Comparing marginal effects of offensive events upon wOBA and OPSby Shane Sanders and Adam Winn
October 01, 2012
It has long been understood that the OPS measure provides a biased valuation of different offensive events. The degree of bias for each event is not well understood, however, because the OPS measure is difficult to understand fundamentally. It is a hybrid measure that features an unwieldy common denominator and elements of double counting. Nonetheless, OPS remains a popular offensive productivity measure for its ability to explain variation in runs scored.
In this article, we'll examine the marginal effects of fundamental offensive events (e.g., a walk, single, double, etc.) upon both OPS and wOBA. As its weights are fundamentally the results of regression estimates rather than of imagination, we believe wOBA is the best benchmark measure. Using this benchmark, we will assess the degree to which the OPS measure biases the relative importance of different offensive events.
Analyzing 2011 data from 305 major league batters, we found that the OPS measure greatly undervalues a marginal walk relative to a marginal single or double and greatly undervalues a marginal single relative to a marginal double. Given that these respective biases differ in degree, a weighted OPS measure that emphasizes on-base percentage would remain biased in assigning relative values to different offensive events.
This outcome likely is due to the artificially fitted nature of the baseline OPS measure. These results, detailed in the following paragraphs, may partly explain why the OPS measure lags behind regression-fitted measures of offensive productivity (e.g., wOBA) in explaining variation in runs scored.
The OPS measure is defined as on-base percentage plus slugging percentage. Though it sounds simple, the math is actually quite complicated. The following equation summarizes the functional form of the measure:
OPS = OBP + SLG = ((H + BB + HBP) / (AB + BB + SF + HBP)) + (TB / AB)
wOBA is fundamentally a regression-fitted model in which the weights of offensive events differ by season. According to Fangraphs, these were the weights assigned in 2011:
wOBA = (0.69*uBB + 0.72*HBP + 0.89*1B + 1.26*2B + 1.60*3B + 2.08*HR + 0.25*SB – 0.50*CS)/PA
(uBB stands for unintentional base on balls)
One already notes that the two measures are fundamentally different by the economy of the wOBA measure—each offensive event is represented one time, and the denominator features a single term—juxtaposed with the redundancy of the OPS measure. The OPS measure features various forms of double counting and a non-linear lowest common denominator.
We obtained elementary offensive data for 305 major league baseball players in the 2011 season (i.e., every major league player who participated in at least 81 games for whom complete offensive data was available). Our data sources include Baseball Reference and Fangraphs. For each player, we calculate the marginal effect of a walk upon OPS as follows:
(ΔOPS / ΔBB) = ((H0 + BB0 + HBP0 + 1)/(AB0 + BB0 + SF0 + HBP0 + 1)) - ((H0 + BB0 + HBP0)/( AB0 + BB0 + SF0 + HBP0))
(Δ is the symbol for change.)
We also calculate, for each player, the marginal effect of a single upon OPS as follows:
(ΔOPS / ΔSingle) = [((TB0 + 1)/(AB0 + 1)) – ((TB0 )/(AB0))] + (ΔOPS / ΔBB)
From these two equations, we calculate the marginal effect of a single upon OPS relative to that of a walk. We do so for each of 305 players using 2011 end-of-season data. In other words, we assume that each player in the data set receives one more plate appearance at the end of the season. In the first case, he hits a single in that plate appearance. In the second case, he earns a walk. This thought experiment allows us to calculate, for each player, the value of one more single at the end of the season relative to the value of one more walk in terms of OPS gains.
The following table summarizes our results.
Mean ratio value Median ratio value Maximum ratio value Minimum ratio value 1.97 1.973 2.25 1.708In the case of OPS, marginal effects vary by player according to their existing OPS value, number of plate appearances, and number of at-bats. Marginal effects for the wOBA measure, on the other hand, are invariant across player. As wOBA is essentially a linear model, one need only divide the coefficient in front of the singles variable by the coefficient in front of the walks variable to obtain the relative marginal value of a single to a walk. We do this calculation using the 2011 wOBA formula to obtain the following result.
Value for all Players
(0.90/0.72) = 1.29
Therefore, we find that a marginal single is worth 97 percent more than a walk in the OPS measure but only 29 percent more than a walk in the wOBA measure. As the wOBA measure is regression-fitted, we estimate that the OPS measure overvalues singles relative to walks by approximately 52.7 percent (1.97 divided by 1.29).
As we will see in a future calculation, OPS gives 10.8 percent more credit (1.97/1.826), on average, for a marginal single relative to a marginal walk than wOBA gives for a marginal double relative to a marginal walk! Using similar calculations and the same data, we find the following relative valuations for doubles and walks across the two measures.
Mean ratio value Median ratio value Maximum ratio value Minimum ratio value 3.625 3.61 4.214 3.211Meanwhile, the wOBA double-to-walk marginal value ratio is...
Value for all Players
(1.26/0.69) = 1.826
From these tables, we find that OPS deems a marginal double as equivalent in value to 3.625 marginal walks, on average, whereas wOBA deems a marginal double as equivalent in value to 1.826 marginal walks. In other words, OPS inflates the marginal value of a double relative to that of a walk by 98.5 percent (3.625/1.826), on average.
The next two tables compare relative valuations for doubles and singles across the two measures. For OPS:
Mean ratio value Median ratio value Maximum ratio value Minimum ratio value 1.842 1.83 2.212 1.658And here's the wOBA double-to-single marginal value ratio:
Value for all Players
(1.26/0.89) = 1.416
From these tables, we find that OPS inflates the marginal value of a double relative to that of a single by 30.1 percent (1.842/1.416), on average.
We could continue with these comparisons. However, the implications of the analysis are clear. The OPS measure greatly biases the relative value of different events. For example, a double has almost twice as much marginal value, relative to a walk, in the OPS measure as in the wOBA calculation.
The degree of bias varies according to which event pair is being considered. The variability of the bias implies that we cannot simply weight on-base percentage by 1.8 or 2.0 in the OPS measure and "unbias" the OPS measure. Such a weighting exercise is akin to an insoluble Rubik’s Cube, in which certain weights decrease some relative event biases only to increase others.
The OPS measure (and weighted OPS measure) is popular because it explains much of the variation in runs scored when plugged into a regression. However, it does not perform as well as wOBA in this regard. The reason for this difference is likely to be substantial internal biases within the OPS accounting methodology. Moreover, it is possible for these biases to influence individual offensive productivity estimates a great deal more than they influence a single parameter measuring the performance of aggregated runs scored regressions.
References and Resources