인공신경망-금융시계열 모형을 이용한 KOSPI 200주가지수의 변동성 예측

As various funds and derivatives are developed using KOSPI 200 index, many investment banks and investment trust get interests on risk management of KOSPI 200 index. Accurate volatility estimation and prediction is the core in risk management in which various portfolio’s pricing, hedging, and option strategy is exercised by estimating volatilities. Up to now, many financial institutions give more values on risk management, but the methodologies of it are not established systematically. Many researchers have tried to forecast volatilities more accurately using financial time series models. Historically, many papers for volatility forecasting have concentrated on the comparison between forecasting models, but these researches focus on improving the predictive power of models by integrating ANN and financial time series models. In this paper, we first show that financial time series models, GARCH, outperforms existing ANN in forecasting the direction of volatility and that ANN model excels GARCH in reducing the precision error of the forecasted volatility by analyzing KOSPI 200 index time series data. Based on these results, this study propose the integrated model between ANN model and financial time series model to forecast volatilities of KOSPI 200 index time series. For selecting input variables for ANN model, new variable can be extracted by the financial time series models through analyzing the KOSPI 200 domain time series statistically. Then, these newly selected input variables can enhance the predictive power in the perspectives of precision error and direction accuracy through ANN learning process. The ANN-financial time series integrated model can enhance the predictive power by comparing the forecasted volatilities by single models in the framework of precision error(MAE) and direction accuracy(hit ratio). Especially, the integrated NN-EGARCH model is proposed as most predictive integrated model in forecasting volatilities in the perspectives of precision error and hit ratio simultaneously. In addition to prediction power, the integrated models can reduce time which it takes to adjust input variables by repetitive trial and error. Most of existing studies have adjusted the weight of raw volatilities by repetitive trial and error of learning process and found the optimal coefficient of input variables to produce the best results. This study finds the coefficients of input variables by financial time series process at once and extracts new variables that greatly influence the results through analyzing KOSPI 200 domain statistically without time consumption. At the same time, we consider market variables as those that adjust extracted variables to reflect true market behaviors. Therefore, integrated models can adjust variables realistically by ANN process and reduce time consumption by financial time series models.