Colin Kaepernick, Nike, and the
Consequences of the 2018 Just Do It
Campaign:
An Analysis of Social Media, Press,
and Stock Reactions
Kwong Yau
Sarah Erbes
Scott Mow
& Jordan Cherry
December 14, 2018
Executive Summary
This project attempts to analyze
the effect of Nike’s use of Colin Kaepernick as the lead athlete for its 2018
marketing campaign on its stock price during that period.
We outlined a three-fold
approach:
·
Examine the social media reaction to determine
overall sentiment about Colin Kaepernick’s partnership with Nike
·
Examine press reactions to determine press
coverage sentiment about Colin Kaepernick’s partnership with Nike
·
Examine financial data to examine possible
correlations between media (both traditional and social) reactions and Nike’s
subsequent financial performance, for simplicity, measured in daily closing
stock price.
By incorporating web and text
mining using Rapid Miner and R to create a corpus of news articles, blogs and
Tweets, our goal is to draw conclusions of the sentiment surrounding Nike
during that period and gain insight into the possible value and peril of
running politically controversial advertising campaigns for companies.
Background
Colin Kaepernick, former quarterback
for the NFL’s San Francisco 49ers began his protest of police brutality by
refusing to stand for the national anthem in the 2016 NFL preseason. When asked
about his decision not to stand for the anthem, Kaepernick stated his reasons
for protesting. The controversy regarding his actions quickly became one of the
biggest news stories of the year, and a dividing line in American politics.
Sports Illustrated magazine awarded him with the honor of Sportsperson of the
Year, while opponents widely and passionately criticized him. Although
Kaepernick was released from his contract with the 49ers at the end of the NFL
season, Nike maintained its sponsorship contract with the quarterback, even as
he remained unsigned by any other NFL team.
On September 3, 2018, Colin
Kaepernick and Nike sent out tweets signaling that Colin Kaepernick would be
the face of Nike’s 2018 “Just Do It” campaign. The reaction on twitter was
immediate; opponents sharply criticized Nike while supporters praised the
company just as enthusiastically. As of the time of this report, Colin
Kaepernick remains the face of Nike’s Just Do It campaign.
The caption of Kaepernick’s Nike
ad reads, “Believe in something, even if it costs you everything.” The question
we seek to answer is what, if anything, did the ad campaign cost Nike?
Methods
We used RapidMiner and R in order
to conduct our analyses. In order to conduct each of our analyses, we
determined that a multi-pronged approach was necessary. While financial data
and media articles were relatively easy to obtain, limits on RapidMiner and R’s
interfaces with Twitter made it impossible to access tweets from September
2018, when the reactions specific to Nike’s Colin Kaepernick ad were
trending. Further to this, Twitter
imposes hefty charges for providing “filtered access” to archived content, e.g.
entry level pricing is $1,250 for a one-time request limited to 1 million
tweets over a 40-day period.
However, both Nike and Colin
Kaepernick are still currently widely discussed on Twitter, and many of the
original hashtags relating to Kaepernick’s protest are active. The tweets
collected were used to gauge overall sentiment regarding Kaepernick and Nike as
well as identify the themes and concepts most often connected to both
Kaepernick and Nike.
Article Analysis
A piece of our research was to analyze the
sentiment of articles released during the time of the Nike Colin Kaepernick
ad. Our goal was to see if the sentiment
of the articles aligned with the stock price movement of Nike. Our article listing is a combination of news
articles and opinion pieces written about the controversial advertisement.
Clustering Article Analysis
The RapidMiner process above was used for a
cluster analysis of the articles and opinion blogs that had been
collected. The sub-processes under the
“Process Documents” operator were Transform Cases, Tokenize, Filter Tokens,
Stem(porter), Filter Stopwords, and Generate n-grams. A limit of 4 clusters for the articles was
imposed. Below is the output for the
word list among the articles showing the top 10 words:
Below is
the output of the cluster analysis:
The two
largest clusters were 1 and 3, the output of these clusters follow. Below is the centroid table output for the top
ten words in cluster 1:
Top ten
words for Cluster 3:
While Cluster 3 contains more words that would be
expected with a marketing ad for a clothing company, there are still
politically driven words at the top of the list. Cluster 1 contains heavily political words as
well in their top ten, especially ‘black_male’ which is at the center of this
political issue.
Analyzing the Sentiment of the
Articles
We used the Aylien
extension to analyze the sentiment of the articles based on the text of the
articles. The results from this analysis
were 11 positive articles, 0 negative articles, and 34 neutral articles with an
average polarity confidence of 0.9. It
was surprising that none of the articles were found to be negative by the
Aylien algorithm. There were 30
objective articles and 15 subjective articles, with a confidence of 1. Interesting to note that the articles that
were marked with subjectivity were all opinion pieces.
