhere between 46 and 63 cents.)

Should we prefer a spread in forecasts?

All other things equal?  Glenn talks about something that Colin brought up.

Let's take a step back first.  Marcel, as we all know, gives an identical forecast for any player who has never played in MLB.  And, as we learned, Marcel has a smaller margin of error than virtually every forecasting system.

Hold on to your hats here.  Ready?  These are the players that are Pure Rookies.  They had no prior MLB history for any system to draw from.  Marcel decided to give a blanket .335 forecast for each player, while the other four systems relied on their minor league stats.

<span>wOBA    Error    System 0.319    0.0000    Actual 0.306    0.0436    Chone 0.335    0.0416    Marcel 0.320    0.0414    Oliver 0.313    0.0430    Pecota 0.307    0.0439    Zips</span>

First off, we see forecasts all over the place.  While the group of Pure Rookies hit .319 wOBA, the other four systems forecasted .306 to .320.  Marcel of course was exactly .335.

But, look at the error term: Marcel nearly won!  And Chone, which was leading in each sub-category took a bit of a hit here.  Chone, along with Zips forecasted the overall mean too low, and the error term were the highest.  Not that any of the systems really redeemed themselves here.

I go on and discuss at length the selection bias issue.  Go over there to read about it.  Anyway, let's set aside the selection bias issue.  Let's presume there isn't one.  And we don't even have to talk about wOBA. 

We can simply talk about a team's pre-season W/L forecast.  And we can do it for NFL.  If I predict 8-8 for every single team, I'll be off by, I dunno, say an average of 3 wins per team.  But, if someone else starts forecasting 11-5 and 7-9 and 3-13, and they ALSO end up with an average error of 3 wins per team, is that better?  What if they forecast one team for 4-12 and they end up 9-7.  Should we be happy that that's the price we pay for predicting a team to go 13-3 and they in fact go 14-2?

Let's say it is.

Let's take another case where I forecast 8-8 for every team, and I'm off by 3 wins per team.  And someone else has forecasts all over the place from 3-13 to 13-3, and they are off by 3.5 wins per team.  Who made the better forecast?

In my case, the standard deviation of my forecasts was exactly 0, while the other guy had a standard deviation in his forecasts of say 2.3 wins.  And let's say we actually observe 2.8 wins as a standard deviation.  So, yes, he was better able to forecast the league-wide spread.  But, was the price too high?  That he was off by 3.5 wins per team in making that kind of spread in forecast, is that necesssarily better than me being off by 3 wins by having zero spread in team forecasts?

Where's the tradeoff here?

?

Tuesday, February 05, 2013

Example of Replacement-Level players

Dave does a good job at identifying replacement-level players.  In order to (partially) combat the selection bias that Dave correctly noted, he should present the 2011 season as separate from the 2012 season.

To combat selection bias, we don’t want to just focus on what these players did last year, as the fact that we’re identifying players who were forced to sign minor league deals means that we’re starting with a group that likely underachieved last year. A replacement level player who overperformed in 2012 likely secured his place on a 40 man roster for the winter, removing him from the pool of players available to sign minor league deals or get passed through waivers. So, we need to adjust for the fact that the 2012 performances are likely a bit below their actual talent levels, and we can simply adjust by looking at a larger pool of data. We don’t want to go back too far, of course, as many of these guys are aging players who aren’t what they used to be, so to try and come up with a balance of a larger but still relevant sample, we’ll simply focus on how these 24 players did over the last two years.

So, everything he said was correct, except the last three words. He should have focused on 2011, as that (mostly) represents an unbiased estimator of the group's talent in 2013.?

I'd also like to see the study repeated for pitchers.  I think we might get some "weirder" results.

Method of Baseball Reference’s madness

In response to discussion of Paul Abbott on Bill James' site:

Paul Abbott: his Wins Above Replacement (WAR) as calculated by Baseball Reference (which I note as rWAR) in 2000 was 2.3 wins, while in 2001 was 1.1.  That 1.2 wins difference is explained in part because Forman has calculated that his opponents in those years were stronger in 2000 than 2001 (by 0.45 runs per 9IP), as well as his fielders were better in 2001 than 2000 (by coincidentially also 0.45 runs per 9IP).  After also considering park factors, Forman figured an average pitcher in Abbott's context in 2000 would allow 5.00 runs per 9IP and in 2001 would allow 4.28.  Since he gave up 4.47 in 2000 (0.53 runs per 9IP better than average) and 4.36 in 2001 (0.08 runs worse than average), we have a 0.61 run per 9IP gap.  Which for the seasons in question is about a 12 run gap, or 1.2 wins.  This is not to suggest that I agree or can support all of Forman's assumptions in those calculations, but simply that there was a method to what seemed like madness.?

***

Paul Abbott was assigned a record of 17-4 for those historical 2001 Mariners, while a 9-7 record for the 2000 Mariners.  For whatever it's worth, Sean's neutralizing function gives Abbott a 9-10 record in 2000 and 7-10 in 2001.

Note that Sean's neutralizing function actually assumes that the pitcher will receive a normal distribution of runs, and recasts his performance as if it were randomly set against that distribution.  This is a bit hard to explain, but:

If a pitcher did in fact face a normal distribution of runs, in a normal park, under a normal environment, his neutralized W-L record can STILL be massively affected, because Sean ALSO presumes that the pitcher's performance (that is, his runs allowed) should have been randomly distributed. 

So, he's not only neutralizing the player's environment, but he's also neutralizing any "timing" component of his runs allowed.  I don't know if I necessarily agree (or disagree) with that approach.

