For balls in play, there are three pieces of information that are potentially useful in determining the value of particular plays and to whom that value should be credited (or debited): (i) the first fielder to make a play on the ball, (ii) the type of hit (bunt, ground ball, fly ball, line drive), and (iii) the location of the ball. The extent to which these three pieces of information are available in
Retrosheet play-by-play data varies considerably through the years.
(i) First Fielder
The first fielder to touch the ball is the most important consideration for determining credit. The first fielder to touch the ball is identified for virtually all plays for the last 25 - 30 years of Retrosheet data. For earlier years, the first fielder to touch base hits is frequently unknown. As data goes back even further, there are even some outs for which the fielder of record is unknown.
For my work, the identity of the first fielder is used for assigning credit whenever this information is available. When this information is not available, credit is allocated across all fielders in the proportion in which fielders get credit across similar plays for which the fielder is known.
(ii) Hit Type
The second level of detail on balls in play is the type of hit: bunt, ground ball, fly ball (or pop up), line drive. This information is available from Retrosheet for all balls in play for the years 1989 - 1999 and for seasons since 2003. For other years, hit types are generally only available on outs-in-play, not hits.
As with first-fielder information, hit-type information is used in calculating Player won-lost records whenever this information is available. When this information is not available, credit is allocated based on the expected distribution of hit type based on the final play result.
(iii) Location
For the years 1989 - 1999, the
location of all balls-in-play are identified in Retrosheet's play-by-play data. I do not use this location data directly in calculating Player won-lost records, however. Instead, I use location data for these seasons to
calculate expected ex ante probabilities for ball-in-play events. That is, based on 1989 - 1999 location data, I calculate what the probability of an out would have been on a play that ended up as, say, a line drive double to the left fielder.
After a great deal of research and consideration, I decided to use location data only in this indirect way even for those seasons for which Retrosheet provides location data (i.e., 1989 - 1999). I made this decision for several reasons. For one thing, using location data only indirectly leads to a more consistent methodology across all seasons for which I estimate Player won-lost records. But also, it was not clear to me, in looking at results from those years for which location data are available, that the location data actually improved the results.
Location data are fundamentally subjective, by its very nature. Relying on individual pieces of subjective data will inevitably introduce errors and possible biases into the valuation of these individual plays. Relying on location data only indirectly, however, and by relying on all of the location data - 11 years' worth - in assessing every play, should allow these individual errors to balance out and offset in such a way as to vastly reduce any potential biases or errors.
Consider, for example, the impact of using
STATS data versus BIS data for calculating
UZR fielding statistics. Simply changing the data source leads to
wildly different stories about some players' defense: was
Andruw Jones the best fielder in baseball from 2003 - 2008 (+112 runs in UZR using BIS data) or a slightly below-average center fielder (-5 using STATS)? If the results are that unstable across different location measurements of the same plays, then it's hard to see exactly how much information location data are bringing to the party at all.
Beyond the question of whether the actual locations being reported are accurate, however, another issue with using location data is that I think that relying too heavily on location data builds on a fundamental assumption that I am not entirely sure is true. This is that balls hit to the same location are more similar than balls that end up with the same end result. That is, a fielding system based on location data treats two fly balls to medium right-center field as equivalent - implicitly assuming that all fly balls to medium right-center field are created equally. My fielding system here treats two fly ball doubles fielded by the right fielder as equivalent - implicitly assuming that all fly-ball doubles fielded by the right fielder are equivalent.
I am not saying the latter of these implicit assumptions is necessarily right, so much as I wonder whether the former implicit assumption is actually more right. And if our focus is purely on player value rather than player talent (as it is in my system), then, in fact, in many ways it makes more sense to me to view one fly-ball double to right field as being equal in value to any other fly-ball double to right field than to view a fly-ball double to medium right-center field as being equal in value to a fly out to the same location.
The Impact of Location Data on my Fielding System
The problem with evaluating fielding systems, in general, is that we don't really know what the "right" answer is - after all, if we knew the right answer, we'd just use that.
