ScholarWorks@성균관대학교

초록

Forecasting dynamic correlations among diverse financial assets, while effectively incorporating changing economic conditions, is crucial for risk management and portfolio allocation. Traditional methods often fall short due to reliance on static historical measures and limited integration of broader economic signals. This paper introduces a deep learning framework using a two-tier Long Short-Term Memory (LSTM) network to overcome these limitations. The architecture models macroeconomic indicator dynamics with dedicated LSTMs and captures their time-varying joint dependencies via an LSTM-based dynamic copula network. Trained via maximum likelihood, the model adaptively learns complex, economically-informed correlation structures. An extensive empirical study shows our LSTM-based model significantly outperforms a 1-year rolling historical correlation benchmark in out-of-sample predictive log-likelihood. The analysis identifies the volatility index (VIX), GDP growth, and inflation-related indicators as key predictive drivers, with ESG factors also adding substantial value, especially during relatively stable market periods. These findings highlight deep learning's potential to integrate economic information for more accurate and adaptive dynamic correlation forecasting.

키워드