Prediction from Regional Angst - A Study of NFL Sentiment in Twitter Using Stock Market ChartingSpeaker: Robert P. Schumaker – New Britain, CT, United States
Topic(s): Information Systems, Search, Information Retrieval, Database Systems, Data Mining, Data Science
A three-game losing streak and Chicago Bears fans are hopeful, whereas a two-game winning streak and Oakland Raiders fans are expecting the worst. Two different fan-bases and two very different ways of expressing sentiment, one optimistic and the other pessimistic. This type of fan behavior makes the comparison of sentiment between geographically different fan bases a non-trivial problem. In an ideal world with 32 NFL teams, there would be 32 identical fan-bases which would have similar experiences and behave in a fairly consistent, predictable manner. However, this isn’t the case, as illustrated by the earlier example. Each NFL market is a unique and ever-changing mix of individuals with differing perceptions of the world around them. Their perceptions of expected team performance can be captured in their writings (in our case, tweets), analyzed for sentiment and aggregated to form a crowdsourced signal that can be used to forecast the winning team. For example, on a scale of -100%, all negative, to +100%, all positive, the Raiders may have a pre-game tweet sentiment rating of -6.04% and the Bears 36.7%. Following our example further, the Raiders had a record of 2 wins and 1 loss while the Bears had lost all three of their contests. The Raiders were favored in the pre-match betting lines, -179 Raiders to +148 Bears. At that point, did the 36.7% positive tweet sentiment for the Bears have greater predictive value than the teams’ record or the betting line? What if Raiders fans are naturally more pessimistic than Bears fans, can we identify a comparable signal in the data? The problem is that making comparisons between teams in an absolute sense and without consideration of their fanbase differences may lead to a less successful prediction. While studies have been successful in using absolute sentiment, we feel that there is a better approach to address this problem.
Our approach is to borrow techniques from technical charting used in stock price analysis. In particular, we analyze sentiment polarity as a time-series signal and examine the position and magnitude of signal change between two temporal windows. In technical charting, a popular technique to analyze price movement is to study the 50 day and 200 day moving averages. Stocks whose 50 day averages cross above their 200 day average are referred to as golden crosses and typify investment opportunity. Whereas stocks whose 50 day averages cross below their 200 day average are referred to as death crosses and typically signal stock price trouble. This type of time-series price signal is similar to that of polarity sentiment prior to a game, which we also can analyze as time-series data.
From our analysis we found a $14.84 average return per sentiment-based wager compared to a $12.21 average return loss on the entire 256 games of the 2015-2016 regular season if using an odds-only approach. We further noted that wagers on underdogs (i.e., the less favored teams) that exhibit a “golden cross” pattern in sentiment (e.g., the most recent sentiment signal crosses the longer baseline sentiment), netted a $48.18 return per wager on 41 wagers. These results show promise of cross-domain research and we believe that applying stock market techniques to sports wagering may open an entire new research area.
About this LectureNumber of Slides: 34
Duration: 60 minutes
Languages Available: English
Request this Lecture
To request this particular lecture, please complete this online form.
Request a Tour
To request a tour with this speaker, please complete this online form.
All requests will be sent to ACM headquarters for review.