We then ran the same articles through the Rosette
sentiment analysis, this produced a different result. There were 30 neutral, 3 positive, and 2
negative articles many more neutral than the Aylien analysis but Rosette was
also able to identify some negative articles.
The sentiment score minimum was .545, maximum of 1, and an average of
0.857. Below is the pie chart output
from RapidMiner.
Both of the prediction models below are
attempting to predict the sentiment based on the text mined from each
article. The sentiment was based on the
Aylien sentiment analysis conducted above.
Predicting
Polarity with Naïve Bayes:
This model took the text and publish date to try
and predict the polarity. As seen by the
results, the Naïve Bayes model did not do a good job attempting to predict the
polarity of the articles. The accuracy
level is only 12.5% and it was only able to predict some of the positive
articles correctly. Next, we tried to use a decision tree to better predict the
outcome.
Decision
Tree Model:
The decision tree model was much better at
predicting the classification than the Naïve Bayes model was. The accuracy level was much higher at 60.83%
but there are still issues with this model.
The model predicted every article as neutral, while the neutral recall
is 100% the others are still 0%.
Article Conclusions:
There were a few key points to take away from the
analysis of the articles. The first, is
the insight the clustering provided. The
word lists and the clusters both point to social activism words being present
in these articles. This confirms our
thinking that this was advertisement was indeed more than just another clothing
ad. Surprisingly, there was not as much negative sentiment for the
advertisement as our team would have expected.
The positive reactions align with the stock market increase that Nike
saw immediately after the advertisement was released. The final take away is that Naïve Bayes as
well as the decision tree were unable to successfully predict the sentiment of
an article. If further research were to
be done, we would need to find a better prediction model.
Establishing Corpora of Tweets
We used RapidMiner in order to
establish our corpora of tweets. Using Rapidminer’s Twitter interface, we were
able to conduct searches for “Colin Kaepernick”, “Nike”, “#boycottNike”, and
“#TakeAKnee.” The two hashtag searches were hashtags used by opponents and
supporters, respectively. The searches specifically excluded retweets and
tweets with URL links in order to emphasize more original tweets. These corpora
were exported to Excel documents for further analysis.
A screenshot of our RapidMiner
process for establishing our corpora can be found in Appendix A.
Sentiment Analysis
Once the corpora of tweets were
collected, we used RapidMiner’s Aylien extension to perform sentiment analyses
on each of the corpora. Unfortunately, we were unable to access Aylien’s
dictionary for positive and negative terms, and thus the sentiment analysis
operator represents a “black box” process in our analysis. However, the Aylien
extension is considered reliable for the purposes of this course.
A screenshot of our RapidMiner
process for analyzing corpora can be found in Appendix A.
Word Could Creation
Using Excel, we created CSV
copies of the corpora that could be imported into R to create word clouds. Our
R script used the TM package for text mining, and the WordCloud package for
creating the word clouds.
The TM package was used to
perform the following steps:
·
Create a corpus from the imported CSV files
·
Transform cases to all lower case
·
Remove numbers from the text
·
Remove stopwords
·
Strip extra white space
·
Stem the terms in the text
After we used the TM package to
clean the text of the tweets, we used the WordCloud package to create word
clouds showing the terms most associated with each search. All default settings
were used for this process. The resulting word clouds can be found in the
“Results” section.
The R script for creating word
clouds can be found in Appendix B.
Stock Price
The daily stock prices were scraped from Yahoo.com. The URL follows a consistent pattern, so we
just had to substitute each company's stock symbol. Using the “Get Pages” operator, we were able
to read the raw HTML for each company's page.
There was a common class name for each row containing the data, so we
were able to parse the close date and value using the “Cut Documents” operator
with an XPath query type. From there,
dummy ids were used to join the data into a table. The full RapidMiner process can be found in
Appendix C.
Sentiment Analysis Results
The Aylien Sentiment analysis of
the search “Nike” showed that of the 550 tweets collected, 291 of them were
neutral, 138 were negative, and 121 were positive.
Graphs of the polarity and
subjectivity of the search results for “Nike” can be seen below:
The Aylien sentiment analysis of
the search for “Colin Kaepernick” found that out of 500 tweets, 243 were
neutral, 64 were positive, and 193 were negative. 375 of the tweets were
subjective.
