Inside the projections process

Mookie Betts may have great talent, but playing time issues can make his stats hard to project. David Butler II/USA TODAY Sports

Predictions -- not promises.

Essentially, that's what projections are. If you've played fantasy baseball for any length of time, you're familiar with them, presumably have used them in your draft preparation. As Todd Zola, who has a heavy hand in generating ESPN's projections, astutely pointed out, projections provide the very foundation for rankings. If you've ever crafted rankings, you've used projections; you don't need a fancy spreadsheet to set expectations of player performance.

But when it comes to very projections themselves, the numbers you see published in our Draft Kit or by many other sources, are you using them correctly?

To help answer, I asked Zola and Dan Szymborski, ESPN Insider and creator of the ZiPS projection system, to share their insight into the projections process.

"Projections help cut down on an overwhelming amount of player data now readily available," Szymborski said. "While there are always things computer projections won't do well, one thing computers do manage well is a dispassionate, 'just-the-facts' guess."

And there's the projections upshot: They are emotionless estimates.

They give you an initial impression of the player's projected, upcoming-season value, and they do it without a driven opinion of that player. Then it's up to you to make two determinations: (1) Do you agree with the projection itself as the player's most probable season outcome, which naturally you should if they're your projections, and (2) how wide is the player's range of potential season outcomes relative to that projection?

Project on your own...

One could craft his or her own set of projections, depending upon the time he or she has available. I've made reference many times to "my projections," as I do believe in the importance of having one's own, uninfluenced set, but at the same time, mine are rough calculations using a system similar to CAIRO's -- we'll get to what that projection system is in a moment -- generated in considerably less time and with less complexity in its formulas.

Think: Manageable, in your terms.

Perhaps you've got time to develop your own, detailed projection system. Or, perhaps you don't; that's why we provide ESPN projections to spare you the time. Either way, a smart projection system should use three-plus years of data weighting recent performance more heavily, address factors like aging and ballparks, while adjusting for outliers caused by "lucky outcomes" -- things like Danny Santana's absurdly high .409 BABIP.

But the most important thing to keep in mind: The more manual adjustment than computer generation involved in the process, the more your personal biases might influence the results. I manually adjust my projections, but I'm perfectly prepared to accept that personal biases might explain some of my more -- ahem -- unexpected rankings that result.

... or pick your projection system?

There are as many as seven other significant projection models out there in addition to ESPN's projections. Each one uses slightly different methodology, explained in brief below:

Marcel: Developed by Tom Tango, Marcel examines a player's performance over the past three seasons, weighting more recent seasons more heavily, while regressing performance toward the league mean and applying an age adjustment to generate a season forecast, both of those at equal rates per player. Major League Equivalencies (MLEs) -- translations of a player's minor league stats into major league terms -- aren't included.

Bill James: Created by Baseball Info Solutions and published annually in the back of the Bill James Handbook, these utilize a formula that is proprietary, but which includes three to eight seasons' worth of data (the strongest emphasis on the past three) plus age, ballpark and playing time calculations.

Oliver: Created by Brian Cartwright and published by The Hardball Times since 2010, Oliver uses weighted averages of the past three seasons, adjusting for age and regression. Oliver's twist, however, is a different method of calculating minor league equivalencies (MLEs): It directly compares each player's minor to major league performance, rather than using a chaining method (Class A to Double-A, Double-A to Triple-A, Triple-A to MLB).

CAIRO: Developed by the website "Revenge of the RLYW [Replacement Level Yankees Website]," CAIRO is a projection system that utilizes the Marcel model but makes several adjustments. CAIRO utilizes four seasons of data rather than three, adjusts for ballpark, utilizes minor league equivalencies (MLEs), adjusts for defense and weights regression toward league average at differing rates depending upon the player's age and primary position.

PECOTA: An acronym for Player Empirical Comparison and Optimization Test Algorithm and one that draws from the name of former utility infielder Bill Pecota, PECOTA is a sabermetrically based projection system developed by Nate Silver and published by Baseball Prospectus since 2003. It utilizes "similarity scores," comparing each player's statistics to a historical statistics database, using the career paths of the most similar players to generate a forecast for that player's future.

Steamer: Created by Jared Cross, Dash Davidson and Peter Rosenbloom in 2008, Steamer utilizes a system of weighted averages to generate a player forecast, following similar methods to the other projection systems, though its exact formula isn't publicly shared.

And last, but certainly not least ...

ZiPS: It's not a hipster's juice box

... we arrive at ZiPS, Szymborski's projection model that is annually -- as well as daily in-season -- available on FanGraphs.com.

