All the Range: An Update to Infield RDA

We launched Range Defense Added (RDA) two years ago, providing us with a defensive range metric competitive in accuracy to SIS’s Defensive Runs Saved (DRS) and MLB’s Outs Above Average. RDA combines with baserunning prevention and framing to provide our overall Defensive Runs Prevented (DRP), but RDA is by far the largest component of DRP for non-catchers.

One of RDA’s distinctive features is Attempt Range, tracking the extent to which infielders are reaching—or not reaching—ground balls otherwise ticketed for the outfield. We theorized at the time, correctly we believe, that this was a distinguishing aspect of infielder performance: ground balls that are harder to reach are also harder to convert to outs, so paradoxically, the less range you have, the better your stats can look. (Sources say you can dive at the ones you miss, to make it look good.) And what we found was illuminating: the better shortstops at the time (like Willy Adames) both reached more balls and created more outs, whereas the below average shortstops either showed below-average range, or above-average range but below-average execution (like Tim Anderson). The point is that measuring infielder skill requires estimating both the numerator and the denominator, which can be challenging and certainly is unusual.

One thing RDA did not initially do was extend Attempt Range beyond the shortstop position. The reason was simple: I could not get it to work in a repeatable manner. There was no shame in this, as we could already match the reliability of metrics based on in-person video reviews (DRS) and actual fielder start positions (OAA). But this deficiency irritated me, and one of my offseason priorities was to address it.

The breakthrough came when I realized that when it comes to baseball defense, sometimes less information is more. Accepting this unlocked our ability to better capture the range of other infielders, including 2024 Platinum Glove recipient Brice Turang. Turang went from slightly above average in our previous ratings to now being the best rated infielder in baseball last year.

Attempt Range is now on our leaderboards (and player cards) for all infielders from 2015 through 2024. The story of how we got there illustrates how different metrics tend to answer slightly different questions, even when providing similar answers. We also discovered anomalies in MLB Gameday pixel coordinates and learned more about the desirable depth of boosted tree models, both of which are discussed below.

The Numbers

RDA was performing acceptably already, but the expansion of Attempt Range across the rest of the infield, as well as some other changes we will discuss, caused remarkable improvement at some positions.

Table 1: Year to Year Reliability, Team-Switchers

Full MLB Seasons from 2016 – 2024

Metric
1B
2B
3B
SS

Mean

OAA
0.33
0.25
0.12
0.38

0.27

DRS
0.13
0.21
0.32
0.25

0.24

RDA
0.18
0.17
0.35
0.24

0.24

RDA Updated
0.47
0.23
0.27
0.47

0.35

UZR
0.21
0.09
0.11
0.15

0.13

Table 1 shows the year-to-year reliability by position for each metric, using players who switched teams the following year, along with their summary statistics, weighted by number of expected plays at each position. Using team switchers helps guard against team-specific effects.

A difference of a few points does not matter, but here the improvements are stark. The reliability of our first base estimates has more than doubled, and we now outpace competing metrics at multiple infield positions. Our mean reliability with team switchers thoroughly exceeds the .3 barrier, breaking new ground for an infield defensive metric.

Let’s look at the effects on individual players by position, with a comparison to other metrics courtesy of our friends at FanGraphs. We’ll focus on 2024, the most recent completed season.

Table 1: Top and Bottom Second Basemen, 2024 MLB Season

The value range for the best second basemen has expanded considerably. RDA’s opinions are still more restrained than DRS or OAA, but we also see more overlap. The best second basemen now have double-digit positive outs value, which seems correct. As we saw originally with shortstops, superior attempt range tends to coincide with the highest ratings. Poor attempt range drags down other players more generally, as RDA punishes them for not reaching ground balls other players do. RDA has knives out in particular for Ketel Marte and Bryson Stott.

Table 3: Top and Bottom Third Basemen, 2024 MLB Season

Third basemen require less range, at least in foul territory, and perhaps for that reason are less affected by / amenable to Attempt Range. Nonetheless, we see again that good Attempt Range coincides with better ratings, although not always, that Matt Chapman remains excellent, that Austin Riley is uniquely immobile, and that RDA now agrees with DRS that Nolan Arenado, has declined from his career heights, and Arenado has fallen out of the leaders.

Table 3: Top and Bottom First Basemen, 2024 MLB Season

The values expand at first base also. The rangier first basemen rate better than the rest, although curiously both extremes show some higher mobility. Less mobile first basemen pay a bit of a price. None of these values are going to overwhelm contributions these players (hopefully) make with the bat. DRS’s outlier rating of Olson is interesting.

Finally, a list of largest 2024 movers by RDA, positive and negative across all infield positions:

Table 4: Biggest Infielder RDA Movers, 2024 MLB Season

The movers at 1B, 2B, and 3B reflect the addition of attempt range to that position. SS already had Attempt Range, but here we also see the effects of our new correction for Gameday’s pixel bias (discussed below), which affects the other positions also. Thanks to this, Missouri stadiums feature in our biggest movers, with the Angels and Diamondbacks stadiums also entering the chat.

Discussion

Different Questions, Sometimes Different Answers

When good metrics disagree, the usual (and frequently ignored) reason is that they are answering slightly different questions. I think that is true with defensive metrics as well.

