How to make accurate NFL spread predictions

This is the script to the second half of The 60 Minute Guide to the 2024 Football Season, my audiobook preview of the season. This was meant for audio, as the Sports Illustrated jokes do not work in the written format. But if you would like to read it, go for it.

You may be able to listen to all of this audiobook preview for free. Learn more about The 60 Minute Guide to the 2024 Football Season.

Kirk Cousins needed a big play. Minnesota was down a touchdown to Kansas City at home. With less than five minutes remaining in the game, they had to convert a 4th and 12 at the Kansas City 24 yard line. 

Cousins dropped back to pass, and the Chiefs rushed six defenders. He only had two and a half seconds to get rid of the ball. Did he throw to Justin Jefferson? No, the ultimate NFL cheat code strained a hamstring earlier in the game and didn’t return. He would miss the next seven games.

Instead, Cousins flung the ball to his right in the direction of rookie wide receiver Jordan Addison. Cornerback L’Jarius Sneed had tight coverage, and the ball fell harmlessly to the turf in the end zone. 

However, a referee threw a yellow penalty flag. The pass interference penalty on Sneed would give Minnesota a first down at the two yard line. Sneed ripped off his helmet in rage to argue with the referees, a violation that didn’t draw an additional penalty.

The referees gathered to discuss the play, and they ultimately decided not to call pass interference. The replay showed that Sneed had a hold of Addison’s jersey with his left hand, but Addison never pulled away to make it an obvious penalty.

The bigger problem was where the football landed. Addison was running towards the back right corner of the endzone, and Cousin’s pass hit the turf five yards to the interior of the field. The referees most likely ruled the pass uncatchable.

The Vikings got the ball again with a minute left in the 4th quarter, but they ran out of time before they could score a touchdown. With the 27-20 loss, Minnesota dropped to 1-4 for the season. 

During the previous 2022 season, Minnesota went 12-0 in one score games. Randomness plays a big role in games decided by eight or fewer points, and NFL teams regress to the mean of winning half of these games. Quants spent the entire season screaming about how overrated the Vikings were.

In 2023, each of Minnesota’s first five games had ended within one score. This season, they had a 1-4 record. Those same quants celebrated like Robert Kraft when the massage parlor opens.

Kansas City surged to a 4-1 record with the road win. More important to some bettors, the Chiefs covered the three point spread on the road. There were many ways to make the decision that Kansas City -3 at Minnesota was a good bet. 

Some of these methods make sense, like data driven models. Data is king, especially from someone like myself: Ed Feng, Stanford Ph.D. My company The Power Rank has been built on custom algorithms that use data.

Some of these methods make zero sense, yet somehow work quite effectively. Let’s look at one of the best shows of the 21st century to explain the basic idea.

---

Bloody hell, all of them. All of them have us finishing last this season. Every newspaper, every television pundit, every lonely middle aged sports blogging loser writing in his mother’s basement.

In the third season of the Apple TV series Ted Lasso, owner Rebecca Welton lamented how the supposed experts viewed her AFC Richmond team. Her soccer team had just been promoted to the Premier League.

Coach Ted Lasso pops into the office and tells Rebecca that the pundits are wrong. However, Ted notices that these pundits haven’t just affected the owner. His AFC Richmond players are also concerned about the unanimous opinion of the team.

To defuse the situation, Ted Lasso does the most Ted Lasso of Ted Lasso things: he takes the team to the London sewer. They enjoy the fumes of raw sewage and discuss the great stink of 1858 that led to the creation of the sewer. Finally, Richmond captain Issac McAdoo demands to know what they’re doing. Ted explains that the pundits are like poop. The team needs to build a tunnel system to let it flow through and not bother them.

The scene stinks. Yet it exemplifies the wisdom of Ted Lasso, who preaches a positive mental outlook to his players. In addition to thinking the pundits are wrong, Ted convinces his team to ignore them.

Like AFC Richmond, I look at preseason NFL rankings the week before the start of the season. These power rankings reflect the subjective opinion of sports writers. Not only do I look at these articles, but I use them to build a model.

I can see you wince. The analytics guy who touts his Stanford Ph.D. resorts to subjective power rankings in the preseason instead of putting in the hard work of building a data driven model. That probably doesn’t make you want to sign up for a membership to The Power Rank for $99 a year.

