How AI Is Turning Sports Forecasting Into a Data Game

Follow those forecasts outward from the field, from run designs to draft-day trades and on to the halftime stage, and a pattern is hard to miss. The farther a question sits from measured play, the less the tracking data helps, and the more the answer depends on who knows what and on how a contract defines the result.

Next Gen Stats Now Reads What A Run Play Was Meant To Be

The first Next Gen Stats stories were about speed and distance, the league's analytics team recalled on NFL.com on September 7. The run game stayed stubborn. A box score says who carried the ball and how far he got, and nothing about the design of the play or which blocker lost his matchup.

This offseason the team took on that gap with the professional services group at Amazon Web Services, using the same transformer architecture that powers large language models. The new run-scheme model labels each run as one of 16 intended designs, sorts them into man, zone and gap schemes, and records the gap the play aimed for next to the one the runner actually hit. Spotting that detail on every play used to take a trained eye and hours of manual charting.

A second new model maps every blocker's assignment, the type of block and the moment each engagement starts and ends, so the league can count how often a defender draws a double team and how long he needs to shed it.

Select metrics now reach the broadcast in real time, starting with the first replay. Prime Vision, the Thursday night alternate stream, is back for a fifth season with on-screen predictive features, and ESPN's MNF Playbook returns after its debut last December, built around a 22-player field view and live probabilities. This is the rung where the AI claim holds up best. Every snap produces the same kind of measurement, and a model can be checked against thousands of plays a season.

Draft Models Score Prospects, Front Offices Model Each Other

The draft sits one step back. The Next Gen Stats Draft Model, in its seventh year this spring, scores prospects on production, athletic testing and other indicators, with position-specific models trained on draft classes going back to 2003, the league's analytics staff wrote on NFL.com before the draft.

The combine is where the league's tracking data meets the prospects. Indiana cornerback D'Angelo Ponds posted the fastest maximum acceleration of any defensive back in this year's W drill, 6.53 yards per second squared. Notre Dame running back Jeremiyah Love drew a draft score of 96, the highest in the class and the best by a running back in 24 years, and Arizona took him third overall.

The team's own summary of its record names hits, among them eventual Defensive Rookie of the Year Carson Schwesinger, who made last year's list. It doesn't give a hit rate, and a list of successes isn't a calibration test. Read the score as a careful profile, not a forecast of where a player goes.

The forecast that moved a pick this spring was about rival teams. Kansas City general manager Brett Veach and the head of the club's data analytics team ran predictive models on how the top of the board would fall, ESPN reported. "The predictive model said that the Saints are going to be the biggest obstacle," Veach said.

So the Chiefs sent picks 9, 74 and 148 to Cleveland, jumped to No. 6, ahead of New Orleans at No. 8, and took LSU cornerback Mansoor Delane. They had also worked to stay invisible. Defensive coordinator Steve Spagnuolo advised against bringing Delane in for a facility visit and suggested a video interview instead, to keep the interest quiet. That's the problem for anyone forecasting the draft from outside a building. Teams model one another, and the sharpest of them hide the very signals everyone else is modeling.

On Draft Night, Information Beats The Model

Mock drafts are the draft's public forecasts, and they get graded. After the draft, Blogging The Boys scored 50 final mocks with The Huddle Report's system, one point for naming a first-round player and two for matching him to his team. Luke Easterling of Athlon Sports named 30 of the 32 first-rounders, and Phil Perry of NBC Sports Boston matched 12 players to the right clubs, the most among the 50. The crowd landed mid-pack. The consensus mock at NFLMockDraftDatabase.com got 26 players and six exact matches, tied for 25th, in a first round where 11 teams changed position through trades.

Some misses were information problems rather than modeling problems. ESPN's Matt Miller heard days of buzz about Arizona taking Love and stuck with chalk after pushback from sources. "Sometimes, the buzz is right," he said afterward. The Rams' pick at No. 13 had the opposite problem. Alabama quarterback Ty Simpson arrived with almost no predraft buzz, as NFL Draft Buzz's review of the class put it, and most boards had ranked him around 40th.

Prediction markets priced the same night. Kalshi listed 127 draft markets before the first round, USA Today reported. Polymarket's market on the first pick drew $1,358,646 in volume and settled on Fernando Mendoza, the Raiders' choice. Its rules name official NFL information as the source and add, "A consensus of credible reporting may also be used."