"The idea of ZiPS came from conversations I had with now-SABR [Society of American Baseball Research] board member Chris Dial during the late 1990s," Szymborski said. "We got the idea of making an open projection system that would get us most of the way to the very secretive projection systems around at the time that you mostly saw in fantasy magazines or behind pay walls, but would be open to the public. This is kind of what Tom Tango did later with Marcel.

"We never did anything with it at the time, but as I started blogging about transactions in 2001, it made sense to project players.

"At the time, Voros McCracken's DIPS was a fairly new concept that hadn't really been dealt with by the projections that existed, so I wanted a system that utilized some of this new research. I wanted to rhyme it with DIPS in Voros' honor. I called it SIPS at first, 'S' for my last name, but that sounds like a juice box for hipsters. So I used the second letter, and the first sound, in my last name and it became ZIPS. But since I was a fan of CHiPs as a kid, I called it ZiPs. Then I mistakenly typed it as ZiPS when I introduced it and, since Jay Jaffe had already plugged it, I left it that way."

ZiPS, like many of the other projection systems, uses weighted averages of the previous four seasons, but uses a PECOTA-like comparison system to help create an age adjustment. With a database that now extends back into the early 1970s, it aims to compare Ben Revere types to Ben Revere and Adam Dunn types to Adam Dunn, using techniques that fall under the umbrella of cluster analysis.

"Mookie Betts' performance of the last few years was compared by ZiPS to the history of all players around the same age," Szymborski said. "Every time ZiPS is run, it compares the desired player to tens of thousands of two-, three- or four-year chunks of historical performance, and ranks all the players at every age compared to Betts."

Who was Betts' top comp? It's a fun one: Andrew McCutchen.

"That, of course, doesn't mean that Betts will be McCutchen, just that his prior experience at similar ages was the most similar to Betts's prior experience," Szymborski said. "Betts' top comps after McCutchen were Mark Carreon, Shannon Stewart, B.J. Upton, Chad Curtis, Ellis Burks, Bernard Gilkey, Nate McLouth, Keith Mitchell, Vernon Wells, Gabe Kapler, Bernie Williams, Milton Bradley, Matt Lawton, Richard Hidalgo, Chris James, Phil Lombardi and so on.

"While comp No. 1 isn't that much bigger part of the model than, say, any of 10th to 20th comps, it's fun to look at the names. People relate to baseball in the context of what they've seen before. And it's fun to look up players that one may not be familiar with. Brad Miller's top comp, for example, was Dick McAuliffe. A lot of people may not have heard of him, but if a few people look up McAuliffe out of curiosity, I think that's pretty cool."

The playing time conundrum

One of the aspects that seems to most trip up computer projections systems is playing time, though it's perhaps the most integral ingredient in fantasy baseball player valuation.

"Playing time is an overlooked element that is probably more responsible for differences of opinions than skills analysis," Zola said.

Liberal playing time estimates are by design in most projection systems. For example, as of Feb. 19, Steamer forecasted non-pitchers to accumulate 185,253 total plate appearances and pitchers to amass 47,141 innings pitched in 2015; but non-pitchers totaled only 178,436 PAs and pitchers 43,613 2/3 innings in 2014. This "over-project" strategy, though, is sensible.

"A computer projection system can answer the question 'How good is Kris Bryant?' a ton better than 'How will the Cubs use Kris Bryant in the majors?'" Szymborski said.

"This also allows me to project more players. If ZiPS only cared about playing time that would actually happen, then it's giving less information to a baseball fan. What's more useful to a fan, a projection of Julio Urias of a 0-0 record and 0.00 ERA, or a projection that shows where he stands in terms of ability?"

Szymborski does, however, craft team-based projections from ZiPS, at which point he must make his own playing-time assumptions -- and that is a key takeaway, that it's a manual rather than automated process. Human error then comes into play; ZiPS projected the Texas Rangers to win just shy of 90 games last season based upon Szymborski's playing-time estimates, which would've mirrored those of anyone's in the industry. Think about it: Who predicted that the 2014 Rangers would lose 95 games?

"If I told ZiPS who actually played for the Rangers in 2014," said Szymborski, "It'd drop 20 games off their projected win total."

Whether a fantasy owner makes his/her own projections or uses another source's, it's a worthwhile task to make his/her own playing-time forecasts, as Szymborski has, while recognizing the influence of human error on the process. For instance, if you're using the ESPN projections as your baseline, why not collect them, but manually adjust -- scaling appropriately -- playing time to fit your own expectations?