Looking at the details of these subjective power rankings might make you even less likely to sign up. Let’s look at an example:

Before the 2023 season, Sports Illustrated ranked the Carolina Panthers 25th. They wrote the following about Frank Reich’s team:

Bryce Young and the Panthers are going to end the season higher up than No. 25… A reasonable win ceiling for this Panthers team is nine games, based on the overall weakness of the division. I don’t think it’s far off to say the Panthers could win this division.

Sports Illustrated nailed one thing: the weakness of the division. Tampa Bay won the NFC South by inching over the .500 mark with a 9-8 record. Sports Illustrated had the Bucs 31st in these preseason rankings.

They failed on every other word about Carolina. Frank Reich got fired after a 1-10 start to the season. Carolina finished with an NFL worst 2-15 record. They would have gotten the first pick in the 2024 draft. However, they traded that pick to draft quarterback Bryce Young as the top pick in the 2023 draft. 

The Sports Illustrated preview is notable for another reason. In early 2024, news broke that Sports Illustrated used AI to generate content. Computer generated articles appeared on the website of what was once the most important publication in sports. Sports Illustrated denied any wrongdoing, but the story accentuated the decline of the sports media dinosaur. 

In addition, there is no picture on the profile page for Conor Orr, the senior NFL writer who wrote the power rankings article. In addition, Orr’s X profile has a cartoon of a wizard. “I don’t think it’s far off to say the Panthers could win this division.”

You are questioning any expertise that I have into sports analytics and betting. Why do I look at subjective power rankings? What could have possibly inspired me to do that?

---

In 2012, Nate Silver wrote about his model to predict the NCAA tournament in The New York Times. As a well known quant for predicting baseball and then elections, Silver used computer rankings like Jeff Sagarin and Ken Pomeroy. However, he had another component that caught my eye: the preseason AP poll.

Maybe this should not have been a surprise. Two years earlier, Ken Pomeroy wrote about his affection for the preseason AP poll. No single ballot to a poll is perfect, as each sports writer has the quirks inherent to all humans. However, collecting a large set of ballots cancels out these errors and leaves a strong signal of team strength. This is the wisdom of crowds. 

Silver noted that the higher ranked team in the preseason AP poll won March Madness games at an astounding rate. I’ve updated the analysis through the 2024 tournament. In the past 20 tournaments, the higher ranked team in the preseason AP poll has won 71.1% of games.

Maybe Ted Lasso was wrong.

The results for the preseason AP poll during March Madness made me think the same idea might work in the NFL.

In 2014, I started collecting subjective NFL power rankings before the start of the season. I don’t think too much when collecting these power rankings, as I go to big media sites like ESPN and Yahoo that pay sports writers to rank teams. I look for evidence that the writer put some thought into the rankings, and a short write up for each team suffices.

In using the wisdom of crowds, I want to rank teams but also make margin of victory, or spread, predictions. To do this, a team’s rank from a subjective power ranking gets mapped to a rating, or an expected margin of victory against an average team. For teams ranked first through 32nd, I get this rating from the past ten seasons of NFL data.

For example, Sports Illustrated had Kansas City first. Based on my historical analysis, the top ranked team in the NFL is 7.6 points better than average. Minnesota was 20th in Sports Illustrated, and this implies a rating of -1.5, or the Vikings are a point and a half worse than NFL average. 

The week before the start of the season, I collect 20 subjective power rankings. A team’s rating is the average over the 20 ratings from these experts. The aggregation of these power rankings is my preseason wisdom of crowds model, and they get published for free every year at The Power Rank the day before the start of the season.

How does the preseason model perform?

To get a prediction, let’s go back to the Kansas City at Minnesota game. In 2023, Kansas City was the opposite of AFC Richmond. Coming off a Super Bowl win over Philadelphia, the Chiefs were a unanimous choice for first among 20 subjective power rankings. This implied a rating of 7.6 points.

Despite a stellar 13-4 regular season record the previous season, the experts did not like the Minnesota Vikings. Maybe it was a playoff loss to the New York Giants. Maybe it was common sense. They were ranked 17th and rated 0.6 points worse than the NFL average.

