[Paper Review] Predicting the NFL Using Twitter

Shiladitya Sinha, Chris Dyer+2 more|arXiv (Cornell University)|Oct 25, 2013

Sports Analytics and Performance18 references29 citations

TL;DR

This paper proposes using Twitter sentiment and volume metrics to predict NFL game outcomes and betting results, demonstrating that simple Twitter-derived features—especially rate-of-change in tweet volume—can match or exceed traditional statistical features. The key finding is that a Twitter rateS feature (based on tweet volume changes) achieves 55.3% accuracy in predicting point spread winners, surpassing the 53% threshold needed for profitability after bookmaker fees.

ABSTRACT

We study the relationship between social media output and National Football League (NFL) games, using a dataset containing messages from Twitter and NFL game statistics. Specifically, we consider tweets pertaining to specific teams and games in the NFL season and use them alongside statistical game data to build predictive models for future game outcomes (which team will win?) and sports betting outcomes (which team will win with the point spread? will the total points be over/under the line?). We experiment with several feature sets and find that simple features using large volumes of tweets can match or exceed the performance of more traditional features that use game statistics.

Motivation & Objective

• To investigate whether social media data from Twitter can predict NFL game outcomes and betting results with higher accuracy than traditional game statistics.
• To evaluate the effectiveness of various Twitter-derived features—such as sentiment, volume, and rate of change—compared to conventional statistical features in forecasting NFL game outcomes.
• To develop and release a large-scale, annotated dataset of tweets aligned with NFL games (2010–2012) for academic research.
• To determine whether real-time social media sentiment and volume trends can serve as leading indicators for team performance and betting market movements.

Proposed method

• Collected 10% of the Twitter stream (garden hose) during the 2010–2012 NFL seasons, processing ~42M tweets/day using the twokenize tokenizer.
• Assigned tweets to NFL teams using manually curated hashtags (e.g., #giants, #jets), discarding tweets with multiple team hashtags to ensure team-specific sentiment.
• Defined three tweet categories: pregame (24–1 hr before), postgame (4–28 hrs after), and weekly (between games), using time-based filtering.
• Extracted features including unigram word frequencies (e.g., 'win', 'loss'), and novel rate features: rateS(vprev, ∆=500) and rateP(vprev, θ), which measure tweet volume change relative to prior volume.
• Used canonical correlation analysis (CCA) to combine Twitter features with statistical game features, and trained binary classifiers for three prediction tasks: winner, winner with spread (WTS), and over/under.
• Implemented dynamic feature selection: for each week, selected the best-performing feature set from the prior two weeks, enabling adaptive model updates.

Experimental results

Research questions

• RQ1 Can Twitter-based features, particularly volume and sentiment trends, predict NFL game outcomes as accurately as traditional game statistics?
• RQ2 Do rate-of-change features in tweet volume (e.g., rateS, rateP) outperform static tweet counts or sentiment features in forecasting game outcomes?
• RQ3 Is there a measurable relationship between social media sentiment and actual game outcomes, especially in the context of sports betting markets?
• RQ4 Can dynamic feature selection based on historical performance improve prediction accuracy over time in a real-world forecasting setting?
• RQ5 To what extent do Twitter features alone, without game statistics, achieve profitability in NFL betting (i.e., >53% accuracy on WTS predictions)?

Key findings

• The Twitter rateS feature (using vprev and ∆=500) achieved 55.3% accuracy in predicting the winner with the spread (WTS), exceeding the 53% threshold required for profitability after bookmaker fees.
• Twitter unigram features alone achieved 54.3% accuracy on over/under predictions, outperforming most other feature sets and suggesting a strong signal in raw tweet volume and word frequency.
• Combining Twitter unigrams with statistical features via CCA improved WTS prediction accuracy to 51.0% (with 4 components), though the best performance came from hybrid models including rate features.
• The best-performing feature set for WTS prediction changed frequently—8 distinct sets in 13 weeks—highlighting the need for adaptive, time-varying feature selection.
• The F5 ∪ F9 ∪ Twitter rateP (vprev, θ=0.2) combination achieved 65.9% accuracy on winner prediction and 51.4% on WTS, both exceeding baseline models and indicating strong predictive power of combined features.
• Despite high accuracy in some feature sets (e.g., 58.2% for over/under with F3 ∪ F4 ∪ Twitter rateS), the dynamic feature selection strategy only achieved 44.1% accuracy on over/under, indicating room for improvement in this task.