Graphs of the polarity and
subjectivity of the search results for “Colin Kaepernick” can be seen below:
The Aylien sentiment analysis of
the search for “#boycottNike” found that out of 77 tweets, 34 were negative, 13
were positive, and 30 were neutral. Most
of the tweets were considered subjective.
Graphs of the polarity and
subjectivity of the search results for “#boycottNike” can be seen below:
The Aylien sentiment analysis of
the search for “#TakeAKnee” found that out of 85 tweets, 46 were negative, 13
were positive, and 26 were neutral. A strong majority of the tweets were
subjective.
Graphs of the polarity and
subjectivity of the search results for “#TakeAKnee” can be seen below:
Twitter Analysis
We focused on the sentiment
analysis statistics in order to gain a high-level view of the tweets, and to
gauge engagement with each of the search terms. The word clouds were analyzed
with the goal of finding patterns in the content of the tweets and the terms
and concepts that were used within the tweet texts.
Sentiment Analysis
The Sentiment analysis of the
tweets reveal several important facts for understanding the social media
reaction to Nike’s use of Colin Kaepernick in the Just Do It campaign. First,
there is an overwhelming trend toward subjectivity rather than objectivity.
This indicates that the campaign not only generated news, but reactions to that
news. Second, the analysis indicated that the most common responses to “Nike”
were neutral, while the responses to “Kaepernick” and the two hashtags were
both negative. It is important to understand that the negative sentiment is
more indicative of negative word choice than specific attitudes toward Nike,
Colin Kaepernick, or the ad campaign, although the content of the tweets will
be discussed further. Third, the results
for searches for Nike and Colin Kaepernick were robust (each search hit its
respective limit), while the results for the hashtags were less than 100 tweets
each. This could indicate that the hashtags themselves have gone out of usage,
the organization of social media regarding Nike and Colin Kaepernick has become
less centralized, or that the discussion of Nike and Kaepernick has become less
political.
Word Cloud Analysis
Word Cloud for Nike:
The word clouds offer further
insight into the concepts associated with each search. The word cloud for Nike
revealed the types of terms that one might normally associate with a shoe
company – a mix of generic words such as “like,” “new,” “wear,” and “buy.”
Adidas, Nike’s main competitor, is also mentioned, along with “brand” and
“black”, which together make the name of a smaller rival, Brand Black. Neither
Colin Kaepernick nor any term associated with his protest is not found within
the Nike word cloud, which may indicate that he is not essential to the brand’s
identity or perception, although the word “just” is one of the largest entries.
This could either indicate Nike’s slogan, “just do it,” or the ad campaign
which features Colin Kaepernick.
This is a word cloud of tweets involving Colin Kaepernick:
The search for Colin Kaepernick
is a mix of both football and protest-related terms. Football terms like “NFL,”
“Redskins,” “Sanchez,” “play,” “team” and “job” are heavily featured, which
would indicate interest in the possibility of Colin Kaepernick joining the
Washington Redskins after their quarterback was injured. This is particularly
interesting because Kaepernick being offered an NFL job again would possibly
affect the main thesis of Nike’s ad campaign, “believe in something, even if it
costs you everything.” In addition to football terms, there are several terms
-- “kneel,” “protest,” “racist,” “Trump” and “police” -- that indicate
Kaepernick’s protest is central to his public perception and identity.
Furthermore, terms like -- “Ruben,” “Foster,” “Kareem,” and “Hunt” indicate a
debate over Kaepernick not being offered jobs while NFL teams employ players
accused of domestic violence like Kareem Hunt and Ruben Foster (It should be
noted that both Kareem Hunt and Ruben Foster were released from their
respective teams; Ruben Foster was claimed off waivers by the Washington
Redskins).
Word Cloud for "#boycottNike":
The searches for the hashtags
return far more openly political results. The search for “#boycottNike”
returned terms like "America,” “liberal,” “conservative,” and “anti,”
which could be combined with any number of the descriptive terms returned by
the search. These terms indicate that
users critical of Colin Kaepernick are using larger political concepts such as
America, liberalism or conservatism in their arguments, and Colin Kaepernick is
part of a larger political argument. This conclusion is further substantiated
by the return of the terms “Keurig,” “Starbucks,” and “Elle Magazine,” which,
while completely unrelated to Colin Kaepernick, are all companies that have
also been subject to political criticism.
