THE BOOK--Playing The Percentages In Baseball

I’ve thought about this, and wondered if projection systems would benefit from the normalization of aggregate W/L records, or total PA/IP based on previous years.  Was planning to track all the major systems’ ‘07 success anyway, and maybe I’ll do their normalized versions too.  Also, even when PAs or IP are baselined, their individual distributions may still not reflect reality well (Marcel has this problem at the lower end at least, but problem isn’t the right word.) Would it be more accurate still to baseline, and then fit the dstribution to something more normal?  The “area under the curve” would be the same, of course.

I was thinking that my second test, where I end up with 83.7 wins, should not be compared to 81 wins.

What if I select the actual 2006 performances, with batters having at least 269 PA and pitchers having at least 54 IP?

This is their actual per 9 IP and per 39.125 PA, of H, HR, R:
hitting: 9.74, 1.22, 5.19
pitching: 9.25, 1.08, 4.72

As you can see, if I set these threshholds, I get a good set of players, certainly above-average.

And the pythag of such a set of players?  88 wins (.545 win%).

Clearly, choosing players, after the fact, based on the quantity of their performance (undoubtedly meaning that their quality of performance was fairly decent) is a selection bias.  I think it’s easy to see how choosing a set of such players, that produced 88 wins, is in-line with expecting a true win total of around 83-84 wins.

Therefore, I think Marcel has it mostly right, that it can serve as a fair barmoter to other forecasting systems, in terms of figuring out how many wins the league should produce.

For those wondering why are the aggregate PA totals so much higher than the aggregate IP totals:

The formula for PA is:
50% of PA in 2006 plus
10% of PA in 2005 plus
200

So, if I have 862 nonpitchers, I immediately start off with 172,400 PA, which is 91% of the whole 2006 (which includes pitchers batting).

The formula for IP is:
50% of IP in 2006 plus
10% of IP in 2005 plus
25

With 915 pitchers, the minimum IP total is 22,875, or only 53% of the 2006 total.

The problem really lies with the starter/relief roles.  If I knew a pitcher was a starter, I wouldn’t start his baseline at only 25 IP.

As well, the equivalent of 200 PA is 46 IP.  But, pitchers collapse far more than hitters, in terms of playing time and/or injuries. 

For these reasons, you’ll always estimate more playing time for the hitter.

And, without knowing who is on the 25- or 40-man rosters, Marcel forecests anybody that has played in 2006, which means 60 players per team.

Just things to remember…

It seems to me another way to assess the projections is to SEPARATELY compare the forecasted RA and RS components against the actual RA and RS for the league. This would give you some insight into where the net RA-RS bias might be coming from.

Could it be the most of the bias is due to underestimating the RA because of inadequate discounting for the risk of injury to pitchers in your forecasts?  Silver gives the Derrek Lee example but there are probably more Lirianos (relatively speaking) who get season-ending injuries. Or perhaps the bias isn’t primarily due to injuries, but to something else in your depth charts or your system for projecting hitting vs. pitching (and defensive) performance.

If you look at the separate RA and RS components over time (the last few years) you can also see whether the amount of bias has increased in each component (RA or RS) of your forecast, or perhaps only in the net.  And of course the bias in each of these components (compared with the actual league RA and RS) can, in principle, be compared across forecasting systems.