Thursday, September 30, 2004

Predictive Playoff Metrics: Tom Tippett,
Components & the Craft of the Small Project  

In the last entry, I talked about a ridiculous attempt at cobbling together ill-considered measures to predict the outcome of playoff series. Any kind of prediction of a short series of games (five or seven) cannot be made, in itself, a science. You can, however, use measures to improve your chances of differentiating the very skilled and a little lucky from the somewhat skilled and very lucky.

Beyond baseball, managers have to do this all the time, what I'll call "the Playoff Challenge". You have a small, quick-turnaround project or effort, you're going to make almost as many decisions in a short period as you would in a fuller, longer project. Successful efforts and projects have a fixed overhead of planning and design and forethought and set-up and clean-up decisions to make, and this foundation is not going to be much different for a quick effort than it is for a larger one. A manager will make more decisions with fewer chances for mid-course corrections. Good decisions are just as necessary, but not as valuable, because the living, operating moments of that good decision are shorter-lived and have less time to accumulate advantages. On the plus side, though, the mediocre ones hurt you for less time, too.

Learning how to use historical information, but also to keep light on your feet to adapt to the changes whizzing by on a short endeavor is a challenge much like a team trying to win a short series. A lesser team has a better chance of upsetting a superior opponent over a 7-game series by beating them 4 times than they do in a 162-game series by beating them 82 times. So measuring overall quality will certainly be a factor, but not the factor.

So how do you isolate out smaller factors that might have disproportionate impact on a short series that will take place soon?


Tom Tippett, the creator of Diamond Mind Baseball, the most sophisticated stat-based simulation of baseball on the field, posted on his weblog in a 9/21 entry called "And down the stretch they come..." some data to help people try to think about which teams might have the best chances for success in the playoffs and World Series. Tippett's work is figuring out how to create mathematical models that reflect reality as closely as possible, so it's never the clean-room Ivory Tower math that is always so interesting to read but usually lacking in context and applied torque. His stuff, when he presents it, is among the most interesting and the most practical sabermetric research we can get our hands on.

Here's his intro:

Two weeks from today, the first pitch of the postseason will be sent plateward, and while we don't yet know who'll be on the mound or in the batter's box at that moment, it's almost certain that the series in question could go either way. That's always true of any short series, but it seems especially so this year, with all of the leading playoff contenders riding a wave of success since the trade deadline.

In an effort to gauge the quality of the teams most likely to survive the regular season, I decided to take a look at how they've performed since the trade deadline. Some of the contenders made important changes to their team at that time, so their records since the deadline might be a better indicator of the quality of those teams than their overall records.

Like other sabermetricians, Tippett applies a momentum model. A general test many use is Win-Loss records in the second half of the season. Based on his own number-crunching, he tweaked his measure of recent-record from second half to the 7 weeks from the July 31st trading deadline to when he wrote the piece. ┬┐Why July 31st instead of the beginning of the second half, around July 1? Because contending teams tend to be active in trading for complementary pieces in the weeks leading up to that July 31 "deadline" (it's not such a rigid barrier; creative general managers can still make significant moves after that, but its administratively simpler before and because many teams act as though it's a hard barrier, many make the moves before then anyway, reducing options for patient G.M.s, making even the patient ones have to pander some to the panic). So the roster a team takes into the playoffs will be a closer to the post-July 31 roster than the roster it carried in July, and these moves can affect teams for good and ill.

So Tippett's approach, measuring the win-loss record of the team as close as it can to being the team it will take into the playoffs is a good way to trim data points and focus down on the more critical. But win-loss records, while they measure actual success, may hide some part of a team's quality because luck plays a little factor, and the smaller the run of games that makes up the record, the more likely there will be some drift, some bit of hiccup between the actual win-loss record during that stretch and the components of winning.

Tippett, like many sabermetricians, quantifies measures of winning, which can also indicate relative strength and potential for future winning. Tippett's measure is Net Runs, the runs a team scored minus the number they allowed. Runs, of course, are results themselves, results of other components, in this case bases earned through hits and walks. So he uses a third metric, TBW (total bases plus walks), and creates another measure for each team, Net TBW (what they created minus what they yielded).