For example, one reader recently asked why ESPN has Betts projected for only 430 at-bats, the answer being that the Boston Red Sox still have a crowded outfield that makes it difficult for any one individual to enjoy a full-time, at-bat projection. If you think Betts is more likely to receive 525 at-bats, however, it'd be smart to adjust our projection accordingly, then move him up in your rankings to reflect the change.

Imports: knowing less means guessing more

A particular challenge for projection systems is foreign imports, especially the recent influx of Cuban talent. This task is no easier for ZiPS; Szymborski admitted he grimaces whenever a player from Cuba signs in the States.

One of the reasons is the difficulty in obtaining detailed statistics from Cuba, which are based upon a shorter season -- and a split one in 2013 -- besides. Another is the limited sample of players who have made the move to the U.S., making translation of National Series statistics especially difficult. Players from Japan, and its Nippon Professional Baseball league, represent a larger -- and therefore more predictable -- sample.

And this season, we have another, entirely different player-import challenge: Jung Ho Kang, who signed with the Pittsburgh Pirates after playing nine seasons in the Korean Baseball Organization. Besides there being precious few examples of players to make the jump from KBO to Major League Baseball, there's this: The KBO saw anywhere from 42-45 percent greater home run production in 2014 than 2013, depending upon whether you're measuring totals or rates. How does one translate a league with that level of variance?

According to Szymborski, using examples of Korean players who subsequently played in Japan is one way; Japanese players have more data from which to estimate big league translations, so comparing performances between Japan and Korea helps generate a translation. ZiPS, says Szymborski, treats the KBO as the rough equivalent of a Double-A team in a hitter-friendly environment in the U.S.

Cuba, meanwhile, is regarded by ZiPS roughly equal to a high Class A league, though like Korea, that's based upon a relatively small sample.

As for Japan, ZiPS estimates the NPB as similar to that of a U.S. Triple-A league, if not slightly greater in talent than Triple-A. Szymborski noted that it's certain statistical categories themselves that translate differently from Japan: Batting average has translated almost identically to big league production, but power tends to have translated even more poorly than it would have for a player making the jump from Triple-A ball to the majors.

"That's why Ichiro translated really well," Szymborski said, "because he wasn't reliant upon power."

Former Seattle Mariners and Cincinnati Reds outfielder Wladimir Balentien represented the opposite end of the spectrum, a power-oriented hitter who had too many holes in his swing to thrive in the majors. Balentien hit 15 home runs in 511 at-bats during a three-year span in the majors (2007-09), but exploded for at least 30 home runs in each of his four seasons in the NPO, capped by a 60-homer-in-439-at-bat campaign in 2013.

Still, don't tell ZiPS it can't take its best shot translating such players' foreign-league statistics to generate upcoming-year projections. Here are those 2015 projections for the imports, Yasmany Tomas and Kang:

Tomas: .267/.302/.464, 520 PA, 21 HR, 60 RBI, 1.9 WAR
Kang: .230/.299/.389, 502 PA, 14 HR, 57 RBI, 1.5 WAR

Incidentally, Jordy Mercer, Kang's primary competition for the Pittsburgh Pirates' starting shortstop job, was projected for more WAR (1.9) and an OPS six points higher (.694-.688). And that got Szymborski thinking ...

"Someone made the joke, 'I voted for Jordy Mercer,' " said Szymborski, referencing "The Simpsons" episode "Treehouse of Horror VII." "So that's going to be one of my fantasy team names this season."

Fun with comparisons

Fantasy owners sometimes become fixated upon player comparisons, or "comps." You'll see the questions in our chats sometimes, and colleague Eric Karabell and I have even addressed the matter at times on the Fantasy Focus Baseball podcast: "Who is 2015's Fill-In-Name-Who-Had-Breakout-2014?"

That player comps are an integral part of the ZiPS (and PECOTA) projections systems doesn't help; fantasy owners construe that as an opportunity to expect a player to replicate, in 2015, what a specific comp did in a past, historical season played at the same age. Ah, but there's a critical difference: Since projection systems eliminate player bias from the equation, they are using a mathematical approach to contrasting players' career paths.

If the aforementioned Betts-McCutchen comp didn't already excite you -- surely some now expect a .286-12-54, 22-steal campaign from Betts, mirroring McCutchen's age-22 stat line from 2009 -- Szymborski nominated Byron Buxton, the Minnesota Twins outfielder who placed second on Keith Law's top 100 prospects list, as another example that illustrates player bias.

ZiPS selected Carlos Gomez as Buxton's top comp entering 2015.

"That's just cruel to Twins fans," said Szymborski. "Mainly because he did succeed ... just not in Minnesota."

