What is there to be skeptical about that part on trades? Seems obvious to me.

Who trades older, expensive, but still good players away in July? Teams that know they are out of the playoff run.

Who trades for good or great players when they’ll be free agents in 2 months? That would be the good teams.

Who trades older, expensive, but still good players away in July? Teams that know they are out of the playoff run.

Who trades for good or great players when they’ll be free agents in 2 months? That would be the good teams.

That blurb was about teams that were projected (pre-season) with fewer than 70 or more than 90 wins. I think the amount of talent that flows from/to those types of teams via in-season trades is not as big as people assume. Usually (I suspect), when these teams on either extreme side do even more extreme than projected, the effect of traded/traded-away players only explains a small portion of the deviance.

Redacted cuz I did some dumb math

I believe the trade scenario is a partial explanation. I would imagine the models might regress to the mean a bit more than they should. The regression’s may be the most accurate for individual players but somehow underestimate the strength and depth of the strongest teams?

One question is, when they say they uncer/over-estimate at the extremes, are we talking half a win or 5 wins? Half a win could easily be trade flows or it could easily be other sources of noise (playing time assumptions, minor-league breakouts, etc.) Or sure, a whole greater/lesser than the sum of its parts.

And I don't know what teams we're talking about. I don't know how the 2021 Nats projected ... but we do know they trades Scherzer, Turner and Schwarber. The 2022 Nats I'm pretty sure projected badly and they traded Soto and Bell. That right there is an awful lot of trade flow. The 2021 Cubs probably projected fine but did move Rizzo, Bryant and Baez. I assume the 2021 Pirates projected badly and while clearly not in the same league, they moved Frazier, Tyler Anderson, Clay Holmes and Richard Rodriguez. Over two months collectively those Pirates probably don't produce more than 1-1.5 wins but the Cub trio should produce at least 3 wins and the Nats trades maybe more than that. 3 stinko teams and 3 not good teams moving 1-3 wins each would produce sufficient flow to the top to explain a difference of a win or two.

Note although maybe it could have been phrased better, this doesn't necessarily require the flow to be directly from <70 to >90 teams just that the flow away from the bottom explains the win shortfall at the bottom and the flow into the top explains the win excess at the top. That might not be true but it doesn't matter where the Pirates players ended up as long as they left the <70 teams.

The trades are a part of it, but so is the "let's try some young guys, might as well get next season going right now" part of it, too.

Also, I'm not dinging BP's efforts, but one can understand how you read this:

In other words, if we say that we think your team is going to win 85.5 games, that means that half the time, they’ll win between 79 and 92 games

and be like, "No s**t." Half the time a team that you think is probably a little better than .500 will win between 79 and 92 games? Did we need a ton of analysis to make that statement? I mean, half the time they will win less than 79 games, or more than 92 games.

isnt the error bar here similar to that of pythagorean runs and its translation into wins/losses? or is pythagorean runs more predictive?

Half a win could easily be trade flows or it could easily be other sources of noise (playing time assumptions, minor-league breakouts, etc.)

Hi Walt: maybe Im being a little too picky, but one would not expect minor league breakouts to occur more often with high win teams and less often with low win teams. It should be more or less random wouldn't it? Playing time assumptions maybe you are right, but wouldn't FG also be able to account for that and say rookie X on the LAD is not gonna play much, and rookie Z on the PIT is going to play more. Maybe they build that in or maybe not?

and be like, "No s**t." Half the time a team that you think is probably a little better than .500 will win between 79 and 92 games? Did we need a ton of analysis to make that statement? I mean, half the time they will win less than 79 games, or more than 92 games.

I agree the math here seems problematical. Your numbers here seem a little off so I'll post the quote from the article:

More specifically, 88% of our projections get within 15 wins of a team’s actual total. Only 7% of teams in our entire sample outperformed by 15 or more wins.

Last year 63% of MLB teams finished between 66 and 96 wins (19 teams). So you could put PUnxatawney Phil up there, claim every team is going 81-81 and get 63%. So every prognosticator essentially gets 63% for writing their name. The added value of FG is analysis is they get 7 1/2 more teams in the right pigeon hole than random guessing. Is that a lot? Maybe it is.

Also why 88 + 7 does not equal 100%. Shouldnt the numbers in the quote add up to 100?

