The Z Files: Pitcher Streaming -- A Cautionary Tale

The best lineups to stream against were presented in the most recent Weekly Pitching Rankings. They were listed by handedness, with each team's year-to-date K%, BB%, HR% and wOBA included.

The implication is if you want strikeouts, stream against lineups with a high K%. If you need run prevention, pick on offenses with a low wOBA. If your pitcher is wild, use him against teams with a low BB% or if he's homer-prone, deploy against clubs with the lowest HR%.

Seems intuitive, no?

Unfortunately, I broke one of the cardinal rules of analysis. For that I apologize. As someone with an advanced science degree, doing so was even more egregious.

You see, I have no evidence that how a team has performed to date reflects current expectations. If the tables had been data from the last "X" weeks, the same would be true. I don't know the proper sample to use to gauge current expectations. Maybe it exists, and I'm just not aware. However, since this information would be the Holy Grail of daily fantasy, DFS and betting, I have to think if it were truly known, the research and data would be publicly accessible.

Making matters worse is, I posted the tables (and provided advice) knowing I can't validate its efficacy. The fact everyone does the same does not make it right.

The problem is, there has to be a reason why the proper time frame isn't public knowledge. At one extreme, everyone just assumes year-to-date stats are the best predictor and doesn't dig deeper. At the other, there is no correct answer.

My gut sense is the answer is somewhere in between. There may not be a perfect sample size, but there is likely a range in which the probability is highest. This has long been on my Toddy-do list, but something else has always taken greater priority. Unfortunately, this is more than an All-Star break project, so once again I'm planning on addressing it in the offseason.

The best I can do now is demonstrate why year-to-date metrics could be misleading. What follows is a team by team breakdown of plate appearances, wOBA, K%, BB%, HR% and BABIP against left-handed and right-handed pitching. The data is pulled through July 11. At that time, there was an average of 87 games played per team. Choosing to break the data into 29-game segments (one-third of 87 games) is arbitrary but serves to show the noise within the sample, and the perils of making unsubstantiated assumptions.

• Period 1: April 7 - May 9
• Period 2: May 10 - June 10
• Period 3: June 11 - July 11

PA versus RHP

PA versus LHP

Standard deviation is included to illustrate "games" may not be the best measure. Perhaps plate appearances is better. As can be witnessed via the standard deviation, the number of plate appearances within the game sample can vary by a lot.

The conundrum is, the proper sample is likely a balance between one large enough to flesh out biases, but short enough so the composition of the lineup hasn't changed enough to affect the team results. Complicating matters is teams accrue fewer plate appearances facing southpaws, so by the time a lineup achieves the requisite number of plate appearances against lefties, the lineup may be significantly different. That said, we're dealing with probabilities, not absolutes, so it will likely be the sample against LHP simply isn't as reliable as it is against RHP.

wOBA versus RHP

wOBA versus LHP

It's a crude measure, but the average standard deviation versus RHP (.019) is less than versus LHP (.028). This makes sense since there is more noise within a smaller sample.

Let's cherry-pick some teams with misleading results. For the first two periods, the Angels and Atlanta were pretty close in wOBA versus RHP. However, for the third period, the Angels were one of the best teams against which to stream right-handers, while Atlanta was one of the worst. After the first period, the Red Sox appeared meek facing southpaws, but they crushed them over the next 58 games.

My takeaway is rest-of-season expectations could be more important that actual team performance, given there is a relationship between the two as expectations are seasoned with what's been done. Further, and this is obvious but impossible to foresee without a crystal ball, judgements should be made on a game-by-game basis, fueled by that day's starting lineup.

Even so, the goal is probability, so we can still determine the optimal sample to decide if a pitcher should be streamed in a two-start week, even if we don't know the Tuesday and Sunday lineups he'll be facing.

Let's finish the story with the tables for the other metrics. BABIP is included to display luck is also a factor, not to mention a driving force for wOBA.

K% versus RHP

K% versus LHP

There is more variance than I expected with strikeouts. As such, one of the ancillary projects will be determining if the trends within team strikeouts are due to the ups and downs of individual players, or the in and out of players within the lineup. That is, how is K% affected by lineup composition?

HR% versus RHP

HR% versus LHP

I don't use HR% as much as wOBA and K% when identifying pitchers to stream, but there are occasions when the hurler in question's main crutch is homers, so there are scenarios in which using an iffy guy against a powerless lineup is a sage play.

BB% versus RHP

BB% versus LHP

Like HR%, I don't use a team's BB% a lot, but there are pitchers whose main issue is control, so knowing the impatient lineups can help.

BABIP versus RHP

BABIP versus LHP

Out of curiosity, I wanted to determine which metric correlates best with wOBA. Here are the results:

RHP Correlation to wOBA

LHP Correlation to wOBA

I'm not exactly sure how to apply this yet, but if I figured if I was curious, you may be as well.

I don't want to give the impression using data of this nature is a waste of time without "proof". That said, I do think it needs to presented without an implied level of certainty. This hobby is based on deciding what is most likely to happen, and this is just another example. Shoot, if we all KNEW what would happen, I'd still be making peptides.

However, there must be a sample size that yields the best answer; we just don't know what it is (yet).