Doubles and Triples Really Aren’t So Different (Off the Bat)

One of the first statistics a new baseball fan might find value in is OPS. While there are limitations to OPS as well as other measures of hitter productivity which better capture total offensive prowess (wRC+, wOBA, OPS+), OPS nonetheless has a lot going for it. It is pretty straightforward to grasp a player’s OPS (no normalizing or park factors, etc.), it is (relatively speaking) easy to explain in a sentence, and importantly it does at least a decent job conferring batter skill.

OPS is, for those unfamiliar, the sum of a player’s on-base percentage (the number of times a hitter reaches base over their number of plate appearances, with a couple caveats) and their slugging percentage (total bases as a result of a hit over total at-bats). Should a player reach base 30% of the time and touch 40 bases as a result of their own hits in 100 at-bats (.400 SLG), that player would have a .700 OPS.

While there are plenty of shortcomings of OPS, the ingredient I wish to focus on here is slugging percentage (SLG). Slugging percentage denotes the power of a player. Those players whose hits are disproportionately of the fence-eclipsing variety collect four bases for a single at-bat. Meanwhile, players predisposed to singles are more often collecting just one base for each of their hits. On average, home runs — and extra base hits more broadly — are struck more forcefully than singles, hence the “slugging” title.

Recently though, I wondered about the validity of giving a player an additional point/base in the slugging percentage formula for a triple relative to a double. If doubles and triples are so similar on average in batted ball profiles, and I hypothesized, why should a triple contributed more to a player’s “slugging” ability? To be clear, this is a wholly theoretical consideration; there is an elegance to slugging percentage simply aggregating total bases from hits. Quite literally, a triple represents one more base than a double. Still, one might suspect the leading differences in doubles and triples to lie outside of a player’s power/slugging abilities, and instead be in fact more about their speed (or age).

To that end, let’s compare doubles and triples to see if there is a meaningful difference in how the baseball is struck between those two hit results. Data was pulled for all regular season doubles and triples hit during the 2022 and 2023 seasons via Statcast. Below are several batted ball metrics grouped by hit result and presented as histograms.

First, projected hit distance.

First, note the floating y-axis values between these two histograms: there have been a lot more doubles in recent years than triples.

Doubles and triples look to travel similar (projected) distances overall. That said, triples appear to be even more drastically skewed toward very long hits. Put another way, a greater fraction of doubles appear to have been projected as sub-250 feet than triples. One takeaway here is that doubles might be more likely to be hit on the ground (down either line?) than triples.

The story is mostly unchanged when it comes to exit velocities: both doubles and triples are hit quite hard on average. If anything, it looks as though there is a tad less variance in exit velocity among triples than doubles, but that could well be in part of function of there simply being far fewer triples.

As a final juxtaposition, below are histograms for launch angle.

At this point the “doubles aren’t so different than triples” hypothesis is feeling fairly reasonable. How far, hard, and high doubles are hit does not appear demonstrably different than triples. While there are some visible variations, none appear to be particularly glaring.

While visualizations are great, sometimes a numerical summary is easier to make real sense of. Below is a table that summarizes all of the visual data from above.

This table highlights some of the key differences between doubles and triples by taking a couple measures of center: the mean and median. Triples do appear to be hit slightly farther, harder, and higher on average than doubles, but not by a glaringly wide margin. Still, it is validating to SLG that triples do entail more robust hits off the bat than doubles — again, on average — even if only by a relatively modest degree.

While there might be slight differences between doubles and triples in terms of batted ball profiles, I still maintain that it would be difficult to correctly guess the outcome of any particular hit based on batted ball metrics alone. There are still plenty of 400+ ft doubles and 10 degree launch angle triples, after all. And if one cannot reasonably differentiate one type of hit from another, should one hit result in more SLG when really that outcome may have come down more to speed, effort, or defensive alignment?

To better capture the difficulty in classifying batted balls as either doubles or triples without the benefit of actually knowing a hit’s result, I applied a couple machine learning algorithms to the data to see how they fared. Specifically, I applied k-nearest neighbors (kNN) and a C5.0 decision tree algorithms to the Statcast data for these hits.

