FG on Fox: The Myth of Second Half Momentum
Back at the end of July, there was one overarching narrative: the Tigers and A’s engaging in an arms race to try and get an advantage for their inevitable October match-up. On the day the Tigers acquired David Price and the A’s acquired Jon Lester, Oakland had the best record in baseball, while Detroit had a comfortable four game lead in the AL Central; since adding their respective aces, both have surrounded their divisional leads, with the A’s falling so far behind the Angels that their chances of winning the AL West now stand a fraction of one percent.
Toss in the quickly fading Milwaukee Brewers — 60-49 on July 31st, 14-22 since — and we have three strong first half contenders limping to the finish line. Like the A’s, the Brewers should basically punt any hopes they still hold about a division title, and are now hoping entirely for a surge that will put them back in the Wild Card race. But even if the A’s, Tigers, and Brewers manage to right their ships, should they really be optimistic about their chances in October? After all, they’ve spent the better part of the last few months playing lousy baseball, and at some point, doesn’t a second half slump become an indicator of underlying problems that would derail a playoff run even if the team managed to get to the postseason?
In a word: no.
Dave is the Managing Editor of FanGraphs.


The lazy sportswriter explanation for the A’s poor play has been that trading Cespedes annihilated team chemistry, leaving the team aimless and undesiring of victory.
Thank you, Dave Cameron, for talking me off the ledge!
Have to make the playoffs first no?
This is what makes baseball great, the ups and the downs. Analyze all the data you want, but baseball will always throw you a perfect curve. I love this game.
The InstaGraphs version of this article is three words: “2007 Colorado Rockies”
Tigers are 22-18 since July 31, not great but not “limping to the finish” either. They lost Sanchez and Soria to DL and Verlander missed a start while they were in the middle of playing 31 games in August. They have a very cushy schedule after trip to KC next week. KC has simply been hot. Oakland on the other hand has utterly collapsed to the point where they could miss playoffs completely. Milwaukee also has collapsed so bad they need to climb over some teams to get back in. Detroit’s position looks not so bad.
What I would love to see is this same study done but using first half numbers instead of second half numbers to see if there is any difference there.
Dave, your articles have been great for a while, but this is terrible. Of course there is a relationship between how a team plays in the second half and what their chances are in the post-season. In fact, the relationship is quite strong. There is also a relationship between how a team plays in the first half, for the whole season, and for the last month, although each of those samples would not necessarily have the same strength of relationship.
The reason you are not getting any results, especially a strong correlation, is simply because of the sample size of the post-season!
Imagine that we know exactly the true talent strength of every team going into the post-season (we are God), and we know the true talent of every opponent in those post-season games, such that we know exactly what every team’s chances of winning each of those post-season games is. What kind of correlation do you think we are going to get with those perfect team strengths and the results in the post season? You can actually do a simulation. Try it!
Very small! Simply because the post season sample size for each team is so small.
The other reason you are getting the results you get when you split the teams into buckets (good and bad second halves) is this:
Teams that have a bad second half and still make the playoffs will tend to have a good first half, because you need a good record to make the playoffs. So a bad second half is NOT an unbiased estimate of a team’s strength going into the season.
The bottom line is this: You CAN estimate a team’s chances in the playoffs, quite well in fact, if you know their seasonal record (you have to regress their seasonal record of course, towards .500 if you know nothing else about the team other than the W/L record for the season).
If you know only their second half record, you can do the same thing, although the certainty and accuracy of that estimation is not going to be as large as using the entire season record (probably weighting the second half more strongly is better because of personnel and playing time changes and actual changes in true talent), and you estimate the true talent of the post-season opponents (if you don’t know that or you don’t know who the opponents are, then you assume some opponent strength which is better than average of course).
Now, if by “little or no relationship between the second half record and the results of the post-season,” you mean the “actual results of the post-season” then your thesis is accurate. But that is obvious and not interesting. Nothing has a strong relationship with the actual results of 4 or 5 games (or 10). That is what Beane means by the post-season being a “crapshoot” (although he is way over-stating the case).
