from Amazon
Tuesday, November 27, 2007
Long-term forecasts
I really ought to come up with a better forecast engine than dropping everyone by 0.5 wins every year. It works remarkably well in most cases. But, sometimes it doesn’t make sense. Where it doesn’t make sense the most is with relievers. Let’s create a basic model for starting pitchers:
win% G WAR Age
0.500 24.0 2.9 30
0.490 21.6 2.4 31
0.477 19.4 1.9 32
0.461 17.5 1.4 33
0.442 15.7 1.0 34
0.420 14.2 0.6 35
0.395 12.8 0.2 36
0.367 11.5 (0.1) 37
win% is the pitcher’s true talent level at age 30
G: is the number of full games (24x9=216IP)
WAR: (win%-.380)*G
What this simple model does is start with the win% and drops it by 10 percentage points the first year, 13 the next, 16 the next, etc. I have no idea if this is accurate or not. Seems reasonable for discussion purposes. By the age of 37, this .500 pitcher of age 30 is out of baseball. The aging function might need to be more aggressive. His IP drops 10% every year.
Anyway, as you can see, the WAR drops steadily by about 0.5 wins, which is why I use that basic rule.
But, what about for relievers? This is the closer model, using GuyM’s approach to valuation:
win% bonus G WAR Age
0.650 0.080 9.0 2.3 30
0.640 0.070 8.1 1.9 31
0.627 0.057 7.3 1.6 32
0.611 0.041 6.6 1.2 33
0.592 0.022 5.9 0.9 34
0.570 0.000 5.3 0.5 35
0.545 0.000 4.8 0.4 36
0.517 0.000 4.3 0.2 37
0.486 0.000 3.9 0.1 38
0.452 0.000 3.5 (0.1) 39
The bonus column is the extra wins above the .570 level.
The WAR column is based on .470 being replacement, not .380 like the previous starter’s chart.
Here we see that for closer’s, the drop in WAR should be closer to 0.4 wins. And if we look at a regular reliever (one who doesn’t benefit from the bonus), and we have this:
win% bonus G WAR Age
0.570 0.000 9.0 0.9 30
0.560 0.000 8.1 0.7 31
0.547 0.000 7.3 0.6 32
0.531 0.000 6.6 0.4 33
0.512 0.000 5.9 0.2 34
0.490 0.000 5.3 0.1 35
0.465 0.000 4.8 (0.0) 36
The drop here is more like 0.2 wins. This may be one reason I’ve been undervaluing relievers, in that I’m applying a far too strong aging model for the future years.