Should you not be particularly familiar with ML algorithms, no worries, this is an exceedingly quick overview. Note simply that kNN and decision trees are known as “classifiers”, the job of which is to classify records based on data available to the algorithm. Will this individual default on their loan (“Yes” and “No” being the two classes to predict)? Is this mass of cells benign or malignant? Is this batted ball a double or a triple? The applications of classifiers range from the vastly important to the obscure blogosphere. Should you have interest in a slightly more lengthy breakdown of either decision trees or kNN as applied to baseball data, I covered them awhile back in a bit more in depth here and here.

For the purpose of this exercise though, their use will be summarized in very broad strokes.

The kNN algorithm makes use of numerical data to make classification predictions and, good news, Statcast has numerical data in droves. The fields which informed the algorithm are as follows: hit distance, launch angle, exit velocity, hit coordinates, pitch location, and pitch velocity. Some cleaning and normalization of the data needed to occur before testing kNN, but before long the algorithm was ready to be applied to the data and… kNN predicted that every single hit was a double.

This outcome is due to the fact that there were so many more doubles in the dataset than triples; it is known as class imbalance when there is such an uneven distribution of classes. This is a common obstacle for classifiers as class imbalances appear in plenty of contexts (most cell masses are benign, most loans are not defaulted on, etc.). Unfortunately, triples are in decline and roughly just ~7.7% doubled and triples over the last two seasons were in fact triples according to Statcast. Given that, the model predicted all records to be doubles and was right roughly 92% of the time. That is a pretty high percentage, but it is completely unhelpful as it simply classified all hits as the dominant class.

There are various options and methods to handle such a situation and I took the most simplistic route: I took a random sample of all doubles from 2022-23 that was equivalent to the total number of triples. Combining that sample of doubles to the entire total of triples made for a smaller dataset evenly split between doubles and triples; where the larger dataset was previously ~92% doubles, the new smaller dataset was exactly 50% doubles.

Take two. Most of the data was used to train the model and the remaining 296 observations were used for testing. The results of the kNN algorithm are presented in a cross table below.

Overall, not great. The actual events are listed vertically on the lefthand side and the predicted events are listed along the top. The upper-left (91) and lower-right (111) represent cases where the predicted hit type matched the actual hit type. The lower-left (44) and the upper-right (50) count where the predictions and actuals diverged. Of the 296 total hits that kNN was tested on, 50 doubles were incorrectly predicted to be triples and 44 triples were incorrectly predicted to be doubles. 202 correct predictions of 296 records makes for a 68% accuracy, not much better than a coin toss when a prediction of all observations as one class or the other would have resulted in roughly 50% accuracy.

Unfortunately, the C5.0 decision tree algorithm didn’t fare much better. In that case, non-numeric data was applied in addition to the fields included in the kNN algorithm above. Specifically, batter and pitcher handedness as well as pitch classification were included this time around. Still, of 296 observations, 70 were incorrectly classified, or 23.6%.

While ~76% accuracy is better than kNN and further from a coin-toss, it still isn’t great. Additionally, decision tree models in R (where this analysis was conducted) indicate which fields were leveraged in what order. This is a meaningful insight because decision trees make predictions in the order of each field that is most predictive of class; the field that is most predictive gets leveraged first, the second-most predictive leveraged second, and so on and so forth. Here, hit distance was deemed most insightful by the C5.0 algorithm, followed by hit coordinate (also a function of distance). Interestingly, though, the third and fourth most predictive fields were the horizontal and vertical locations of the pitch that gave way to the double/triple. This is significant to me given that common batted ball metrics (angle and velocity) are so similar across doubles and triples that pitch location is more insightful to the algorithm.

In summation, doubles and triples are in fact different off the bat, but only slightly. On average they vary a bit in distance, trajectory, and velocity. Still, they are not so different that a couple out of the box ML algorithms can reliably identify differences and classify accurately. So, will SLG as a statistic change? No. Should it? Surely not, regardless of my qualms. Still, given how much time is spent in baseball circles classifying pitches, it was an interesting exercise to investigate the classification of hits in this case.