Step with me into the future. We’re covering the most advanced, the most puzzling, the most complicated plays in NFL history today.

The two quarterback offense.

What I’m about to show you cannot be unseen. Your third eye is about to open. Join me in football enlightenment.

We haven’t seen a team split time between two starting quarterbacks un-influenced by injury circumstances since the Los Angeles Rams in 1950:

But that didn’t stop individual snaps from occurring with 2 QBs on the field simultaneously. Those of you who have subscribed for a while know that I love whacky offensive plays.

The GUT

Given NFL snap participation data since 2016,

Use simple SQL queries

To find and aggregate snaps where two or more quarterbacks are on the field, and see who’s performing best on these snaps.

October 28, 2018

The most pivotal day in the multi-quarterback ideology’s history. Saints are playing the Vikings in Minnesota. Sean Payton breaks out a mind-bending THREE quarterback formation. Drew Brees is lined up on the bottom left of the screen as a traditional x receiver, Teddy Bridgewater is at the top of the screen as the y receiver. Taysom Hill is under center.

This is it. The day the NFL changes forever. Blastoff, offenses will never be the same.

Taysom Hill is stuffed in the backfield for no gain. Failure to launch Progress is halted.

Speaking of Taysom Hill

Are we experiencing a boom in 2 QB formations?

Maybe not. Perhaps I’ve over-estimated the impacts of this futuristic offense. There were only 7 2 QB snaps in 2025 that didn’t involve Taysom Hill.

It’s truly Taysom Hill saturated, especially when we look at 2 QB pairings:

Back to 2D for this one. A lot of these are one-off seasonal experiments that were abandoned, like the Joe Flacco-Lamar Jackson shared snaps in 2018 and the 2020 Jalen Hurts/Carson Wentz junk. And a lot of these plays didn’t do too well in terms of efficiency. Another scenario where the positives and negatives are so tiny that I considered un-redding and un-greening the EPA/Play values.

Is the 2 QB Offense Back?

Short, heartbreaking answer, no. For something to be back implies that it was here to begin with in the first place. Taysom Hill is currently a free agent, there’s not much 2 quarterbacking going on outside his sphere of influence. The 2 QB era was here and we didn’t appreciate it. And now it’s gone before we noticed. Like a flash in a pan. Perhaps the league isn’t ready yet, maybe in another 100 years.

That’s it for today, thanks everyone. Will be back in your inboxes tomorrow with another data engineering tutorial.

Sorry for the delay in the usual Friday tutorial. As promised, here’s the code I used to create 3D bar charts in R on our jumbo formation write-up from two weeks ago. I have two tangents to go on first though.

Football Data Engineering Tutorial Update

Thanks to my friend (who you all should subscribe to for great football content) for reporting a bug to me in my football data engineering tutorial repository. Apparently none of you can open the football.duckdb file I’m creating as our database because DuckDB only allows one reader and one writer profile per database. To see the data, I recommend finding where I write the database, re-name the .duckdb file, and run the script this will now give you your own local copy. I will also be uploading a SQLite file in the same database for next week’s tutorial to make this more accessible. And let the record show this is now the second time I’ve tried to show DuckDB off in a more production-ish setting and the second time it has resulted in me returning to SQLite instead for a second time. I’m beginning to think DuckDB’s applications aren’t production ready yet, but it is a cool way to read large files quickly on your local.

Anyways, just wanted to provide an update there and let you know if you’re having issues with that repository, you aren’t crazy. Just level setting because I didn’t have the time to return to it this week. Appreciate the patience and here’s some R code for making 3D bar charts using ggplot, thanks everyone.

3D Bar Charts

Read more

Everyone’s focused on last Thursday’s NFL schedule release for the 2026 season so that’s what I’m going to cover today. There’s tons of great material already out there on the league-wide outlook so today I’m going to focus on two teams who finished in 4th place in their divisions, and how different their 2026 seasons look. Hint: both teams have employed Anquan Boldin.

Beyond that, we’re going to analyze the so-called “fourth place advantage” in football and see if it actually exists. And if it exists, how strong is it? Because when your team finishes in fourth place, they’re supposed to receive an easier schedule the following season. Is this actually the case? This write-up is all about fourth place. I feel like Jackie Moon promoting the Flint Michigan Mega Bowl.

Let’s hop in.

The GUT

Given past season NFL data and past Vegas win total odds…

Use SQL queries and some simple data analysis…

To see if 4th place division NFL teams actually receive easier schedules than their 3rd place counterparts.

The Data

Nothing fancy here. I feel like I say that every week. I pulled schedule data from nflreadR and some historical Vegas win totals information from covers.com just to compare strengths of schedules.

This is a line chart I tweeted as a sneak peak to this write-up:

So this is the main question I’m looking to answer. Does the 4th place advantage exist? And if it exists, how does it hold up?

The Anatomy of an NFL Schedule

Here’s a quick little bar showing the breakdown of an NFL schedule:

There’s only 3 games that are completely different from your division-mates. So not much wiggle room really.

Luck of the Draw: Lions and Cardinals

Who knew this answer based on the Anquan Boldin hint? Kudos if you did. According to pre-season Vegas win totals, the Detroit Lions have the easiest schedule in the NFL, and the Arizona Cardinals have the hardest.

The 4th place advantage vanishes based on the rotating division schedule. The league did try their best though. If you average the win totals of the 3 same-place opponents of 4th place finishers vs. 3rd place finishers, the 3rd place finishers face teams with 1.62 more Vegas pre-season wins totals. But let’s return to the line graph above. Across 17 years of data, the average 4th place finishing team gets a full-season schedule that’s .07 wins easier than the average first place finisher. Just look at the y-axis again. Not as dramatic as it initially seemed, huh?

Final verdict: a spread of .07 wins isn’t worth spending more time noodling on. League parity is at an all-time high, and your team’s season isn’t destined to shake out one way or another based on sportsbook odds and schedule releases.

Would love to hear all your thoughts on this in the comments section below. Shorter one today.

A Pep Talk

The tech industry in general is…experiencing a downturn to put it lightly. This avalanche of layoff announcements has led to nothing but virulent pessimism seeping into every aspect of my digital life. Twitter’s ‘For You’ page is a nightmare, Reddit is even worse. I feel the need to cut through all this negativity and offer a different perspective. Hang in there and stay positive. As someone who’s wasted days, months, and years of his life marinating in negativity, I’m here to tell you that it doesn’t lead to anything good. Log off if you feel yourself gravitating towards the doomsday crowd. Please reach out if you need anything, via comments, emails, or Twitter DMs. In my experience, everything shakes out the way it’s supposed to.

If we must, let’s tie this back into the NFL:

Be You! That’s all anyone can ask of you.

That’s all for today, thanks everyone - RC

Another Friday, another tutorial. This week we’re back to data engineering, as I’m going to show you exactly how I queried jumbo plays in nflfastR. In case you missed it, I’ve published a series of football data engineering tutorials to go along with my data viz tutorials and give you the full scope of how I work with football data. This week’s tutorial includes another contribution to this GitHub repository.

This is part 3 of a data engineering tutorial series using nflfastR data.

Here’s the intro tutorial if you missed it where I discuss exactly what we’re building together.

Here’s part 2 if you missed it where we create an ETL pipeline.

This is the code needed to generate the datasets I used on Tuesday’s jumbo write-up.

Let’s dive in.

Where we left off

We’ve got our data pipelines running in the form of dbt models (SQL code orchestrated with YAML) into a DuckDB database of NFL play-by-play data, all hosted nicely in a Docker container so it works across machines. The only real transformation we’ve done to the data is change the abbreviation for the Los Angeles Rams from LA to LAR. Keeping things light for now. Tonight, we’re going to bring in another data table from nflfastR to go alongside the play by play data, participation data.

a

Step 1: Python script to download participation data

We’re using this script to download the participation data:

import sys
from pathlib import Path
import requests

  DEFAULT_SEASON = 2025
  RELEASE_URL = (
      "https://github.com/nflverse/nflverse-data/releases/download/"
      "pbp_participation/pbp_participation_{season}.csv.gz"
  )

  season = int(sys.argv[1]) if len(sys.argv) > 1 else DEFAULT_SEASON
  url = RELEASE_URL.format(season=season)

  dest = Path("data/bronze")
  dest.mkdir(parents=True, exist_ok=True)

  resp = requests.get(url, timeout=120)
  resp.raise_for_status()

  file_path = dest / f"participation_{season}.csv.gz"
  file_path.write_bytes(resp.content)

  print(f"downloaded {len(resp.content) / 1_048_576:.1f} MB to {file_path}")

This should look similar because it’s just the same script as the pbp data but for pbp_participation. If you need help running this script in the command line, please refer to the other tutorials.

Step 2: Staging Model

Read more

If you’ve had a casual conversation with me in the last year that happened to delve into the subject of football, I’ve brought up jumbo formations to you. They’re the football plays where there’s more than 5 offensive lineman on the field. And without any further rambling, that’s the topic we’re going to look into today. We’re going to answer questions like, “Can we query jumbo formation data? How reliable is this data? Which teams use jumbo the most? Which game situations call for jumbo formations more? Is the NFL getting bigger?” Let’s dive in.

Why?

There’s been a lot of talk lately surrounding NFL teams playing bigger. This coincided with an abundance of tight ends taken in this year’s NFL Draft. If you’ve checked raycarp.com lately, you’ll notice that league-wide trends of 12 personnel (offensive formations with two tight ends) and 13 personnel (three tight ends) plays are up year over year. But does this same trend exist for jumbo formations? I haven’t seen anyone talking about it lately so I’m taking it upon myself to dig into the data.