To get a spread prediction at a neutral site, you take the difference in the rating of the two teams. Since Kansas City and Minnesota are better and worse than the NFL average respectively, the predicted point spread is 7.6 plus 0.6, or that Kansas City wins by 8.2 points. Minnesota played at home in week 5, and an estimate of uniform home field advantage is 1.7 points. Hence, the prediction becomes Kansas City by 6.5 points. 

Did you miss some of that? Don’t worry about the details of the math. The take home message is that the wisdom of crowds model predicted the Chiefs by almost a touchdown, which suggested betting the Chiefs to cover 3 points on the road.

I have put together this wisdom of crowds model since the 2014 season. Let’s look at all games in which the predictions of the model differs from the closing market by two or more points during the first 6 weeks of the season. We will talk about the remainder of the season later. 

In these games, the wisdom of crowds wins 54.6% against the market. This is a record of 296 wins and 224 losses. There were also 13 pushes, or games in which the game results were the same as the closing market. For example, if Kansas City is favored by 3 at Minnesota and Kansas City wins by 3, it is a push. A sports book would return your money for this bet. 

The wisdom of crowds model made a prediction in 506 out of 922 games during the first six weeks of the season. In the remaining games, the prediction was within two points of the market, and the model does not suggest value in these games.

I was shocked at these results when I found them in 2023.

In betting an NFL spread, you most often wager $110 to win $100. Betting slightly more than you win means you have to win more than 50% to beat the house. After working out the math, a bettor must win more than 52.4% of the time to make a profit. The betting community believes winning is particularly difficult in NFL spreads, one of the most massive sports betting markets. A win rate 54.6% of the wisdom of crowds for the six weeks of the season is a little shocking.

Let me stop to point out something about this wisdom of crowds model. It is not rocket science. It is not using the stationary distribution of a Markov chain to rank teams based on success rate. If I gave you the average rating for first through 32nd NFL teams from the past ten seasons, you could easily reproduce this model in a spreadsheet. You grab the preseason NFL rankings from 20 or so different media sites. Then map the rank for each team to the rating I gave you. Then take the average of 20 numbers for each team.

Even with these results, I do not recommend blind betting the prediction of the wisdom of crowds. Let’s look at an example.

---

Smarter sports betting in less than 5 minutes

Want to get these NFL wisdom of crowds preseason rankings? They will go first to The Power Rank email newsletter. This is a free service that strives to be:

• Valuable
• Concise
• Entertaining

In addition, you’ll get 5-Nugget Saturday, my curated list of sports betting tips for that weekend as well as analytics and humor.

To get this free service delivered straight into your inbox, enter your best email address and click on “Sign up now!”

 

 

 

---

In week 6 of the 2023 season, Arizona traveled to NFC West division rival the Los Angeles Rams. Arizona came into the season with all kinds of question marks, as quarterback Kyler Murray would miss the first part of the season in recovering from an injury. Josh Dobbs would start for the Cardinals against the Rams.

NFL experts like Adam Chernoff of Right Angle Sports thought Arizona had the worst roster in the NFL, a remarkable achievement in the salary cap era. In addition, the Cardinals traded away key players like linebacker Isaiah Simmons before the start of the season. Like AFC Richmond, the Cardinals were the consensus worst team among the twenty subjective power rankings.

The Los Angeles Rams didn’t look that much better before the 2023 season. After winning the Super Bowl in the 2021 season, the Rams suffered through a 5-12 season in 2022. Injuries to quarterback Matthew Stafford, receiver Cooper Kupp and defensive tackle Aaron Donald played key roles in this decline.

The crowd didn’t think the Rams would bounce back in the 2023 season. My wisdom of crowds preseason rankings had the Rams 29th out of 32 teams. At home against Arizona, the model predicted a 4.4 point win for the Rams.

However, the early season returns showed promise for the Rams. Stafford was healthy. Even though he didn’t have his favorite target in Kupp, rookie Puka Nacua shocked the league with 572 receiving yards in the first five games of the season. Aaron Donald returned from injury to anchor the defense. 

At The Power Rank, I also make predictions based on data from the current season. In the NFL, passing is king, and I use statistics like yards per pass attempt. This metric includes positive yards from plays downfield but also negative yards from sacks. Then, I adjust this quantity for strength of schedule. This is where that stationary distribution of Markov chains comes in.

