When Can News Sentiment Improve KOSPI's Short-Term Direction Forecast?

뉴스감성의 KOSPI 단기방향예측 성과개선에 관한 연구

초록

Short-term market direction forecasting often yields modest and unstable out-of-sample improvements. With focus on the KOSPI index, this study compares the predictive performance of a Two-Branch Cross-Attention DNN and Histogram-based Gradient Boosting (HGB) under a leakage-controlled forecasting design, and examines whether the predictive contribution of daily news sentiment becomes more pronounced under high-intensity sentiment conditions. Daily sentiment features are constructed from Korean financial news and aligned to ensure that only information available at time t-1 is used for prediction. The comparison is conducted under a consistent feature set and evaluation protocol. Performance is assessed using an out-of-sample ROC-AUC with bootstrap 95% confidence intervals. The results indicate that the fusion architecture's advantage is most pronounced at the one-day-ahead horizon, and that gains diminish across longer horizons. To diagnose this conditional pattern, rolling Z-score gating was applied to sentiment intensity. Findings from event-subsample AUC, coverage, and lift analysis suggest that sentiment-related predictive gains are more concentrated during high-intensity periods than in the full-sample average.

키워드

News Sentiment; Two-Branch Cross-Attention DNN; Histogram-based Gradient Boosting; Out-of-sample ROC-AUC; Rolling Z-score gating
제목
When Can News Sentiment Improve KOSPI's Short-Term Direction Forecast?
제목 (타언어)
뉴스감성의 KOSPI 단기방향예측 성과개선에 관한 연구
저자
석방민; 유시용
DOI
10.5762/KAIS.2026.27.8.508
발행일
2026-08
유형
Y
저널명
한국산학기술학회논문지
권
27
호
8
페이지
508 ~ 514