(You also have the issue of the number of games in the post-season being biased toward teams that win which also “screws up” the correlation.)
But, if you mean that we cannot infer with pretty good accuracy a team’s chances of winning in the post-season from their second half records, and I think that is the only reasonable and/or interesting interpretation of the question at hand, then you are completely wrong. You just aren’t going to get any results to support the relationship using the kind of analysis that you did.
If you think I am wrong, I will be more than happy to take the team with the best record in the second half of the season in every post-season series. Care to make that wager with me?
In fact, go back and look at the record in each playoff series in history and tell me the overall record of the teams with the better second half record.
Bah, me thinks you are overthinking this. In the playoffs, the team with the better record barely advances more often than the team with the worse record. The key goal is making the playoffs, followed by avoiding the wild card game if possible and go straight to the division series. This is way more important than being “hot” in the second half.
It has nothing to do with being hot or cold or momentum. Any subset of performance in the regular season tells us something about the strength of the team. The larger the subset, the more it tells us (and the less we have to regress towards .500 in order to estimate the true strength of the team).
Once we know the strength of the team going into the post-season then we simply use a log5 or odds ratio method to figure each team’s odds of winning a game or series.
One caveat is that a team’s regular season true W/L record (their “strength”) is not equal to its post season true talent because of changes in personnel and personnel usage and its use of starting pitchers and bullpen (i.e, the best 2 or 3 starters pitch the bulk of the post-season games).
However, if you think that the post-season is a “crapshoot” as Beane says, simply look at the Vegas post season lines for games and series. If it were a “crapshoot” you would see roughly even money lines for all games and series, adjusted for HFA. You don’t. You regularly see teams that are 3-2, 3-2, and more for games and series. So SOMEONE is able to estimate each team’s true talent in the post-season and there is a significant spread of that true talent.
No, you obviously don’t understand the way betting lines work. They aren’t designed to predict who WILL win,but rather to counter the changing ratios of money being wagered. If people were ALWAYS rational, those two things might correlate a little more. But people AREN’T always rational. They make bets with emotion and hope ,random chance…etc So, while the “favorite” (where all the money is going) may win more often than the underdog, the ratio of wagered money to odds of winning are not reflective of the % chance to actually win. Profit for the bookie is included, and the record of the teams involved are ONLY reflected in handicapping of the initial betting odds (which are subject to change as more money is wagered setting it’s own level)
TL,DR: I wish to stick to my pleasing narrative even though the data do not support it.
Actually, I think you’re kind of proving Dave’s point. Maybe he isn’t saying anything revolutionary, but he’s showing that being “hot” in the second half doesn’t mean you’ll stay “hot” in the playoffs (yes, partly because the playoffs are a crapshoot). Obviously the better true-talent team would be slightly favoured, but do second-half records indicate who is the better true-talent team? Not nearly as much as the full-season record (which still doesn’t tell the whole story)…..add a small sample to a smaller sample, and you have Dave’s miniscule correlation.
I propose a counter bet…I get the Tigers and Athletics, you get the Orioles and Royals. Most combined playoff wins this year wins the bet. How does that sound?
Haha, you surely don’t get to choose a team or teams once you know something else about them other than their second half records.
And yes, their second half records do give us a pretty good idea (not as much as the full season records, but not half as much either because the second half gets weighted more than the first half) of the true talent of each team. That is my whole point. Correlation tells us nothing in this case.
You are making my argument for me. You are saying that we can’t tell the true talent of the team by looking only at second half records. My point is that even if we KNEW every team’s exact true talent, if we correlate that with the record of a few out of sample games, that correlation will be exceedingly low because of the tiny sample size of the y variable.
Dave is claiming that BECAUSE the correlation is low that means that those second half records do not tell us much of anything. That is patently untrue. The low correlation tells us nothing about what the second half records suggest. It merely tells us that the random variance in the W/L of a few games is large. We already know that. (Well, it doesn’t really “tell” us that – the low correlation is a result of that and NOT a result of a poor signal in the second half records.)
