Kospi Index Prediction using a Financial-specific Sentiment Analysis and Deep Learning-based Time Series Prediction Model

This paper presents a methodology for predicting the KOSPI index using a news data-based sentiment analysis model and a deep learning-based time series prediction model. The closing price of the KOSPI index was used as a target variable, and macroeconomic indicators such as the gold price and market sentiment indicators such as sentiment scores were used as independent variables. We collected and preprocessed the KOSPI-related news data and used them in calculating the sentiment score by using the title or the summarized article. Subsequently, the KLUE-BERT model-based sentiment score by date and the KoFinBERT model-based sentiment score by date were extracted. LSTM, GRU, CNN-LSTM, and CNN-GRU were used as time series prediction models. As a result of conducting an experiment by combination of variables and models, the best performance was achieved when KLUE-BERT is applied on the summarized article and the CNN-GRU model were used.