That clause is where draft night turns into an information race. In a March 29 letter, the NFL asked Kalshi, Polymarket and other companies not to offer contracts whose outcome can be known in advance, live pick-by-pick markets included. It repeated the request on April 22, because picks often leak on social media and news sites before they're formally announced. A model can't outrun a leak. Much of the answer already sits in 32 draft rooms, so the most accurate forecast is often whatever escapes first, and that's the kind of contract the league asked the exchanges not to list.

A Third Of Kalshi's Super Bowl Volume Never Touched The Game

Super Bowl LX showed how far from the field the forecasting money can travel. DeFi Rate, which tracks exchange data, counted $833.2 million in Super Bowl-related trading on Kalshi, and $279.1 million of it, a full third, sat in halftime, entertainment, advertising and guest-appearance markets rather than the game.

The contract on which song Bad Bunny would open with drew $113.5 million, second only to the game-winner market at $358.8 million. The market on the commercials took in $72.2 million, halftime performers $47.3 million and guest appearances $39.8 million. A year earlier, Kalshi's entire Super Bowl menu had been five markets.

None of those contracts asks who wins a football game, and none can be read off a tracking feed. They ask about the show wrapped around the game, from the setlist and the guest list to the commercial breaks, and DeFi Rate's February breakdown filed them under Kalshi's entertainment marketplace rather than its game lines.

The game-winner contract settled on a final score. A contract on a setlist, a cameo or a broadcaster's words has no scoreboard to read, and a guide to pop culture prediction markets from GamingToday counts celebrity and mention contracts as the narrowest of its categories, warning that a market on whether a phrase gets said on live television can hinge on a judgment call about wording, timing or context.

Transcripts Can Price A Phrase, Not A Cameo

Some culture contracts can still be modeled from data. Kalshi ran announcer mention markets, contracts on whether the broadcast crew says a given word, on 64 NFL games last season, with more than $47 million staked on them by the week before the Super Bowl, ESPN's David Purdum reported.

One trader he profiled, a business intelligence analyst who majored in statistics, found a public archive of NFL announcer transcripts covering hundreds of games and wrote code to count how often each phrase turned up. His estimate put "what a catch" in about 10% of games, while the market had the No side at 40 cents.

That's the closest thing culture markets have to tracking data, and it still cost him on a single sentence. In the January 18 playoff game between the Rams and Bears, Cris Collinsworth watched a replay of a 17-yard catch by a Bears receiver and said, "What. A. Catch." One game proves little either way. At Super Bowl LX, DeFi Rate counted 37 announcer strikes, the individual phrases on offer, worth $9.0 million in all. Only 16 resolved Yes, and the most-traded of them, "What a Catch," resolved No.

A performance market has no transcript to mine. Cardi B joined four other guests on the porch set during Bad Bunny's halftime show and danced, but it wasn't clear whether she sang, the Associated Press reported. Kalshi's market on who would perform, which had drawn more than $47.3 million, settled at the last traded price, 74 cents for No and 26 for Yes, citing ambiguity over whether her appearance was a qualifying performance.

A Yes holder took a complaint to the Commodity Futures Trading Commission, seeking $3,700. No transcript archive could have priced that, because the uncertainty sat in one word of the contract rather than in the show.

The League Objects And The Courts Are Split

Whether these contracts get listed at all isn't a modeling question. The NFL told a congressional committee in written testimony last December that it had no plans to participate in prediction markets, and executive vice president Jeff Miller singled out announcer contracts on phrases such as "roughing the passer" as objectionable. The phrase still appeared as a Super Bowl strike on Kalshi, where it drew $724,000 before resolving No.

The law is less settled than the league's position. Federal appeals courts split on sports event contracts this year. In April the Third Circuit backed Kalshi in its fight with New Jersey. Nevada won its round in the Ninth Circuit in August, and Ohio and Tennessee won theirs in the Sixth Circuit in September. New Jersey's petition now sits with the Supreme Court. Until the justices act, a sports contract's legal footing depends on the state, and on which circuit's ruling binds it.

Next spring the draft moves to Washington, D.C., with the Draft Theater on the National Mall in front of the Capitol. The publisher of that culture-market guide also runs GamingToday on Facebook. Which number to believe on draft night is still the one call no model makes for you.