The GUT

Given nflfastR’s play by play and play participation data,

Use SQL queries with regex

To see if consensus hunches about league-wide offensive trends of more jumbo personnel plays are true or not.

Placing another offensive lineman on the field requires them to report as eligible to the referees so they can serve as another receiver on a trick play. You may remember this example of Taylor Decker lining up on the left edge of the Lions offensive line for this overturned trick 2 point conversion. This play was overturned because allegedly Decker did not report as an eligible receiver to the referee correctly. Some really interesting pre-huddle stuff occurs on these plays in terms of keeping this eligible/ineligible reporting concealed to the defense yet visible to the referees. This game within the game stuff is why I enjoy watching football so much.

A player with a jersey number of 50-79 and 90-99 has to report in to the referee as eligible on offensive plays, unless they line up in an ineligible position. Ineligible positions include lining up between two offensive linemen (guards on every play, for example). The reporting in as eligible is only required when a player in those number ranges lines up in an eligible spot, like at either end of the offensive line or in the backfield.

The Data

From nflfastR we have play participation data from the 2016 NFL season until today. That helps with our cross-validation for real jumbo plays because we can cross reference plays where the desc column describes someone as reporting in as eligible with plays where there are six players with the position ‘OL’ on the field. Before the 2016 season, it’s tougher to validate so those will be excluded from the report. Ideally I’d be able to give a more comprehensive trend analysis but we’re going to work with the 10 seasons of data we’re given. I got some funky numbers when trying to go further back in time. And I suspect that trend analyses that open up the dataset to the pre-2016 analysis may not actually indicate upwards or downwards trends, rather an improvement in actually tracking formations.

Disclaimer established, I still made a chart for this data since the 2009 season, so I will go ahead and share that with you now:

Footnote: These numbers are wrong. They just show you the total counts when only conducting a regex text parsing filter on NFL play by play data for instances of players reporting in as eligible.

Messing around with some new fonts and 3D bar charts in R recently if you can’t tell. The mango tango colored dotted line indicates the pre-participation data era and the post-participation data era. I suspect that my text parsing only queries on the play-by-play data just aren’t correctly pulling in all the jumbo formation data, but look at 2016-2025. Let’s now try and better verify this trend. Did teams really use more jumbo plays than ever this season? We can get this number by using the play by play data, the participation data, and the roster info data in unison. First, we filter the participation data down to where six or more players listed in that column have the position ‘OL’ on their respective rosters. Then, we join those to the pbp data using play ids and game ids and filter play description columns again for the reports in as eligible to get rid of noise like extra points and punts. And this is what we’re left with:

Now we have correct data. Or as correct as we can possibly get. This now debunks the hypothesis of an aggressive jump in jumbo formation plays for 2025. Like we saw in the pace of play write-up, we’re returning to norms. Still though, the 2016 NFL season’s large number raises some questions.

This is another 3d bar chart showing the different methods I used to filter this data down, and how I arrived at this all blue bar chart above from 2016-2025.

This is a breakdown of where each play came from with the filtering between the raw text string regex, the participation data/roster info filtering, and both at the same time. And as you can tell I’ve been into 3D bar charts lately. I also messed around with expanding the roster filtering based on jersey number since any player with a jersey number of 50-79 and 90-99 has to report in as eligible. But two things are true as it stands. First, yes, the league saw an increase in jumbo plays from 2024 to 2025. And no, this isn’t indicative of any sweeping trends.

But we’re talking about two distinct things here

Every play with someone reported in as eligible is a jumbo formation play, but not every jumbo formation play has someone report in as eligible. Check out this play from last season where Vita Vea comes in as a fullback for the Buccaneers on the goal line. This is the middle of the venn diagram. This is a jumbo formation play with an extra nose tackle in there too just for fun. The Buccaneers needed two players to report in as eligible to the refs here, Vea in the backfield and Charles Heck all the way on the left edge of the offensive line. This is supposed to provide ample warning for the Falcons defense, but in a game theory way, wouldn’t this confuse them more? The offense is letting the defense know they need to account for a potential throw to Heck, while signaling a run play with their lineup. The result of this play was a Sean Tucker touchdown run. It’s almost as if you were sitting at the poker table and someone across from you announces they have pocket aces. How would this affect your approach to the hand?

Adopters of Jumbo

Here’s a diverging bar chart (back to 2D) comparing 2024 jumbo formation play counts to 2025 jumbo formation play counts in the regular season:

The Texans and Steelers really enjoyed using jumbo formations last season. This is another chart that throws a wrench into the sweeping conclusions of the league playing bigger. For every team adopting the jumbo formation, there’s an almost equal mirror team of a jumbo abandoner.

And it would make sense that the two most reported in as eligible players were offensive linemen on the two leading jumbo formation teams:

Although these numbers are high, they’re not the highest season play counts of reporting in as eligible. Those two honors belong to the same player on the same team in back to back years, George Fant in 2018 and 2019 on the Seattle Seahawks. And Fant actually had 2 targets during this jumbo snap phase, 1 of which turned into one reception for 9 yards. 331 Jumbo snaps for Fant turned into 1 catch for 9 yards. Personally, I feel as if he deserved some more targets in that span.

When do teams use jumbo?

I wanted to see distributions of jumbo plays across situations based on down and score differentials. It seemed like field position didn’t provide much insight, teams who used jumbo used it all over the field. But different score differential brackets showed a slight uptick in jumbo formation plays in every bracket:

The more your team leads by, the more jumbo formation plays they call. Only a slight increase, but still thought it was worth sharing. I’m sure you’re asking, isn’t this just a chart showing when teams run the ball more? And my answer to that would be mostly yes, but with an important distinction. A run play distribution by score differential chart shows steady increases as the lead increases. But jumbo formation plays decrease when a game is essentially out of reach (leading by 17+ bar) while run plays still increase. Perhaps teams don’t want to fatigue another offensive lineman when they don’t have to to seal a victory.

And for these teams who use jumbo often, how do the run/pass splits look?

More often than not, a jumbo formation means a run play for the offense. This comes through when looking at both splits together like so.

So is the league playing bigger? Some teams are, mainly 8 teams are spamming the jumbo plays. But these aren’t exactly offenses I’d categorize as prolific or ones that teams are chomping at the bit to emulate. The 49ers and Rams actually ran zero plays with extra offensive linemen last season. As we look forward into 2026, I’ll be watching for any new adopters. I also think there are other ways to “play big” in the NFL, like using more 12 and 13 personnel that I didn’t get into this week. We’ll save those for another write-up. That’s all I have for today though, I’ll be back in your inboxes on Friday with another tutorial. Thanks everyone, talk soon - RC

We’ll get back to data engineering with next Friday’s tutorial, but today I wanted to give you the SQL query I used in an R script to get pretty close to play clock numbers, as well as a tutorial for doing something I found annoying: Putting your own custom font on a table using embedded HTML.

Remember, I’m not trying to say I’ve re-invented the wheel or solved the play clock remaining field being 0 issue in nflfastR. I’m just showing you what I tried that yielded interesting results.

Let’s dive in:

Play Clock Query

Read more

And congratulations on making it to the real NFL off-season. The NFL Draft is over so we must find a new focus. This week’s write-up is going to focus on pace of play in the NFL. I hope you enjoy it, and let me know your thoughts in the comments below.

Why?

Last weekend I attended the American Statistical Association’s Virtual Sports Analytics Conference. The conference’s first speaker was Michael Lopez, the head of data at the NFL. Dr. Lopez mentioned that the NFL is averaging 10 fewer plays per game than it did a decade ago. Fewer plays complicates data analyses slightly by reducing sample size. But the motivation of this write-up is to dig into this statement by Lopez, try to find out why, and to spark some conversations about pace of play in the NFL.

Hopefully, this write-up resonates with you as a conversation starter. I’m not trying to propose that I’ve solved this problem or that I’ve solved quantifying pace of play. I just have a lot of questions and write a lot of SQL. But pace of play dominates almost every other sport. It’s the reason why the MLB implemented the pitch clock. It’s why Ryan Hammer includes tempo in his college basketball trapezoid of excellence. It’s why your favorite NBA players keep rupturing their achilles tendons. It’s why you hate the group in front of you on the golf course with four carts.

The Data

We’re using our reliable free data source for this one again. We’re turning to nflfastR and its wide array of time columns, both real times and game times. I’ll explain later.

You’ll recognize one of these metrics if you read my Football’s Trapezoid of Excellence piece back in March. Also, I didn’t invent these. Here’s a table of two of these three metrics I’m going to be looking into today from Sharp Football Analysis.

The GUT

Given nflfastR play by play data from 1999 to 2025…

Use SQL queries as strings in an R script…

To analyze 2 different pace and tempo metrics for NFL teams, and propose a new one.

What’s a play in this context? A play is an offensive snap without a penalty. I’m filtering out punts, kickoffs, field goals, and extra point attempts. I’ve kept in spikes and QB kneels. The key distinction here between an offensive play and an offensive snap is the penalty portion. Snaps count the penalty nullified events, while the plays count I’m calculating here doesn’t.

Metric 1: Plays Per Game (PPG)

This one was quick because we have the query from March. Just have to edit some date filters to replicate it across seasons. Here are the results:

League average in 2025-26 was 61.5 plays per game. This is the lowest total since 1999. The highest across the timespan was 2013 with 65.2 plays per game.

