Tesla Stock Price Prediction

Tesla’s stock price is predicted over some months using an LSTM (Long Short-Term Memory) model. Tweets about Tesla are used to improve prediction accuracy.

First, the stock price is predicted over some months using an LSTM Multivariate Time-Series Prediction model. Then, tweets about Tesla are cleaned and their daily average sentiment scores are computed using TextBlob. Lastly, the daily average sentiment scores are added as a feature in the LSTM model and used for prediction.

Disclaimer: The LSTM model cannot be used to predict stock prices in real life because the stock market is highly unpredictable. In this project, the validation phase is used to test the model's performance. The purpose of the project is to implement Multivariate Time-Series Prediction using LSTM.

Problem Description

The task is to investigate the impact of tweets about Tesla on its stock price.

Data

The csvs zip file and model states can be accessed from the data folder.

Solution Approach

Prediction Without Sentiment Scores

Adj close price is used as the response and the following are used as the features:

• High price
• Low price
• Open price
• Close price
• Volume

The features are normalized because an LSTM model is sensitive to the scale of data, and then converted to tensors.

LSTM model parameters:

• input_size=5
• batch_first=True
• num_classes=1
• optimizer=Adam
• loss_function=MSELoss()

LSTM model hyperparameters after tuning with Ray Tune using Grid Search Algorithm:

• hidden_size=3
• num_layers=1
• learning_rate=0.001
• num_epochs=8000

GPU is leveraged.

MSE (Mean Squared Error) results:

• Training and Validation: 108.98853599149807
• Training: 0.37573209398539453
• Validation: 429.7357225013402

Cleaning Tweets

The tweets are cleaned and pre-processed in the following ways:

• Adding whitespaces to the ends to join them
• Lowercasing
• Removing http links
• Removing special characters and numbers
• Removing stop words
• Removing unnecessary words and characters

They are then lemmatized, and a frequency analysis is conducted on the words.

Sentiment Analysis

Sentiment scores for the tweets are calculated using TextBlob. The polarity range is [-1.0, 1.0], with -1.0 as the most negative polarity, 1.0 as the most positive polarity and 0.0 as the neutral polarity. Then, a frequency analysis is conducted on the sentiment scores. Lastly, the daily average sentiment scores are computed.

Prediction With Sentiment Scores

Finally, the daily average sentiment scores are added as a feature to our LSTM model.

LSTM model parameters:

• input_size=6
• num_classes=1
• optimizer=Adam
• loss_function=MSELoss()

LSTM model hyperparameters after tuning with Ray Tune using Grid Search Algorithm:

• hidden_size=5
• num_layers=1
• learning_rate=0.002
• num_epochs=8000

GPU is leveraged.

MSE (Mean Squared Error) results:

• Training and Validation: 35.028627722561374
• Training: 0.11838243676105603
• Validation: 138.12294583219045

Conclusion

From the MSE results of Prediction Without and With Sentiment Scores, it is clear that adding the daily average sentiment scores of tweets as a feature to the LSTM model improves its prediction accuracy. This means that tweets about Tesla have a level of impact on its stock price.