About a week ago, I detailed my process for generating hitting projections. Now it’s time to do the same for the pitchers.
Like with hitting, skills are expressed as a rate stat. Hitting used plate appearances, as does pitching, though I’ll express it as per innings.
Using strikeout and walks as the example, K% and BB% are better than K/9 and BB/9 to get the true skill level. For those unaware, K% and BB% use batters faced (essentially plate appearances) as the denominator. It’s subtle, but K% is a better indicator than K/9. A pitcher allowing more runners faces more hitters, availing more chances to punch them out. Think of it this way. Two pitchers carry an identical 8.1 K/9, whiffing 180 in 200 innings. One faced 800 batters, akin to about a 1.10 WHIP while the other faced 844 hitters, equating to about a 1.30 WHIP. The first posted a 22.5 K% while the second registered a 21.3 K% mark. This is like two batters each garnering 160 hits, but one needing 550 AB (.291 average) while the other required 580 AB (.276 average). Which is the better hitter? Of course, the former. Well, the difference between .291 and .276 is the same as 22.5% and 21.3%.
The engine projects K% and BB%, but I also project batters faced per inning. K% and BB% can easily be converted to K/9 and BB/9. It’s easier for me to project innings when doing pitching playing time, so while technically I use K% and BB% in the projections, the final projection takes K/9 and BB/9 out to K and BB using innings.
Pitching projections utilize the same three-year stat spread and weighted average as hitting. Similarly, MLEs are used to fill in the blanks for prospects. Finally, composite park factors and aging adjustments are incorporated in the same manner.
A common theme with pitching will be regression, even more so than with hitters. The sample size of the different events associated with throwing a baseball is small, even for a workhorse starter. Outcomes don’t always sync with skills. Thus, almost all the components require regression to best frame what’s likely to happen.
Please keep in mind I’m a bit of a obstinate stickler with respect to the term regression. It’s come to mean “play worse” in the fantasy lexicon. In my Utopia, regression would have the specific meaning of correcting for outcomes out of the pitcher’s control. Admittedly, with improvement in data collection and analysis, we’re learning more and more about the proverbial luck versus skill delineation, but we can only go by what we know at the current time. My default level of regression is 50 percent. That is, the projected number is the average of expected and actual. I’ll then massage as appropriate, but always with a reason.
With that as a backdrop, let’s go through the four basal skills intrinsic to a pitcher’s projection: home runs, strikeouts, walks and hits. From there we’ll move onto the standard roto categories then hit some of the stats used in points-leagues scoring.
While hitters set their own home run per fly ball baseline (HR/FB), pitchers cluster around the league average. As you know, this is on the rise:
As an aside, hit types are still determined subjectively. Soon, they should be designated via objective criteria. Until then, some number, like HR/FB may differ between data sources. The key for analysis is using the same source for the research, of in this case, projections.
Based on the number of fly balls a pitcher allows, an expected number of homers can be determined. After some park neutralization, the actual and expected number of long balls are averaged. After being divided by batters faced, the park-neutral HR% is calculated. This will eventually be converted to HR/9 for the final projection.
Some elegant studies show strikeouts are proportional to Swinging Strike Rate (SwStr) with an influence of First Pitch Strike Rate (FpK). I have a formula using this data to generate an expected K%. Again, a park neutral K% is computed then regressed with actual K%.
The great research team at Baseball HQ demonstrated a similar relationship between the number of balls thrown and BB%. I have developed an expected BB%. You know the rest.
There’s a reason hits are discussed last. Similar to how I project hits for batter, I use batting average on balls in play (BABIP) for pitchers. As you can likely surmise, more specifically, expected BABIP. As discussed in the hitting essay, I have data breaking batted balls into multiple classifications: groundball, infield line drive, outfield line drive, fly ball, bunt and popup. All but bunt and popup are broken into soft, medium and hard hit. The league average for each subset is determined and they employed to calculate an expected BABIP. After the usual park neutralization, a park-neutral BABIP is determined and plugged into this formula to derive hits.
Hits = (AB – HR – K + SF) x BABIP + HR
There’s a couple of components needed not discussed yet, namely hit by pitch (HBP) and sacrifice fly (SF). They’re just a three-year weighted average like what was done with hitting.
So, now I have expected hits and actual hits, all that’s left is to regress, blah, blah, blah.