Making a line chart for a y axis that goes from 60 to 66 feels silly, so I’m going to comment out the restrictive filters for the query and make the viz again. When we add special teams and penalties back in, the chart looks more like this:

Same high (2013) with 175.4 plays per game. The lowest total was actually in 2000 with 161.5 plays per game. This year’s total was 165.7 plays per game, and we can see the downward trend more. When we look at 2025 only, here’s how teams rank:

Interesting tidbit, both the highest plays per game and lowest plays per game all time teams had Chip Kelly on their staffs. Chip Kelly was the Eagles head coach in 2013 and the Raiders offensive coordinator for part of 2025.

Metric 2: Seconds Per Play (SPP)

Seconds per play is more aimed at quantifying offensive tempo. I’m using nflfastR’s game_seconds_remaining field here in a SQL LAG query, which is a fancy word for putting the current row and the previous row in the same row so that we can compare. Taking the game seconds remaining before the play and after the play can give us an idea of how long each NFL team takes before and during plays. For these results, I filtered out 1999 because there were huge gaps in this column being filled on a play by play basis. Here are those results:

While this y-axis is silly again, Putting spikes, kneels, and touchbacks back into this would just mess with it even further as those plays take weird amount of seconds off the clock. If a QB kneels with 38 seconds left, that play would technically take 38 seconds off the clock. And spikes ideally take 1 second or less. I don’t including these outliers would help inform us.

On this chart, a higher dot on our line means a slower pace for league offenses. Teams were much faster in the early 2010s. Over 60 plays, the 2025 seconds per play pace can eat up 2 more precious game clock minutes than the 2013 pace would.

When we look at offensive seconds per play on a per team basis, here are our results:

So fast offenses still exist in 2025, like the Saints and Cowboys. When we look at PPG and SPP on a scatter plot, here’s how teams rank in percentiles:

Good offenses can measure poorly in both categories. The Patriots and Seahawks are in the slow/low percentile quadrant after all. There’s a cluster of last year’s bad offenses (Browns, Raiders, Jets, Titans) in the bottom right quadrant, however. The league’s worst offenses were playing at a fast tempo but didn’t have a lot of plays per game. This just means they didn’t see the field too often, and when they did see the field, they were trailing and had to do a lot with a little. I actually don’t think this scatter plot is insightful at all, since we can see these trends with fewer measurements.

Metric 3: Estimated Time Remaining on the Play Clock Using Eligible Plays

As you can tell the creative name juices are flowing. The play clock is usually 40 seconds long. After stoppages in play, it’s 25 seconds long. There is a play_clock field in nflfastR but all its values are zeroes. The issue here is probably the same one that I encountered. There are a few seconds unaccounted for in between the stoppage in play and the referee blowing their whistle to resume play.

So as my contribution to the pace of play conversation, I filtered to a clean dataset where I know for certain the play clock was 40 seconds long. This is a subset per team of in-bounds rushing plays and completed passes with no penalties, timeouts, or turnovers, as well as excluding kneeldowns. Using these filters, here’s what a line chart of league-wide averages of estimated seconds remaining on the play clock looks like from 2000-2025:

So time remaining on the play clock is trending downwards on these 40 second play clock plays. 2025 was the first season in the dataset that had a number slightly below 8 seconds. Here’s how the team bar chart looks from the 2025 regular season:

I checked this for correlation with no huddle rates and the Saints ranked first in both. The leaderboard is also extremely similar to the seconds per play leaderboard. It just took way more lines of SQL to calculate. Unsure if that’s a good thing or bad thing at the moment.

My smoke test for this metric was Chip Kelly’s 2013 Eagles team, and they ranked first averaging 15 seconds remaining on the play clock in these situations while the rest of the league averaged 9 seconds remaining.

I spot checked this metric on 30 plays using the film room on Next Gen Stats and the average error was around 1 second per play. No huddles proved to be the biggest challenges for this and they had the widest margins of error. So even with this narrow pool, I’m still facing challenges on directly isolating how much time is left in the play clock using nflfastR data. But it’s pretty cool to get within 1ish seconds on around 16,000 plays.

B-List of metrics

• Plays per drive stats weren’t really that different from plays per game. Same teams at the top and bottom.

• I checked for rule changes around that time or increases in specific penalty frequencies. Nothing conclusive there.

• Real time drive duration (RTDD): Looked into the time_of_day nflfastR column to see if we could pull out any insights in exactly how long each drive takes per team. Also checked to see if 1pm, 4pm, and primetime games experience different RTDDs.

  

  These y axes feel like a bit at this point. Interesting to see that there’s no noticeable difference in real time drive durations across the different broadcast slots. I would’ve assumed that primetime = more commercials = higher RTDDs, which is true here but only by a few seconds.

• Yards per real time minute: Trying to borrow from soccer, combining RTDD and yards gained as an absolute value but there weren’t any meaningful insights.

My Guesses

If someone asked me why there are fewer plays in the NFL now than there were 10 years ago, my answer would include:

• There were roughly 9 more passing attempts per game back then than there are now. Maybe we experienced a passing renaissance with the peaks of Peyton Manning, Phillip Rivers, Aaron Rodgers, Tom Brady, and Drew Brees at the same time, and those are actually outlier years. More pass attempts = more incomplete passes = more clock stoppages = more plays.

• More pre-snap motion. Motion rates are up year over year since 2019, as far as Next Gen Stats team offensive splits go back. I don’t see a source for pre-2019 motion data, but my hunch tells me the modern game has more pre-snap motion than we had 10 years ago. On a team by team level, more motion doesn’t relate to more play clock used though.

• There are fewer turnovers now than there were back then. In 2015, a fumble occurred every 123 plays in the NFL at a .81% rate. Now it’s a fumble every 150 plays (.67%). Same with interceptions. In 2015 2.15% of all pass attempts were intercepted, now it’s 1.92%. Fewer turnovers means fewer drives, which means fewer plays.

• The game itself is getting more advanced, which leads to small increases in calling individual plays. It’s taking longer for a play caller to decide which play to call, and also taking longer for quarterbacks to read the defense. Over time, these extra seconds spent in the huddle compound into more clock chewing. Downstream copycats of Mike MacDonald’s Seahawks defense will only magnify this.

• It’s death by a thousand cuts. There’s no specific thing we can point to (at least with nflfastR data) that is the one reason why we have fewer plays than we did ten years ago in NFL games. The dynamic kickoff was also researched, as well as different touchback field position yardlines and those results were inconclusive too. It’s a mix of more runs, fewer turnovers, more motion, and tiny increases in the amount of play clock seconds used by offenses.

Here’s one last table of almost every metric I looked into to try and find trends:

Yes, I even looked at punt frequency. I’m deep in the weeds.

Leaving this one open-ended, I’d love to hear your thoughts on this.

Constraints

My wishlist includes metrics like huddle break times and play clock snap times for all plays but those are either paywalled or don’t exist yet. I remember Deniz Selman referencing the time it takes to break the huddle with tweets about the fun Eagles 2025 offense. These look walled behind TruMedia though and I’m not sure how they’re getting this data unless they’re manually charting or have an advanced huddle break computer vision model for football game broadcasts.

If anyone has a lead on how to make or obtain this data, I would be very grateful. Speaking of the Eagles, this play clock displaying tenths of seconds rule proposal from 2023 would’ve been an interesting addition to NFL games.

That’s all I have for day. Talk soon everyone - RC

I’ve got some more minty data visualizations for you to take a gander at today. But first, some announcements:

1. I’m speaking at a conference on Friday April 24 at 4:50 PM

Interested in hearing my thoughts on the college football transfer portal? Then I have just the event for you. At 4:50 PM Eastern this Friday, I’ll be giving an in-depth talk about the transfer vulnerability score (TVS) formula I’ve been working on for the last year at the American Statistical Association’s Virtual Sports Analytics Conference. Here’s a link to register. I am very humbled and excited to listen to other colleagues talks on Friday and Saturday. I hope to see you there!

2. Next week’s write-up is a doozy

The only hints I’m giving are that a) we’re back using NFL related data and b) the dataset has over 50 million rows. I had to one-up this week’s dataset somehow.

3. I’m busy consuming amazing NFL Draft content

I’ve been largely checked out of NFL Draft coverage this year because I’ve been busy working on other projects, but I wanted to give some shoutouts to people whose content I’ve enjoyed reading in an attempt to get caught up on all prospect info I may need:

• ’s Grinding the Mocks site is a staple for anyone trying to make sense out of this deep ocean of mock draft content. Check it out here.

• I’ve really enjoyed Sam Teets’ prospect analyses. They’ve been a great source for me while trying to cram prospect knowledge before this weekend’s Draft. Check it out here.

• This NFL Draft content season was defined by an assortment of vibe coded websites gunning to replace PFF as the internet’s top mock draft tool. I enjoyed messing around on the Big Board Lab, Stick to the Model, and NFLIQ.

• ’s taken data-driven prospect comparisons further than I ever could have and I’ve really enjoyed his positional group write-ups this draft season. Check it out here.

• I’ve been tuned into every episode of The Evaluation on the SumerSports Youtube channel. Check it out here.

