Posts in social media often cause changes in the stock market that may lead to unpredictable losses in one's investment portfolio. In this project we analyse which words are more powerful and what influence they may have. Will they bring losses or profits to our portfolio? 💰
As an example, we take tweets posted by Elon Musk and investigate how they may impact Tesla's stock prices.
Using Twitter API we collected all posts made by Elon Musk in 2021 along with counts of likes, replies, and retweets. We merged it with another dataset of Elon's tweets for 2011-2020 found on Kaggle. All Elon's reposts and replies were excluded from analysis.
Yahoo_fin library was used to get historical Tesla's stock price data for the same period.
Classification and LDA models were built to analyse datasets.
There are five members in our team. The role of each team member will remain the same each week to ensure that we had an expert for each topic of the project. A Slack channel was created to support communication amongst the team, and will be used to assign tickets, provide updates, and discuss any issues. The team will additionally have meetings twice a week to go over project progress and next steps.
- Getting and storing data
- Twitter data
- Pull data from Twitter API and clean it
- Pull tweets data from Kaggle for 2011-2020 and clean it
- Merge 2 datasets and export to tweets_data_2010_2020_ungrouped.csv for Tableau dashboard
- Preprocess data for machine learning (ML)
- Upload cleaned data to PostgreSQL database and export to tweets_data_2010_2020.csv
- Stocks data
- Pull data from Yahoo fin library
- Clean data
- Upload cleaned data to PostgreSQL database and export to tesla_stocks.csv
- Twitter data
- SQL Database
- Create ERD schema
- Create tables to store tweets and stock data
- Merge 2 tables and export to twitter_vs_stocks.csv
- Machine Learning Model
- Latent Dirichlet Allocation (LDA) Modelling
- Preprocess data
- Build LDA model
- Classification
- Preprocess data
- Build classification model
- Latent Dirichlet Allocation (LDA) Modelling
- Tableau dashboard
- Create presentation in Google Slides
437 posts were pulled from Twitter API directly for the period January 1 - July 18, 2021. We extended the dataset by adding tweets from 2011 till 2020 found on Kaggle. All replies were excluded, so only 4,629 tweets were included for analysis.
During analysis we revealed that Elon Musk posted significantly more tweets since 2018.
"Tesla" was the most commonly tweeted word, followed by "model".
Tesla stock prices soared since September 2019 and reached a peak in January 2021.
The most Tesla shares were traded on February 4th 2020.
The greatest increase in closing price was seen on January 8th 2021, while the greatest decrease was seen on September 8th 2020. Elon Musk posted a few tweets these days.
A correlation between the number of likes and stock trading volume can be seen at some periods of time.
We used SQL query language to upload the datasets to a Postgres database. Then, using the INNER JOIN, we merged the two datasets to create a third dataset called twitter_vs_stocks. The twitter_vs_stocks combines the data from both datasets using the date as the ID. This table displays the tokenized_text versus the close amount for each date. In addition, the change column shows, for each date, whether the stock price has increased or decreased in comparison with the previous day’s amount after Elon Musk has posted the tweet.
For topic modelling in this project, we used Latent Dirichlet Allocation (LDA). The modelling procedure is divided into 4 stages:
- clean the data
- create a bag of words
- identify the number of subjects
- run the LDA algorithm
As LDA is an unsupervised ML model, the model uses a whole dataset as input; It shouldn't be split into training and testing subsets.
After obtaining data we removed NaN values, duplicates, and any unnecessary columns, as well as formatted data types. The pandas library was used to clean the data and join the twitter datasets.
To preprocess the data for LDA, we applied a few text cleaning techniques:
- Ignoring case; We made all words lowercase.
- Ignoring punctuation; We tokenized text.
- Ignoring words that don’t contain much information; We used the
nltklibrary to get English stop words and extended the list based on available data. - Reducing words to their stem; We used the
PorterStemmeralgorithm.
Following the cleaning process, the bag of words stage was used to extract the most popular words in tweets with their counts.
Elon Musk used 8,197 words in his tweets; an enormous number of features to analyse. However, Latent Dirichlet Allocation (LDA) is a probabilistic transformation of bag-of-words counts to a lower-dimensional topic space. It classifies the text data on a topic-by-topic basis, which means disregarding its original position in the text while maintaining its frequency. In terms of coherence score, the degree of semantic similarity between a topic's high scoring words is used to quantify its coherence. The method chooses the most frequently occurring words in each topic. The coherence score for a specific topic is then calculated by computing and aggregating all pairwise scores (UMass) for each of the words.
The average coherence score per topic for a variety of models trained with varying numbers of topics is shown below. The number of topics for which the average score reaches a peak point is the perfect balance we seek.
As a result, our best guess for the number of topics is around two. We should keep in mind that a corpus of very short documents (Elon's tweets in our case) is typically more difficult to apply to coherent models than a corpus of longer documents.
Although there is "Tesla" in topic 0, the remaining words are related to SpaceX and its components have a strong relationship with one another. Topic 1 illustrates that Tesla and its components are clustered together. So, we can admit that Elon Musk posts mainly about 2 topics: Tesla or SpaceX.
The LDA model maintains a significant number of features, however it doesn't take into account the position of a word in the sentence. It's hard to define the right number of topics as the coherence score will vary from tiny changes in the dataset. Words are often mixed between a few topics and this makes the LDA model less trusted.