Based on data from the current season, The Power Rank had the Rams rated 2.8 points better than NFL average heading into the Arizona game. Five games is a small sample size. However, the markets had already caught on to the Rams and shifted off its preseason expectations. The markets favored the Rams by a touchdown at home over the Cardinals.

Arizona led 9-6 at halftime, but the Rams scored 20 unanswered points in the second half. They ran the ball with second year running back Kyren Williams and easily covered the seven points.

Again, do not blindly bet with the preseason wisdom of crowds model during the first six weeks of the season due to a historical accuracy of 54.6%. The NFL is too hard to beat with a model alone during any time in the season.

Instead, use the wisdom of crowds model as a resource in your handicapping. Let’s go back to Kansas City at Minnesota, a game in which the preseason model suggested value in Kansas City to cover 3 points.

In my analysis before the week 5 game, many other factors favored the Chiefs. They brought back quarterback Patrick Mahomes, the best in the NFL, and long time coach Andy Reid. With this duo since the 2018 season, the Chiefs had played in every single AFC championship game. All five of these games were at home.

Before the 2023 season, the Chiefs had questions about their wide receivers or whether elite defensive tackle Chris Jones would sign a deal before the start of the season. Still, the Chiefs were the unanimous choice for number one.

In contrast, Minnesota ranked 17th in these preseason wisdom of crowds rankings. The Vikings have a superstar in wide receiver Justin Jefferson, a player that can make any quarterback look good. However, no other player on the roster stands out. The metrics for Kirk Cousins scream average among NFL quarterbacks. 

Kansas City covered the three points at Minnesota, giving the wisdom of crowds a nice win. But it wasn’t easy. The game shows the thin margins in winning an NFL spread bet. The referees could have never picked up the flag for pass interference. Minnesota scores a touchdown from the two yard line to tie the game. With a tie game, NFL teams tend to opt for a field goal since winning is the main objective. If Kansas City wins by a field goal, then the Kansas City -3 is a push. The best outcome is you get your money back.

With the excellent win rate of 54.6%, it would be nice to identify a general trend among games that the wisdom of crowds picks in the first six weeks of the season. For example, let’s suppose that Minnesota won by a 52-10 score in week 4 of the season. The markets would overreact to this one game, and the wisdom of crowds model, which sticks to preseason expectations, would identify this overreaction. 

However, after poring over these games, there is no general trend that I can find. Here is my conclusion: the wisdom of crowds has a track record of accuracy in the first six weeks of the season, although the 54.6% win rate could still be the result of small sample size. Still, there is enough evidence to use the model as part of our handicapping, although we reserve the right to ignore it.

How does the model perform later in the season? Let’s get into that next.

---

The audiobook version of this story is in The 60 Minute Guide to the 2024 Football Season. In addition, the audiobook has another story on whether Josh Allen, the gunslinger, is interception prone.

You may be able to listen to the audiobook for free. Learn more about The 60 Minute Guide to the 2024 Football Season.

---

Patrick Mahomes needed a hug.

Week 16 started with all the promise in the world for Kansas City. They would clinch the AFC West division with a win against Las Vegas at home. In addition, the Chiefs would not have to beat veteran quarterback Jimmy Garappolo. No, they only had to deal with rookie quarterback Aiden O’Connell, a 4th round pick. The markets closed with the Chiefs as an 11 point favorite. Easy enough.

Then disaster struck.

With 5 minutes remaining in the 2nd quarter, Kansas City got cute. With both Patrick Mahomes and running back Isiah Pacheco in the backfield, they snapped the ball to Pacheco. He then handed the football off to Mahomes for what was most likely a downfield pass. However, Mahomes botched the exchange. A Las Vegas defensive linemen scooped up the football and ran six yards for an easy touchdown.

After a kickoff, the Chiefs had the ball at the 25 yard line. Enough of that cute stuff. Kansas City decided to throw the ball to get the offense back on track. Throwing the football for the Chiefs is as safe as it gets in the NFL as Patrick Mahomes excels at not putting the football in dangerous situations. However, on this play, Mahomes throws it right to a Las Vegas defender who jogs into the endzone for a touchdown.

Las Vegas scored two defensive touchdowns in 12 seconds. Aiden O’Connell couldn’t generate a touchdown drive for the Las Vegas offense, but it didn’t matter. Las Vegas won 20-14.

