본문 바로가기
  • Home

Big Data Text Mining-Based Stock Prediction System

  • Journal of The Korea Society of Computer and Information
  • Abbr : JKSCI
  • 2026, 31(1), pp.51~58
  • Publisher : The Korean Society Of Computer And Information
  • Research Area : Engineering > Computer Science
  • Received : November 17, 2025
  • Accepted : December 29, 2025
  • Published : January 30, 2026

Tai-Sung Hur 1 Ariunjargal Amintsog 1

1인하공업전문대학

Accredited

ABSTRACT

This study presents an end-to-end system that transforms multilingual financial texts (Korean news and English forums) into quantifiable trading signals using domain-specific language models FinBERT and KR-FinBERT. The key innovation is an adaptive volatility threshold mechanism where trading signals are triggered when the 5-day EWMA of sentiment indices crosses dynamic bands defined by ±1 standard deviation of a 60-day rolling window. This approach automatically adjusts sensitivity to market volatility, enhancing strategy robustness across diverse conditions. Through rigorous backtesting on KOSPI stocks and Apple (January 2020 - June 2025) with realistic transaction costs (0.1% one-way) and execution delays, the active strategy consistently outperforms passive buy-and-hold benchmarks. Notably, Apple and Kia achieved Sharpe ratios of 1.42 and 1.13, respectively, with significantly lower maximum drawdowns. Event studies and parameter sensitivity analysis confirm the statistical and economic significance of this NLP-based algorithmic trading approach.

Citation status

* References for papers published after 2024 are currently being built.