And if you’re interested in some prospect comparison cards, I’ll be posting them as the draft unfolds here. You can also generate them yourself using this Streamlit App:

I love NFL Draft coverage because it’s quixotic by nature. There are 257 NFL Draft picks happening this weekend. Let’s be modest and give a shallow estimate that each of these 257 selected players only played one full-length 12 game college football season. If you fast forward through commercials, every college football game has 60 minutes of play. That’s an extremely large sample size, and almost all of them played multiple seasons for multiple teams according to Transferportl.com. That’s a lot of film study to do in the 95 days between Indiana beating Miami and Fernando Mendoza’s name being called on Thursday night. That’s roughly 2,280 hours. If someone’s getting 8 hours of sleep a night, that’s 1,520 hours leftover to grind film. And we haven’t even touched on the Senior Bowl, the Shrine Bowl, the Combine, and every single Pro Day. And that’s just for the players who get drafted, there’s at least another 200 players declared who will not hear their names called. How much film can Mel Kiper get through in one pumpkin pie’s worth of time? We may never have an answer.

Let’s dive into The Beast now

What is The Beast?

Not to be confused with the dog from The Sandlot. The Athletic’s Dane Brugler publishes an NFL Draft Guide every season that’s worth the annual subscription to the site itself. It’s oftentimes the source of new information on prospects when it releases and is a huge feat every year. It increases in size annually too. This is an ode to the apex of NFL Draft content.

This was inspired by Luke McCartney’s similar project on Sam Vecenie’s NBA Draft analysis. I highly recommend you follow him on Substack if you don’t yet. Like Luke, I’m using cosine similarity to find common players and clusters of players across these different reports. But I also have the fun added ripple of creating a PDF parser.

The GUT

Here’s a new section of each Substack write-up I’m going to give you. The GUT stands for the given, use, to. And it goes like this for this week’s topic:

Given every copy of Dane Brugler’s ‘The Beast’ NFL Draft Guide since 2019…

Use cosine similarity and Python text analysis…

To find similarities between NFL Draft prospects.

All text analyses are up for feedback and correction here too as I just recently learned how to run them through a series of grad school lectures. Here’s some jargon on what I’m doing to these PDFs:

• Data pre-processing: English is a hard language, especially for a computer. I’m scanning these documents for stop words and for opportunities for stemming. I’m parsing out words like football, NFL, etc. to get rid of the fluff, then running a stemmer to combine words like running, run, and ran to somewhat de-clutter the text. Some examples of stop words are articles (a, an, the).

• Bag of words: I’m turning each prospect profile into a vector by mixing up the order of words, keeping counts of which tokens appeared how often, so our text analysis has an even base per prospect.

• TF-IDF: Term Frequency + Inverse Document Frequency re-weighs the bag of words counts so words that show up everywhere score low. Things I didn’t filter out like ‘play’ and ‘player’, just another way of reducing noise.

• Vector Space Model: Once every profile has been turned into a TF-IDF vector, each player becomes a point in a gigantic multi-dimensional space. Similar to any crazy futuristic network diagrams you may see. Two players will be closer to each other in this space if their summaries contained similar words.

• Latent Semantic Indexing/Truncated SVD: That huge map made by the VSM above is too complicated, so we’re running a singular value decomposition to narrow down the top 100 directions that carry the most signal. This helps the script realize the similarity in phrases like “elite pocket passer” and “stands tall in the pocket” are similar even if the words don’t match 1:1.

• Cosine Similarity: Now that every player is a point in this compressed SVD space, We’re measuring the angle between player points.

• Uniform Manifold Approximation and Projection (UMAP): The SVD 100-dimensional space is still too big, so UMAP shrinks it down to a 2-D scatter plot while maintaining similarities.

• Last but not least, K-Means Clustering: Finding cluster centers across the 100-dimension LSI space and assigning each prospect to the closest one.

Again, I just learned most of these phrases 2 weeks ago. If you see anything out of whack, please let me know. I knew about cosine similarity because of the NFL Draft similarity cards, but things like SVD and TF-IDF are new.

The Results

Here’s an average character count by prospect by year in The Beast:

The shortest summary since 2019 is a three-way tie between Demani Richardson of Texas A&M, Emany Johnson of Nevada, and Beanie Bishop Jr. of West Virginia. The longest summary of all time belongs to Marvin Harrison Jr. at 4,217 characters.

Here are the most common two-word phrases in The Beast:

Lacks ideal made me laugh. Very polite. Most of these are expected though, especially ball carrier and pass rusher. Brugler also always shouts out whenever someone has long arms. Here’s a cluster map for the class of 2026:

You can see the clusters separating themselves by position group. It’s cool to see clusters you’d expect to overlap actually overlap, like linebacker, safety, and cornerback. Here’s how these clusters look if we just analyze the words instead of player names:

For the top-ranked player at every premium position, here are their 3 most similar players:

When I first ran this, Caleb Downs’ #1 comparison was actually Emeka Egbuka. That made me backtrack to exclude school names in the comparison.

Like I said, still wrapping my head around text analysis. But this was a fun exercise on my favorite piece of draft coverage year in and year out.

Happy draft week, everyone. Talk soon - RC

I am back in your inbox with more systems design and data pipelines. Let’s dive right in

ICYMI: Here’s a link to part 1 of the football data engineering tutorial series. Here we covered the repository setup, downloading everything, our extract Python script, and dbt models.

Today we’re going to be transforming some old data from the nflfastR dataset and making it cleaner.

Why this tutorial series?

Wanted to field a question I’ve received here, and I’ve got two answers for you. The first is that data engineering is my expertise more than advanced metrics are, and the second is supply and demand. As of my most recent search on Team Work Online, the sports job postings website, I’m seeing only 3 listings for sports data scientists and 5 for sports data engineers. That’s almost double! Kidding aside, data engineering is so crucial to data science but I look around and see tutorials more saturated on the model training side of things than I do on the incremental data loads side of things. So I’m trying to show all of my readers why data engineering is important and how they can learn to be one, or learn some best practices from a data engineer to implement in their daily lives.

Read more

Players are still transferring, but spring games are occurring so I think it’s time to turn the page on the 2026 transfer window. What a wild ride it’s been. Before we turn the page to next season, I wanted to organize some thoughts on the window and what I’ve learned about how college football teams operate in the past few months.

1. We’re in the Entourage Era

I’ve been on this soapbox for a while now, but it’s worth repeating. When a program hires a new coach, because of the portal that coach can now bring every good player along with him to his upgraded program. Don’t believe me? Here’s a bar chart of the top 10 most popular transfer routes from this window:

James Franklin’s move is the only one here where multiple players followed a fired head coach. And this is just the top 10 too. It’s interesting to see how many of these were sequential. Like Franklin to Matt Campbell to Jimmy Rogers. I wrote about the transfer portal chain reaction here a few weeks ago.

2. Is the portal growing?

Out of the roughly 20,000-ish college football players in the United States, it looks like a very similar number entered the 2026 transfer portal as they did in 2025:

The 2026 portal isn’t technically over yet and I’m sure some withdrawals will trickle in as the weeks go by, but you get the idea.

3. The Portal’s Biggest Makers and Takers

Here are some bar charts showing departures and arrivals by team:

Oklahoma State lost basically half their roster after firing Mike Gundy, and replaced it with Eric Morris’ North Texas roster. Lane Kiffin needs one more transfer to Baton Rouge to break even this window, and towards the bottom of the viz you can see new South Florida head coach Brian Hartline adding more players than he lost. Colorado continues to be extremely active in the portal too, and Coach Prime’s team is probably the definitive team of the transfer portal era, since his move from Jackson State kicked off this frenzy.

4. Positional Breakdown

Here’s a position by position breakdown of the number of players in the transfer portal:

Everyone who’s worked with football data before knows positional filtering is a nightmare, but I did my best. It’s interesting to see OL as the highest group since they’re my biggest data blindspot right now, since they don’t tend to create box score stats.

5. Frequent Flyers

Here are the college football players who have made a habit out of the transfer portal so far:

It’s interesting to see how portal entries don’t always correlate with more transfers. Syrus Dumas has entered the portal in 2021, April 2023, December 2023, and December 2024 but he’s stayed put at UTSA since transferring there in 2021. And I’m seeing more cases like Charles Brantley across the nation occurring, where a player starts at a school, goes somewhere else, and returns back.

6. FBS Transfer Carousel

A lot of players swapped conferences this portal window:

78 players went from the Big 12 to the Big Ten. Matt Campbell strikes again.

7. Benedict Arnolds

Were there any rivalry transfers? Let’s see:

Do rivalries even mean anything anymore? Players are crossing Oklahoma like the Bedlam Series means nothing. Chapel Hill to Raleigh? I can’t believe Bill Belichick let that happen. Maybe Davion Gause will complete the research triangle next window with a transfer to Duke. Kentucky to Louisville here 6 times is probably the most eye-opening occurrence.

Edit: I ran this viz on the 2025 portal originally. Just a silly typo I didn’t catch. Edited at 8:43AM on 4/16.

8. Withdrawal Leaders

Just because someone enters the transfer portal doesn’t mean they’re guaranteed to transfer. Here’s a list of the schools with the highest withdrawal rates:

Highest withdrawal rate in the land is 15.2%, players are on their way out if they’re in the portal. One foot out the door, makes sense.

9. What About TVS?

Yes, let’s check in that metric I’m trying to create to predict portal entries, the transfer vulnerability score (TVS):

These predictions were made with pre-portal contextual data and validate after. Interpret this bar chart as saying, out of every FBS player whos TVS was 30 or above before the transfer portal, 31% of them entered the portal. A TVS of 3+ means there’s a 1 in 3 chance that player enters the portal. Higher TVS equals a higher actual portal rate, but there’s still some improvement to be made as I get more data in. Players flagged as low-risk transferred at a 3% rate. Players flagged as high risk transferred at a 31% rate.