Also why 88 + 7 does not equal 100%. Shouldnt the numbers in the quote add up to 100?

Probably 5 percent underperform by 15 or more wins.

yeah you're right my bad

More specifically, 88% of our projections get within 15 wins of a team’s actual total. Only 7% of teams in our entire sample outperformed by 15 or more wins.

I remember the first (and only) time I played the Century Club game in college, where you have to drink one ounce of beer every minute for 100 minutes. It sounds pretty doable, and for the first 20 minutes, it's not too hard. Then, you get about 35 minutes in, and you realize you are in the process of drinking over eight cans of beer in an hour and 40 minutes. By the end of it, you are in bad shape. Ergo, I played it only one time.

Maybe trying to guess every team's win total, and get them all with 15 in either direction, is a lot tougher than it sounds. Consider:

1) What is the highest number of wins a team will have in 2023? 110? If you wanted to cover a handful of the best teams in case they have a historically good regular season, then guess they'll win 95. That way, you're covered for anything from 80 to 110 wins.

2) What is the lowest number of wins a team will have in 2023? 50? If you want to cover a handful of the worst teams in case they have a historically poor season, then guess they'll win 65. That way, you're covered for anything from 50 to 80 wins.

3) Most teams are, of course, in neither bucket - but then again, if you wager "81 wins" for most teams, you'll be covered for 66-96 wins.

I think it should be pretty easy to get all 30 right, or darned close - but maybe it is like the Century Club, and is much tougher than it looks.

If somebody gave you 5-to-1 odds - you put $200 in, and if you get all 30 teams within 15 wins of your guess for them, you get $1000, but if you get even one wrong, you lose the wager - would you take that bet? Or would the odds need to be longer?

I remember the first (and only) time I played the Century Club game in college, where you have to drink one ounce of beer every minute for 100 minutes. It sounds pretty doable, and for the first 20 minutes, it's not too hard. Then, you get about 35 minutes in, and you realize you are in the process of drinking over eight cans of beer in an hour and 40 minutes. By the end of it, you are in bad shape. Ergo, I played it only one time.

Somewhat similar is the "give me a penny and then double the amount the next day, and then so on for 30 days", and you end up with over $10M on day 30 (plus all the money accumulated in the previous 29 days).

but maybe it is like the Century Club, and is much tougher than it looks.

I think it is like that, yes.

#8: Sure but maybe there is some signal there. Again, are we talking about them missing by 0.5 wins on average or missing by 5 wins on average? Maybe good teams are less patient with under-performing vets than bad teams (this is probably true) and reduce their playing time. Presumably bad teams are more likely to take that kid with a hot start from AA, projected to 75 PAs at season's start, promote him early and watch him flop even worse than the guy he replaced. Maybe at age 26 or whatever, Max Muncy and Connor Joe project about the same but the Dodgers only give Muncy 450 PA if he actually hits a 120 OPS+ while the Rox give Joe 450 PAs on a 88 OPS+ or a 120 OPS+. Those things will mostly even out but I won't guarantee there's not a half-win of signal in there.

Good teams have options (they have more talent) and if their GMs and managers and analysts are a cut above (as evidenced by being a good team) they might make better decisions. Bad teams pretty much play the hand they're dealt, certainly won't paper over injuries/flops by acquiring ML talent. It may be more a variance thing than a bias thing in the end ... a "genuinely" good team probably should only very rarely underperform by 10+ wins because they are not just gonna sit there and watch their season collapse and they have things to try. If the Pirates underperform their 65-win projection, they don't do anything about it really ... while early overperformance won't likely change their intention to trade their decent players at the deadline.

In short, you have a lot of options if you go into the season with Muncy, Hernandez and Taylor with Lux waiting in the minors. If somebody flops, you shift the playing time to the others and don't miss a beat. That won't improve your performance over your projection but it should go a long way to keep you from underperforming badly. When you've got Adam Frazier and three replacement level players, you're in a very different situation if Frazier flops.

