CAJ | 학술논문

条件VaR模型的正确设定检验等价于检验均值化的“撞击序列”是否服从鞅差分序列，然而通常的反馈检验方法只检验了该序列的部分性质。采用对该鞅差分性质进行直接检验的广义谱检验方法，全面考察中国股票市场（香港恒生指数、上证综合指数和台湾加权指数）上各参数、非参数和半参数共22个VaR模型在采用滚动窗口预测机制时的样本外预测绩效。鉴于条件VaR模型正确设定检验无法反映超过某VaR水平的尾部风险信息，为避免极端损失的发生以及增加结果的稳健性，同时采用模型置信集检验方法。研究结果表明，采用通常的反馈检验方法常会得出错误的结论；在1％和5％置信水平，与历史模拟法、极值理论模型、CAViaR模型和CARE模型相比，误差项为t分布的GARCH模型族在金融危机期间具有较好的样本外预测绩效；涨跌停板制度对于选取预测绩效最优的VaR模型具有重要影响。
조건VaR모형적정학설정검험등개우검험균치화적“당격서렬”시부복종앙차분서렬，연이통상적반궤검험방법지검험료해서렬적부분성질。채용대해앙차분성질진행직접검험적엄의보검험방법，전면고찰중국고표시장（향항항생지수、상증종합지수화태만가권지수）상각삼수、비삼수화반삼수공22개VaR모형재채용곤동창구예측궤제시적양본외예측적효。감우조건VaR모형정학설정검험무법반영초과모VaR수평적미부풍험신식，위피면겁단손실적발생이급증가결과적은건성，동시채용모형치신집검험방법。연구결과표명，채용통상적반궤검험방법상회득출착오적결론；재1％화5％치신수평，여역사모의법、겁치이론모형、CAViaR모형화CARE모형상비，오차항위t분포적GARCH모형족재금융위궤기간구유교호적양본외예측적효；창질정판제도대우선취예측적효최우적VaR모형구유중요영향。
Conditional VaR models′correct specification test is equivalent to testing the de-meaned hit sequence following a mar-tingale difference sequence (m.d.s), however the commonly used backtesting techniques only test some properties of the se-quence.Using generalized spectral test which directly tests the m.d.s property of the de-meaned hit sequence, we evaluate the out-of-sample predictive performance of various parametric, nonparametric and semi-parametric VaR models with a total of 22 models calculated by using rolling predictive method for China′s stock markets including Shanghai Composite Index, Hang Seng Index and Taiwan Weighted Index.Because conditional VaR models′correct specification test can not reflect the tail risk infor-mation exceeding one specific VaR level, in order to avoid the occurrence of extreme losses as well as to increase the robustness of the results, we adopt MCS ( model confidence set) test simultaneously by selecting the asymmetric loss functions proposed by Koenker and Bassett and the magnitude loss function proposed by Lopez.Comparing with SPA( Superior Predictive Ability) test, the main advantage of MCS test is that it does not require a benchmark model to be specified as is the case for SPA tests.It char-acterizes the entire set of models that are not significantly outperformed by other models, while a test for SPA only provides evi-dence about the relative performance of a single model ( the benchmark) . The empirical results imply the following three conclusions:①it would cause wrong results using the commonly applied backtest-ing techniques such as Kupiec likelihood ratio test, Christoffersen likelihood ratio test and Engle and Manganelli dynamic quantile test.However adopting generalized spectral test and MCS test with Lopez loss function simultaneously would give us more accu-rate and robust results.②Comparing with historical simulation models, extreme value theory models, CAViaR and CARE mod-els, the out-of-sample predictive performance of the GARCH family with student-t distribution is the best at 1%and 5%signifi-cant level during the financial crisis for these three stock indexes.This implies that the risk characteristics of mainland stock mar-ket is getting closer and closer to the mature stock markets of Hong Kong and Taiwan after more than 20 years development.③At 1%significant level, the optimal VaR predictive models of Hang Seng Index include one of the CARE models which can be used to measure extreme loss situation with small probability.This implies that price limit system implemented by Hong Kong yet not by mainland and Taiwan will make Hong Kong′s stock market face more risk during the financial crisis.