The neutral H/9 and BB/9 are treated with the aging and park adjustments to get projected hit and walk rates. This is a bit circular, but based on the projected innings, projected hits and walks are determined, which are then summed and divided by projected innings to generate projected WHIP.
I use a modified expected ERA formula, using the aforementioned described skills to derive expected runs. This is regressed to actual neutralized earned runs to land on projected earned runs which gets the aging and park alteration for the final numbers. Using projected innings, the projected ERA follows.
This isn’t perfect, but I’ve been using it for over a decade and it works as well as any other method I’ve seen. Many years ago, Bill James came up with a manner to estimate team wins using what’s now knows as Bill James Pythagorean Theorem. It incorporates runs allowed and runs scored to calculate an expected winning percentage. To get wins, the winning percentage is multiplied by the number of decisions.
Let’s start with runs allowed. Above, earned runs are projected. Using a team defense factor, I generate the number of runs. Next is estimating a bullpen component. The number of runs allowed while the pitcher in question is in the game Is added to the bullpen projection. I now have total runs allowed.
Runs scored is simply an estimation, based on previous season’s numbers and how the team has improved or declined.
Decisions are proportional to the number of projected innings, using 9 x 162 in the denominator. It’s not perfect, but correlation studies show it’s reasonable.
Calculating wins for starting pitchers plugs all this into the standard Bill James formula. Relievers are tricky, since set up men and closers have a greater chance to lose games than win them, based on their usage. As such, I flag all relievers projected for holds and saves and apply a modified Bill James formula.
Based on some research I’ll present in an upcoming Z Files, available for Platinum subscribers, there’s some science involved with projecting saves. The short version is percentage of wins that are saved correlate best with team ERA. Using this, I generate a projected team saves total. A percentage of saves projection is made for relievers, which is multiplied by projected team saves to yield the saves projection.
Currently, I haven’t found any relationship between team wins or saves and holds. It looks to be a matter of how each manage deploys his bullpen. Some use more lefty specialists, some rely on their best setup reliever to work more than an inning. As such, holds are projected manually, on a player-by-player basis.
There’s a couple of formulas available on the web to derive quality starts (QS). Each year, I look to see which did the best job of back-projecting the previous season’s number of QS and I’ll use that. I’ll be interested to see how well these hold up with the current trend of pitchers throwing fewer innings. There are coefficients in each that could need tweaking with the changing landscape. With many leagues incorporating QS into their scoring, I want to make sure I provide a usable number.
Compete Games, Shut Outs, No hitters
Yes, some leagues give points for no hitters. No, I don’t project anyone to toss a no-no. I will project CG and SO using historical data, but it’s more a guess than scientific.
Singles, Doubles and Triples
Some points leagues score this so I need to project it. Homers are done as discussed while singles are hits minus extra base hits. That leaves doubles and triples. These are park-neutralized then adjusted via BABIP before the usual aging and park changes to yield the final projection.
All that’s left are innings. For starters, I use the past three seasons to derive an innings pitched per start number. It’s not always the three-year weighted, but that’s the starting point. I’ll tweak as I do each pitcher’s projected games started. Relievers are done on a pitcher-by-basis, based on past and expected usage.
As with hitter’s plate appearances, I try to keep each team reasonable, but I no doubt over-project some staffs. Most of the time, there’s a sixth and seventh starter pushing the total team starts over 162. They’re very likely to pick up starts as in injury replacement, I just don’t know who will get hurt. All I can do is give an honest appraisal.
That should do it for the Mastersball projection process. It’s a fluid process, constantly undergoing changes as more data is available to refine regressions. I’ve been asked on several occasions over the years how well it stacks up against other models, as well as wondering if I back-test against the previous season. The answers are I don’t know, and no. This usually disappoints the person asking, but it’s the truth. The primary reason is I have yet to see a grading system that adequately scores the components of projections. Some use rate stats, but that ignores the diligence of playing time estimations. Some use raw numbers, which are also influenced by playing time as well as luck. I suppose the obvious follow-up is why don’t I devise a system that scores playing time and basal skills. It’s a fair question. The answer is I know intuitively if there’s a deficiency in a specific area; I don’t need to quantify it. Early on, I could sense where the projections were faulty. Over the years, I’ve refined the process to the point my time is best spend boning up on the new research and incorporating the results into the engine, usually to further refine regressions.
Next up is bringing the valuation methodology out from behind the firewall.
Questions? Concerns? Criticisms? Hit me up in the comments, or preferably on the newly renovated site forum.