For this supervised learning model, we used the preprocessed data from the SQL database found in the tables tesla_stocks and tweets_data_2010_2020.csv. We used a SQL query to assemble a dataframe called tweets_price. In this dataframe, we have the tweet, tokens of the tweet, prev_day_close, next_day_close columns, where the data can be easily viewed. One last column called close_price_diff was created. It is the calculation between the day before and the day of closing stock price for the date the tweet was posted. This column takes into consideration weekends for which we do not have a closing stock price value in the dataset, and the randomness of Elon Musk's posting of tweets. We also made sure to remove from the training models any tweets that had none or less than one tokens inside them, as it ensures for more successful training of the machine learning algorithm. This table can be referenced below:
For the Classification model, we decided to use LogisticRegression to predict the probability that a tweet of Elon Musk will have an impact on Tesla stock or the probability this event occurs. The advantage of the LogisticRegression model is that it can provide us with statistical information on the tweets' significance in relation to the change of stock price. However, a limitation to this model is the assumption that there is LogisticRegression between the data and it is prone to overfitting.
We created a Pipeline that used the tokenized tweets via CountVectorizer, then TfidfTransformer to take into consideration the frequency of each token in the tweet compared to its frequency in the corpus/dataset, and LogisticRegression to learn the link between the tweets and the change in stock price. The data was split into training and testing sets to minimize discrepancies and ensure a good performance when it encounters unseen data. This was done by using sklearn's train_test_split method with a random_state of 1 for reproducible results.
The overall accuracy of the model is 57% which shows that the model is not able to accurately predict a link between a tweet and the change in Tesla stock price. Considering the number of variables that influence the change in stock price, and that not all of Elon Musk's tweets are directly influencing investors, this is an expected result.
However, it is to be noted that when observing the results_test data frame, we can see that the model is able to correctly identify the most positive tweets. This shows the model is able to make good Natural Language Processing (NLP) deduction calls on tokens and is able to learn from the data.
A dashboard was created using Tableau to showcase the exporatory analysis of the twitter and stock datasets. An interactive component was created whereby users can filter the graphs by year to show the changes in Tweets over time. An additional interactive component will be created with findings from the machine learning model whereby users can filter the change in stock price by a key word.
A presentation had additionally been drafted in Google Slides.
Twitter, a social media platform that enables its users to broadcast their thoughts in “tweets” ranging from 140 to 280 characters (2021) with around 186 million users as of 2020, has become a widely used platform for many prominent figures to intimately engage with the world.
Elon Musk, the founder and CEO of Tesla, is an active Twitter user himself (@elonmusk) and has consistently increased the number of tweets he sends out each year, increasing since 2015.
The purpose of our project was to try to find a correlation between the change of stock price and the influence of Elon Musk’s tweets, and test whether or not they influence the stock market. This was tested by gathering data on Elon’s tweets from 2011 to 2020 as well as Tesla’s stock data for the same time frame. This data was used to train a machine learning model to test for a correlation between Musk’s tweets, and any positive or negative change in stock. Also, we tested to see if a machine learning model was able to understand tweets and learn if they were positive based on the fluctuation in stock value.
The results of the machine learning model demonstrate the weak correlation between Elon Musk’s tweets and the fluctuation of stock price, as shown by the 57% accuracy result. As expected from the stock market, a system that is heavily influenced by an array of factors ranging from demand and supply, interest rates, dividends, investors and many more. The results further reinforce that this relationship is weak, as they are not isolated events. For Elon Musk’s tweets, we ran an LDA model that showed they are heavily focused on his interests on subjects regarding SpaceX, Tesla, and space exploration. While the world becomes more heavily influenced by social media, it’s also getting overloaded with data, as we saw from most of Elon Musk’s tweets, which were found to have no significant impact on stock prices.
Based on our findings, it is recommended to use the Logistic Regression ML model to assess the influence of likes and retweets on changes in stocks, as they appeared to show some correlation with trading volumne during our exploratory analysis. Furthermore, as articles have indicated Elon Musk's influence on cryptocurrencies, we could use ML to predict whether Musk's tweets have an influence on cryptocurrencies instead of Tesla stock prices. Finally, this model could be applied to other individuals with greater public influence to determine whether their tweets are more strongly correlated with stocks.
If this project were to be completed again, the team would have set different deadlines for each task of the project versus following the deadlines provided for a segment deliverable. Following this approach would have ensured certain portions of the project, such as data preprocessing, were completed prior to selecting the database and creating the ML model. During the current project, the data preprocessing was completed in tandem with the database and ML model being built, resulting in code having to be updated more frequently. By setting deadlines for each task, work could be more efficient.
Furthermore, the team should have spent more time investigating datasets prior to preprocessing the data, creating the database, and training the ML model. Half-way through the project the dataset changed and the database had to be re-done as a result. Likewise, the team should have used a database more compatible with text data, such as MongoDB, versus using SQL Alchmey and postgres which the team was more comfortable with.
Finally, the team would have liked to have more variety in their tasks so that each member had equal opportunity to work with the code, as during the project some team members had more code-heavy roles than others. The original intent was to have individuals who were experts in their portion of the task, however, this prevented some team members from fully comprehending components of the project. Having each team member work in all aspects of the project would have additionally allowed for greater understanding within the team.