DRS provides the quintessential “eye test,” as one might expect for a system that relies upon human review of each fielding play. From its inception two decades ago, DRS insisted that a perennial Gold Glove shortstop for the Yankees in fact had terrible range, and it was right.

OAA has unique access to actual fielder locations, permitting a more objective approach. Thus, OAA can measure how far the fielder had to go from where they started, how far the fielder was from the runner’s base if they intercepted the ball, how far the runner was from reaching base when the ball was intercepted, and how fast the runner was going. The downside is that OAA tacitly assumes the fielder plays no role in selecting his starting position. To the extent that is untrue, conditioning on the fielder’s initial location obscures their responsibility for the play’s outcome.

RDA also takes an objective approach, relying on a more limited but holistic set of inputs. RDA cares very much about the launch speed, launch angle, and lateral direction of the ball off the bat, and expects there to be angles at which, conditional on the infield’s alignment, the infielder is going to make a given play. RDA does not know exactly where the fielder starts, but there is justification not to care: if a player’s starting position isn’t good enough to make the plays other fielders do, then that fielder needs to reposition himself or find another position. For example, a shortstop needs to be able to field ground balls bearing -20 degrees off center, full stop. If the shortstop can’t do that from wherever they are setting up, then they are simply not doing their job and there is nothing wrong with penalizing them for it.

If an infielder’s RDA and OAA values differ greatly, you can reasonably wonder if team positioning is the problem. But at a minimum, RDA supplies a strong prior belief of what it means to field your position well.

Brice Turang, last year’s Platinum Glove winner, illustrates this divide. DRS, with the benefit of the eye test, says he was the most valuable fielder in baseball last year, with Andrés Giménez a close second. OAA rates Turang as slightly above-average, but nothing special. OAA, however, loves Giménez and Marcus Semien. RDA rated Turang similarly to OAA before incorporating Attempt Range, but now sides with DRS, rating him the best infielder in baseball.

Why the stark disagreement between OAA and RDA? It is possible that Turang’s instincts influence his positioning, such that measuring from his final starting position does not fairly account for his contribution. If so, this creates a selection bias, somewhat like the bias that can infect measurements of infielder arm strength: fielders who get to balls quicker do not need to throw as hard to get outs. The Brewers’ insistence on evaluating Turang at shortstop this spring suggests that, at least in their view, MLB’s metrics sell Turang short on both measurements.

Of course, it is possible that the metrics’ differences are driven by something else entirely. We’re certainly open to a different view, but in the meantime, unfortunately only individual teams know where their fielders are specified to be, and the extent to which individual fielders deviate from those locations.

Coordinate Bias

In its Gameday system, MLB displays an approximate landing location for each batted ball, estimated by stringers from a system of horizontal and vertical pixels. These coordinates, typically called hc_x and hc_y, are available to the public through the Savant system. Standard practice is to presume their rough accuracy and convert Gameday pixels to relative distances in feet using an arctangent function.

Unfortunately, the pixel system is primitive, and even when converted, these coordinates turn out to be distorted for certain stadiums, particularly the two stadiums in the State of Missouri. Even worse, the distortions vary within stadiums for each position. Taking them at converted face value, lateral angles can be off by 5 degrees or more, enough to misrepresent a ground ball to the “5.5 hole” as a routine out for the shortstop, which it most certainly is not. Our new correction for this bias is one reason for RDA’s further jump in accuracy. It also explains why Masyn Winn has improved in fielding value while Nolan Arenado has taken a hit.

This bias in the Gameday coordinates has not been widely discussed, and it is a problem for public defensive analysis. If MLB could provide us with the actual batted ball direction as measured off the bat, it would be much appreciated, and render this a moot point.

The Paradox of Model Depth

We’ll end with a technical note, as the manner in which we ended up modeling defensive structures generated surprising results.

Often, modern baseball models consist of two parts: (1) initial, “physical factors” models, typically employing some boosted tree architecture, that predict the general likelihood of the outcome of interest, here the probability of an out; and (2) a shrinkage model, which uses random effects or analogous structures to parcel out remaining variance, conservatively, to the players of interest. Typically, the goal is to make the physical factors model as explanatory as possible to ensure that the only credit (variance) left for players is that likely due to them.

With infield defense, though, we found that formula may not work. In fact, if you try to make your physical factors model as accurate as possible out of sample, typically through increasing tree depth, you remove variance that properly belongs to the players. (We know this because the player values compress down to nothing and yearly reliability numbers go into the tank). Thus, if you want to discover the right amount of player responsibility, you need to make sure your physical factors model explains some, but not too much, of the out probability: the model needs to be accurate but not too accurate. This is completely contrary to typical model tuning practice, and quite notable.

If you use a boosted tree system that favors shallow trees (like BART), you are protected from this trap. But if you use frequentist options like xgboost, which otherwise works quite well, be careful. (We use BART, specifically, stan4bart, and keep xgboost’s depth under control when it is needed).

Wrap Up

The new values have been pushed for all MLB seasons from 2015 through 2025. As always, let us know if you see something amiss, have unanswered questions, or otherwise have additional insights to share.