One thing that I can compare, however, is how my results compare to what they would have been had I used location data for those years for which it is available, 1989 - 1999. For those seasons, I calculated Player won-lost records both ways. I then calculated a
weighted correlation of winning percentages by fielding position between the two methods. I calculated correlations two ways: for overall (context-neutral, teammate-adjusted) fielding win percentage and for
Component 5 win percentage, which is based purely on whether a ball in play becomes a hit or an out and corresponds most directly to other location-based fielding systems. The results were as follows:
Weighted Correlation,
Fielding Winning Percentages:
Location Data v. No Location Data
|
Total |
Component 5 |
| Pitcher |
82.64% |
82.06% |
| Catcher |
75.63% |
72.69% |
| First Base |
86.08% |
84.18% |
| Second Base |
77.07% |
71.45% |
| Third Base |
89.82% |
89.00% |
| Shortstop |
80.11% |
75.48% |
| Left Field |
92.65% |
89.04% |
| Center Field |
88.71% |
83.53% |
| Right Field |
93.24% |
88.56% |
Note: Catcher figures exclude SB, WP; Totals calculated using the same pitcher-fielder splits for both sets of numbers.
These correlations are extremely high, which is quite encouraging.
It occurs to me that one way in which fielding records based on location data (e.g., fly balls to medium right field) can be compared to fielding records based on event data (e.g., fly-ball doubles fielded by the right fielder) is by comparing the extent to which player win percentages
persist.
As I explain in the article linked in the previous sentence, I
divide credit on balls in play between pitchers and fielders based on "persistence equations" which measure the extent to which player winning percentages on even-numbered plays can be explained as a function of player winning percentage on odd-numbered plays within the same season: i.e.,
WinPctEven = 0.500 + b*(WinPctOdd - 0.500)
The coefficient b in the Persistence Equation measures the persistence of fielding winning percentage between the two samples (even plays v. odd plays). Broadly speaking, a higher value of b in a persistence equation suggests that more of a real skill is being measured. Hence, if the values of b in persistence equations based on location data were consistently higher than the values of b in persistence equations based on event data, this could suggest that location-based data were capturing more fielding skill than event-based data.
The next table, then, compares the values of b from persistence equations for
Components 5,
6, and
7 for both pitchers and fielders based on location-based and event-based data.
Persistence Coefficients,
Location Data v. No Location Data
|
Component 5 |
Component 6 |
Component 7 |
|
Pitcher |
Fielder |
Pitcher |
Fielder |
Pitcher |
Fielder |
|
Location |
Event |
Location |
Event |
Location |
Event |
Location |
Event |
Location |
Event |
Location |
Event |
| Catcher |
22.88% |
23.82% |
33.10% |
6.40% |
72.07% |
-2.92% |
25.20% |
9.59% |
7.14% |
16.05% |
7.87% |
14.19% |
| First Base |
9.41% |
8.66% |
46.56% |
42.91% |
19.26% |
-18.69% |
20.78% |
18.99% |
72.64% |
70.95% |
33.95% |
34.26% |
| Second Base |
9.27% |
13.97% |
41.40% |
28.68% |
66.98% |
80.44% |
34.87% |
62.68% |
7.55% |
7.83% |
29.06% |
29.53% |
| Third Base |
3.64% |
5.90% |
35.41% |
41.14% |
73.76% |
38.69% |
23.24% |
14.80% |
91.22% |
88.32% |
36.17% |
33.44% |
| Shortstop |
19.03% |
12.03% |
48.99% |
31.59% |
90.35% |
90.44% |
34.77% |
24.43% |
-2.92% |
-2.16% |
16.64% |
15.40% |
| Left Field |
14.51% |
16.20% |
37.93% |
38.94% |
4.38% |
-0.09% |
40.42% |
23.96% |
|
|
|
|
| Center Field |
3.19% |
12.18% |
37.15% |
39.51% |
9.88% |
11.58% |
16.94% |
23.26% |
|
|
|
|
| Right Field |
15.88% |
18.55% |
39.19% |
36.81% |
12.79% |
11.01% |
36.94% |
25.07% |
|
|
|
|
There are a lot of numbers there. Let me try to walk through some of the numbers to give an idea of what they are saying. I will then provide some summary data.
