Data-driven Research On Volatility Prediction And Application For Multiple Financial Assets

Accurate measurement of volatility is of great significance to market participants in asset pricing,investment portfolios,hedging,and risk management since it measures the magnitude of financial risk.At present,the development and innovation of artificial intelligence and big data have brought new challenges to forecasting tasks in traditional financial research.The complex changes of emergencies,such as the COVID-19 pandemic,affect the global economic pattern significantly,which increases the instability of the financial market day by day.In such an environment,research on how to effectively characterize investor behaviour in a complex financial environment and connect to financial risk can facilitate one to understand what role investors play in the operating mechanism of financial markets.Besides,research on effectively mining the value information from financial big data can improve financial asset volatility prediction accuracy.Moreover,research on effectively applying the relationship between the external environment and asset volatility to enhance risk management levels helps maintain and promote financial stability.Considering multiple financial assets,and using popular methods,including text sentiment analysis,machine learning data-driven forecasting techniques,hedging models,and risk contagion models,this paper investigates the problem of volatility forecasting and its applications.More specifically,we focus on the following three aspects,including(i)the characterization of investor behaviour and its impact on market volatility forecasting,(ii)the innovation of volatility forecasting method with considering a large set of predictors,and(iii)the application of research results in hedging and risk contagion.This paper contributes to the literature from the following aspects.Firstly,we propose three types of investor behaviour measures using data-driven technologies and study their impacts on financial asset volatilities.(i)using the partial least squares method,we propose a new investor sentiment index that conforms to the characteristics of China’s market.It is found that the constructed index reflects the Chinese stock market better and has more powerful and more robust predictive power on stock volatility relative to extant sentiment indices.(ii)We propose volatility technical indicators for the energy futures market based on the features of volatility.Using the improved conditional sure independent screening(CSIS)method,we investigate the predictability of technical indicators to high-frequency volatility of crude oil and natural gas futures.The results show that the constructed technical indicators have significant in-sample and out-of-sample predictive power.Besides,compared with classical data-driven methods,the improved CSIS method has more outstanding and robust prediction performance since it can effectively extract the prediction information from predictors.(iii)We propose a relatively fixed global composite uncertainty index via capturing uncertainty from market,investor,and policy levels by using the data-driven-based scaled-PCA method.This index predicts stock volatilities well in the global 23 markets.The in-and out-ofsample results show it has better prediction performance than the classical dimensionality reduction method.Furthermore,we reveal the reasons for its superiority.Secondly,we investigate the predictive power of text sentiment,which is constructed using the data mining and text analysis methods,on the Chinese stock market volatility.The text is downloaded from the articles about stock analyses in the investor forum,namely Xueqiu.com.We construct text sentiment indices with investors’ memory based on the sentiment dictionary method,where the dictionary that we propose combines platform characteristics,financial characteristics,and users’ habits.In the empirical analysis,we show that the text sentiment affects Chinese stock market volatility through the non-linear channel,which contrasts with the evidence from developed markets.Furthermore,based on the long short-term memory network model,we find that the text sentiment significantly predicts volatility.In addition,the results show that the pessimistic text sentiment has more substantial predictive power,and show that the text sentiment only has a significant effect on short-term volatility during the financial crisis.Thirdly,we investigate the new method for volatility prediction when considering a large set of predictors.Based on the background of financial big data,we use 86 potential volatility predictors from inside and outside the financial market to simulate the complex economic system.A new forecast combination method is proposed using the R-square and t-statistic of explanatory variables as weight.The R-square and t-statistic capture the economic and statistical significance,respectively.In the empirical analysis,we show the proposed method outperforms extant methods in predicting the target variable and is more suitable for forecasting problems with a large set of predictors.In addition,when there are significant differences in the predictive power of variables,the advantage of the proposed method is more pronounced.What’s more,based on the proposed model,we find that the prediction accuracy decreases with the number of predictors when there are strong predictors in the variables.In comparison,the prediction accuracy increases firstly and then decreases as the number of predictors increases when the predictive powers of the variable are similar.Finally,as an application,we study how to improve risk management capabilities by applying the relationship between external factors and volatility prediction to hedging and risk contagion across international stock markets.Based on the conclusion that external factors can improve the prediction accuracy of volatility,this study further explores the impact of enhanced accuracy on hedging efficiency.We find that the hedging efficiency based on the highfrequency data volatility model is significantly higher than the traditional low-frequency one.Besides,considering external factors in the high-frequency volatility model further improves the hedging efficiency.To this end,the findings have significant application values in improving the risk hedging ability and economically benefiting investors.For another,this paper discusses the channel of risk contagion across global stock markets and further analyses the sources and mechanisms.It is empirically found that volatility risk is mainly contagious from developed countries in Europe and the United States to Asia-Pacific.Moreover,we show that the implied volatility index and market liquidity measures from risk spillover markets have a significant and robust explanatory power on risk contagion,which is more prevalent during financial crises.Thus,our results provide meaningful decision-making reference for market participants to carry out risk pre-warning and control.