What's weird is that he doesn't do that for hitters, with respect to runs scored and RBIs.  In that case, a leadoff hitter like Rickey will maintain his massive gap in R and RBI, thereby preserving some part of that context.  I didn't check to see if a guy's "clutch" performance (i.e., driving in far more runners than his seasonal lines would suggest) is preserved.

Anyway, no one ever talks about it, so I'm putting it out there.

Discussion of the stats landscape

A decent effort to lay out the discussion.  First the old school, then some of the basic new school stats.

There are some errors, for example:

Lots of people like to use the statistic OPS (OBP+SLG) as a quick, shorthand way of combing all of these stats.  The caveat to this is thus; is a “point” of on-base percentage equal to a “point” of slugging?  No, it is not; the slugging point is worth more because of what it represents.

?Unfortunately, that's not true. As I've shown in the past, the best weighting of OBP and SLG is roughly 1.7 OBP for every 1 SLG.

PITCHf/x park effects

Good job by Jon to present a basic version of park effects.

You can of course refine it further, but I think his basic version provides enough information to show that there is some park-to-park variance in the data.

Ubermodels are…

?models that capture the involvement of various entities in games to estimate each of their impact toward winning

Involvement: participation, without necessarily attribution of responsibility, skill, or luck

Entity: player, manager, umpire, park, weather, loving hand of god, cruel hand of fate

Estimate: approximate calculation whose rough value can be derived in multiple ways

***

Can we all agree on this?

Monday, February 04, 2013

Match-fixing in Europe

The story.?

Fangraphs Plus

Make your annual five dollar donation to Fangraphs to say "thank you", and get ?plenty of "you are welcome" gifts from the good group of analysts there.  I think it's obvious that David has done more than anyone to making wOBA, FIP and other things I've dabbled in as ubiquitous as possible, not to mention the other 99% of the things that make his site great, and you should do your best to show your appreciation to what David is doing over there.  It's really insane that he does all that he does without charging subscription fees, and yet paying his writers.

Spread of offense v defense

Phil makes the point that while the spread in offense and defense is roughly the same in MLB, the NHL has a larger spread in defense than offense.  And the reason should be clear: goalies.  I agree with basically everything Phil said in there.?

When do QB metrics stabilize?

Great stuff here!  Love to see the techniques I use on baseball being applied to other sports.  So, yards per attempt has an r=.50 after about 800 attempts.  Basically, that means you need nearly two years for half the metric to be considered signal and half to be noise.  He goes through some metrics, and determines:

Note that because he took a QB's numbers in the same year, then something like TD rate does not necessarily reflect the QB's skill, but rather may have a bias in his receivers.

Interception rate being so low is also interesting.  That may be because in order to survive, then you can't throw alot of interceptions to begin with.  In order to find a signal in something, you need to have samples that have a large range in talent to begin with.  If everyone has low interception rates, then it's harder to find out who is really really low in interception rates, and who is really low, and who is low.

This is why something like save percentage for goalies has low correlation: if you aren't saving pucks, you aren't going to be in the league long enough to be part of the sample.  And this is why strikeout rates have high correlation: you CAN survive in the league if you have a very low and very high K rate.  So, with a large range in talent, then  you need less sample for the signal to get through.

Anyway, so that explains the sack rate stabilizing so fast: it's highly biased on the team's line, and staying away from sacks is not a primary requirement to being a QB (though it is a secondary one).

WAR hammer for nail question

Dave does a good job of contrasting WAR to the other metrics, and why it has particular uses.  My favorite part though is when he dismisses the Pitcher Wins as an answer, since the question is one that no one asks to begin with:

How many times did that pitcher complete at least five innings, leave the game with his team having outscored the opponent through the point at which he was removed, and then watch his relievers finish the game for him without surrendering the lead that his teammates helped create in the first place??

Sabermetric Super Bowl

I don't watch enough football to remember to have Brian's live win probability chart up.  One thing that I like to know is if a team that is trailing has more than a 50% chance of winning.

It's more obvious to see in baseball, where if you are down by 1 in the bottom of the ninth, no outs, and bases loaded, it's the defense that's sweating bullets, not the offense.  But, when it does it flip over to the offense?  Well, it just tips over with runners on first and second, no outs, down by 1, or runner on third, down by 1.  (It's often the case that having runners on first and second is equivalent to just having a runner on third.)

Brian's site is (now) blocked at the office.  Can someone tell me if there was ever a point where the trailing team had more than a 50% chance of winning?

The other fascinating play was at the end, with the safety.  I think the broadcast team did a great job to bring it up at all, and it was interesting to hear their off-the-cuff unprepared analysis for it, saying they wouldn't do it.  But, they totally didn't consider that they could run eight seconds off the clock.  Even without the clock-running, it would seem it might have been more than breakeven to go for it.  I'll wait for Brian's analysis on that too.  But, I thought the director blew it by not showing us the defense formation? from a high view, and showing what the defense was going to do about it.  I think in that case, it demanded a bird's eye view.

Any other high-leverage strategy plays?

Mozeliak talks the numbers game

Blogger Anna gives us his insight:

"There is no perfect stat, but when you look at trying to define Wins Above Replacement, it is a very simple place to grab information and get a feel for it," he said.

There are several versions of WAR out there, and while he looks at all of them, he does have a favorite.

"I use our own internal system, because that's what I'm most familiar with and also in the past five, six years I feel like we've made very good decisions based on it," Mozeliak said. "My confidence in it is very strong."

?