시계열 데이터의 교차 검증을 활용한 융합적 모델 기반의 시계열 특징에 의한 코스피지수 예측

In this study, an experiment was conducted using the closing price of the KOSPI index to improve prediction performance based on time series features. Denoising and prediction were performanced based on the time series features using the time series decomposition method and seven types of wavelets. And based on time-series cross-validation, the performance and evaluation of prediction were measured by averaging the measurements by interval. In order to improve the performance of prediction by convergence and linear combination of two different models, simple average and weight estimation methods were compared. Based on this, the ridge regression weight estimation method was selected, and the comparison target model and prediction error were measured, and then the independent sample t-test was conducted. Looking at the results of the empirical analysis, the individual model and the linear combination model produced statistically significant results with respect to the comparison target model. In this study, the superiority of time series element decomposition based on time series features, seven types of discrete wavelet transformation processes, cross-validation, and method of linear combination of different models was confirmed. In particular, the ARIMA individual model showed better prediction performance than the bidirectional LSTM model.