By breaking building blocks into smaller sub-assemblies, he builds a foundation for measuring overall quality using multiple tests. His table for the contenders as of September 21 looks like this:

American        W   L   Pct   GBL  Runs   TBW
Boston         33  12  .733         +86  +222
New York       28  17  .622   5.0   +39  +130
Minnesota      28  17  .622   5.0   +41  +129
Oakland        28  17  .622   5.0   +18   +91
Anaheim        27  17  .614   5.5   +41    +3

National       W   L     Pct  GBL Runs   TBW
St. Louis     30  14    .682       +53  +165
Houston       30  15    .667  0.5  +50   +82
Atlanta       30  16    .652  1.0  +55  +114
San Francsco  27  16    .628  2.5  +62  +197
Florida       26  17    .605  3.5  +41   +26
Chicago       25  17    .595  4.0  +36  +119
Los Angeles   25  20    .556  5.5  +29   +53

The Tippett analysis is flat-out interesting. The won-lost records over the seven weeks show that none of the teams with a chance at the playoffs is lugging -- all have good records. The Cards had the best record in this time slice, just as they would the rest of the season if you removed this set of games from the sample.

In the American League, the won-lost record lines up really smoothly with the TBW, and reasonably well with Net Runs, with Anaheim clearly overproducing runs relative to their components. In 2002, the Angels who won the World Series had underperformed most of their component measures Here, for the Angels and As to compete with the others, they'll need more than their share of good bounces and breaks, and while in a short series that can easily happen, chance tends to even out for most teams.

In the National League, there are more seams, some bigger differences between actual records and components. The Giants have the best Net TBW by quite a bit, the best Net Runs by a noticeable step, and a good record, but not quite as good as three other very good teams. The senior circuit's match-ups, as measured by momentum and components of recent efforts, is pretty even, though the Cards and Giants are the leaders.

Analysis can't end here. There are a lot of other factors you want to stitch into your judgement of which is the better team in any individual match-up.

Team tendencies: Some teams have a strong ability to beat up on pitching of one kind or another (left-handed, or sinker-slider, or power pitching, or ambidextrous). Others look feeble against that. Since most human systems are self-amplifying, organizations tend to produce (and sometimes even go to the trouble of acquiring) players of similar aptitudes over and over, sometimes adhering to the pattern over decades (Dodgers at 3rd base, Athletics at shortstop from about 1901 until Bert Campaneris). Boston's home park favors hitters who bat from the right side, so the team will load up with right-handed pull hitters so you don't want to put up too many lefty pitchers. The patterns are well-known even when they aren't true; it's an automatic thought that you throw left-handed pitchers against the Yankees because they build their team around batters who hit from the left side to take advantage of the reachable right-field fence, so lefty pitchers have the potential to do a better job of supressing their left-handed hitters. This year, that isn't true; the team is very balanced, with a microscopic statistical edge against left-handed pitchers (their acquisition of Gary Sheffield means the Bronxians' marquee offensive player is right-handed this year), tweaking the chemistry. But it goes ebyond "teams" because at key junctures, the smaller components, the players, will decide outcomes.

Individual tendencies: Each game itself is broken down into a series of pitcher versus batter & batter versus pitcher matchups. At many ordinary moments, and at a few of the critical ones, one side or the other will get a match-up that favors them strongly. You can examine historical pitcher-to-batter match-up info, and sometimes you get the roughly 20 plate appearance history that crosses the line for a pretty good indicator. It beats the heck out of the model I pointed to in the last entry -- matching up, for example, 3rd basemen head-to-head, because rarely will one team's 3rd sacker interact in a focused way with the opposition's.

If you add these to the momentum factors like Tippett's, you can better gauge the probabilities of one playoff team winning over the other. The shortness of the series makes the random factors more likely to weigh in than they would over a whole season, but better teams usually win more series, even the short ones.


In non-baseball organizations, you can apply these models to your own "Playoff Challenges".

Your short, intense projects or efforts or campaigns will be more highly affected by random factors. High-quality achievement will be more diluted by a smaller series of events because "Quality will out in the end" is a long-term, glacial, process.

That doesn't mean you ignore quality, but it does mean as a manager you have to attend to more small details and respond more quickly, being more aggressive about matching people up in ad hoc teams to complement skills, being more willing to accept certain kinds of imperfections in exchgange for harvesting the unforeseen advantages that pop up.

It's a partially different skill set from being successful over the long haul where long-term probabilities as known by "The Book" are likeliest to yield the highest returns. The short, intense effort requires knowledge of the book, but a greater willingness to run against it at the "right" time, and those "right" times will dominate more of the outcome in the shorter series of events.

This difference is why some managers, even ones with long records of mediocrity, seem to have a knack for helping their teams win playoff series, while others who manage well over a long season have a knack for lower yields in playoff series.

To succeed at your own Playoff Challenges, embrace metrics, but embrace the knowledge that randomness will have a proportionately higher effect on the outcomes than on a longer effort. Grab the edges the environment offers you, even the ones you wouldn't normally pursue, make the most of every event, every choice, every affordance. The playoffs can be fickle, but you increase your chances if you are flexible and bold.

This page is powered by Blogger. Isn't yours?

free website counter