Earticle

This study aims to forecast Portfolio VaR using Noise-reduced Correlation Matrix based on Denoising Autoencoders, which helps anticipate extreme losses in the future. The dataset comprises the US stocks with daily adjusted closing prices. The portfolio weights for each asset were calculated by applying the Risk Parity method to the entire asset group using the Noise-reduced Correlation Matrix. Forecasting VaR through Delta-Normal, Historical, GARCH, and Denoising Autoencoders methods, and subsequently conducted Backtest VaR. Empirical analysis revealed that DAE-GARCH-VaR (Student’s T Distribution) exhibited a significant noise reduction effect on the correlation matrix and the superior performance.

This article investigated the relevance of the skewed Student-t distribution innovation in analyzing volatility stylized facts, namely, volatility clustering, volatility asymmetry, and volatility persistence, in three individual Korean shares. For this purpose, we assessed the performance of RiskMetrics and two long memory Value-at-Risk (VaR) models (FIGARCH and FIAPARCH) with the normal, Student-t, and skewed Student-t distribution innovations. From the results of the empirical VaR analysis, the skewed Student-t distribution innovation provided more accurate VaR calculations, in capturing stylized facts in the volatility of three sample returns. Thus, the correct assumption of return distribution might improve the estimated performance of VaR models in the Korean stock market.

본 연구는 실무에서 많이 사용되는 리스크 측정 지표인 VaR(Value-at-Risk)를 이용해 화재보험 사고 손해액 분포를 분석하고, 지수적 가중치 이동평균(EWMA), 정규(Normal), 경험적 시뮬레이션(HS), 그리고 GARCH-EVT 모델 중 어떤 모형이 가장 적합한지를 검증하였다. 모형별로 VaR를 도출하고 사후검증(Back- testing)을 통해 그 결과를 비교한다. 이를 위해 일별 자료를 월별로 합산한 자료를 분석함과 동시에 원자료를 사용한 분석을 시도한다. 분석 결과에 따르면 일별 합산 자료의 경우 ARCH 검정에서 변동성의 군집화를 발견하지 못했으며 극단치이론(EVT)과 경험적 시뮬레이션(HS) 모형에 의한 VaR 추정치가 상대적으로 좋은 모형으로 판명되었다. 반면 원자료의 경우 ARCH 검정에서 변동성의 군집화 현상이 발견되어 GARCH모형으로 필터링을 실시한 이후 VaR를 추정한 결과 GARCH-EVT 모형에 의한 VaR 추정치가 가장 좋은 모형인 것으로 판명되었다.

This study applies VaR, which is a representative risk measurement indicator, to the given data and verifies which model is most suitable with insurance damage amount. We derive VaR for each of the models employed, perform backtesting, and compare the results. To do this, we sum up the given data on the daily basis and then analyze it and at the same time try to analyze it by the raw data. According to the results of the analysis, we did not find the clustering of volatility in aggregated daily data and it was proved that VaR estimates by extreme value theory (EVT) and historical simulation (HS) models are relatively preferred to the other models. On the other hand, in the case of the original data, the clustering of the volatility is found and the VaR is estimated after filtering by the GARCH model. As a result of the backtesting, the VaR estimate by the GARCH-EVT model proved to be the best model.

Value at Risk(VaR) has been proven useful in finance literature as a tool of risk management(cf. Jorion(2001)). This article is concerned with introducing VaR to various Korean financial time series. Five daily data sets with sample period ranging from 2000 and 2004 such as KOSPI, KOSPI 200, KOSDAQ, KOSDAQ 50 and won-dollar exchange rate are analyzed using GARCH modeling and in turn VaR is obtained for each data.