The Beginner's Guide to EPA in the NFL

EPA is becoming a more and more common stat. This guide will help you understand how it works and what it's saying to you.

Anybody who's read anything I've written understands by now that I view one stat as crucial when evaluating teams, games, and Quarterbacks. Expected Points Added (EPA) is the most important stat on the football field. Using it, you can analyse past performance (as I've done plenty on here), you can predict future performance, and you can even test whatever opinions you may have, as I'll show you how to do below.

I use EPA rather flippantly in my writing, but it's come to my attention that many who read are not as well versed in it as I am. For deeper reading on it (and other advanced football subjects), I recommend rbdsm.com and their terrific 'nerd to human translator' page, which you can find right here. On there, you'll find whole articles dedicated to introducing these subjects. I don't deem it necessary to go that deep. Today, I only intend to provide a working understanding on EPA, mostly for people not well versed in stat talk.

What are Expected Points?

It starts with an Expected Points model. You may hear the word 'model' and become intimidated, but I promise it's actually quite simple. You can actually make your own basic one in just a few minutes using the following graph:

The graph shows the probability of the three scoring outcomes (field goal, touchdown, or not scoring at all) based on your field position. The probability of a touchdown is the height of the blue line. The probability of a field goal is between the blue and the purple, and this all makes sense. For example, all drives starting way in your own territory (close to 100 yards from the end zone) have around a ten percent chance of scoring a touchdown, and about a further ten percent chance to score a field goal, for a total chance of scoring between 20 and 25 percent as you can see to the left side of the graph.

As teams get closer to the end zone, naturally their probability of scoring goes up, all the way down to the one yard line, where the roughly 90 percent chance of a touchdown plus the roughly five percent chance of a field goal gives about a 95 percent chance of scoring.

You can use these probabilities to make a rudimentary Expected Points model of your own, and I'm going to give you an example. Take for example the fifty yard line. You can see in the graph the probability of scoring a touchdown is about 30 percent. The probability of scoring a field goal is about 25, which means the probability of not scoring at all is 100 subtract 30 subtract 25, for 45 percent.

Therefore, your Expected Points from possessing the ball at the fifty yard line is your 30 percent chance at seven points, or 0.3*7, plus your 25 percent chance at scoring three points, or 0.25*3, plus your 45 percent chance at scoring nothing, 0.45*0.

In this case, EP = 0.3*7 + 0.25*3 + 0.45*0.

Not complicated at all. You can plug this into your office calculator at home to get the result that possessing the ball on the 50 means you should expect to score about 2.85 points. Repeating the same calculation, but with the goal line probabilities gives EP = 0.9*7 + 0.05*3 + 0.05*0. Again, plug this into your own calculator if you want to, and see that possession on the one yard line means your team should expect to score about 6.45 points.

Both of these results make sense. More points are scored the closer you get, and so your Expected Points should be higher on the one than the fifty. You can do this same calculation for every place on the field to create your own rudimentary model if you wish. In fact, this is exactly what the NFL did when they invented this statistic all the way back in 1971.

That's right. EPA is not new. It's ancient.

There are some problems with this basic approach however. The biggest and most obvious is that it doesn't know anything about the game scenario. What down is it? First and ten on the 50 is obviously different than fourth and ten on the 50. How much time is left? Obviously the above probabilities don't hold when we're starting a drive with 13 seconds left in the first half. How many timeouts do we have? A two minute drill with two timeouts is much more likely to score than one with none.

This is where the stat nerds come in. People far smarter than you or I have been able to augment that basic model we built above to include things like down and distance, time left on the clock, number of timeouts remaining, and all kinds of things. It'd take ten minutes just to mention them all.

You don't need to know the specifics of this fancier model or its conception in order to understand its implications. Doing so essentially requires being a statistics major, something neither I nor almost anybody reading this can claim to be. All you need to know is that it tells you the amount of points a team is expected to score given the down, and field position (measured below by distance to the end zone). It's summarized in the following basic graph:

There's nothing unexpected in here. For example, a team with first down from the one yard line is expected to score more than six points. A team with first down at the fifty is expected to score roughly 2.5 points. I think these are both fairly standard football knowledge, and are approximately the same as the answer any football fan could've given. Notice how similar these results are to the ones we derived ourselves. Don't let the fancy statistical modelling fool you. This is very simple stuff.