The whole greater than the sum of its parts might be operationalized as something like:

Muncy, Herndandez, Taylor and Lux are projected to a total of about 1400 PA and 7 WAR. But if Muncy performs at 1 WAR below his projection, the Dodgers (a) have the luxury of deciding if maybe Muncy isn't quite as good as they thought and (b) shifting PT to the other guys so that the end result is 1400 PA of 6.5 WAR -- i.e. Muncy isn't given the PT to hurt the team by a full win, only half a win. Similarly if Mancy overperforms by 1 WAR, they can shift PT to him -- in theory make him full time and get 3.5 WAR out of him and pick up a full 1.5 wins. So if Muncy flops, they lose only half a win; if he excels they pick up 1.25 wins (to be conservative).

Adam Frazier and three (or 8) replacement-level players project to about 1400 PA and 2.5 WAR ... Frazier and Muncy might project the same, say as 2.5 WAR players in a full-time role. They are equally likely to over/under-perform by a win. But while Muncy is projected to just 450 PA and 2 WAR, Frazier is projected to a full 650. A difference is that if Frazier performs 1 win below expectation -- well he's still the best option the Pirates have. So he still gets 650 PA at 1.5 WAR and the Pirates a full win worse. If Frazier over-performs, they can't give him more PT and pick up just the 1 win.

Don't take any of that math nor projections seriously, it's just meant conceptually -- genuinely good teams obviously have better Plan As but they also tend to have better Plan Bs and will make trades if necessary for Plan C. Genuinely bad teams have bad Plan As and they aren't really deviating from that. A team projection picks up all of the difference in Plan A (ideally) and certainly some and maybe most of the difference in Plan B. I suspect it doesn't pick up all of the difference in Plan B (which may be Plans B1 through B10 depending on what goes wrong) and there's really no way it can pick up the difference in Plan C. (Possibly they could just use a long-term average like "good teams usually pick up 1 win over 2 months of talent at the deadline" or similar.)

The fundamental problem is one that MGL used to regularly point out. God him/her/it/themself could tell you that a player is a true 350 OBP hitter and that he will get 600 PA in the upcoming season and that, on average, he will face an average pitcher. But God still lets randomness run. That player will, on average, reach base 210 times but the standard deviation on that is about 11.5. So that produces a 95% confidence interval of +/- 23 times meaning that even if we know a player is a true 350 OBP and will get exactly 600 PA, the error bars around that 350 "projection" (it's the god-given truth) is about 40 points in each direction -- 310 to 390.

That's not very useful and, even worse, God doesn't speak to us, that 350 is based on the last 3-4 years of data only. So there are error bars around that 350 OBP too, maybe a standard error of about 15 points. I'll save you the boring stat detail (and save myself the work of doing it right) and guesstimate that when we put thsoe two bits of info together that we end up with a 95% interval of about 300 to 400. Well, no duh. An implication of the second bit (15 point error around OBP projections which I made up) is that two player projections would have to be about 40 points apart before we would conclude with 95% confidence that the player with the higher projection actually was better.

For teams -- a 500 team playing, on average a 500 team, 162 times is (give or take) 162 coin flips which has a 95% confidence interval of about +/- 13 wins. (This has been roughly mentioned already.) But again, that's if God told you they are a true 500 team that will, on average, play an average team. Add in the uncertainty of our projection they are a 500 team, add in the uncertainty around injuries, trades, flops, breakouts, etc. and obviously it gets worse. But even with perfect knowledge, the range is 68 to 94 wins for a 500 team -- I'm pretty sure there have been seasons where every team was in that range ... and, on average, if every team was really 500, we'd expect 1.5 teams to fall outside that range every season. 2022, when there were 11 teams outside that range, is one of the few seasons when we can conclude there were clearly some major differences among teams.

So I sympathize with the projectors from the statistical perspective that a precise projection is impossible. On the other hand, in this case, I think the proper course of action is to admit the projections are useless and stop doing them. I mean "half the time your team will be within +/- 6.5 wins" is not gonna do you a lot of good with the over/under bet at the sportsbook. At the very least, reduce projections to something like top 5, bottom 5 and the rest.

I think most baseball fans who would approach the challenge with some thought and a plan could crack at least 70%. A simple plan like grouping teams into 3 piles of I think they'll better than average, worse than average, don't know what to expect. Predict something like 88 wins for better than average teams pile, 81 for don't know what to expect pile, and something like 76 for below average pile.

If I read those charts correctly, they project the Athletics to score 3.80 Runs/game, with the only other team projected to be under 4 the Tigers at 3.97. Even still, they project the A's to win 70 games, 10 more than last year, and hence to be better than the Reds, Nationals, and Rockies, though last in the American League - I guess since the Washington Senators aren't around anymore.

