자연어 처리 모델을 활용한 주가지수 예측 연구

Abstract

The field of natural language processing (NLP) has made remarkable progress with the development of Transformer-based deep learning models. This study examines the applicability of NLP models, including RNN, LSTM, and Transformer, for predicting financial time series using KOSPI200 and S&P500 data. The findings indicate that greater model complexity does not necessarily lead to better predictive performance. Complex models, such as Transformers, often overfit noise, resulting in unstable training and higher prediction errors. Furthermore, financial time series possess unique characteristics, such as continuous values, high volatility, and nonlinear dependencies, which distinguish them from natural language data. This study underscores the importance of selecting models that are specifically tailored to the unique attributes of financial data.