As Szymborski pointed out, perception influences the perception of Gomez's value, at least to Minnesota Twins fans. They might remember him more for his early-career failures, forgetting that Gomez was a 64-steal (2005), .345-OBP (2006) performer in the minors who earned top-100 rankings on every prospect rankings list. Drawing Gomez as a top comp might appear like a criticism; one could argue it's quite the compliment.

Comps for prospects tend to inspire the most player bias, especially on the pitching side. Fantasy owners always seek the "next Sandy Koufax," and Szymborski pointed out that many top pitching prospects draw Koufax as a comp at a young age, Jonathan Sanchez a prime such example.

But, considering the high rate of failure of pitching prospects, practically zero pitchers will ever become Koufax during their prime. Remember, Koufax himself had a sub-.500 record (36-40), a 4.10 ERA, 1.43 WHIP and 1.69 K-to-walk ratio in his major league career at the time of his 25th birthday. It wasn't until his final six seasons, during which time he won 82 more games than he lost, had a 2.19 ERA, 0.97 WHIP and 4.16 K-to-walk ratio, that Koufax really became the "Sandy Koufax" we remember, and that transformation represented about the most advantageous breakthrough a pitcher could ask.

"That's why I love Clayton Kershaw's projection [this year]," Szymborski said. "He's the first player ever in ZiPS to get real Sandy Koufax as his top comp."

Player comps need not always be scientific, either. Further characterizing the exercise as fun as much as science, Szymborski pointed out that ZiPS listed Marlo Thomas as Tyler Moore's and Tom Skerritt as Burt Reynolds' top comps. (Reynolds, for those unfamiliar, is not the actor but rather Robinson Cano's cousin, who played in the Seattle Mariners' farm system last season.)

Bold ZiPS statements

Speaking of fun, ZiPS can be used to make bold statements, depending upon how you're using it. Szymborski shared a few of the more interesting ZiPS findings of 2015, each of which might help better inform your player valuation:

The Washington Nationals' combined Nos. 6-10 starters project better than the combined Nos. 1-5 of several teams.

Bryant, according to Szymborski, might have the best projection ZiPS has ever given to a player who had no major league experience.

Kershaw isn't the only Los Angeles Dodgers pitcher with an inner-circle Hall of Famer as his top comp; prospect Urias has Randy Johnson as his. Urias' professional career sample might be small, making a projection for him more challenging, but ZiPS translated his 2014 as better than Dwight Gooden's 1983 season in Class A Lynchburg.

ZiPS projects 30 WAR remaining -- that's a remaining career total -- for Troy Tulowitzki despite never thinking he has a mean games projection of better than approximately 115.

Even with a big projection adjustment for Tommy John surgery, ZiPS still loves Jose Fernandez. One can only wonder how lofty Fernandez's projection might have been had he enjoyed a full, healthy 2014.

Wherefore art thou, range of outcomes?

Besides playing time, if there's a significant drawback to projections, it's that it provides a black-and-white answer to a question best answered in gray. Projections cannot account for random variance, but that is often the reason that an individual player projection winds up off the mark. This is why projections are wiser used as a starting point rather than your conclusion.

Let's use batting average to illustrate. From 2012-14, the top 10 qualified hitters by individual season combined for a .323 average, and there were four instances of a .333 batting average or better during that span. But from 2012-14 combined -- meaning the three-year average -- the top 10 hitters combined for a .317 average, and no individual batted better than .329. And if you're at all familiar with projections, you know that a three-year average is going to be a lot closer to the final number than a player's single year.

Taking that point to ZiPS, Szymborski said no specific individual is forecasted to bat .325 or better in 2015, but ZiPS anticipates an average of six players to bat .325, with Troy Tulowitzki (31 percent), Miguel Cabrera (26), Jose Altuve (25), Victor Martinez (20) and Mike Trout (15) the six with the best odds.

Tulowitzki, meanwhile, has a 9 percent chance of hitting .350. That's as good an answer as any for those who question how, in the ESPN rankings, we have Tulowitzki valued a second-rounder in spite of his injury history.

Shifting the focus to home runs, Giancarlo Stanton has a 0.4 percent chance of hitting 60, and he, Jose Abreu, Mike Trout and Chris Davis have the best odds of hitting 60 homers. Those might be especially poor odds, but they identify those four players as the ones with the most expansive power ceilings for 2015, even if the leader of the pack struggles to breeze far past 40.

Conversely, and in the interest of fun, the player with the worst odds of hitting 60 home runs was the aforementioned Burt Reynolds. Szymborski termed Reynolds' odds as "one in 60-with-300-and-something-zeroes, kind of like fewer years than the end of the universe."

"Nothing's impossible," Szymborski said. "It's just exceedingly improbable."