Predicting the post-season: Part 1
Posted by Andy on July 7, 2009
Most teams have played half their season now so it's a good time to predict playoff winners.
I'm going to make a few posts over the next few days to show you some of the indicators I use to determine which teams have the best shot of making the playoffs.
The best indicator is, of course, those teams that are leading their divisions right now. They have the best shot of making the playoffs, and odd are that at least half the current division leaders will win their division.
The first thing I like to look at is run-scoring differential.
Let's start in the the American League. Click that link to see the standings, and go down to the "AL East Detailed Standings" as well as the sections that follow for the other divisions.
Focus in on the following columns:
Tm W L R RA pythWL Luck
BOS 49 32 5.3 4.3 48-33 0
NYY 48 33 5.6 4.8 46-35 1
TBR 44 39 5.4 4.5 48-35 -4
TOR 42 41 5 4.5 45-38 -3
BAL 36 46 4.7 5.5 35-47 0
Sorry about the formatting.
Here I've reproduced actual Wins and Losses, average Runs scored per game and average Runs Against per game, as well as the projected W-L record and something called "Luck".
The projected W-L uses the R and RA values to determine what the team's record should be. This was developed by (you guessed it) Bill James and is described here. Turns out that it is amazingly accurate in almost all cases. I've been using it the last several years and nearly 100% of the time at this point of the season, a team that is ahead or behind in its actual record as compared to the projected record catches up in the second half, in other words has its luck reversed. That Luck column shows the difference in actual wins from projected wins. So the Red Sox are right where they are supposed to be, the Yankees have 1 more win than expected, the Rays and Jays have fewer wins, and the Orioles are spot on. A number as large as 3 or 4 (positive or negative) is quite unusual and is very unlikely to be maintained.
So what does this mean? If you take away one victory from the Yankees and give 4 to the Rays, it is Tampa Bay and not New York that should be in second place. Either way you look at it, Boston deserves its 1-game lead.
Look at the AL Central, Detroit should have 1 fewer win while Minnesota should have 2 more. That would give the Twins a 2-game lead as opposed to the same lead held by the Tigers right now.
All of the teams in the AL West currently above .500 have had some luck and it should be Texas with a 1-game lead.
Over in the National League, here is the story:
The Phillies are the only NL East team outscoring their opponents and deserve to have a whopping 6-game lead right now. The Phillies are a pretty good team while nobody else is a serious playoff contender. This race is actually over already, in my opinion.
In the NL Central, the top 3 teams all have matching records with the projections. The Reds have luck to the tune of +4 and are absolutely done for the year. They deserve to be 11.5 games behind and have 3 teams ahead of them. They have no chance to win the division.
The NL West is the only division that doesn't need this analysis to be crystal clear. The Dodgers are by far the best team.
So, by this method, here are my playoff predictions for 2009:
AL East: Red Sox
AL Central: Twins
AL West: Rangers
AL Wildcard: Rays
NL East: Phillies
NL Central: Cardinals
NL West: Dodgers
NL Wildcard: Giants
Tomorrow I'll look at a different indicator and the predictions might be a little different too.
July 7th, 2009 at 11:20 am
I'm a huge fan of pythWL, and a big fan of comparing it to actual W/Ls. However, I hate calling the differential "LUCK". As Branch Rickey said, "Luck is the residue of design." In other words, if you have good fundamentals, good things will happen. Maybe it isn't LUCK, but sheer guts, determination or heart (or lack there of) that makes these differentials.
I'd prefer calling LUCK something else, like Intangibles, Fundamentals, Guts or Heart.
July 7th, 2009 at 11:28 am
I think the 4 suggestions you made are actually a lot worse than the term 'luck'. After all, it's not necessarily fundamentals that won or lost a game for a given team that has a disparity between its actual and projected records. It could be a ball that happened to bounce over the wall for a double and prevented a runner from scoring.
The bottom line is this--it's been pretty well established that records in close games are fairly random. In other words, most teams finish fairly close to .500 in games decided by 1 run. Find a team that, so far, is 12-3 in 1-run games and I'll bet my bottom dollar that they go 4-10 in 1-run games the rest of the season. Teams that consistently score more runs than their opponents win more games (obviously) and the bigger the differential, the more games they win. When the differential is small, especially in any one game, anything can happen, and it's more down to normal random distribution of a fairly small number of events. If the season were 10,000 games long, I doubt any team would vary from its expected W-L by more than a few games.
July 7th, 2009 at 11:31 am
Kingturtle is right, it isn't just luck, although luck is part of it. (A better word than luck would be statistical fluctuation.) But also scoring runs when they are needed has something to do with it, which is more related to clutch than luck.
July 7th, 2009 at 11:32 am
Whiz, you said what I said but in many fewer words and more eloquently. Thanks.
July 7th, 2009 at 11:39 am
Andy, one quibble. A team that goes 12-3 in 1-run games may be playing over its head, but I would expect them to go 7-7 the rest of the way, not 4-10. Remember, if flipping a coin yields 10 heads in a row, the odds are that the next 10 flips will be 5 heads and 5 tails, not 10 tails.
Oops, another quibble. For 10,000 games, the actual win percentage should get closer to the Pythagorean win percentage (the difference is proportional to the inverse of the square root of the number of games, crudely speaking), but the absolute win difference grows roughly as the square root of the number of games.
