HardHit/9: Statcast's Best Hard Contact Metric | pitcher list (2023)

In World Series Spiel 5,Tyler Glasnowhad a rough day. Of the 13 balls in play he allowed, 11 were hit at 95 mph or harder. The Dodgers' at-bats were so loud that MLB.com's Mike Petriello did some research and found that Glasnow's tough day in terms of hard hit rate was actually the worst postseason performance ever in the Statcast era.

Highest hard hit percentage allowed, 2015-20 postseason, by a pitcher with at least 10 balls batted in a game. (There were 404 such games.)pic.twitter.com/J0ux5HtySS

— Mike Petriello (@mike_petriello)26. October 2020

As Statcast defines, HardHit% is defined as the percentage of balls in the game that are hit at 95 mph or more. And it matters at all because 95 mph is the start when the uphill wOBA trend starts - that's when exit speed actually starts to matter. When we speak of this as a percentage, we pretend that the only "whole" sample that counts is the balls in play.

That's why I felt something bad when I saw Petriello's tweet. Of the six other names on the list, three were from pitchers known for their flashy strikeout totals. Glasnow himself batted out seven to five innings. Was he pitching poorly, or just knocking people out instead of inducing weak groundballs?

Instead of focusing on one game, I decided to take a closer look at Glasnows 2020. And I decided to compare him to someone I consider his polar opposite:Dallas Keuchel.

It only takes one bad statistic to tell the wrong story. Despite Keuchel's reputation for squelching hard contact, he actually gave up more hard-hit balls both per batter and per inning than Glasnow.

If you're not familiar with HardHit/TBF or HardHit/9, that's because Statcast doesn't show them. But you should be: HardHit/9 is more meaningful than any traditional hard contact statistic while also reaching confidence thresholds faster than BB/9.

One of the best and easiest tools for evaluating contact suppression had been hiding under our noses the whole time.

What is the problem with HardHit%?

Let's start with a look at Glasnow's Statcast "sliders" for 2020.

The bottom row - all speed and spin - are on a per pitch basis. Whiff% is per swing. And all expected stats and K% are per PA. And the three "Contact" stats - Barrel %, Exit Velocity, and Hard Hit % - are the only stats scaled to balls in play.

Barrel rate in particular is an odd case. Barrels/BBE% is displayed on the player pages. But Baseball Savant also lists Barrels/PA% in its Exit Velocity & Barrels Leaderboard, making it the default sort. The Pro PA version is not shown anywhere else on the entire site.

The difference between these two stats is somewhat similar to the difference between breath rate and swing rate. A pitch with a good whiff rate might be hard to hit, but that might not be a good thing—many curveballs with tremendous whiff rates end up far below the zone too often, leaving pitchers behind in counts. But a pitch with a high swing strike rate is usually successful in both convincing batters to swing and inducing puffs. That makes it the preferred statistic in almost all cases: it tells more of the story.

Data is only as good as what we ask of it. And in the event of a hard-hit rate, Statcast urges us to ask the wrong questions. Both hard hit rate and barrel rate — two of the most cited Statcast contact stats — cause us to think about contact management solely in terms of the amount of weak contact that pitchers give up. If we try to compare two pitchers using these stats, the pitcher who hits more batters - arguably the most efficient way to quell hard contact - will be at a disadvantage. And that leaves a big gap in both our descriptive and predictive toolkits.

What should we ask?

Data can answer two questions: What happened and what will happen? And it's very rare that one number helps answer both.

Descriptive statistics answer the first question. And among them, FIP has been among the best for over a decade. It attempts to gauge what a pitcher's past ERA should have been by cutting out the noise of team defense and keeping only what it thinks a pitcher can control: strikeouts, walks, and home runs. It bakes in some noise - park factors affect home runs, for example - but its strong correlation with ERA makes it extremely useful for spotting pitchers whose results don't quite match their inputs.

But as much as the Red Sox pitchers of 2020 have tried to show us otherwise, home runs aren't exactly common occurrences. The FanGraphs statistical library tells us that a pitcher's home run rate only after 1,320 batters is more likely to predict his future home run rate than just using the league average. This is the reasoning behind xFIP replacing home runs with flyballs and assuming a league average HR/FB%. Common events like the fly ball rate are much more stable, making them much more useful for predicting the future. The problem with this approach, however, is that these common events often have little impact on previous results. FIP massively outperforms xFIP in describing, but xFIP beats FIP in predicting the future.

