April Hitting Stats Mean Nothing… Except When They Kinda Do

As part of my exhausting shtick, I like to respond “April!” to questions in my chats involving player performances in the season’s early going. This is effective shorthand when someone wants to know if, say, George Springer is a bust because he’s put up a .480 OPS in his first two weeks in the majors. It’s also dead wrong. April stats, in their proper context, are meaningful.

“But Dan, a few weeks of baseball is a tiny sample!” That’s correct, but you have to take into consideration the underlying reasons projections can prove to be inaccurate. It’s not just that things change, though they do — pitcher X learns a sweet knuckle-curve or batter Y realizes that not hitting everything into the ground might be good — it’s that it’s challenging to gauge where players stand in the first place. Players’ stats themselves aren’t even perfect at this. Tim Anderson hit .322 in 2020, but that doesn’t actually mean his mean batting average projection should have been .322. We don’t actually know if a theoretical player was “truly” a .322 hitter, a .312 hitter who got lucky, a very unlucky .342 hitter, or a .252 hitter who made a deal with a supernatural or extraterrestrial entity. A .300 hitter isn’t observed, they’re inferred.

The way most, if not all, in-season projections (or any projections, really) function is by applying what we call Bayesian inference. We won’t get into a full-blown math class, but in essence, it simply means that we update our hypotheses to take new data into account. And for players, data comes in all the time: every pitch or swing of the bat is new information about a player. It’s valuable information, too, as only the last handful of seasons have much predictive value and recent performance is the most useful.

We only have about two weeks of baseball in the hopper, but projections for some players have already shifted by considerable margins. I’m using ZiPS here — it’s very convenient to use the projection system I own and run — but this will likely hold true for other projections that update in-season, though the magnitude of shift may vary:

Probably the most fun name in the improvement category is Yermín Mercedes of the Chicago White Sox. A solid hitter in the minors but one with a poor defensive reputation behind the plate, Mercedes wasn’t considered a top prospect (he was 15th on Eric Longenhagen’s White Sox list this year). But the injury to Eloy Jiménez opened up an opportunity for someone to seize the DH job, and Mercedes has. Starting seven of the team’s nine games at DH, he’s gotten off to a blistering start, going 15-for-28 with five extra-base hits. It’s easy for people to dismiss comic book numbers, and a 298 wRC+ definitely fits into that category. But it’s those outlandish performances that are most likely to suggest a performance improvement. A .300 batter hitting .320 for a week seems “normal,” but it’s also less likely to represent a change in abilities than the same player hitting .800 for a week.

In this case, ZiPS and Steamer both give Mercedes a wRC+ boost of 13 points based on just these seven games of data. It sounds like a lot, but that’s a product of our uncertainty in making projections. Coming into the season, ZiPS was only 50% sure that Mercedes would have a wRC+ between 77 and 102. Think of a hurricane being tracked in the Caribbean, where even the slightest change in direction or timing can be the difference between the storm running smack-dab into the Atlantic coast or churning (mostly) harmlessly out into empty waters. In Mercedes’ case, with only a single plate appearance a year ago, his seven games in 2020 make up roughly 10% of his rest-of-season projection.

While the weights ZiPS uses are derived from baseball history, sometimes recent examples are the most effective illustrations. So let’s go back to 2019 — the last season of normal length — and look at how the players with the biggest changes in their projections at this point in that season fared the rest of the year. I looked at the preseason OPS projections and the RoS OPS projections through April 8, 2019 and compared them to the actual results for the 233 players who qualified for our splits leaderboards. From April 9 to the end of the season, using the updated OPS projection rather than the preseason one improved the average absolute error from 68 points of OPS to 62 points. The in-season model “expects” to come closer to the actual result than the preseason projection about 57% of the time. In 2019, that was 55.8% (130 of 233).

But the larger question is: does the model actually work well for the outliers? Is it too generous to players like Mercedes and too miserly to players like Marcell Ozuna? Again, let’s look back at 2019’s largest projection changes at a similar point in the season.

In this case, ZiPS RoS did a little better than expected with the overachievers (closer on 13 of 20) and worse than expected with the underachievers (seven of 20). Rather than hitting on the 22-23 players the model expected, it only hit on 20 of the 40 outliers. However, if we expand our search to the top 20 positive and negative outliers from 2004-19, the RoS model was closer to what actually happened for 362 of the 640 players (56.6%). In other words, yes, these big leaps forward and backward for individual players have real meaning.

Yermín Mercedes will not finish the season hitting .500. He’s unlikely to get MVP votes or make people forget about Eloy Jiménez. But if you’re not a little more optimistic about him than you were two weeks ago, even after just seven games, you’re doing it wrong.