Ok now I’m confused…you, Dave, and I all agree that second half records don’t really tell us much about the true talent of a team. We all agree that the playoffs are basically a crapshoot, where the better true-talent team has a slightly better chance of winning.
I understand Dave’s point to be that for these two reasons, second half records don’t really tell us anything in terms of who is more likely to be successful…and your point is that a team with a better second half record might have a slightly better chance of being successful because a half season sample tells us they might be a better true talent team?
Sounds like 12 of one and a dozen of the other…
I ran a correlation between true team WP in the “post-season” and that team’s W/L record in 5 games.
I assumed there were 5 teams in the post-season and that each team had a true talent of .54, .55, .56, .57, .58, and .59. These numbers are for illustration purposed only.
I also assumed each team played against a .500 team in the post-season which is obviously not a realistic scenario since each team would play the other teams and both teams would likely be in that .54 to .59 true talent range.
However, assuming that each of those 6 playoff teams played a .500 team in the post-season 5 times, we expect that team 1 would be have a .540 record and team 6 would have a .590 record. If we were setting “lines” for each game we would have team 1 a small favorite and team 6 almost a 3-2 favorite. So these games would not be “crapshoots.”
If I run a correlation (over thousands of post-seasons) between what we know for a fact is each team’s true talent and their result in 5 games against a .500 team, what do you think the “r” would be?
It is a lowly .056! Correlations do NOT tell us the relationship between two variables as David is suggesting in this article, such that because the correlation coefficient is low that means that a team’s W/L record in the second half or the whole season does not tell us what their chances are in a post-season games. They do tell us that to a large degree. To a very large degree. Even if I change the spread in talent to .5 to .6, which means that each post-season game has one team anywhere from even money to a 3-2 favorite or dog, the correlation for 5 games in the post season is only .144.
If I increase that post-season to 50 games, guess what happens, even though nothing changes? The teams still have the same spread in true talent, and each game still has a team favored by anywhere from even money to 3-2. You would not notice any difference in each games’ odds. The correlation goes to .444. 500 post-season games and we have .836 or near perfect correlation. 3 games, and we are back to .083 correlation even though nothing observable about each game and each team changes.
What the low correlation tells us is that one variable cannot explain much of the variance in the other variable. Well, of course not! That is because in 5 games (or 3 games of 1 game), there is going to be huge random variance because of the sample size which will overwhelm the variance associated with the spread of team talent.
As I indicated in my last post, this is a misuse of correlations and regressions.
somebody needs to get out and touch a boob.
This really is a bit catastrophic. MGL does an excellent job of showing there was never any chance of finding a strong correlation between second half winning percentage and postseason records no matter whether second half record is predictive or not.
It is also worth pointing out that the distribution of records in the postseason does not even remotely resemble the distribution of records from the regular season. In the postseason there are a very finite number of possible records (the data is discreet), and teams get removed from play when they have losing records. Every year 50% of playoff teams finish the playoffs with records of 0-3, 1-3, or 2-3! So, in addition to the sample size issue, there is also the issue that only teams that win series are allowed to keep playing, so the playoff records are not estimated from the same number of games for all the teams. The teams with the worst estimates (fewest games) are guaranteed to have losing playoff records because of the elimination bias.
So, Dave’s analysis was basically guaranteed to produce the answer he wanted, however the test was not appropriate to the question.
The question is “Is second half record more predictive than first half record or full season record?” Even a playoff super team that is expected to win 70% of its playoff games loses a 5 game series 16.3% of the time. So, given the small advantages that the best teams enjoy, and the wide variances in expected outcomes, Dave’s question is not difficult to test in theory, but in practice would require massive sample sizes (in terms of teams) to distinguish between the alternative hypothesis, since there is so much overlap between the predictions of each.
You’re wrong about what the “question is.” Dave is simply trying to show that predicting results or betting on teams in the playoffs based on a team’s late season results is a poor strategy. He is correct in that, though a larger sample size is required to prove it.
player rest, player fatigue, future injury risk:
Is there a difference in playoff performance of teams that win a division by 5+ games vs 1 game, or with a week or more remaining in the season vs the final day of the season vs teams that need an extra 1-game playoff.