ESG ASPECT-BASED SENTIMENT ANALYSIS ON NEWS FOR STOCK PRICE PREDICTION USING TIME SERIES METHOD

• Main Author: Ariq Prathama, Ubaidillah
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/82360
• Institution: Institut Teknologi Bandung

Over the last decade, Environmental, Social, and Governance (ESG) aspects have become a focal point in investment analysis. ESG highlights the importance of sustainable and ethical practices in predicting market performance. This Final Project aims to integrate ESG sentiment analysis into a stock price prediction model. This could provide new insights for investors who wish to incorporate sustainability into their investment decisions. To achieve this goal, a multi-model approach is used in sentiment analysis and stock price prediction. First, a model based on Bidirectional Encoder Representations from Transformers (BERT) is used to classify ESG-related aspects from news. Next, a Large Language Model (LLM) is applied for sentiment classification based on ESG aspects of the text. Lastly, Bidirectional Long Short-Term Memory (Bi- LSTM) is applied to predict stock prices based on the time sequence of generated ESG sentiments. This pipeline has successfully outperformed stock price predictions using only historical stock price data. The pipeline uses the best model based on experimental results. ESG classification using post-trained IndoBERT achieved an f1-score of 0.82. Sentiment classification using Merak achieved an f1-score of 0.92. Stock price prediction using Bi-LSTM achieved an RMSE difference of 0.001 compared to without sentiment. This indicates that ESG news sentiment impacts stock prices, and the system created successfully leverages this.