What’s Next for TVS and Transferportl?

Building a website alone while having a full-time job and being in grad school at the same time is proving to be quite a challenge. Nevertheless, it won’t stop me from trying. Now that the portal is about done and dusted until next season, I have from now until December to work on things on the backlog, like:

• Cleaning data again and again and again. I’ve relied on CFBD’s API in the past for transfers but I’m having to source data from elsewhere and combine players to one common uuid in the system now, which makes things messier.

• Reducing bot traffic, which is extremely high right now. If you’ve been blocked from the site lately it’s because I’m trying to use Cloudflare to counteract these, sorry about that.

• Figuring out a way to capture TVS accuracy scores while also manually tanking a player’s TVS once they reach their new school in the data. Just need to set up more data snapshots.

I’m excited to keep improving the model and the site and share the progress on here. Before we kiss the 2026 transfer portal window goodbye, here’s one more map routes viz from the front page of TransferPortl:

I’ll be back in your inbox soon - RC

The Spade is a weekly football analytics newsletter covering topics across the NFL and college football. If that sounds interesting to you, I’d love to have you along for the ride as a subscriber for free:

Subscribe now

I wanted to apologize for the lack of a tutorial Friday email in your inbox last week, I was too busy to craft one. But this week’s should make up for that.

I’ve been sharing data visualization tutorials through The Spade and those won’t stop. I just needed to come clean with you all that those aren’t my strong suit. Instead, I’m more well-versed in the back of house work, the data engineering work. I’ve seen some talk online lately, and sports teams posting more data engineering jobs. I’ve also received lots of feedback from my faithful subscribers asking questions about what a data engineer does. This is my attempt to answer all of these questions and teach you how to be a data engineer yourself using sports data as my moat.

So I’m going to start a tutorial Friday series publishing what I believe to be the first football data engineering tutorials out here on the internet. I’m going to walk you through an enterprise-level tech stack for both batch and stream processing of data in anticipation to have these pipelines run automatically as we have more data coming in as the 2026 NFL season starts in September. That gives us a lot of time to mix in some more data viz tutorials here and make lots of progress on the DE side of things.

So without further interruption, here’s my first attempt at building a gigantic scalable data pipeline for free NFL data. I hope you enjoy it.

What exactly does a data engineer do?

Data engineering work is not as attractive to most as being a data scientist and working with models all day is, pun intended. But I really love it. It’s a profession that attracts people that are extremely organized and detail oriented, and at its best feels like gardening in a way. The primary role of a data engineer, from what I’ve discovered in my 5ish years doing it, is to source and clean data. I’m the point guard, the distributor, the shelf stocker, comparison you want to use. I’ve also used comparisons to archaeology in the past and being the prep chef. My co-workers, whether they be data analysts, data scientists, machine learning engineers, project managers, product managers, or any combination of these, need data. And it’s my job to provide it in a clean and interpretable way.

This means setting up two styles of processes, batch and stream processes. These names explain themselves, batch is data processed in batches and stream processes are live. Batches are cheaper and by all accounts easier to maintain since they run on a schedule and then they finish. Stream processes are more like trying to fix a car’s engine while it’s going 60mph. So in this initial tutorial, we’re going to focus on batches.

My Toolkit

Read more

I have swapped out my 2025 weekly mock draft content for exploratory data analyses across college football and the NFL. But don’t worry, mock drafts are back now. We have a little less than a month until the 2026 NFL Draft in Pittsburgh, that really snuck up on us.

Background on these cards

This was one of the first projects I posted online back in 2023. It started before I knew how model building worked, and I was comparing raw stats to raw stats via cosine similarity to find the most similar players. In 2024, I learned about the k-nearest neighbors algorithm for finding similarity and changed these measurements to use percentile rankings to measure similarity. I didn’t make any updates to it for this season as I’ve focused less on draft content lately and more on other things. Not for any particular reason, I just think the draft content space is very crowded with some amazing material.

I added the caveat of “Go Watch Film” in the subtitle to offset the wonky comparisons that might pop up, especially for positions like offensive linemen that are more combine measurements based since they don’t have many free stats available online.

If you’d like to generate these cards for any prospect since 2006, you can use this Streamlit app. The original dataset I used coalesced missing combine data with pro day data, so it makes for a pretty comprehensive dataset. I’ve varied how I’ve gathered the data over the years, it started as straight typing and various web scrapers duct taped together. But it came from a lack of speed when gathering combine data, I wanted to run some live comparisons against the set during the combine.

This write-up definitely isn’t as verbose as mine usually are because this week has been incredibly busy. The data driven portion is meant to be tongue in cheek. I’ve been creating some stuff I’m really excited to share with you all soon, but can’t just yet. More on that later. That’s all there is to it though, I hope you enjoy the mock draft and talk soon - RC

Round 1 Mock Draft

1. Las Vegas Raiders: Fernando Mendoza, QB, Indiana

Won’t waste your time with this one. This one’s a lock across mocks. A mock lock, if you will.

2. New York Jets: Arvell Reese, DE, Ohio State

I wanted to give the Jets Reese’s teammate Sonny Styles, but they made way too many free agency linebacker signings for me to include that in a realistic mock.

3. Arizona Cardinals: Francis Mauigoa, OT, Miami

The Cardinals take Mauigoa who should be their franchise cornerstone right tackle for years to come with the third overall pick. Francis’ brother, Francisco, is currently a linebacker in the Jets organization.

4. Tennessee Titans: David Bailey, DE, Texas Tech

A lot of experts have Notre Dame RB Jeremiyah Love going here to Nashville but I think they lean defense here instead.

5. New York Giants: Sonny Styles, LB, Ohio State

A good Giants defense gets even better with a game breaker like Styles. Abdul Carter, Dexter Lawrence, and Sonny Styles playing together would be insane. Note: I wrote this before Lawrence publicly requested a trade. They could still use Styles here though.

6. Cleveland Browns - Spencer Fano, OT, Utah

7. Washington Commanders - Jeremiyah Love, RB, Notre Dame

Classic conundrum here of placing too many players in where needs are, then seeing a way too talented player fall too much. I can see the Commanders wanting to upgrade on their running back position, and Love might be the most talented player in the draft.

Reminder that my model and Daniel Jeremiah got the same comp for Love, Demarco Murray. That was cool. I have a half-baked idea to write a CFB unsolved mysteries article about how no one knows how many fumbles Demarco Murray had in college, but we’ll save that for another day.

8. New Orleans Saints - Mansoor Delane, CB, LSU

9. Kansas City Chiefs - Carnell Tate, WR, Ohio State

I mean, come on. A Brian Hartline comp to Carnell Tate? How can you not be romantic about KNN?

10. Cincinnati Bengals - Rueben Bain Jr., DE, Miami

11. Miami Dolphins - Makai Lemon, WR, USC

Almost everyone has Arizona State WR Jordyn Tyson ranked higher than Lemon, but I prefer Lemon. Dolphins soften the blow of losing Tyreek Hill by getting Lemon as Malik Willis’ #1 target.

12. Dallas Cowboys - Caleb Downs, S, Ohio State

I think the Caleb Downs foot injury stuff is probably smoke from a team hoping he drops. But I see Downs going in Kyle Hamilton territory in the first round.

13. Los Angeles Rams (from Falcons) - Jordyn Tyson, WR, Arizona State

Seems like everyone is mocking a wide receiver to the Rams. This would be an embarrassment of riches. Davante Adams, Puka Nacua, and Jordyn Tyson? Good luck, NFC.

14. Baltimore Ravens - Monroe Freeling, OT, Georgia

15. Tampa Bay Buccaneers - Jermod McCoy, CB, Tennessee

A Travis Hunter comp? The first I’ve seen so far. This is only comparing his CB half of stats though.

16. New York Jets (from Colts) - Kenyon Sadiq, TE, Oregon

Geno Smith finds his new Brock Bowers here with Oregon TE Kenyon Sadiq. I think this would be very exciting to see.

17. Detroit Lions - Olaivavega Ioane, OT, Penn State

18. Minnesota Vikings - Dillon Thieneman, S, Oregon

Don’t make a Harrison Smith comp. Don’t make a Harrison Smith comp. Don’t make a Harrison Smith comp.

19. Carolina Panthers - Max Iheanachor, OT, Arizona State

20. Dallas Cowboys (from Packers) - Keldric Faulk, DE, Auburn

21. Pittsburgh Steelers - Akheem Mesidor, DE, Miami

22. Los Angeles Chargers - Kadyn Proctor, OT, Alabama

I’ve liked Proctor ever since I saw him take a goal line carry last season. He didn’t score, but it showed grit.

23. Philadelphia Eagles - Emmanuel McNeil-Warren, S, Toledo

The Eagles desperately need a safety. McNeil-Warren helps with that. He’ll probably go before 23 although Thieneman tested better at the combine. That’s 3 safeties I have going in the first round.

24. Cleveland Browns (from Jaguars) - Denzel Boston, WR, Washington

25. Chicago Bears - Blake Miller, OT, Clemson

26. Buffalo Bills - Omar Cooper Jr., WR, Indiana

I tend to always have an WR-heavy round 1. Why do I feel like the Bills always need a wide receiver?