The only thing that can sometimes trip people up about this is toward the left side of the graph, when the team gets backed toward their own end zone, and their expected points turn negative. Our basic model could not produce negative results. The worst that could happen was zero. Obviously you can't score negative points on offence, so what gives?

The answer is fairly simple. A team in this position, first down inside their own ten yard line, is more likely to improve the opponent's position than their own. For example, say a team gets first down on their own seven. From here they run three plays consisting of two runs for seven total yards and an incomplete pass. Now it's fourth and three from the 14 and it's a must punt situation. A punter standing on his own goal line is likely to generate a punt of 50 net yards or so, giving the opposing team the ball on their own 36. According to the graph above (look at the blue line 64 yards from the end zone), the opposing team is now expected to score just north of 1.5 points.

This also explains why Expected Points are negative on fourth down all the way up until field goal range at around the 40. Obviously, a punt does more for your opponent's chances of scoring than yours, because you're giving the ball away. Don't get caught up with negative expected points. All it means is that your opponent's chances of scoring on their ensuing possession are much higher than yours with your current offensive possession, given the circumstances you’re in.

The biggest benefit of using expected points as your metric of choice is that it knows the context of the game. For example, not every eight yard pass is the same. An eight yard pass on a third and seven is worth over a whole point on its own. An eight yard pass on third and 18 is a bad play. Not all incompletions are the same either. An incompletion on third down hurts the team much more than an incompletion on first down. The basic stats involved in these scenarios, passing yards and completion percentage, can't understand the difference, but Expected Points can.

How do I Make Expected Points Into EPA?

Now that we understand the concept of expected points, Expected Points Added (EPA) is very simple. For any play, subtract the offence's expected points before from their expected points after, as in the figure below, and you've generated the EPA from that play.

What EPA does is assign a value to each play in the form of points. This is no different than Passing Yards or Receiving Yards, which would each assign the above 20 yard completion a value of 20. 20 yards on the box score for the QB, and 20 yards on the box score for the receiver. EPA is no different. Each player would be assigned 1.36 points in their EPA column for this play.

What is vastly different between these two stats is that yards don't go on the scoreboard. Points do, making the 1.36 expected points gained by this play much more important than the 20 offensive yards gained. Every play (and penalty) has an associated EPA value; positive meaning it made your team more likely to score, and negative meaning the play made your team less likely to score. Add up all of the values for every play of the game, and you have the total EPA for each team and (offensive) player in the game.

This is where I reach an important point to make. Do not use EPA to judge individual players that aren't quarterbacks. Since the quarterback holds so much control over the offence, it is reasonable to assume he gets all the credit for the team's gains in Expected Points on pass plays. For receivers on the other hand, this is not true. Take for instance a receiver who on first and ten gets behind the defence on a go route. If the quarterback hits the pass, this receiver gets credit for big EPA on that target (likely a touchdown). However, if the ball is overthrown, and falls incomplete, the receiver gets credit for negative EPA for the target (moving from first to second down), despite having done nothing different.

The NFL season is not long enough for these plays to even out across receivers. As a result, the NFL leaders in EPA for receivers tend to be really wacky names that are nowhere near any top ten receiver list. There are other metrics that better correlate to actual performance for receivers (perhaps I'll touch on them in another article).

Now that we've got that out of the way, and we know that individual players (aside from QBs) are not the strong suit of EPA, let's use it to analyse teams.

Using EPA to Analyse NFL Teams:

Let's start with teams. There are six main ways to divvy up EPA. Again, there is no difference between the EPA and yards gained stats, as you can divvy EPA into pass, rush, and total offence, as well as pass, rush, and total defence.

Take note that (just like yards gained) more EPA is better on offence, and less is better on defence. If you're the defenders, you obviously want the offence to not be adding expected points to their tally against you. You want their plays to generate negative EPA. If you're on offence, it's the opposite. You want your plays to be generating positive EPA all the time.

