VaR模型在金融风险管理中具有重要地位。本文引入一种基于高频日内收益和隔夜收益的条件分布抽样合成日收益分布的新框架以预测VaR。参数的估计和预测基于最近流行的广义自回归得分模型（GAS模型），GAS模型的参数主要依靠分布密度的得分函数来驱动。基于2014-2020年的上证指数和深证成指的30分钟日内收益和隔夜收益进行VaR滚动预测，采用MCS检验和常见的回测方法如Kupiec检验、DQ检验来综合比较预测效果。

实证研究表明，隔夜收益表现出了比日内收益更显著的尖峰性和有偏性，隔夜收益的平方与日内收益平方和比值的平方根最高可达到7，说明了隔夜收益对已实现波动的影响不可忽视。在正态分布、t分布等7种常见分布中，基于偏t分布的GAS模型对于隔夜收益具有最好的拟合效果。GAS模型基本通过了所有的VaR回测检验，而已实现GARCH模型在上证指数VaR回测检验中被拒绝。根据MCS检验的结果，隔夜收益的确能提升VaR预测效果，综合表现最好的是考虑隔夜收益的GAS模型，比加入隔夜信息的已实现GARCH模型更加稳健。

VaR forecasting models have an important position in financial risk management. A new framework of forecasting dynamic Value-at-Risk (VaR) is introduced based on Bootstrap sampling from the conditional distribution of high-frequency intraday returns and overnight returns to synthesise daily return distribution. The estimation and prediction of parameters are based on the recently popular Generalized Autoregressive Score model (GAS model), and the parameters of GAS model are mainly driven by logarithmic density score function. The rolling forecasting of VaR is made based on the 30-minute intraday and overnight returns of important stock indices from 2014 to 2020. Data sources are Shanghai Stock Exchange and Shenzhen Stock Exchange. MCS test and backtest methods such as Kupiec test and DQ test are used to evaluate the prediction effect comprehensively.

The empirical study shows that the overnight returns are more leptokurtic and left-skewed than the intraday returns. The square root of the ratio of the square of overnight returns and the sum of the squares of the intraday returns is up to 7, indicating the importance of the overnight returns. Among 7 common distributions such as normal distribution and student-t distribution, the GAS model based on skewed-student-t distribution has the best fitting effect for overnight returns. The GAS model has basically passed all VaR backtest methods, while the realized GARCH model has been rejected in those of the Shanghai Composite Index. According to the results of MCS test, overnight returns can indeed improve the VaR prediction effect. The GAS model with overnight returns has the best overall performance, which is better than the realized GARCH model with overnight returns.