27. San Francisco 49ers - Caleb Lomu, OT, Utah

28. Houston Texans - Kayden McDonald, DT, Ohio State

29. Kansas City Chiefs (from Rams) - Colton Hood, CB, Tennessee

30. Miami Dolphins (from Broncos) - Peter Woods, DT, Clemson

31. New England Patriots - Cashius Howell, DE, Texas A&M

32. Seattle Seahawks - Jacob Rodriguez, LB, Texas Tech

32 is always a weird spot and I feel like someone always makes it to day 2 without us understanding why. KC Concepcion is that person for me this year. It’s hard to find needs for the defending Super Bowl champion Seahawks so I have them taking a flyer on the leader of the mustache crew.

The Spade is an entirely reader-supported weekly football analytics newsletter covering topic across the NFL and college football. If that sounds interesting to you, I’d love to have you along for the ride as a subscriber for free:

Subscribe now

For those who have been around for a whole year, you know that April Fool’s Day is the only day in the year where The Spade doesn’t cover football. Last year we covered basketball in an attempt to find the NBA’s most efficient shooters from beyond halfcourt. Check it out here. Today we’re turning to baseball for this special article, and we’re unearthing a metric that is going the way of the dodo this season. In light of the automated ball/strike (ABS) system taking hold in the MLB this year, I wanted to analyze the almost extinct act of an umpire’s makeup call. A makeup call is defined as another errantly called ball/strike right after one occurs.

Phillies performance aside, I’ve been enjoying the first week of ABS challenges. My initial kneejerk reaction is that there should be more ABS involvement, to relieve the pressure off of home plate umpires so they can focus on making correct calls in other non-automated areas of the game. Umpire Scorecards have added ABS challenges to their tweets too, to improve on an already flawless data visualization. Today we’re going to dig deep into 9 seasons of baseball pitch data to see if we can find old opportunities for makeup calls in baseball games, things that for either emotional or technological reasons may be leaving this world soon. Umpires know that a mistake has happened, they called a strike a ball or a ball a strike. To set things even, did they flip flop immediately after to make the count ‘right’ again? Let’s find out.

This is a data analysis purely rooted in anecdotal evidence. I wanted to see if the data matches what my eyes have seen multiple times. Happy April Fool’s Day, and I hope you enjoy this.

The Data

Holy cow, this much baseball data has been free and accessible this entire time? I didn’t even realize. To me though, that’s no fun. I’ll be returning to the football data mines next week but this week we’re diving into the deep well of baseball stats.

There’s a whole Python package directly synced with Baseball Savant called pybaseball. I found some pitch by pitch data that allows me to create some visualizations similar to what you see on MLB broadcasts, with a classic strike zone and home plate look. I’ve also grabbed every home plate umpire using statsapi - the MLB’s official public data source. Given data on every called ball or strike from 2017 to 2025, I’m using SQL LEAD() and LAG() functions to locate makeup calls, which we’re defining as a ball called as a strike immediately after a strike was errantly called a ball, or vice versa. We’re looking for moments where hitters took two pitches in a row, an umpire made a mistake on pitch 1, then made up for it on pitch 2. This is slightly flawed, of course. I may have accidentally queried moments where home plate umpires are just flustered. But let’s still see where this goes. It is April Fool’s Day after all, I have the cloak of the joke to fall behind if these numbers are wonky.

So here’s our strict criteria for looking into potential makeup call territory. We’re looking for:

• Two taken pitches in a row from the same pitcher to the same batter. Where…

  • Pitch 1 was errantly called a ball or strike. And…

    • Pitch 2 was errantly categorized in the other direction by the umpire. (ball to strike or strike to ball sequences)

The Results

Here’s every instance of a makeup call within 1 inch of error:

Made a nice little baseball edition of The Spade logo for this one. This plot is busy, here’s what happens if we filter these makeup calls down even further to within 2 inches of error:

This one here is showing the frequency of makeup calls when an umpire misses by 2+ inches. If they called a ball a strike that was over 2 inches outside the zone, they made up for it on the next pitch with an opposite ruling.

Apparently the larger the flub, the more likely there is to be a make-up call. When comparing each error distance, that makes itself apparent:

The gray bars show how often you’d expect a makeup call to occur just by chance. The blue bars show how often they actually occur grouped by error margin. Slight differences, but still.

Is there any correlation between the number of errant calls and makeup call rate by umpire? Because technically a makeup call is a stacking of errors, right? Let’s see:

R-squared of .15? That means no.

Individual Umpire Scorecards

Here are some makeup call scorecards for names you’ve heard before:

CB Bucknor:

Angel Hernandez:

And Pat Hoberg, who you might remember for calling that perfect game in the World Series back in 2022, then subsequently being banned from umpiring for betting:

That’s about all I have for today. I hope you enjoy April Fool’s Day, have a great day, everyone - RC

ØÑĖ MØR£ THÏÑG

I lied. I have a Cam Newton Google Translate clone for you to mess around with today too. This is technically writing about football, right? Check it out.

The Spade is an entirely reader-supported weekly football analytics newsletter covering topics like the NFL, College Football, and all things in between. Recently, we covered college football recruiting/portal splits and NFL Draft Combine/Pro Day Shapeshifters. If any of that sounds interesting to you, I’d love to have you along for the ride as a subscriber. Thanks for all the support.

Subscribe now

Here’s my code to make this chart here from this week’s recruiting turf war write-up.

I’m calling them needlepoint belt bar charts because to me they look like these:

Let’s jump right into it.

Read more

We’re back looking into college football, but instead of looking at the portal itself this week we’re going to take a step down to high school recruiting. An outstanding question I’ve had for this new era of transfer portal recruiting is: are teams switching their strategies? On a more dramatic level, are colleges still recruiting high school players? Or are they just waiting for good players to fall through the cracks in their redshirt seasons and swooping in through the portal? We’re going to get to the bottom of this today.

The Data

Same data we usually turn to when we analyze college football, CollegeFootballData.com. There’s tables available to us detailing each college football team’s activity on the recruiting trail and in the portal. It’s the same primary data pool that TransferPortl.com is built on. I’ve had to get slightly creative for TransferPortl lately, aiming to have data that’s more live than CFBD provides. But the recruiting trail is done and dusted for the 2026 season, so it works here.

We’re assigning territory to college football teams here in the United States on a county level. And we’re keeping things relatively explainable. This is directly in response to finding 247 Sports’ composite scores to be convoluted. So I’ve created the super advanced metric we’re going to call stars squared. Stars squared just squares the value of recruiting stars the recruit has and adds that to a points tally per college. So 5 stars get 25 points, 4 stars get 16, 3 stars 9, 2 is 4, and 1 is 1. There’s probably a better way to quantify this, but I enjoy making easily explainable stats.

We’re going to assign territory to each college football team based on the highest stars squared per county. For example, getting 4 5 star players, one 4 star player, and 2 3 star players will give a team 134 stars squared for that county.

The Results

Let’s look at the recruiting battle on a state by state level, ranked in order of 5 star recruit supply. First up is Texas:

The two football pipeline hotbeds you see on the inset map in the top left corner are The Houston metro area in the Southeast, and the Dallas-Fort Worth metro area in the Northeast. The conclusions you can draw from these maps are so intriguing to me, like how UTEP was being poached in their backyard until the class of 2024. You can see how different administrations change their strategies over time. Houston just recently gained more recruits in their own backyard, and Texas State is making more of an emphasis on Texas local recruiting as of recent too.

The next hottest state for recruits over the past 5 years is Florida:

The thing that sticks out to me here is the battle for Miami-Dade County in the very Southeastern tip of Florida, and Broward County directly above it. The Class of 2024 belonged to FAU? And Florida State’s winning the county in 2026 with 41 stars squared, but Syracuse is actually tied with them. Over the years, you can see Miami’s stronghold on its area dwindling. And as of recent, the Florida Gators are re-focusing on their backyard in Alachua County. There was only one 4 star recruit from the class of 2026 there, but he chose the Gators.

Next we’ll look at Georgia:

We’re focused on that Northwest-ish cluster of talent in the Atlanta metropolitan area. Looks like Georgia really swooped in on the talent in 2024 in that area, but Alabama has moved in during the 2026 recruiting trail. Clemson has almost been entirely eliminated from is once-dominant stronghold in 2021.

Now let’s look at Ohio:

3 clusters of talent in the Cincinnati (southwest), Columbus (center), and Cleveland (northeast) areas. It’s interesting to see only one Ohio State held square in 2026. The state of Ohio didn’t have a single 5 star recruit this year.

Now let’s take a look at California:

LA County belonged to the Campbell Camels in 2023, over UCLA and USC right in their backyard. Out of the 11 players Campbell took from Los Angeles in 2023, only two of them have transferred. Rashawn Carr to Marshall in 2025 and Aiden Valdez to New Mexico in the same year. It looks like Cal Poly is making waves across California, and Oregon’s 2024 action has died down in the last couple of years.

Those are the top 5 states in order for stars squared values, but I wanted to include my home state of Pennsylvania too just for good measure:

This is a Penn State power deterioration chart. James Franklin brought his Delco pipeline with him, and they’ve lost ground in Philly to Temple, in Allegheny County to Pitt, Montgomery County to Notre Dame, and Chester County to Syracuse. The 24 Iowa State transfers will certainly help as a 2026 band-aid, but the Nittany Lions have their work cut out for them on the 2027 recruiting path.

Who’s Succeeding in their own backyard?