Much more than yards gained, a stat often left in raw form, EPA is normally dealt with on a per play basis, which you get by dividing the total EPA in the above six categories by the number of plays in those categories. Here we arrive at the stat I use all the time: EPA/Play.

For example, say a team generates 15 total EPA in a game, and they had 65 offensive plays in that game. This means their average EPA for a play in that game was 15 divided by 65, for 0.23 EPA/Play.

Note: EPA/Play stats crystallise around about the 11 game mark, so if you're using them midseason before the eleventh game to judge the quality of an offense or defence, I'm happy you've come to the correct stat, but proceed with caution.

EPA/Play on all offensive plays, EPA/Play on passes, EPA/Play on runs, EPA/Play on all defensive plays, EPA/Play facing passes, and EPA/Play facing runs are all valid ways to use EPA, but for now let's just look at totals.

By EPA/Play, here were the top three total offences in 2023: 49ers, Cowboys, Bills

Bottom three total offences: Jets, Patriots, Giants

These seem to check out. They agree with what your eyes tell you when you watch these teams, but the most eye opening EPA stats (in my opinion) are on the defensive side. Defense is harder to judge with the naked eye, and traditional defensive stats such as yards and points allowed are very much functions of how long a defence is on the field, so EPA/Play is a very good measuring stick on which to judge defences. Everybody knows I think EPA/Play allowed is the best barometer of defence there is right now.

Top three total defences in 2023: Browns, Ravens, Jets

Bottom three total defences: Commanders, Cardinals, Seahawks

Again, these are reasonable results. There are no wild results coming from the different methodology here. Cleveland may be a shock to the casual fan, being middle of the road in points allowed, but they led the league in yards allowed per drive. They just had to face a lot of plays with that horrendous offence they had. Everybody knows the Browns, Ravens and Jets were the league's best three defences.

Linking EPA to Winning Using EPA Difference:

The same exercise can be done using any component of defence or offence you want, but I'm going to move on to how this actually affects winning. Analysing a team's EPA/Play difference can be a great barometer of team performance. The way you do this is by subtracting their EPA/Play allowed on defence from their EPA/play gained on offence to get a net measure of how much more they're scoring than their opponents, on a per play basis.

Generally, if a team's EPA/Play is more than their EPA/Play allowed, they're good. If they're equal, the team is okay. If their EPA/Play allowed is more than their EPA/Play, they're generally bad. This is exactly what is being shown on those 'team tiers' graphs that are always going around on Twitter. The ones that look like this:

For an example of how this works, anybody doing this basic EPA/Play analysis could have seen the falloff coming for the 2023 Eagles, who as of week twelve had a record of 10-1 despite their 0.088 EPA/Play generated on offence and 0.003 EPA/Play allowed on defence not being that much different from each other. This difference indicates a playoff level team, because it is positive, but not an elite one. By the time time the season ended, the Eagles had generated 0.058 EPA/Play on offence, and allowed 0.061 EPA/Play on defence. These are basically exactly equal, indicating the average team they were. They were bound to come down eventually.

This analysis can also go the other way. It reveals the 2023 Packers, with their 0.075 EPA/Play on offence compared to their 0.009 EPA/Play allowed for a fairly substantial 0.066 difference as probably being a bit better than their record says, and we saw that in a big way come playoff time.

The team with the biggest positive difference between their EPA/Play and their EPA/Play allowed? The San Francisco 49ers. The team with the biggest negative difference? The Carolina Panthers. These are who EPA thinks were the best and worst teams in the league. EPA thinks the worst team of all time is the 2008 Detroit Lions, who went 0-16. The best team of all time? The 2007 New England Patriots, the only 16-0 team ever. Again, it all checks out. These are also the teams I had in all these spots. Even if you disagree, it's probably only a slight disagreement.

To tell approximately how good a team is using their EPA difference, the formula is very simple. Begin with a .500 team, that is 8 wins if you're looking at years gone by, or 8.5 wins now, and add to this .500 team their EPA difference multiplied by 24. The result will obviously not give you a team's real record, but it will complete the '___ quality team' sentence, as in ten win quality, 12 win quality, etcetera.