With this loss to Las Vegas, Kansas City finished an eight game stretch with a 3-5 record. Losses to Philadelphia and Buffalo, two preseason Super Bowl contenders, might have been acceptable. However, losses to Denver and Las Vegas were not.

Kansas City had strayed far from the consensus number one team in the preseason. Before the start of the season, our friend Connor Orr at Sports Illustrated wrote the following in ranking Kansas City first:

The Chiefs are among a group of teams in the “elite” range. Patrick Mahomes then serves as a scale-tipper because, as we saw in the Super Bowl, when all else is at a stalemate, he can still create and advance.

Create and advance? Patrick Mahomes gave the game away to the Raiders with two defensive touchdowns in 12 seconds.

The preseason wisdom of crowds model had predicted Kansas City to beat Las Vegas by 13.9 points. The model suggested that the Chiefs should cover as an 11 point favorite, but they lost outright. This was a loser for the model.

This should not be surprising, as even a strong preseason model should lose its predictive power over the course of the season. In contrast, the market should get stronger with more data from the current season. 

Let’s look at how the preseason wisdom of crowds performs from the 2014 through 2023 season. From week 7 through the end of the regular season, the wisdom of crowds model wins at a 51.3% rate. This is a record of 595 wins, 564 losses and 37 pushes in games in which the prediction differs from the market by two or more points. This record is not as good as the 54.6% win rate during the first six weeks of the season.

A win rate of 51.3% is still respectable, especially for a model that didn’t make sense at first. However, the model does get worse by another metric. In evaluating models, I also look at error metrics that consider how the prediction differs from the actual game result. For example, the prediction was Chiefs by 13.9 points over the Raiders. The Raiders won by 6, so this gives a difference of 19.9 points.

In building my models, I look at error metrics based on the averages of these differences. While there are many different error metrics, I use something called the root mean square, or rms error. I won’t focus on the details here, but Wikipedia has an explanation. For our purposes, when I refer to error, I mean the rms error of the model from the actual game result.

To evaluate the preseason wisdom of crowds model, I will report the error during various times of the season. I’ll compare this with the error in the closing market, the gold standard in making predictions.

During the first six weeks of the season, the preseason wisdom of crowds model has an error of 13.2 points. For the closing market, the error is 13.0 points. The wisdom of crowds is close to the closing market during the first six weeks of the season, a result both thrilling and surprising.

However, for weeks 7 through the end of the regular season, the error of the wisdom of crowds model drops to 14.0 points. This might not seem that much worse than the error of 13.2 during the first six weeks of the season. However, these small differences matter in making predictions. The preseason model is less useful as the season progresses.

How does the model do in the playoffs? We’ll get to that soon. But first, let’s get back to the Chiefs. 

---

Kansas City was lucky to play in a weak AFC West division. After the loss to Las Vegas, they clinched the division the next week with a win against Cincinnati without Joe Burrow. With their playoff seeding locked into place, the starters rested the final week of the regular season.

What was wrong with Kansas City?

The problem was the offense. In the 2023 regular season, the Chiefs had scored 21.8 points per game, 15th in the NFL. This was worse than the NFL average of 21.9 points. Mahomes had tight end Travis Kelce, and wide receiver Rashee Rice was having a solid rookie season. However, there were questions about the other receivers. 

The only change the Chiefs made during the season was getting Mecole Hardman Jr. back. While the wide receiver had won a Super Bowl with the Chiefs the previous season, he decided to sign with the New York Jets. However, Hardman couldn’t get on the field behind two rookies. He asked for a trade and got sent back to Kansas City.

Surprisingly, the defense was the better unit for Kansas City. Led by stalwart defensive tackle Chris Jones and the emergence of second-year cornerback Trent McDuffie, they allowed 17.3 points per game, second best behind Baltimore in the 2023 regular season. 

Once the playoffs started, Kansas City started to play well again. In a Wild Card game at home, the Chiefs cruised to a 26-7 win over Miami. In the Divisional playoff, Kansas City won a tough 27-24 game at Buffalo.

In the AFC conference championship, the Chiefs would make their sixth straight appearance. But unlike their first five appearances, Kansas City had to go on the road. In addition, they faced Baltimore, the team that had been the best in the NFL for the 2023 season. The preseason model liked the Chiefs by 2.8 points at Baltimore. In contrast, the markets closed with Baltimore as a 4.5 point favorite.