Let's start with catchers. For
Component 5, which measures whether a ball-in-play is converted into an out or a hit, on plays made by the catcher, the persistence in pitcher win percentages is very similar for location-based data (22.88%) versus event-based data (23.82%). For fielders (i.e., catchers), however, the persistence is much stronger using location-based data (33.10%) than for event-based data (6.40%). This is clearly a vote in support of the superiority of location-based data.
Plays made by the third baseman, on the other hand, suggest that event-based data are superior, with somewhat greater persistence coefficients for both pitchers (5.90% vs. 3.64% using location-based data) and fielders (41.14% vs. 35.41%).
The next table combines the results across fielders. The numbers here are weighted averages, with the total share of fielding decisions by component used for weights (these weights differ slightly between location-based and event-based records; the numbers used here are the average of the two).
|
Component 5 |
Component 6 |
Component 7 |
|
Pitcher |
Fielder |
Pitcher |
Fielder |
Pitcher |
Fielder |
|
Location |
Event |
Location |
Event |
Location |
Event |
Location |
Event |
Location |
Event |
Location |
Event |
| All Positions |
10.94% |
12.41% |
40.92% |
36.08% |
11.76% |
8.15% |
33.24% |
24.04% |
8.12% |
8.55% |
23.90% |
23.59% |
| Location minus Event |
-1.47% |
|
4.85% |
|
3.61% |
|
9.19% |
|
-0.43% |
|
0.31% |
|
So, on average, location data do not seem to provide any additional information in evaluating Component 5 performance for pitchers. This makes sense to me. Pitchers have some control over what happens to balls in play, but at a much more generalized level of detail - e.g., fly balls vs. ground balls, perhaps how hard a ball is hit - than by location - e.g., I'm skeptical that a pitcher can control whether a ground ball is hit in the 56 hole or more directly in the 6 hole. Hence, it makes sense to me that focusing on events rather than locations provides somewhat more information than focusing on location (although the difference here, 1.5%, is not really enough to favor either of these two over the other with any degree of certainty).
In contrast, location-based data does appear to provide some additional information in evaluating Component 5 performance for fielders, although the difference between the two persistence coefficients, on average, is relatively small (less than 5%).
For
Component 6 - whether hits are converted into singles, doubles, or triples - location data seems to provide more persistent measures for both pitchers and fielders. This makes sense as, for example, the depths of fly ball hits likely makes a big difference as to how many bases they go for. Still, the differences here are not enormous (less than 10% for both pitchers and fielders).
For
Component 7 - whether ground ball outs are converted into double plays - on the other hand, location data appears to add essentially no value.
Overall, the results seem generally supportive of the idea that calculating fielding records based on events rather than locations is probably not much, if any, worse than location-based fielding records.
Ultimately, the proof of the pudding is in the eating. The top 25 defensive players, measured by net Fielding Wins (
eWins minus
eLosses), of the Retrosheet Era are as follows:
Career Fielding Won-Lost Records
Tying my results back to the UZR data discussed above, my Player won-lost records rank
Andruw Jones as the
14
th-best fielder (measured by
net fielding eWins), across all fielding positions, of the Retrosheet Era (career record of
81.4 eWins and
72.7 eLosses), although they also see him as being below average in 2007 and 2008, so that from 2003 - 2008,
Andruw Jones has a fielding record of
32.2 eWins against
31.7 eLosses (0.504 winning percentage,
0.5 net wins).
The 25 worst defensive players by the same measure are as follows:
Career Fielding Won-Lost Records
Nothing necessarily jumps out of those two tables as especially unreasonable to me.
Article last updated: August 20, 2019
All articles are written so that they pull data directly from the most recent version of the Player won-lost database. Hence, any numbers cited within these articles should automatically incorporate the most recent update to Player won-lost records. In some cases, however, the accompanying text may have been written based on previous versions of Player won-lost records. I apologize if this results in non-sensical text in any cases.
Home
List of Articles