That may seem complicated in words, but the formula is very simple:

Win Quality = 8.5 + 24*(EPA/Play - EPA/Play Allowed)

My expected wins formula I use on this page that took me days and weeks to build is approximated scarily well by this math equation that includes just two variables. To me, that's a little disheartening, but to you it's a wonderful thing, because plugging this equation into your calculator will take 30 seconds, and you can get some very useful insights out of it.

For the San Francisco 49ers, this formula gives 8.5 + 24*(0.179 - (-0.054)) = 14.04, making the 49ers a 14 win calibre team according to basic EPA analysis.

For the Carolina Panthers, the formula gives 8.5 + 24*(-0.160 - 0.029) = 3.96, meaning they probably should've won more than twice, but with rookie QBs sometimes these things happen.

For the Super Bowl Champion Kansas City Chiefs, we get 8.5 + 24*(0.021 - (-0.076)) = 10.82, meaning Kansas City (according to a basic EPA analysis) should have won about 11 games. Guess how many they won?

I'd love you to find the EPA/Play and the EPA/Play allowed for your favourite team, and plug them in. Did your team underperform like the Carolina Panthers? Did they overperform like the Philadelphia Eagles? Or did they perform just about as they should have, like the Kansas City Chiefs? You can do this for any team in any season since the start of the play tracking era in 1999, so maybe go back and see how well your favourite edition (or one of your favourites if the data can't quite cover it) of your squad stacks up. Let your imagination run wild. I promise you'll learn something, and maybe have a little fun in the process.

Why Do This Analysis?

Comparing Kansas City to San Francisco using this metric explains why the 49ers were favoured by all the super smart bettor people to win the Super Bowl. They don't build those buildings in Vegas for nothing folks. They didn't give points on Kansas City because they're advanced stat junkies. They really thought the 49ers were going to win, and they were going to make some money.

They spent the whole two weeks before the Super Bowl trying to come to this result, and we found it with a math equation that has three terms. Nothing even as complicated as the math we all did in high school has gone into any of this. All you have to do is look up a team's EPA/Play and their EPA/Play allowed on rbsdm.com, and plug them in, and you can come up with a passable metric of the quality of any given team. Just recall what I said above. I don't recommend using any EPA based metrics until about week 11 and things around he league start to settle. As long as you heed this warning you can be the smart one in your football discussions from here on using EPA analysis.

Before moving onto QBs, I will address one more thing: I have constantly compared the results of EPA analysis to the eye test throughout this section, and more or less always gotten the same results. The eye test thinks SF was the best team in the NFL, so did the numbers. The eye test thinks Carolina was the worst team in the NFL, so do the numbers. This begs the obvious question: Why don't you just watch the games and come to the same result?

The answer is simple: Yes, if you have time to watch every play of the game, you will come to the same conclusion as these numbers do. Football fans are smart about sniffing out frauds and picking a good 6-6 team from a bad 6-6 team, but most people don't have time to watch every play of the games, and it doesn't take three hours to do an EPA analysis. That's what makes these numbers so useful.

Analysing Quarterbacks Using EPA:

This is sometimes a touchy subject with people. Just because I'm okay with giving the QB 100% percent of the credit for what happens in the passing game doesn't mean you have to be. There are other stats that you can use for them, but before you tune out, here is the top five QBs in career EPA/Play since the play tracking era began in 1999: Patrick Mahomes, Peyton Manning, Aaron Rodgers, Tom Brady, Drew Brees, in that order. That's a pretty good top five list to me. If that's the result, I'm willing to look into the methodology.

The best thing about EPA/Play analysis for QBs is that it cuts through the process of watching the play unfold, and purely gives you the results of the play, in terms of the value given to that QB's team. Some people will debate whether this is a plus, but I think it is. I'll give an example. Say it's first and ten from the 25, and it's a Bears game. Justin Fields drops back and takes forever to find an open man, gets into a scramble drill, but eventually does find one for a 15 yard gain. Compare this with another game going on at the same time down in Miami, with Tua Tagovailoa at the helm, who takes his drop, hits his back foot and gets the ball out immediately on a slant pattern for a 15 yard catch and run.