July 7th, 2009 at 11:43 am
Yes---again you wrote what I meant to say. What I wrote above what a classic stupid error that a stats-dumb person makes. Of course for independent events, past history doesn't matter, so teams should go roughly .500 in any future 1-run games (assuming they truly are independent, which of course they aren't actually....for starters, sometimes a team wins a 1-run game by using 6 relief pitchers in a single game. Then these guys are less fresh, and if used a lot that way in close games earlier in the season, aren't as available or effective later in the season, so it might be a bit tougher for the team to win close games.)
July 7th, 2009 at 12:46 pm
Andy, using pythWL at the all-star break, as you've done here, to predict post-season qualifiers, does history support it? how often has all-star break pythWL successfully predicted which teams would make the playoffs?
July 7th, 2009 at 12:48 pm
This is the 4th year I've done it, and it has worked well the previous 3 years. Of course it is far from the only factor.
July 7th, 2009 at 1:40 pm
kingturtle- You can evaluate the quality of PythWL as a predictive tool by using the PI standings tool. It gives the team WL expected WL and WL for the rest of the season for any date in history.
July 8th, 2009 at 3:37 pm
I've done a study of how well the mid-season record predicted the second half record, and compared it to how well the mid-season Pythagorean record did as a predictor. Using data from 1901 to 2008, the second-half record was more likely to match the first-half Pythagorean record than the actual first-half record. It wasn't a huge difference -- the standard deviation win percentage difference was .07707 versus .08065, or about 0.3 wins per team per half-season.
The interesting thing was that using a different Pythagorean exponent (closer to 1.3 than 1.8 or 1.9) gave even better predictions, by another 0.3 wins per team per half-season. A smaller Pythagorean exponent is essentially saying there will be a regression to the mean, i.e., many extreme first-half performances are unlikely to be repeated in the second-half. I may write this up as a Dugout Central article soon.
July 8th, 2009 at 10:17 pm
Interesting Whiz. I'm pretty sure B-R just uses the traditional Pythag formula of squaring the runs. The most recent research indicates one should use a floating exponent based on the scoring level; maybe closer to 1.5 for low-scoring league and closer to 2 for a high-scoring league. (Look up "PythagenPat" for more on that.) Certainly have never seen anything with an exponent as low as 1.3.
July 8th, 2009 at 11:19 pm
Yes, I use the usual Pythagenpat version for the standard expectation, which is basically a fit to full-season data. But it looks like a lower exponent is better for predicting the second half of a season. As I mentioned, it seems to be a regression to the mean thing.
July 15th, 2009 at 7:12 pm
In addition to the Pythagorean formula, I use a weighted average to project final standings. This is similar to the way the Favorite Toy weights past seasons for the EPL, except that it weights the games in the season. Game #1 has a weight of 1, game #2 has a weight of 2, and so on. To calculate the expected winning percentage for the remaining games of the season, you multiply the outcome (1 for a win, 0 for a loss) by the weight, then divide that by the sum of the weights [= n(n+1)/2, where n = number of games played]. This puts more emphasis on current trends and helps to account for changes in the rosters. By mid-season, early winning or losing streaks don't count for much, and teams playing well currently will pick up a few games over a simple extension of the current W-L% (and vice-versa). Anyway, here are the projected final standings (as of the All-Star break) based on this method:
ALE:
BOS 99-63 --
NYY 97-65 2 (WC)
TAM 92-70 7
TOR 75-87 14
BAL 73-89 16
ALC:
DET 88-74 --
CWS 86-76 2
MIN 82-80 6
KCR 66-96 22
CLE 65-97 23
ALW:
LAA 95-67 --
TEX 90-72 5
SEA 85-77 10
OAK 69-93 26
NLE:
PHI 90-72 --
FLA 83-79 7
ATL 80-82 10
NYM 74-88 16
WAS 50-112 40
NLC:
STL 85-77 --
HOU 83-79 2
MIL 82-80 3
CHC 80-82 5
CIN 77-85 8
PIT 69-93 16
NLW:
LAD 100-62 --
SFG 94-68 6 (WC)
COL 90-72 10
ARI 71-91 29
SDP 61-101 39
These seem to me to be intuitively reasonable results, though I haven't tested it on past seasons for accuracy (yet). I have used this method for the past several years, and I suppose I should go back and see how they fared at various points in hte season.
As to this year, by these projections Phillies finish with the largest lead (7 games over the Marlins, the only team in the division less than 10 games out), but clearly they do not have the best record. In fact, the Rockies and Rangers finish percentage points ahead of the Phils (but all three teams round to 90-72), so the Phillies aren't even one of the top eight teams. The Rays, along with the aforementioned Rockies and Rangers, are deserving but left out of the playoff picture due to playing in the "wrong" division or league. In addition to the Phillies, the Tigers and Cardinls benefit from being in weak (competitive?) divisions and get their post-season tickets punched despite having worse records.
July 15th, 2009 at 8:36 pm
Note, btw, that because of the way I set up my spreadsheet, the infamous WAS-HOU game suspended May 5 and completed July 9 is weighted as game 83 (WAS) or 84 (CIN) when it was completed on July 9, not game 25 (WAS) or 27 (CIN) as the "official" record would indicate. All games in between were accordingly weighted by one less than their "official" game number. (I just checked, and it doesn't make a significant difference, the number of wins and losses comes out the same, only the WL% is slightly changed.)
This also points out that because the same game can be a different game number for each team, it can have a different weighting for each team. That, combined with rounding errors, can lead the final total wins to not equal the final total losses. (Currently, the league W-L totals are 2431-2429).