Projection Retrospective Retrospective

In looking through my recent retrospective posts I noticed what appeared to be some trends with the error in the current system. I wanted to take a few moments here to yell into the void about those observations and what I plan to do about them.

I’m not too worried about the “0” “Nailed it”s, because that’s an incredibly narrow window to hit.

Starting Pitchers is fairly balanced, especially because several of them were really close. If I re-did the analysis to look at how many SP squads finished within +/- 4 WAR of their projections (relatively speaking, these would be “coin-flips”), I wouldn’t be surprised if the source of the imbalance was due more to luck than a systemic failure in the projection system.

Hitters and Relievers though are a different story.

Relievers aren’t a complete loss, and I only have so much time at my disposal, so I will shift my focus to addressing the failure of the “Hitters” projections.

I have two theories about why the current approach consistently over-estimates WAR from hitters.

Theory 1 – ESPN Projections are inherently optimistic

Now, I do sincerely believe this is the case. But, my opinion may have little basis in reality.
In trying to justify this stance, I identified that in 2022, 132 batters reached 500+ Plate Appearances. The current ESPN projections have 103 players projected to break that number in 2023. Without doing more investigation, that surface level analysis suggests that the ESPN projection system may actually be pessimistic!

I highly doubt that’s the case, and this superficial investigation hardly proves anything one way or the other, but, at least for trying to identify why my projection model is so consistently overvaluing of hitting squads, this theory should probably be de-prioritized in favor of other theories.

Speaking of which…..

Theory 2 – My model overvalues bench roster depth on the 30 Man Opening Day Rosters

I’ve mentioned in previous posts that one significant weakness of the projection model is it’s current inability to take into account roster depth beyond the initial 30 man pro roster.

However, one possible source of the perceived over-valuing could, in a seeming contradiction, be placing too much credit in the value derived from projected 30 man bench pieces. The way I calculate projected contributions from bench pieces relies on two assumptions that may both be faulty.

Assumption 1 – There are an equal number of games played on each day of the week

My projection model assumes an equal number of games played each day of the week, and therefore that bench pieces will have a fair chance to be playing on the off days of the primary position holder.

To explain why this matters for my projection model, some quick math.

Scenario: Primary position holder (“starter”) is projected to play 120 games and the backup is projected to play 90 games and the MLB calendar is 180 days long.

If the backup always had a fair chance to play and the starter always had a fair chance to have a day off, you could calculate the backups projected number of fantasy starts using the following formula:

(Games Played by Backup/Days per Season) X (Days where primary position holder is not playing/Days per Season)

(90/180) X (60/180) = 30 Fantasy starts in a season by the backup

So, why does this matter? Because, as I mentioned at the beginning of this paragraph, this approach assumes that the primary position holder is just as likely to be off on a Monday as they are a Tuesday, Wednesday or any other day of the week. And that the backup is equally likely to play on those days. As we all know, this is not how the MLB schedule works. Below is a quick breakdown of a random week of the 2022 playing season (June, 5th, 2022), and how many MLB teams played per day of the week.

So, if the schedule is not equally distributed, and if the bench player has a higher chance of not playing on the same day as the starter because both teams have the day off, how can we account for this?

Honestly, I’m not 100% sure.

My best guess right now is that to account for the schedule inconsistency, we can multiply the bench players stats by the number of actual game slots per week divided by the number of game slots if every team played every day.

So, if there are only 180 games a week rather than 210, we can adjust all bench probabilities accordingly by simply multiplying the bench contributions by ((180/210) = ~85.7%).

Will this measly ~15% reduction in bench contributions make a significant difference? I dunno. But, maybe in tandem with my second assumption, there can be some significant improvement.

Assumption 2 – 4.16 Plate Appearances/Game for all players

ESPN projections do not provide the projected number of “Games” per player for which to calculate opportunities for backups.

Instead, to calculate this indirectly, I take the number of Plate Appearances (estimated as At Bat’s plus Walks, which in itself is a crude estimator at best, but does capture the overwhelming majority of Plate Appearances) and divide by the “Average number of plate appearances per game per player”.

Average Number of PA/G/player is calculated using “League Plate Appearances” divided by “Games Played” multiplied by “Hitting Slots per game”.

(League PA = 182,052) / (League G = 2,430 X 18 Hitting slots/game) = 4.16 PA/hitter slot/G

4.16 Plate Appearances/game may be fair to the average MLB player, but is probably overly generous to a typical #9 hitter and probably penalizing a typical #1 hitter. And in the CPL, the distribution of hitters by lineup spot is almost definitely favoring those higher in the order. As a result, using this average value that represents all MLB hitters probably does not properly represent the average hitter in the CPL.

To accommodate this, I will be shifting the Plate Appearances/Game assumption from reflecting the average of all 9 hitting spots, and use the average of the first 6 hitting spots instead.

(126,179 Plate Appearances / (2430 X 12 Hitting Slots)) = 4.33 PA/G.

This is far from perfect, but for the purposes of my projection model, there needs to be some representation for Games Played and I’m far too lazy to dig into the ESPN projection model to find their Games Played projection models.

Conclusion

So, I’ll be increasing the number of games played by starters, decreasing the number of opportunities for bench pieces, and on top of that automatically reducing all bench contributions by ~15% to account for the unevenness of the MLB playing calendar.

Will this be enough to properly weight hitters contributions?

Time will tell!

Thank you all for indulging me if you’ve gotten this far.

Best of luck in 2023 to all!