Some watching these two plays will say that Justin showcased much more individual talent than Tua to get the 15 yard gain, and therefore his play was better, but others will say that if Justin had seen the pattern right off the bat (like Tua did in our hypothetical scenario) none of the other stuff would have ever had to happen, and therefore Tua's play was better. The only difference is your school of thought as to what you think the QB is there to do.

Every scheme has the QB there to do slightly different things. Therefore, apples to apples comparisons at this position are almost impossible, which makes a stat like EPA/Play invaluable in comparing the success of different people. Perhaps Justin did showcase more talent than Tua, but in terms of their team scoring points, Tua's efficient 15 yard catch and run helped the Dolphins just as much as Justin Fields' magic show.

This is the crux to understanding what EPA/Play is trying to tell you at the Quarterback position. It's not a measure of skill. A QB with a higher EPA/Play does not mean he's better physically at the position, or has more natural talent for the position, than a man with a lower EPA/Play. What it does tell you is the QB's efficiency, in terms of scoring points, compared to his opposition. If you believe (as I do) that a QB's primary job on a football field is to score points, this makes EPA/Play the prime stat to measure QB value in the NFL.

It's this logic that allows me to tell you that 2023 Tua Tagovailoa and 2023 Jordan Love, in terms of value to their teams, were essentially the same player. Tua generated 0.16 EPA/Play (about six points above expectation per game) for the Dolphins, and Jordan did the same for the Packers. This does not tell me anything about who is the better player going forward. It doesn't tell me anything about their talent level. It tells me that in terms of scoring points (the primary job for a QB in my opinion), there's nothing to choose between these two, and that in itself is an interesting insight.

Because of its obsession with results and lack of aversion towards catch and runs, EPA/Play can tend to favour game manager types such as 2022 Jimmy Garoppolo or 2023 Patrick Mahomes. However, it also gives you leeway to outscore your mistakes. As a result, EPA/Play also tends to look favourably upon mistake-prone gunslinger types, like Jameis Winston or Carson Palmer. What EPA/Play does not like very much are players that frequently give value to the other team by failing in big spots or in big ways. Being average on third down (CJ Stroud), having a bad tendency to take sacks (Joe Burrow), or having a turnover problem without the sufficient ability to make up for it by scoring in bunches (Trevor Lawrence) are all reasons this specific stat doesn't like certain players, despite them certainly passing the eye test.

In short, EPA/Play views every yard as the same, meaning that it is immune to falling for the hype that's generated by players like Michael Vick, Cam Newton, or Lamar Jackson, which is why it doesn't much like players in this mold either. Yes, watching Jackson avoid a sack to rip off a 26 yard scramble is exciting, but at the end of the day, this is no different than a Brock Purdy 26 yard catch and run in terms of the ultimate goal, scoring points.

Why Use EPA/Play to Value QBs?

Since, unlike above, EPA/Play grading of quarterbacks tends to disagree harshly with the eye test, I feel the need to provide some justification as to why you should use this metric instead of others to grade them with. Thankfully I have this justification. It comes in the form of the following:

Every study determining the correlation between QB stats and winning uncovers that EPA is the best way to predict a winner. Note that it is not EPA/Play. It's total EPA, but the meaning in general is the same. In general, the best QB stat in terms of its predictive power on who wins the game is EPA. This makes sense. If your value is higher in terms of scoring points than your opponent, of course you're going to win the game, but it is sometimes difficult to conceptualize things that way.

Also note that only 40% of variation in winning is able to be explained even by the best QB statistic, because even when we're dealing with the best QB stat, it's a team game folks. A good example is the most recent Super Bowl, where the QB with the higher EPA/Play (Brock Purdy) did not win, mostly because of turnovers he himself did not commit. These things happen, but it's important to keep in mind that just because EPA/Play is a great stat, it cannot tell you everything.

Drawbacks for Evaluating QBs

Now that we're through all the pros for using EPA/Play to grade QBs, it is now time to get into the cons. As you've noticed if you've read anything of mine, I almost never advocate for evaluating a QB strictly with EPA, preferring to exclusively use it beside a stat that's more geared towards quantifying individual ability (either CPOE or PFF grade. Take your pick). This is because one of EPA's fundamental shortcomings, even at the position it's best at quantifying, is quantifying individual performance.