Word Cloud for #TakeAKnee:
The search for “#TakeAKnee” also
yielded heavily political results. Terms like “Black Lives Matter,” “shot,”
“white,” “black,” “racist,” and “flag” have the same nods toward wide political
debates and broad ideologies. However, there are more specific terms found in
this search, particularly individuals. “OliverLNorth,” “TalbertSwan,”
“realdonaldtrump,” and “daddydaddymac,” are all mentioned. Oliver North is a
Reagan-era conservative political figure. Talbert Swan is a bishop often
affiliated with police reform. @realdonaldtrump is the Twitter handle of US
President Donald Trump, and @daddydaddymac is an active liberal Twitter
user. This suggests that while critics
of Colin Kaepernick speak in broader political terms, supporters of Colin Kaepernick
are engaging more with individual users of Twitter, including the President.
Additionally, it should be noted that the term “Kareem Hunt” also appears in
this word cloud. This could indicate that the users tweeting about Kareem Hunt
in the “Kaepernick” search would be sympathetic toward Colin Kaepernick.
In conclusion, the word cloud
data indicates that the tweets about the broader subjects of Nike and Colin
Kaepernick are less politically charged than the tweets using the hashtags
#TakeAKnee and #BoycottNike. Additionally, the number of tweets from each
search suggests that the number of users actively tweeting using the hashtags
is far fewer than the number of users engaging in the wider discussions of Nike
and Colin Kaepernick. The data also suggests that while Colin Kaepernick’s public
perception is intimately tied to his protest movement, and political discussion
more broadly, Nike’s public image has not become intimately tied to the same
protest even though Colin Kaepernick is the face of their “Just Do It” ad
campaign.
Stock Price
It appears that Nike saw some level of success
with the Kaepernick campaign. Nike saw a
small surge in stock price for about month following the September 3rd
announcement ($3.35 increase from $82.20 on September 3rd to $85.55
on September 21st. The
effects did not last however, and the stock price has ultimately fallen back
down to pre-campaign levels seen in June ($74.32). When comparing Nike with competitors Adidas
and Under Armour, it appears that Adidas has seen a similar change in stock
price. Under Armor doesn’t appear to
have been affected but has actually grown in the last month. There does seem to be evidence of a
significant event around September 3rd for both Nike and Adidas. The
following graph shows the change in price over the last five months. Both companies see an increase prior to
September, a small dip early in the month followed by an increase, and a slow
decline through mid-December.
More interesting is the overall trend, both Nike
and Adidas have seen an average loss in stock price of around twelve cents per
day since September 3rd
It would be interesting to see if the sentiment
surrounding Nike could be used as a predictor for both company’s stock
prices. Twitter is a superior for
sourcing sentiment, when compared to articles, due to ease of use and size of
corpora. Unfortunately, due to the
restrictions discussed earlier (limited free queryable history / expensive
one-time fees), we were unable to procure tweets from early September.
Conclusions
Within the scope of this project,
our group recognizes that we are assuming that public sentiment is the only
force driving stock prices of Nike and competitors. However, in reality there are a plethora of
other factors that can affect the stock prices and financial performance of a
company e.g. leadership changes, earnings reports and so on.
The sentiment from the news
articles dating to around September 3rd were strongly neutral. The stock prices near that date increased in
price, and our group agrees that there is a higher than average possibility
that a sentiment analysis on Tweets around that date would show a higher number
of positive tweets compared to neutral or negative. As we used the text mining tools more, we
started to suspect that the sentiment analysis algorithms were not as robust as
we initially thought, as the analyses frequently resulted in higher “neutral”
results than a strong positive or negative.
We feel that our search queries were polarizing enough to expect results
that would return fewer “neutral” results.
If we were able to recreate this
research we would do a few things differently.
First, having the funding to access historical tweets would allow us to
analyze the sentiment during the time of the ad release and compare that to the
sentiment of the articles and stock prices.
Access to that data would allow us to build a predictive model to allow
us to apply this research to future marketing campaigns. Finally, doing a more robust search around
other drivers of the stock market changes would allow us to see if it can be
narrowed down to simply the ad causing the movement in stock prices.
We learned that the Nike decision
to use Colin Kaepernick as the star in their ads continues to drive social
media discussion around larger social issues such as police brutality and the
Black Lives Matter movement, but that the stock prices of Nike and its
competitors seem relatively unaffected by the ongoing social media coverage.