I bring up these shortcomings to emphasize what good contact statistics will do for us. Holding FIP and xFIP up to the light we can see the path to what we are looking for in the space left by what they miss:

1. To reduce variance, it should use input that occurs more frequently than home runs.
2. It should be something we can confidently show pitchers are in control of.
3. To eliminate noise, it should not depend on parking factors or team defense.
4. We shouldn't need much context (other than sample size) to use to compare two pitchers.

How will we know if we are successful?

One of the tricky things about predictive statistics is that there is no universal ruler by which to measure their success. We're not trying to delete any R²number (although I'll set some arbitrary benchmarks later). We're looking for better than what we have. My approach was pretty blunt — I just drew every pitcher's season with at least 100 innings since 2015 and started checking the correlations that datasets gave me. I'll recap the R²at the bottom (for those of you using tiny mobile screens, that should make things easier!), but you can hover over the trend lines to get the details. Darker trend lines indicate stronger correlations, and near-invisible trend lines mean weak ones.

So to get our bearings, let's first look at a few things we feel good about: strikeouts and walks.

You'll find that the average distance from the trendline is much shorter for strikeouts than for walks -- that's about what correlation measures. The R²of .273 for strikeout rate would mean that, on average, 27.3% of a pitcher's ERA in a given season can be explained by his strikeout rate. Put these skills to work and some interesting observations surface early on: K% looks a lot more informative than K/9, but BB/9 seems to be more important than BB%. K-BB% seems to be a solid indicator while K/BB is not. And overall, it looks like strikeouts are far more important than walks. Nothing groundbreaking here; We are currently using this dataset as a guide.

With that in mind, let's start with home runs. We know they are not very forward looking. But we do know that they have a lot of explanatory power. But how much?

The most effective measure was HR/9 with a staggering 0.427 R². So if we're trying to develop a contact statistic, the bar for a good descriptive statistic is pretty high. Homers per nine do the job pretty well.

This particular stat against HR/PA which I suspected would be an effective measure. Why could that be? Well, ERA and HR/9 both have the same denominator, so they tend to correlate. It's also possible that the runs-to-outs ratio contains some implicit information about a pitcher's ability to generate outs that HR/PA does not contain. If you look at BB% and BB/9, we see a similar trend. Worth watching.

So we have some guidelines for descriptive statistics. But if we want to know how meaningful a statistic is, we have to perform two tests. First, it's important to know how well a statistic predicts itself. And secondly, knowing how well it correlates to next season's ERA is helpful if we want to bake it into prediction statistics. Here, too, we start with deletions.

We can say that the strikeout rate is relatively stable, even one season into the future. Your 0.567R²is our highest mark for predictive power. Walk rate of 0.358 R²with next year's results aren't as stable, but compared to what we'll see next at home runs it's still decent. We knew pitchers change from year to year. This data only tells us how and by how much.

As for the correlation with ERA, the results are not all that strong.

Wondering why you can't see a trend line on walks? It's because functionally there aren't any. While the .209 R^2 is weak for K%, it's still palpably there. But the R^2 for the walking rate? Only 0.029. However, K-BB% beats K%, which tells us that walks deserve inclusion in ERA indicators, but that they're not a good way to compare two pitchers as a standalone stat. We would be happy to see a contact stats rival of K% for predictive power, but more realistically we would be happy to find one that is more correlated than BB/9.

And in case you're curious, home runs don't even come close.

This is the diagram that explains why xFIP exists. We should have little confidence that a player's home run rate in a season - measured in any way - will mean anything going forward. The best formulation of this, HR/9, has only 0.062 R²with himself. If you're wondering why these lines are flat, it's because the formula ignores the inputs and simply uses the league average instead. In other words, put little to no supplies in a house that's allowed, even for an entire working season.

As for correlation with ERA? It gets worse.