For example, if a QB is very inaccurate, but somehow continuously finds ways to be very good at scoring points, it's usually a case of a mediocre QB being carried by an amazing offensive supporting cast. EPA/Play has no measures of accuracy, no measures of how good the supporting cast is, no measures of how good the opponent was, or anything of the like. What it is good at is quantifying results, and that's all. Luckily, for a football statistic, that's an extremely important thing to be good at, and gets it a seat at the QB evaluation table. Generally, if you use EPA to be your results measure, and either CPOE or PFF Grade to be your individual ability measure, and combine the two into one number (because no matter how good you are as a QB, results have to come into the evaluation in some capacity. I'm looking at you Justin Herbert), you can get a very good indicator of QB performance.

It's regrettable that there is no one number that's quite good enough to judge QB performance with, but getting away with using just two is pretty good most of the time.

EPA Benchmarks:

Now that I've said everything I have to say, I'll provide some benchmarks to compare your team against. I can tell you everything I know about EPA/Play and all that it means, but ultimately if I tell you that the New York Jets allowed -0.11 EPA/Play against them, what does that mean? When you're out on your own looking at EPA numbers, think back to these benchmark values I'm going to give you to assign your own meaning to what you're seeing.

In 2023, the median defence was the Chicago Bears, who allowed teams to generate -0.022 EPA/Play on them, roughly 26 points below expected over the course of the whole season. Remember since we're talking about defence, lower is better. This year, it took -0.052 EPA/Play to be a top ten defence. If you can get to -0.08, that's 80 points less than expected over the whole season. You've got a defence that can carry a team. Any better than that is quite dominant. See the levels to NFL defence below:

0.091 EPA/Play Allowed: Worst defence in the NFL in 2023 (Washington Commanders)

-0.022: League average defence (Example: Chicago Bears)

-0.052 or lower: Top ten defence (Example: San Francisco 49ers)

-0.08 or lower: Defence that can carry even a quite bad offence to a great season (Example: Baltimore Ravens)

-0.155: 2023's best defence (Cleveland Browns)

-0.221: Best defence of the play tracking era (2000 Baltimore Ravens)

When looking at your team's EPA/Play allowed, see where they fall compared to these benchmarks, and decide from there whether they were great or bad or somewhere in the middle.

The same approach is valid for offence. In 2023, the median NFL offence (Tennessee Titans) generated -0.024 EPA/Play, roughly 24 points below expectation over the whole season. Looking at your team's EPA/Play numbers, if it's higher than the median, it's a good offence. The criteria to crack the top ten is 0.031 EPA/Play (34 points over expectation over a whole season). If your team's offence meets this threshold, it's a very good offence. If your team can get to 0.1 EPA/Play, it's a Super Bowl calibre offence. Keep all of these thresholds in mind as you're looking at your team's EPA data next season:

-0.258 EPA/Play: NFL's worst offence of the play tracking era (2012 Arizona Cardinals)

-0.23 EPA/Play: Worst offence in 2023 (New York Jets)

-0.024 EPA/Play: League average offence (Tennessee Titans)

0.031 EPA/Play: Top ten offence (Example: Detroit Lions)

0.1 EPA/Play: Super Bowl calibre offence (Example: Dallas Cowboys)

0.179 EPA/Play: Best offence in 2023 (San Francisco 49ers)

0.256 EPA/Play: Best NFL offence of the play tracking era (2007 New England Patriots)

Much like the above on defence, when you see a team's offensive EPA/Play values, be it because you're reading something I've written or just out in the football world in general, compare it to these benchmarks. It'll help you get a gauge for what you're actually reading.

Since QBs are much more valuable than average players, it's easier for them to generate higher value. As such, their tiers are higher than the offence as a whole. For a singular game, EPA/Play is subject to the empirical rule. This is a statistical concept that you don't have to know. All you need to understand is the tiers.

68 percent of all QB performances fall between -0.211 and 0.211 EPA/Play. The average game will fall in this window. This is the window in which your QB hasn't won or lost you the game. They obviously have good and bad games within this window, but they cannot have had a truly great or truly awful game if they fall inside this window.