There is one monkey wrench that I want to throw in here. There are teams (I am painfully aware of this because the Cardinals are one of them) who deliberately try to manipulate their talent and payroll to 1) be good enough to make the playoffs, but 2) not so good that they make the World Series. The Cardinals, to use the team I know best, know perfectly well that, in order to make a run at the WS, they need a lefty cleanup hitter to put in the middle of their righty power core and also a proven ace starter. Josh Bell would have been a perfect sign for them, but they told him that he would have to play himself into a full-time role, as if they had any other lefty cleanup candidates (they have lots of righties, including hot prospect Jordan Walker). And their response to signing an ace starter was the same as it has been for years: They cost too much, since salaries have gone up this offseason (as if salaries have ever gone down since 1916 or so). So, they will make the playoffs, but they are too vulnerable to a platoon of righty relievers and ace starters opposing them, to win any short playoffs. I have no idea how a mathematical system for projecting wins deals with this or if it deals with it at all. just a monkey wrench. This concept was first expressed by Branch Rickey, back in the 1930s. Rickey said that, if your goal is to maximize profit, you should try to finish second. The fans will keep coming, because you are SO close, and you won't have to pay the stars what it would take to actually win. Cynical, but financially true.

There is one monkey wrench that I want to throw in here. There are teams (I am painfully aware of this because the Cardinals are one of them) who deliberately try to manipulate their talent and payroll to 1) be good enough to make the playoffs, but 2) not so good that they make the World Series

This is not a thing, for several pretty obvious reasons.

Indeed. There's no absolute proven blueprint to get to the World Series other than be in the AL in 1939 and have all the talent.

#7 Several other issues first. Assuming you get playing time perfect with no trades or injuries, every single projection has a standard error in the general range of 14 runs per full time player. So that adds to an extra level of uncertainty in the runs scored and allowed.

Then of course there's the injuries and trades.

All in all, team win projections are noisy.

in this case, I think the proper course of action is to admit the projections are useless and stop doing them. I mean "half the time your team will be within +/- 6.5 wins" is not gonna do you a lot of good with the over/under bet at the sportsbook. At the very least, reduce projections to something like top 5, bottom 5 and the rest.

That's my feeling as well.

I'm trying to apply the same thinking to playoff odds . . . if those are a better reflection because they look at relative levels of talent/divisional comparisons, or if they're simply a function of w/l projections and thus just as subject to massive uncertainty as to make them relatively equivalent to basic best-guess thinking . . .

#23 ... I will say that of course fans seems to like the projections and they generate reaction so, for those reasons, fangraphs et al should keep doing them.

On the idea of "playoff odds." Yes, those should be more stable in a sense. That is similar to the idea of projecting teams into "clearly above-average" "average-ish" and "clearly below-average" buckets which obvsiously gives you a lot more wiggle room than "89 wins." The Dodgers are very, very likely to make the playoffs -- and you've saved yourself any concern whether they do so with 88 or 111 wins.

An interesting project I was involved in addressed analysing data that has been probabilistically matched. That's where we have two datasets covering the same population but we don't have a common ID. So, based on common variables between the dataset, you take informed guesses and link observation A in Dataset 1 to observation B in D2. This is a source of uncertainty and bias -- you can kinda correct for the bias. Anyway, the whole issue is a lot easier when the variable you are trying to predict is categorical because, even when A shouldn't be linked to B, there's a good chance B has the same value on the outcome as A.

First of all this is a really fun thread with some really interesting posts from all the usual subjects. So thanks everyone. I have a question and Im totally not sure what the answer is but since some of you are really good with stats I will try to explain. Something Walt mentioned in his last post got me to thinking there might be some more "stable" or more predictable aspects of the pennant race. Not simply the "clearly above average" bucket etc. but something similar. I dont know quite how to explain but I will try...

Lets say we take the LAD and we project them to 92 wins but its plus/minus 12 wins or so. Ok lots of room for error there as we all seem to agree upon that.

But the LAD are not generating their record in a vacuum. Their final record is related to the record of Padres, as well as SFG as well as every team in the league in some way. And of course every team's record is somewhat related to every other teams record. So UNLIKE the projection of a single ball player these projections are interrelated. In the case of one ball player he could have a great year or a bad year and it really wont change how the rest of the players did that year (not by much anyways).

