Abstract
Gaining an edge in algorithmic trading has always been a lucrative endeavour. As such, any new data to help predict better stock prices are always welcome. Reddit data from the forum wallstreetbets is one such avenue. The possibility of better stock price prediction, a trading signal, could help create better trade execution. Over the years, time series prediction has gotten better with more sophisticated models. However, there is a lack of good comparison between the various models for stock price data. In this thesis, we will compare the various time-series prediction models to identify the best model from a list of five time-series prediction models in predicting the closing adjusted stock price. They are compared using 2012-2021 daily stock prices and Reddit data. These models include Arima, N-beats, Long-short Term memory, DeepAR and Transformational Fusion Transformer(TFT). We will also select the best model using covariates such as open, high, low, volume and moving average data. This model will then predict stock prices using our newly created wallstreetbets text dataset. This stock price dataset is created by scraping Reddit data and matching each post with stock tickers whenever they are mentioned in the post or the submission title. Sentiment and embedding data are extracted from this text data to incorporate into our models. The N-beats model was the best model in a setting where the percentage change was used as input for our models. The TFT model was found to be the best among models that can make use of covariate data. Hence, we expanded the TFT models to incorporate them with embeddings and sentiment data. The results were similar for two metrics, MAE and RMSE for the models. Though the TFT model with sentiments was the best compared to the MAPE.We decided to compare TFT with and without sentiments using the raw data instead of percentages and found that the model with sentiments fared noticeably better than the one without. The predictions from the model with sentiment Reddit data and non-Reddit data as input are also shown to be statistically different from one another. Hence it can be concluded that the Reddit data did help the model create a better model in predicting stock price data. The framework discussed in this paper to test the usefulness of Reddit data can be applied to other new datasets. This, in turn, can create a marketplace for various datasets. The beneficiaries of such a marketplace will be both trading firms and alternative data suppliers, as it will create a more open market for such data sources.