95 percent of all performances lie between -0.422 and 0.422 EPA/Play. This is where the great or the awful games lie. A performance between -0.422 and -0.211 EPA/Play means your QB has played truly badly. Bad enough to lose the game for his team. Likewise, a performance between 0.211 and 0.422 EPA/Play means your QB had played very good. He's played well enough to win even despite a bad team performance.

99 percent of all performances lie between -0.633 and 0.633 EPA/Play. In these tiers is where all reasonable games lie. The 4 percent of games yet unaccounted for in the -0.633 to -0.422 or 0.422 to 0.633 zones are the horrendous and the elite. You'll see games like that a few times per week.

The remaining one percent lie outside the -0.633 to 0.633 bounds. This is where the all time greats and the worst games of all time lie. You will see only a few games outside these bounds in a given season. Probably less than ten. The 2022 season had nine. Think back to my piece about about 2006 Rex Grossman, where he himself had four of them (three bad, one good), and it's yet the more evidence how much of an enigma that man was. When analysing a QB's performance in a game, use the below benchmarks.

-0.633 EPA/Play or worse: Unfathomably bad. If a QB were to run a QB spike every play all game, it would not result in an EPA/Play this bad. Most QBs will not have a game this bad in their entire career. (Example: Infamous Rex Grossman performance against the Cardinals in 2006)

Between -0.633 and -0.422 EPA/Play: Truly awful QB play. If your QB ends up here on a given day, it's truly disappointing, no matter how bad they are, but most QBs will have a game this bad in their careers. Having a QB play this badly makes it almost impossible for your team to win (Example: Kyler Murray vs Los Angeles in the 2021 Wild Card)

Between -0.422 and -0.211 EPA/Play: Very bad play. Bad enough play for the QB to have lost the game for his team, independent of how well they played. (Example: Tua Tagovailoa in the 2023 Wild Card)

Between -0.211 and 0.211 EPA/Play: In this range, the QB has either generated slightly positive or slightly negative value for his team. Whatever side of zero he's on, he's not done enough to decide the outcome of the game on his own. He's not won or lost the game for you. Most QB performances fall in this window (Example: Both participants in the 2023 Super Bowl)

Between 0.211 and 0.422 EPA/Play: A very good game. Generally, a team with a QB who's performed in this range will win, regardless of anything happening around them, although it is still possible to lose if the opposing QB does the same thing. (Example: Both participants in the Lions vs Rams 2023 Wild Card matchup)

Between 0.422 and 0.633 EPA/Play: An elite game. Not seen very often, a QB playing in this tier virtually guarantees their team a win, barring a miracle on the other side. (Example: Patrick Mahomes in the 2022 Super Bowl)

Above 0.633 EPA/Play: Rarified air. If you see a QB above this benchmark, you've just seen one of the greatest games of all time. Leaguewide, this mark will be hit approximately five times per season. Most QBs will finish their careers without a game this good. (Example: Jordan Love in the 2023 Wild Card)

Now when you're looking at your QB's EPA/Play stats (either season long or for one game), you can understand how great it was by placing them in one of these tiers.

Hopefully, now that we understand the benchmarks for EPA/Play on offence, defence, and for the QB, it will ease our understanding of looking at the EPA numbers from now on.

Conclusions and Future Reading:

We've now learned everything their is to know about EPA, and how it should be used as a stat when evaluating the NFL. We've created our own Expected Points model. We've gone over how that translates into EPA, a metric that shows the value of any individual play in terms of points.

We've learned how to use this fancy metric to analyse teams, and their QBs, gone over its pros and cons, and we've also learned roughly what numbers we should be looking for. Now that we know how to do all of this, my hope is that we can be the informed ones when either trying to predict outcomes of NFL games, or seeing how good our teams truly are, or just chatting with our buddies about what's going on around the league. In addition, if you ever see any advanced football writing, either mine or anybody else's, using EPA flippantly as part of their analysis, I now hope that you would feel more comfortable reading that article on your own. If that's the case for even one of you reading, this introductory guide was well worth my time writing.