Appendix A: ScreenShot of the RapidMiner Process for Assembling Tweet Corpora
Appendix B: R Code for Word Cloud
Visualizations
### Creates Word Cloud Visualizations from imported CSV files
### Used for MSA 8225 Final Project
### calls packages ###
require(twitteR)
require(RCurl)
require(tm)
require(wordcloud)
require(SnowballC)
### Reads in CSV List of Tweets
boycottnike_csv<- read.csv(file =
"boycottnike.csv", header = TRUE)
kap_csv <- read.csv(file = "kaepernick.csv", header
= TRUE)
nike_csv <- read.csv(file = "nike.csv", header =
TRUE)
nikekap_csv <- read.csv(file = "nikekap.csv",
header = TRUE)
takeaknee_csv <- read.csv(file = "takeaknee.csv",
header = TRUE)
### Create Corpus from Tweet Vector ###
boycottnike_corpus <- Corpus(VectorSource(boycottnike_csv$Text))
kap_corpus <- Corpus(VectorSource(kap_csv$Text))
nike_corpus <- Corpus(VectorSource(nike_csv$Text))
nikekap_corpus <- Corpus(VectorSource(nikekap_csv$Text))
takeaknee_corpus <- Corpus(VectorSource(takeaknee_csv$Text))
### Prepares Text###
boycottnike_clean <- tm_map(boycottnike_corpus,
content_transformer(tolower))
boycottnike_clean <- tm_map(boycottnike_clean,
content_transformer(removeNumbers))
boycottnike_clean <- tm_map(boycottnike_clean, removeWords,
stopwords("english"))
boycottnike_clean <- tm_map(boycottnike_clean,
stripWhitespace)
boycottnike_clean <- tm_map(boycottnike_clean, stemDocument)
boycottnike_clean <- tm_map(boycottnike_clean, removeWords,
"boycott")
boycottnike_clean <- tm_map(boycottnike_clean, removeWords,
"nike")
kap_clean <- tm_map(kap_corpus, content_transformer(tolower))
kap_clean <- tm_map(kap_clean,
content_transformer(removeNumbers))
kap_clean <- tm_map(kap_clean, removeWords,
stopwords("english"))
kap_clean <- tm_map(kap_clean, stripWhitespace)
kap_clean <- tm_map(kap_clean, stemDocument)
kap_clean <- tm_map(kap_clean, removeWords, "kap")
kap_clean <- tm_map(kap_clean, removeWords,
"kaepernick")
kap_clean <- tm_map(kap_clean, removeWords,
"colin")
nike_clean <- tm_map(nike_corpus, content_transformer(tolower))
nike_clean <- tm_map(nike_clean,
content_transformer(removeNumbers))
nike_clean <- tm_map(nike_clean, removeWords,
stopwords("english"))
nike_clean <- tm_map(nike_clean, stripWhitespace)
nike_clean <- tm_map(nike_clean, stemDocument)
nike_clean <- tm_map(nike_clean, removeWords,
"nike")
nikekap_clean <- tm_map(nikekap_corpus,
content_transformer(tolower))
nikekap_clean <- tm_map(nikekap_clean,
content_transformer(removeNumbers))
nikekap_clean <- tm_map(nikekap_clean, removeWords, stopwords("english"))
nikekap_clean <- tm_map(nikekap_clean, stripWhitespace)
nikekap_clean <- tm_map(nikekap_clean, stemDocument)
nikekap_clean <- tm_map(nikekap_clean, removeWords,
"kap")
nikekap_clean <- tm_map(nikekap_clean, removeWords,
"kaepernick")
nikekap_clean <- tm_map(nikekap_clean, removeWords,
"nike")
nikekap_clean <- tm_map(nikekap_clean, removeWords,
"colin")
takeaknee_clean <- tm_map(takeaknee_corpus,
content_transformer(tolower))
takeaknee_clean <- tm_map(takeaknee_clean, content_transformer(removeNumbers))
takeaknee_clean <- tm_map(takeaknee_clean, removeWords,
stopwords("english"))
takeaknee_clean <- tm_map(takeaknee_clean, stripWhitespace)
takeaknee_clean <- tm_map(takeaknee_clean, stemDocument)
takeaknee_clean <- tm_map(takeaknee_clean, removeWords,
"take")
takeaknee_clean <- tm_map(takeaknee_clean, removeWords,
"knee")
takeaknee_clean <- tm_map(takeaknee_clean, removeWords,
"takeakne")
### Create Wordclouds
wordcloud(boycottnike_clean, max.words = 100)
wordcloud(kap_clean, max.words = 100)
wordcloud(nike_clean, max.words = 100)
wordcloud(nikekap_clean, max.words = 100)
wordcloud(takeaknee_clean, max.words = 100)
Appendix C: Financial Data Scrape
Appendix D: Article Links
Appendix E: Article RapidMiner Processes
[1]
Removing the specific search terms from the text allows for a better word cloud
that is not dominated by the search terms themselves.
Comments
Post a Comment