BUt some leagues/years the disparity from top to bottom is skewed more than others. Take the 1910 NL (Cubs won: .675) or 1915 AL (BOS .669) those leagues had a large talent disparity from top to bottom. Versus say 1975 NL East (PIT .571) a very closely matched division from top to bottom MON in last place only 17.5 games behind. The next year the PHI won 102 games and MON was 47 GB. Similar teams only one year apart and a much greater disparity.

So lets try the following thought experiment. We drop the 2023 LAd into various leagues in various years. On paper today lets say the LAD project to go 95-67. But we drop them into say the NL of 1910 or the NL east of 1975. Now of course Mookie Betts doesnt hit 35 HRs in the NL of 1910, and Bellinger doesnt steal 14 bases in 1910 and Julio Urias doesnt pitch 175 inn. he pitches like 285 inn. and has 15 CG. Right? So we dont project their actual numbers into 1910 but rather their abilities. Betts has power, he hits maybe 10 HRs, Bellinger is fast he steals 60 bases etc.

So we drop the LAD into 1910 and they win 105 games. And we drop them into 1975 NL east and they win 92 games. Exact same team, with the exact same relative ability to the league, but the with wildly divergent final records. You could have a pennant winner win 108 games in a skewed league and you could have another winner win 92 games in a really tight league.

So right there we could make these projections tighter. Or could we?

Could we cancel out that "noise" and derive projections much closer to the final actual records of these teams? Say we have a contest where everyone projects all their team records, but at the end of the season we do an adjustment for "league skew" the disparity mentioned above. Could be get more accurate projections?

SO say someome projections BOS to be 85-77, but at the end of the season, the league was more skewed than usual, and we "correct" that projection to be 88-74. COuld we get better projections?

But many of you are way better at statistics than I am. Is there a name for this sort of thing? and can it be done?

Figured I'd bring up the method Bill James used to use before projections became reasonably good.


Just to summarize, the key factors were:

Teams that outperformed their pythag tended to decline (and vice-versa) -- Also, teams that scored more runs than you'd expect given their base stats tend to decline.

Teams that had a big single year improvement tend to regress (and vice versa)

Teams tend to regress to .500

Young teams tend to improve. Old teams tend to decline.

Teams with strong AAA tend to improve (this is one are where the current form of projections clearly helps -- It's not so much whether the AAA team is good but whether there are impact players in the minors that matters)

Late season performance. Teams that play significantly better in the second half tend to improve.

He'd actually just keep track of how many of these indicators apply to any given team.

Hardball times article

Figured I'd bring up the method Bill James used to use before projections became reasonably good.


Just to summarize, the key factors were...

God knows, he couldnt have done it with win shares.

Versus say 1975 NL East (PIT .571) a very closely matched division from top to bottom MON in last place only 17.5 games behind.

two years earlier, the Mets won the same division with a .509 winning percentage (82-79).
the 5th-place Cubs finished 5 games out and the 6th/last-place Expos came up only 11.5 games short.

the Mets were in last place as late as Aug. 30 (61-71), and were in 4th place on Sept. 18 (74-77) with only 10 games remaining. they then went 8-2, and everything else just magically fell into place.

the Cubs were in 5th but only 2.5 games out when the Mets had only 8 games left.

at 77-77, the Mets suddenly were in 1st place just 3 days after being in 4th. on the final weekend, the Mets and Cubs were rained out on Friday and on Saturday, so they played a doubleheader on Sunday. the Mets lost the first game to fall to 80-79 - so on the scheduled final day of the season, they still had yet to clinch a winning record.

but they won the nightcap as well as another makeup game on Monday to reach that 82-79 mark.

then they faced the Big Red Machine dynasty and beat them to reach the World Series and then took the Oakland A's dynasty (in the middle of back-to-back-to-back World Series titles) the full 7 games because - well, that's baseball.

The 73 Mets are fascinating. Even though game 7 was not very competitive, they almost won the world series before that. Losing game 1 despite playing well was probably the biggest factor. Cleon Jones throwing up in LF, Staub underhanding the ball in RF, their OF was decimated.

#27 Predates Win Shares by quite a few years. When he first came up with these methods he hadn't really finished his offensive methods.