Keep in mind as you move along that this was just a surface level introduction to what EPA is and how it can be used, intended to build a basic understanding. There are countless things that can be done with EPA that I did not mention here, such as opponent adjusted EPA (particularly helpful for college football analysis), or EPA adjusted for quality of supporting cast for QBs (intended to help rectify the 'needing a second stat' problem). If you want to read further into what those are and what they're useful for, I recommend the Substack page Unexpected Points, who does a good job explaining lots of advanced metric topics from around the football world.

This concludes the analysis, and you are clear to stop reading from here, but if you still feel like reading even after all this, below are some footnotes about the practicalities of using EPA for analysis, and some fun questions you can answer with it.

Footnotes:

Which EPA Model Do I Use?

If you've done any looking around the internet, either during this reading or beforehand, regarding EPA, you will have noticed there's more than one version of the stat. How can this be?

Regrettably, due to the complicated nature of building an Expected Points model, and different people's desire to make different choices in the process, there are at least four versions out there: ESPN's version; Next Gen Stats' version; PFF's version, and NFLFastR's version, and those are just the major ones. Which one do you use?

First and foremost, who am I to tell you which to use? You are allowed to use whichever version of Expected Points you'd like, but I have reservations about all of the major corporate ones, because none of them are reproducible.

All the graphs I've shown in this piece are based on NFLFastR's Expected Points model, because the other three are all private, meaning ESPN, Next Gen Stats, and PFF all own the copyright to their respective mathematical formulas. They don't share their formulas with us, and therefore we don't know what they are. Although NFLFastR's model is dense and extremely difficult to understand, it is in fact public, meaning technically if you wanted to and somehow had the statistical abilities, you could remake it yourself.

What this means is that they can't mess with it without anybody catching on. The results are the results. The same is not true for the other three. What if ESPN didn't like their EPA/Play leaderboard and decided to just change the formula midstream in order to change the results? There's nothing stopping them. None of us would ever be able to tell, because we didn't know the original formula.

Do I think this would ever happen? No, but I know it certainly could, and the possibility alone makes it so that I have to recommend the NFLFastR version.

Where Do I Find EPA Data?

If during the course of the exercises I prescribed above you went looking for your team or QB's EPA data, you will have noticed it's not on Football Reference or StatMuse or anywhere more basic stats are kept. To find EPA Data, you have to dig a little deeper.

If you've chosen above to use any of the corporate data models, the only way to obtain it is to pay for their services and download it from there, but there is a free way to go about using NFLFastR.

My most comfortable answer is to tell you to use the NFLFastR R package, but doing so requires a slight knowledge of computer code (R being a programming language). Not a big one, but enough of a barrier to make me uncomfortable with making that my only answer.

Thankfully, there exists the excellent resource rbsdm.com, on which you can view box scores for single games going back to 2003, view and download full season EPA stats for both teams and QBs currently going back to 2012, and likely back to 1999 once the website is optimized during this offseason, and lots of other fun things too. This is the exact same website I use to source all the data for almost everything I do on this page, so you'd be going right to the top of the food chain. Anything you see in any of my articles can likely also be found there. If anybody would like, I can give a tutorial on how to use the site.

What Fun Questions Can I Answer With EPA?

I'll give a couple examples below that just scratch the surface of things you can answer with EPA analysis. With these answered, perhaps you can think of some more questions to answer yourself.

Why Has the NFL Become so Pass-Heavy?

It comes down to the average value of the play. The average pass makes teams more likely to score. The average run makes a team less likely to score. Therefore, the average play should be a pass, and not a run. There are non-scoring benefits to running, such as running more time off the clock, but as far as scoring points, passing is the way to go.

How Many Yards Should a Play Gain to Be Positive?

As you can see, almost everywhere on the field a second and two is about as good as a first and ten, meaning first down plays should aim to achieve eight yards or more in order to increase their likelihood of scoring. On downs that are not first, the answer to this question is generally how many yards are left to the first down marker.

Additionally, you can see that nowhere on the field is a third down and short as good as a first down and ten, making the concept of 'third and manageable' not as true as it may seem on the surface.

There are other questions I would recommend looking into for yourself, such as why are teams throwing to backs so much these days? Why are fourth down tries on the rise? Why do teams not take shots on second and one? There are plenty of things you can look into, and all of it can be done with your new knowledge of EPA.