The best correlation here is only 0.016 R²for HR/9. That's actually the only relationship on this plot with a p-value below 0.05, and given the weakness of the correlation and context, that's not enough for me to strongly reject the null hypothesis either. In layman's terms, we should doubt that a pitcher's home run stats are related to his future ERA.

To recap, here's the R of each stat²the correlation for ERA in the season, ERA in the next season and for itself in the next season. Values ​​in red have a p-value greater than their correlation, and statistics in green represent the leading indicator in that column.

We can see how big the gap is between HomeRuns/9 and K% - they clearly have different uses. K-BB% tops the S2ERA correlation sweepstakes and also shows the importance of considering multiple inputs simultaneously.

The completely arbitrary R²The goals I have set for a new contact status are as follows:

1. For S1 ERA: better than K%.
2. For S2 ERA: better than BB/9.
3. For himself in Season 2: midway between BB/9 and HR/9 (greater than .202).

Which Statcast Numbers Perform Best?

Choosing which numbers to look at is the most difficult task. I knew I wanted to take a closer look at both hard hit rate and barrel rate, but I also know that raw exit velocity is quoted often enough to be worth a look.

After some early testing, I decided not to pursue exit velocity any further - there were both correlation and prediction issues. I'll include it in my summary at the end, but I'm not suggesting building around it. We use the hard hit rate because wOBA gains in exit velo are not linear, so calculating an average can provide an unhelpful measure. The expected results of four balls going 90 mph aren't the same as three balls going 85 mph and one going 105 mph, but that would give the same average.

So what types of hard hit rate and barrel rate were most useful?

HardHit/9 was the most telling pitcher input stats I found. It is 0.459R²will not beat ERA estimators like FIP, but this statistic requires multiple inputs. As a single indicator, it explains ERA better than home runs. For that reason alone, it deserves to be among the top stats we use to evaluate pitchers. Even its slightly less descriptive cousin, HardHit/TBF%, likely plays a role. K-HH% is worth considering as a stat on the back of the napkin.

As for barrels, they look pretty much like home runs, don't they?Six Man Rotation's Connor Kurcon has had great success with Barrels/BBE% as part of pCRA, but this arrangement was actually the least individual description of the six I include here. Again, using per-nine stats seems to be the best way to capture variance, probably because it contains more implicit information about a pitcher's ability to get outs. Whether this is good or not remains to be determined. But anyway, it looks like kegs are less effective than home runs in describing a pitcher's past ERA. And that makes sense: Not every barrel has scaled a fence, but every home run has.

As we look forward to next season, HardHit/9 and HardHit/TBF not only outperform barrels across the board - they're almost as stable as BB%.

Starting with the self-prediction, HardHit/TBF% - which, as a reminder, has the same denominator as K% - narrowly beats HardHit/9 as the most consistent stat. I don't think the margin makes sense - both should be useful. This is further proof that K-HH% probably has a bright future (as long as you use HH/TBF%).

The gap between barrels/9 and barrels/TBF% is also small, but the bigger news is that they just aren't that sticky. A0.12R²is twice the year-over-year correlation for home runs, but still comically lags behind the walk rate. In other words, we should have some doubts about whether looking into the future can help us. Given that home runs surpass barrels in terms of vividness, it's not clear what role barrels should play in our pitcher rating - there's always a better option.

When we look at S2ERA correlations, these lessons play out again.

This corresponds exactly to our expectations. As it stands, HardHit/9 seems to be the second most predictable input we can find, only behind Strikeout Rate. The gains versus HR/9 or the traditional barrel rate are gigantic in relative terms.

So, to recap, here's what we've learned so far:

HardHit/9 achieved all three of my totally random goals. Not only is it more meaningful than HR/9 — it's an order of magnitude more sticky and predictive of future ERA. It's head and shoulders above the field in everything except when predicting itself where it's at the top along with HardHit/TBF%.

This data questions whether HardHit/BBE% and Barrels/BBE% deserve to be prominently featured by Baseball Savant. They're both just a hair better at explaining past ERA without predicting the future too well. They may still have a place in formulas that put them in the right context, but as standalone stats they're more of a distraction than a help.