How do power programs fare in their closest recruiting battles? Let’s find out:

Georgia Tech can’t compete in Atlanta with the SEC. And they’re feeling pressure from their ACC conference-mates Clemson too. Any Alabama locals want to shed some light on Phenix City, AL and if that city leans more Roll Tide or War Eagle? At the moment, just based on crow flies distance, I’m giving the high school to Auburn, but I might be wrong. I’m surprised SMU is losing ground in all of these areas, based on the influx of cash I was assuming would have the opposite effect.

I thought this table did an insufficient job of showing what pipelines had been established already. For example, I’m listing UCF as being raided by Georgia at IMG Academy, but UCF never truly established an IMG Academy pipeline. That’s also an outlier high school since most kids come from out of state to attend there, so hometown connections may be weak. Here’s another table showing more established high school to college pipelines:

I think this table shows how UCF never stood a chance at IMG Academy, and that they probably recruit with this in mind. The only 6/6 stronghold since 2021 has been Kirby Smart’s UGA from IMG Academy. One other observation from this table emphasizes the return to a local shift in talent by USC after letting it falter for a few seasons. It also highlights how Florida State is moving in on Miami’s territory in South Florida. Hawaii’s pipeline from Bishop Gorman is something to watch for the future. I’m calling it The Ninth Island Effect.

One more gigantic table for you to view. How are teams dividing up their transfer and recruiting pies? Does a lack of recruits indicate more of a turn to the portal for teams? Let’s look:

This table is sorted by the count of player turnover since 2021. The shift percentage is used to show the percentage of incoming players per school come from the portal. There’s no team getting fewer than 20% of their new players from the transfer portal.

The numbers all the way to the right indicate how much new player action has shifted towards portal recruiting instead of high school recruiting for schools. I will say it’s interesting to look at the bottom of this table more than the top. Dabo Swinney really stuck to his principles on hating the transfer portal until a small secession this year. Georgia, Ohio State, Notre Dame round out the bottom of the list, all considerably ‘successful’ programs, blue bloods if you will. This table shows stability to me. Fewer transfers out signals that fewer transfers in are needed, a luxury that your average college football team doesn’t have. Penn State didn’t really surrender to the portal game until Matt Campbell got to town.

Incoming Player Strategy Axis

See Colorado, an influx of portal talent instead of high school talent isn’t correlated with success. College football rosters are very delicate and there’s no correct way to build a winning team right now. But it’s interesting to watch teams, coaches, and staffs create their own plans in the portal era.

Here’s a diverging bar chart per FBS team with the service academies excluded since they’re not allowed to participate in the portal frenzy:

This shows the different diverging paths teams are taking when it comes to the 2026 window. North Texas is our heaviest portal team, probably in reaction to their heavy transfers out to Oklahoma State. San Diego State is has the most balanced 2026 strategy in the land. And I was surprised to see Northern Illinois up there with Notre Dame, Georgia, and USC. I had the running hypothesis that recruiting-heavy classes were a luxury of the blue bloods, but the Huskies are throwing a wrench into it.

And here’s one more map for good measure, if we had to assign a winner of stars-squared on the state level:

Things that stand out to me on here:

• UTEP sweeping up Texas recruiting indicates a recruit-focused strategy shift as opposed to the top dogs in the state focusing on the portal, like Texas Tech. Same for Tulsa with regards to Oklahoma, Temple with regards to Pennsylvania and Northern Illinois(?) with Indiana. The Huskies are marching on Hoosier territory?

• Some expected names on here, like Oregon owning Oregon, West Virginia owning West Virginia, Minnesota owning Minnesota, LSU owning Louisiana, and Miami still sweeping the state of Florida. Although the county-specific maps looked like they were being beat for local talent, perhaps they focused on a state-wide search for recruits instead of just Miami-Dade and Broward counties.

• There are only a few states on here owned by new head coaches. LSU and Lane Kiffin cover Louisiana, which feels like the largest talent monopoly in the land. James Franklin and the Hokies swooped in on Virginia, and Ryan Silverfield kept Arkansas under Razorback control.

• Fresno State head coach Matt Entz has marched directly into Cornhusker country and pulled the rug out from beneath Matt Rhule.

Let me know what you all think in the comments below. I think I’ll publish some more recruiting strategy write-ups through the summer. Talk soon - RC

I wanted to share with you that my new website, raycarp.com, is now live. This has been the result of over 4 years of NFL exploratory data analysis and research. In my years of surfing the net, I looked for a few tools and couldn’t find them. This is the result of outstanding data questions I couldn’t find anyone answering. I hope you enjoy the site and find it to be an insightful resource.

Tools

On raycarp.com, you’ll find:

• A run gap tool, so you can create your own run gap charts. This data goes all the way back to 1999.

• An attempt at a free personnel tool. This tries to extract how often each NFL offense lines up in different personnels (11, 12, etc.) and defenses line up in different personnels (base, nickel, dime) depending on the positions of players on the field per snap. This is only as accurate as nflfastR’s definitions of position. I’d say the offensive personnel definitions are better than the defensive ones since defensive players are harder to place in a box (DE/LB, LB/DB definitions get murky). I’ve included a disclaimer at the bottom though to account for these

  

• A golf leaderboard tool, to generate similar graphics that I make and post on The Club, a golf analytics Substack I’m trying to run alongside The Spade.

• An On/Off stats tool for all NFL teams. I’m really excited about this one. You can toggle all players on or off to see how your team performs, as well as choose a “Random 11” which randomly selects an offensive or defensive lineup that team fielded at some point in the season. There’s also options for the “most efficient” lineup and the “most successful” lineup, based on EPA/Play and success rate respectively.

  

• Links to my other projects, like transferportl.com (which will have a hefty update soon) and my two Substacks. As well as to golfastR’s GitHub page, which I still need contributors for! There’s a lot to be done in the mines of free golf analytics data.

And did I mention that there’s a download PNG button that works from your mobile browser too?

Tech Stack

All data comes from free sources nflfastR and golfastR. The website is built using TanStack Start and TypeScript, deployed through CloudFlare (which I discovered while trying to prevent bot traffic on TransferPortl). I’ve stored all filters on the client side so they’re loading lightning quick. The calculations done in R are pre-stored as JSON files for the same effect. Web development is relatively uncharted territory for me, usually my work stops at the data handoff. So if you see anything wonky, please let me know and I’ll fix it.

Please Submit Feedback and Use Responsibly

I’m hoping that this will serve as a conversation starter between me and other sports analytics professionals, numerically inclined fans, or newbies to the world of analytics. I also love hearing feedback on how to improve tools, and above all I’m hoping to boost transparency into football numbers. I think it’s my nature as a data engineer to always have my mind on exactly where I sourced this data from, and exactly how I edited it, rounded it, or transformed it to reach the final numbers. I’m not marketing myself as the most accurate football analyst out there, just as someone trying to make the best free tools I possibly can for folks to use.

Also, I was extremely disappointed to see some folks online using my run gap charts to say hateful things towards NFL players. Under no circumstances do I condone using my graphs as a vehicle for hate speech towards any NFL player. Please use these to win arguments against your friends, trick fantasy football leaguemates into lopsided trades, for teaching your math-minded friend about what nickel defenses are, or teaching your scheme-minded friend about what expected points added are. Behind each and every number is a person. By downloading and sharing, you’re agreeing to use these numbers wisely. Please don’t use these numbers for evil.

I hope you enjoy the site, and I can’t wait to hear what you think. Thanks everyone, talk soon - RC

March Madness begins today, but before we fully focus on basketball I have one more college basketball themed college football project for you to analyze. Scrolling on any social media lately will unearth a bunch of “What if college football had march madness?” infographics like this one, but no one’s actually crunching the numbers. So I’m going to. What we’re going to do today is simulate 100,000 2025 CFP bracket situations and analyze the results. Let’s dive right into it.

We’re running a Monte Carlo simulation again here in Python to generate 100,000 simulations of a 68 team College Football Playoff. Each game outcome is decided by a single random draw centered on the point spread between the two teams.

The Data

How am I ranking the top 68 college football teams? Ranking college football teams is an extremely complex venture that I am not taking up in one week, so I’m borrowing Brian Fremeau’s 2025 FEI rankings. The Fremeau Efficiency Index is one that quantifies opponent-adjusted efficiency, ranking teams based on how they play with some adjustment based on who they play. Indiana led the pack last season with a 1.86 FEI, meaning they’re approximately 1.86 points per drive more efficient than the field. For each of the 100,000 sims, we’re deciding the outcome of the game with one line of code that takes the difference in FEI to choose a winner.

The key variable here is variance. Running the model with real-world football variance, Indiana wins about 45% of the time which is boring. So I calibrated the model’s variance to match real March madness upset rates using this page on bracket odds from the University of Illinois as my target. The goal isn’t prediction, rather to import basketball’s level of chaos, entertainment and fun into college football to see what happens.

The Results

Here’s how our bracket turned out after 100,000 simulations:

Still with Indiana as the undisputed champion, but a 16-0 season translates well when you 100,000x it. Makes sense. Oregon came in ‘second’ here with the 2nd most championships, winning 16.7% of the time. Out of the top 10 teams that won championships, the only two I was surprised to see were Utah and Vanderbilt. I may have given Vanderbilt fans more gas to stoke the flames of missing the playoff. Specifically, it was interesting to see Utah fare better over time than Ole Miss. This was a slightly limited study though. Next March I want to build this with multiple rating systems, maybe add in SP+ and KFord alongside FEI. The first 4 games weren’t cooperating with this matplotlib layout, sorry for excluding those.

If we expand the top 10 teams table to include every team from our 100,000 tournament run who won at least once, we get 64 unique championships. Any crazy Cinderella runs?

