ScholarWorks@중앙대학교

초록

We investigate whether technical indicators extracted from one stock can predict the subsequent movement of other stocks in the KOSPI 200 and how static versus dynamic feature selection influences forecasting performance. We design two complementary frameworks: the first evaluates cross-stock predictability by applying static selection within a rolling window and then a dynamic ensemble that updates feature relevance over time; the second integrates dynamic selection into deep learning architectures, including multi-target models that explicitly exploit cross-stock relationships. Across extensive experiments, dynamic features consistently outperform static ones in both accuracy and stability, reducing dimensionality, curbing overfitting, and improving interpretability. We further find that multi-target approaches deliver stronger predictive performance than single-target baselines, underscoring the importance of modeling interdependence among stocks. Our evidence provides an empirical blueprint for capturing cross-stock linkages and demonstrates that indicators derived from one security can yield robust, generalizable signals for others. The current analysis is limited to technical indicators from the Korean equity market; future work will extend the framework with sentiment and fundamental variables and assess portability across asset classes and international markets. By combining dynamic feature selection with multi-target prediction, this study advances financial econometrics with AI and offers practical guidance for portfolio construction and factor investing.

키워드