BORAS (po) 2.0: Improving Upon the Original BORAS Position Player Contract Model

A few years back, I introduced two models to predict contracts terms: BORAS (pi) and BORAS (po).  That acronym, like many in the data science world, came first and then the words later, but it means Baseball Observation-based Renumeration Assumption System.  The idea being that the bias in determining contract value (even in comp model systems) could be removed, perhaps, by throwing all the recent contracts together in a big pot and regression modeling them with certain performance parameters.

The BORAS (pi) system, which is for pitcher contract, worked remarkably well.  The cumulation r2 of the model is 0.86 and the median miss is 1.7 MM.  Some of the misses are great, such as being off on Ian Kennedy by 11 MM AAV.  Simply put, highly encapsulated pitching metrics were rather effectively in projecting what a contract would be for pitchers.  Traditional metrics were batched together and more "advanced" metrics were batched together.  The previous three years of performance were considered.  It worked out.

I used the same approach with the BORAS (po) system for position players.  Encapsulation of traditional and those more "advanced" metrics appeared to be the wrong move.  Where the pitching metrics for starting pitchers largely all met needed confidence points, the data for position players was more varied.  For instance, fielding metrics for position players needs about two to three years to really stabilize and I was batching this information into more certain offensive data.  In other words (and this is very simply stating it), I was valuing a great fielding season the same as a great offensive season which is a no good, terrible, oh so awful way to do things.  It yielded an r2 of 0.47 and median value off of 3.9 MM, which is decent if you know nothing.  However, a passing interest in contract terms would make you far more away of probable terms than this model would.

This performance has led me to think about other ways to utilize data and come up with, perhaps, a better performing model.  First, I wanted to thin down data sources and noted that, unlike the pitching model, the hitting model valued similar statistics regardless of their traditional nature or more "advanced" nature, so I decided to just go ahead and use less encapsulated "advanced" metrics and ignore the traditional ones.  I also separated out fielding metrics and put them in a more appropriate bin with similar certainty.  So, now we have BORAS 2.0 (po).

Below is how the models compare.

	BORAS 1.3		BORAS 2.0
	Yr	Total	Yr	Total
Alonso	2	18.7	3	34
Duda	2	19.6	2	21.4
Holliday	NRI		NRI
Hosmer	4	61.2	5	93.8
Lind	1	7.9	2	17.5
Moreland	2	21.4	2	16.1
Morison	3	34.8	4	59.8
Reynolds	1	9.5	1	7.7
Santana	3	45.2	3	51.4
Kendrick	1	10.6	2	20.4
Utley	NRI		1	6.8
Walker	2	22.3	3	41
Cozart	4	59.3	5	83.7
Escobar	NRI		1	8
Frazier	3	50.6	3	37.7
Moustakas	3	34.2	3	35.2
Reyes	NRI		1	8.7
Nunez	2	20.8	4	53
Cabrera	NRI		1	8
Dyson	2	25.2	3	32.5
Granderson	1	12.7	2	25.9
Jay	1	9.2	2	17.6
Martinez	4	64.8	4	85.1
Maybin	2	21.8	2	20.5
Upton	5	83.9	5	100.8
Cain	5	92.2	4	64.7
Gomez	2	21.2	3	31.3
Jackson	2	21.8	2	20.6
Bautista	NRI		NRI
Bruce	3	30.1	3	37.5
Ethier	NRI		1	7.2
Gonzalez	1	8.6	2	15.7
Smith	NRI		1	8
Werth	NRI		NRI
Sogard	2	17	2	14.1

BORAS 2.0 appears to do better with the higher end players.  Justin Upton at 5/100.8 made more sense than 5/83.9.  Cain makes more sense as a 4/64.7 than 5/92.2 player.  However, BORAS 1.0 seems to be more into handing our NRIs.  It will be interesting to see how this offseason plays out and whether BORAS 2.0 is an actually improvement on the original model.