I promise I didn’t plan on Penn State being the first team on this table that was usurped by lower ranked FEI teams. That is funny though. Sorry to Houston, Boise State, Memphis, Minnesota, Rutgers, Michigan State, and Maryland, our field’s biggest goose eggs. 100,000 chances and not a single championship. Brutal.

Upset Rates vs. Basketball

Did our 100,000 2025 College Football Mania simulations have more upsets than college basketball?

The upset rates from BracketOdds are for every NCAA tournament since 1985. It looks like our model has more second round upsets in ⅝ seed matchups. The 12-5 matchup one is the most intriguing to me, since that’s always my favorite upset to throw into an office bracket. We simply couldn’t re-create that magic with college football data. The one surprise: our sim actually produced more 16 seed over 1 seed upsets than real basketball. That probably reflects how compressed the bottom of college football’s talent pool is. The gap between a 1 and 16 seed in football might not be as wide as we think. Looking at our bracket, though, I think the only real 16 over 1 upset I see happening in real life might come from AAC champions Tulane beating Notre Dame.

What Did We Learn?

Even with March madness-level variance injected into the model, college football is still more top-heavy than basketball. Indiana, Oregon, Ohio State, and Notre Dame still took home over 70% of the 100,000 championships. Across actual March Madness basketball results, that number is more in the 50 to 60% range. Good teams win more often in college football.

That’s why the 12 team playoff is so interesting to me. Its concept is great, I think the dam will break once we have a marquee upset happen. Football has some work to do, national championship games are still blowouts. Consider this my proposal for a 68 team CFP. How many guaranteed spots would the SEC get? I think we’re heading down a rocky road in college football but there’s parity at the end of it. They just have a lot of ironing to do to get there.

Happy March Madness. May the Bracket Gods be forgiving. Talk soon - RC

The Spade is an entirely reader-supported weekly football analytics newsletter covering topics across the NFL and college football. We’re on a college kick right now, but will most likely pivot back to the NFL Draft soon. If that sounds interesting to you, I’d love to have you along for the ride as a subscriber for free. Paid subscribers get access to the code I use to build these things:

Subscribe now

I’m not usually in your inboxes on the weekends but today’s Pi day, 3/14. Or Pie day, depending on which one you’d rather celebrate. At the expense of revealing my age, here’s a meme I share every year on Pi(e) day.

The Legend of the Pie Charts

I’m in a class right now in grad school called Data and Visual Analytics. The professor is a highly decorated academic and professional data master. In one of his slides, he specifically says ‘under no circumstances should you ever use a pie chart. EVER!!!! Especially a 3D pie chart.’ Pie charts are notorious in the data visualization world for not conveying information correctly due to their circular nature distorting slice sizes. Everything that’s a pie chart would be better served as a bar chart instead. I talk about this more in my “Top 10 Tips for Data Visualization” article from a few months ago.

I started a weekly football analytics Substack before becoming familiar with the notoriety of pie charts, so I experimented with those and donuts charts for a while before setting a hard rule of no more pies. I’m much more akin to the professional trends of data warehousing than I am of graphs, but now I know. So I’ve shelved my pie chart ideas for now, and they will emerge once a year.

On March 14th.

So without further interruption, here’s a gallery of ridiculous pie chart ideas I’ve had over the last year. Look at these as if you are walking through a modern art gallery, instead of the usual tone you read my more analytical NFL and college football write-ups with.

Here’s a 3d pie chart I spotted during season 4 of The Wire.

I can’t help but love pie charts. Here’s what I’ve got for today.

4th Down Courage Index

Taken by grabbing the counts of 4th down attempts and the count of successful 4th down attempts.

5 Star Recruit Hoarding

This one was made with data from CFBD.

The EPA Compass

Two dots above the dotted line indicate two positive EPA/Play numbers on passes and rushes.

The Red Zone Compass

The Penalty Wheel of Shame

I love how this one turned out.

Playcalling Donut

This, of course, doesn’t factor in audibles. And the 4th down Hail Mary heaves are skewing the numbers there.

Quarterback Chaos Wheel

Sorry, Nine.

3D Touchdowns Pie

Now we’re really cooking with gas. One more 3d pie:

The Footaballcci Sequence

Really testing the limits of ggplot with this one.

The Arabesque

Sundial

When do NFL teams score? In Eastern time.

The Orb

It’s spherical. Spherical!

That’s all for today. Happy pie day, everyone. Have a great day - RC

Here’s the tutorial for this week’s write-up on the trapezoid of excellence. I’ve began to ask for some outside help for these coding tutorials to gauge if they’re good enough to follow. I’ve gotten some good feedback so you’ll notice this script has way more comments in it. I hope it inspires you to download R on your computer and give it a try! We will see. Please share anything I’ve inspired you to build in the comments, on twitter, or feel free to email me too. I love hearing about all of it, even if it’s still a work in progress. Thanks again for reading and talk soon - RC

Full Code

Read more

We’ve got March Madness coming up and I’ve got college basketball on the mind, but I also run a weekly football analytics newsletter. So today those worlds are colliding. We’re going to apply the theory of Ryan Hammer’s Trapezoid of Excellence to the NFL and college football. Can we find two metrics that predict future playoff tournament success throughout the regular season for football teams? Let’s jump right into it.

The Spade is an entirely reader-supported weekly football analytics newsletter covering topics across the NFL and college football. If that sounds interesting to you, I’d love to have you come along for the ride as a subscriber for free:

Subscribe now

Let’s talk trapezoids now.

What is the Trapezoid of Excellence?

Coined by Ryan hammer, the Trapezoid of Excellence is a shape that ‘all future NCAA tournament champions fall into". It combines two college basketball metrics, pace and net rating adjusted for opponent. Here’s how it looked on February 18th:

Good college basketball teams from this season fall within the trapezoid. Michigan, Duke, Arizona, Florida, Illinois are all tournament favorites and heading for a high seed. But it also helps uncover some gems that could potentially cause upsets during March Madness. Louisville is currently in 6th in the ACC, but their appearance in the trapezoid should put them on the upset radar if they get a tournament bid. The goal of this exercise is to try and fit a trapezoid around good NFL teams in the simplest way possible using two metrics. We’re going to create trapezoids that include all 4 conference championship participants for the same metrics across multiple seasons.

In the original, it seems like the y-axis limit is set to 36, and the x-axis limit is set to 75. Y-axis labels are in increments of 5 and x-axis labels are in increments of 1. These are not best practices but I’m subscribing to it for this exercise to preserve the trapezoid. We very well could just drag the shape over our ggplots on Canva and group teams that way but that’s no fun. Instead we’re using math to calculate the trapezoid. The x-axis range (horizontal lines) are the 5th and 95th percentiles of plays per game, and the y-axis ranges are similar with some diagonal padding to make the trapezoid instead of a square.

Calculating Football Pace

Pace of play applies to other sports and this was a fun one to untangle. We' have a few fields available to us to run some analyses on but I eventually settled on nflfastR’s time_of_day field. I attempted other measures like calculating plays per game minute, drives per game, drives per quarter, etc. but I thought that the actual real world clock worked better than football’s game clock. This was probably because of the play_clock field in nflfastR containing all zeroes.

So our final pace metric is actually stupid simple. It’s a count of offensive plays per game.

NFL’s Adjusted Net Rating Equivalent

Keeping this one simple too. We’re just going to calculate a difference between offensive EPA/Play and defensive EPA/Play. This isn’t situationally adjusted like pace is, this is just total EPA/Play. This is very similar to the team tiers graph on rbsdm.com.

The Trapezoid

Without further adieu, here are the trapezoids of excellence for the NFL:

It’s interesting to see that the Cowboys were the ‘fastest’ offense according to this pace metric last season. Fast doesn’t always equal efficient though. And pace of play can be better in the future. I ran a quick correlation for time trailing in a game and plays per game and got a very low r-squared value which I was surprised to see. But a full pace of play metric is definitely on the schedule to create soon.

How early do future conference championship teams show up in the trapezoid of excellence? Let’s make some season checkpoints and look:

The same teams are appearing in the same spots year over year. A lot of appearances within the trapezoid from the Rams, Bills, Eagles, and more. This trapezoid looks a little different than the previous image’s because it’s based on the 5 seasons of game data instead of just 2025’s.

What about college football?

After all, this is a college metric. The trapezoid of excellence is similar for college football data, we’re using net EPA again on the y-axis and offensive plays per game on the x-axis. Except to try and filter out noise, I’m only using Power 4 vs. Power 4 games as the sample size.

The spread is smaller for the power 4, hence the tinier chart. It has Indiana and Miami in the trapezoid, so we’re calling it an excellent trapezoid.

When we take off the power 4 filter, this is how the trapezoid looks:

Holy Toledo. They had a great season metrics wise last year, didn’t they? But if you’ve been checking transferportl.com lately, you know that they lost 17 players and head coach Jason Candle to UConn this off-season. North Texas, JMU, Memphis, and Southern Miss are other trapezoid participants who lost their head coaches too. ECU would be in our ‘too fast’ category.

Any teams in both the CBB and CFB Trapezoid?

Which teams are both elite in football and basketball?

Maybe this is a proposal to the CFP Committee to transition to a 68 team playoff. Imagine how far San Diego State would go as a Cinderella. Arizona and Miami share the title for ‘best’ teams across both sports. I’ll monitor as we progress through the conference tournaments.

That’s it for today. Good luck making your brackets last week. Let me know what you think in the comments below and have a great day - RC