This market prediction made sense from a data perspective. At The Power Rank, I put together a model based on data from the current season. From the play by play, the model grabs data like success rate, a metric that puts context into the result of every play. Success rate knows that a two yard gain on 4th and 1 is much different from a two yard gain on 1st and 10. Metrics like success rate get adjusted for strength of schedule, my specialty given my math background. 

This is the same model that I discussed earlier for the Los Angeles Rams. This data driven model predicted Baltimore by 4.3 points. This is close to the market that favored the Ravens by 4.5, and it is far from the preseason wisdom of crowds model that liked Kansas City to win straight up.

At Baltimore, Kansas City didn’t score in the second half, not a winning formula. However, Baltimore had three turnovers to zero for the Chiefs. In the biggest play of the game, wide receiver Zay Flowers fumbled as he was about to score a touchdown. Perhaps a little on the lucky side, Kansas City won 17-10 outright for another trip to the Super Bowl. However, it never seemed like Baltimore would win by five or more points to cover the spread.

In the Super Bowl, Kansas City faced San Francisco at a neutral site in Las Vegas. The preseason model again favored the Chiefs, this time by two points. In contrast, the market closed with San Francisco as a two point favorite.

Again, the market made sense based on data from the current season. My model based on metrics like points, yards per pass attempt and passing success rate favored San Francisco by 2.3 points, very close to the market that favored the Niners by 2.

Kansas City started the game slow, as they scored their first three points within a minute of halftime. They got a break in the second half when San Francisco had an extra point blocked. The Chiefs tied the game with a late field goal, and the Super Bowl went into overtime.

San Francisco scored a field goal on their first possession. Since the score wasn’t a touchdown, Kansas City got an opportunity with the football. On 4th and 1 from their own 34, Patrick Mahomes scampers for a 1st down. Later, on 3rd and 1 from the San Francisco 32, Mahomes runs for another 1st down, this time a 19 yard gain.

Finally, two yards from the end zone, Mahomes throws a touchdown pass to Mecole Hardman Jr., the same player that couldn’t find any playing time with the Jets. Kansas City wins 25-22 for their second straight Super Bowl.

Let’s think back to Connor Orr of Sports Illustrated in ranking Kansas City first in the preseason. “Patrick Mahomes then serves as a scale-tipper because, when all else is at a stalemate, he can still create and advance.” Maybe AI isn’t so bad.

With the spread winner in this Super Bowl, the wisdom of crowds model has gone 52.9% against the spread in the playoffs. This is for games in which the prediction differs from the market by two or more points. This is a record of 46 wins and 41 losses, with two pushes on games.

In most situations, I would not tout a result from a small sample size that includes less than 100 games. In fact, I think that’s my high school math teacher Bob Fielding at my door, ready to give me an earful.

However, the error is smaller in the playoffs than in the latter part of the regular season. The error is 13.3 points in the playoffs, much smaller than the 14.0 from week 7 through the end of the regular season. The error in the playoffs of 13.3 is comparable to the error of 13.2 during the first six weeks of the season.

Again, do not make NFL spread bets solely because of this wisdom of crowds model. The market is too strong. In addition, the closing market has an error of 12.4 points in the playoffs, smaller than the 13.0 during the regular season.

However, the preseason wisdom of crowds model is a useful resource in the playoffs. For Kansas City, it ignores a 3-5 stretch during the latter part of the regular season and notes that the Chiefs still had two key assets: quarterback Patrick Mahomes and head coach Andy Reid. A surprisingly good defense also helped the cause. With less than twenty games, the entire NFL season is a small sample size. Oftentimes, a team will revert to their true skill level, and the preseason wisdom of crowds model sheds light on this talent for the entire season.

Smarter sports betting in less than 5 minutes

Want to get these NFL wisdom of crowds preseason rankings? They will go first to The Power Rank email newsletter. This is a free service that strives to be:

• Valuable
• Concise
• Entertaining

In addition, you’ll get 5-Nugget Saturday, my curated list of sports betting tips for that weekend as well as analytics and humor.

To get this free service delivered straight into your inbox, enter your best email address and click on “Sign up now!”