SENTIMENT ANALYSIS ON STOCK MARKET DISCUSSION AND STOCK PRICE PREDICTION USING FINANCIAL WORD2VEC

• Main Author: Zhafransyah, Fabian
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/76662
• Institution: Institut Teknologi Bandung

Abstract—Online stock discussion forums is a place for investors to discuss various events that occurred within the stock market community. The general sentiment of online stock discussion might reflect the changes of a company’s stock price. This paper reports the development of a financial Word2Vec word embedding, integrated with an LSTM-based sentiment analysis model, and an LSTM-based stock price prediction model utilizing the sentiment of stock discussion forums generated by the previously mentioned sentiment analysis model. Financial Word2Vec is built by integrating the Loughran-McDonald financial dictionary with Word2Vec, which was then trained on financial news data. This financial Word2Vec is then integrated to an LSTM-based seniment analysis model to be trained on online stock discussion texts in order to extract more sentiments from more stock discussion forum data. The sentiment analysis model then generates sentiments for the LSTM-based stock price prediction model to train and evaluate from. Three separate experiments with similar and connecting conclusions were carried out to each model. Experiment on the Word2Vec word embeddings is designed to compare the performance of our financial Word2Vec with the already established Indonesian Word2Vec. This experiment demonstrates the ability of our financial Word2Vec to understand words with financial context, but in general still falls short to the Indonesian Word2Vec. Experiment on sentiment analysis models aims to find out how the different Word2Vec word embeddings fare against each other when used to predict sentiment from stock forum discussion texts. The sentiment analysis model shows lower accuracy when integrated with our financial Word2Vec than the Indonesian Word2Vec. The last experiment is conducted on stock prediction models to find out the best sentiment to be used for predicting stock prices. This experiment demonstrates the superiority of sentiment extracted by sentiment analysis model with Indonesian Word2Vec, compared to model without sentiment and model using sentiment from financial sentiment analysis.