I have more general questions about whether we should concern ourselves with kegs for jugs. Even in their more helpful standalone form, they don't give much improvement in home runs. I think they're probably helpful when we're trying to rate a pitcher who's going to a new team. But unless a pitcher's home park and defense changes, there probably isn't much to be gained by using barrels, especially considering HardHit/9 has beaten Barrels/9 at every step.

And as for exit speed? Somewhere it might be helpful. But since exit velocity returns are not linear, we know there are holes in this statistic. The numbers prove that.

When can we trust these statistics?

Earlier, when I quoted numbers about the reliability of home run rates for pitchers, it was based on a statistical test that can be reproduced. Statistically, reliability is a somewhat difficult characteristic to measure well. With the help of Pitcherlist's in-house data guru, I decided to test it myself.

The method I chose was a Cronbach's alpha test. You can read more about it at Fangraphs, whereJonah Pemstein and Sean Dolinar have provided an extraordinary amount of information,including part of the code used. However, one of the limitations of this method is that it is exceptionally difficult to calculate reliability for per-9 statistics. I decided to only find the stats per hitter. Since the per-nine stats were about as efficient at predicting themselves, we should be able to use the data to draw inferences about both.

So how did HardHit/TBF and Barrels/TBF hold up against more traditional stats?

The order in which the stats become reliable the fastest is pretty much the same as the stat with the highest R year-over-year². The exception? HardHit/TBF% actually becomes reliable faster than BB%! Assuming about 20 hitters per start, which is a little more than double the order, it would take less than 10 starts to estimate what a pitcher's HardHit/9 will be for the rest of the year. And if we assume that the per-nine stats take about that long to even out - think of the annual R²Values ​​- then HardHit/9 is probably also useful in about 10 starts. This means that we get meaningful results almost ten times faster when using HardHit/9 instead of HomeRuns/9.

A side note on my data: I suspect the drop in reliability we're seeing around 300 batsmen face is because I also included data for 2020. I suspect I wouldn't be seeing this if I only found the numbers for 2015-2019. The same effect shows up after about 750 batters are faced, which is about the maximum. So the individual outputs become very messy due to the lack of additional data. Pemstein describes the effect pretty well - check out his work if you're curious.

results at work

The first and simplest application is K-HH%. But what I thought was a significant step up doesn't seem to be. It's just a little less helpful than K-BB%. That doesn't sink it as a metric, but I had higher hopes.

However, my first attempt at an ERA estimator is extremely promising. I call it HHERA (I'll let you guess what that stands for). It's pretty simple - I used K%, BB/9 and HardHit/9 in a linear regression to estimate S2ERA.

HRERA 1.0:4.875703– (7,991393×K%–0,06004585×HardHit/9–0,1678067 ×BB/9)

The good news? After running several k-fold cross-validations, I found that the average R²was 0.238 with an RMSE of 0.838. That's right behind itDan Richards' market leading FRA, Andpotentially before pCRA(It's a black box so I can't make fair comparisons, but HHERA beats its public numbers). I still need to work to assess whether I should include other variables or use nonlinear models.

I plan to give this project some additional work and with the help of some people who know how to get an ERA estimator working. If that's you, make sure you sign up.

Below you can explore HHERA, HardHit/9 and HardHit/TBF%.

Conclusions

The simplest and most obvious conclusion is that HardHit/BBE% doesn't deserve to be the stat we call "Hard Hit Rate". The same goes for barrels/BBE%, but at least we have the ability to easily and quickly check these numbers on the Statcast leaderboards.

In addition, the doors are open for people to take this data and work with it. There is so much room to apply what I have started here to different projects e.g. All the data I used is from Fangraphs leaderboards which is a great way to import reasonably sized files with exactly what you need.

In terms of analysis, I hope this provides compelling evidence that pitchers have some control over the type of contact they give up. Exit velocity itself might not be the best way to measure it, but it seems we can learn a lot by treating hard hits as events rather than descriptors. I'm curious to see if other statistics benefit from a similar perspective.

In any case, it seems that HardHit/9 deserves its place among the few core stats we use when discussing how effective a pitcher has been and how effective he will be. It tells us more about past and future achievements than walking speed, while becoming reliable about as quickly. What's not to like?

Photos by Cliff Welch, Zach Bolinger / Icon Sportswire | Adapted by Jacob Roy (@jmrgraphics3 on IG)