Learning About Football Fandom Using Social Media Data

Millions of tweets referencing NFL teams turn out to hold two genuinely different kinds of information: where fans actually are, geographically, and how a crowd’s collective mood moves in near-real time as a game unfolds — neither of which shows up in a box score.

Locating a Fan Base From Where the Posts Come From

A team’s local ticket sales and market size say something about its fan base, but they miss fans who root for a team without living anywhere near its home city — the college football expatriate, the transplant who never switched allegiances, the fan who simply adopted a team that does not play anywhere close to home. Geotagged social media posts referencing a team offer a way to measure that wider reach directly, since each post carries both a location and a declared (or inferred) rooting interest.

A study by Elisabeth Kollrack, Michael Arnold, Peter Sheridan Dodds, and Christopher Danforth — researchers associated with the University of Vermont’s Computational Story Lab, known for building large-scale sentiment-tracking tools — analyzed game-referencing tweets across the 2011 through 2014 NFL seasons and calculated a “fandom radius” for every team: a measure of how far a team’s geotagged fan engagement extends beyond its home market before dropping off. The results varied enormously across the league. The Baltimore Ravens posted the smallest radius in the study, around 121 kilometers — a fan base heavily concentrated close to home. The Seattle Seahawks posted the largest, around 782 kilometers, reflecting engagement stretching well beyond the immediate Seattle metropolitan area.

Why Fandom Radius Varies So Much Across the League

The study’s explanation for that spread ties back to geography and competition rather than anything about team quality or success. Seattle sits with the Pacific Ocean to its west and the Canadian border to its north, with no nearby NFL franchise competing for fans across a wide swath of the surrounding region — a structural advantage that lets a team become the de facto choice for a large geographic area regardless of how it happens to be performing in a given season. Minnesota fits a similar pattern. Teams in geographically dense NFL markets, by contrast, compete directly with one or more nearby franchises for the same local pool of potential fans, which tends to compress their measured fandom radius even when the team itself is popular.

This reframes what a “big market” team even means. Two teams can have similar total fan engagement while one draws almost entirely from its immediate metro area and the other pulls meaningfully from hundreds of kilometers away — a distinction a simple market-size or attendance figure cannot make, but a geography-aware analysis of social media data can.

How a Crowd’s Mood Moves During a Game

The same research team also tracked sentiment expressed in NFL-referencing tweets before, during, and after games, and found a consistent shape to the emotional arc regardless of which teams were playing: sentiment runs positive in the lead-up to kickoff, dips as the game actually begins (uncertainty replacing pre-game optimism), recovers somewhat around halftime, and then splits sharply based on how the second half unfolds. Fans of the team that ends up winning see sentiment climb back up toward the end of the game; fans of the team that ends up losing see sentiment stay depressed through the finish. Notably, both groups — winners and losers alike — tended to end the game with sentiment below where it had started pre-game, suggesting that even a win does not fully restore the optimism a fan base carried in before kickoff.

Perhaps counterintuitively, the study found only a weak positive relationship overall between a team’s actual on-field success and the sentiment its fans expressed on social media. Winning correlates with happier fans, but not as strongly or as simply as might be assumed — in-game events, pre-game expectations, and other factors evidently shape the emotional signal in ways that final score alone does not capture.

Using Sentiment as a Predictive Signal, Not Just a Descriptive One

A separate line of research has tried to push social media sentiment further, from describing fan mood to using it as an input for predicting game outcomes. A 2017 paper in the journal Decision Support Systems, by Robert Schumaker, Chester Labedz, A. Tomasz Jarmoszko, and Leonard Brown, borrowed a technique from stock-market technical analysis — comparing a short-window moving average of sentiment against a longer-window baseline, looking for a “golden cross” pattern where the short-term signal crosses above the long-term one, mirroring how the same pattern is read as a bullish signal in stock charting. Applied to game-referencing tweet sentiment around NFL games during the 2015–2016 regular season, the paper reported a $14.84averagetheoreticalreturnpersentiment-basedwageracrossall256gamesofthatseason,comparedwitha$12.21 average loss using an odds-only baseline over the same games, with underdog wagers specifically exhibiting the golden-cross pattern performing best of all in the study.

The explicit motivation behind the technical-charting approach was a genuine measurement problem: raw sentiment differs systematically by fan base — some fan bases run more pessimistic or more vocal than others regardless of team quality — so comparing two teams’ absolute sentiment levels directly risks mistaking a fan base’s baseline personality for a real predictive signal. Treating sentiment as a time series and measuring its change relative to its own recent baseline, rather than its raw level compared against a different team’s fan base, was the paper’s answer to that confound — a reminder that a single sentiment snapshot compared across different fan populations, without accounting for those populations’ different baselines, is likely to mislead more than it informs.

What This Kind of Data Cannot Tell You

Every finding above shares a limitation worth naming directly: it describes people who post about football on a specific platform, not football fans in general. Posting behavior is not evenly distributed across a fan base — younger fans, more online-active fans, and fans with stronger opinions in either direction are all more likely to be captured in a dataset built from public posts than a fan who watches every game quietly and never posts about it. A fandom radius built from geotagged posts is really measuring the geography of posting fans specifically, which is a reasonable proxy for the wider fan base but not a direct census of it — the same caveat that applies to essentially any research built on social media data rather than a designed survey with a known sampling method.

This does not make the underlying findings uninteresting or unreliable for what they actually measure. A fandom radius genuinely describes something real about how far a team’s most vocal, most online-active supporters extend geographically, and an emotional arc genuinely describes how that same population’s collective mood shifted around a real set of games. The caution is narrower than “the data is wrong” — it is closer to “the data describes a specific, identifiable subset of fans, and generalizing findings about that subset to an entire fan base requires the same care any survey researcher would apply when reasoning from a non-random sample to a wider population.”

The same discipline applies to the sentiment-as-predictor research above: a reported return over one season of games is a single sample, and the uncertainty and small-sample cautions covered elsewhere apply just as much to a sentiment-based signal as to any other football metric estimated from a limited number of games. Grading a predictive claim properly after the fact — rather than taking one season’s reported result at face value — is the same standard any forecasting method, sentiment-based or otherwise, should be held to.