Research On The Features Of Tail Risk And Tail Dependence In Stock Markets

The frequent attacks of financial crisis always bring pestilent disaster to investors,financial system and even the real economy. During these crises, the financial marketsalways display special features, which would not be observed in normal periods. On theone hand, investors would suffer extreme loss in crises and hence "tail risk" exists fromthe perspective of univariate case; On the other hand, the cross-market and cross-assetrelationships during crisis periods usually become stronger and hence "tail dependence"exists from the perspective of multivariate case. How to successfully describe these twospecial features and further cope with them properly is a challenging task, no matter forinvestors and risk managers, or for policy makers and regulators. Therefore, this thesisemploys various flexible econometric tools to capture the features of "tail risk" and "taildependence" appearing in crises, and then examine their influences on risk management,asset allocation and asset pricing.Firstly, for the tail risk of individual assets, this thesis introduces a more powerfulsaddlepoint technique backtest to strictly re-evaluate the VaR and ES forecasts accuracyof various risk models, and re-explore what kinds of models can capture the tail risk ofunivariate case best. The empirical analysis shows that, simple GATCH-Normal cannotcapture the risk characteristic in China's stock market, and the best risk model should beGARCH-EVT. The further analysis based on more mature stock markets and more riskmodels confirms the necessity of employing Extreme Value Theory (EVT) to fit the tailsseparately from the central portion of financial returns. The asymmetry in the tail riskscannot be correctly modeled by attempting to fit a single distribution, even the skewedStudent t distribution advocated by previous literature, to both small and large financiallosses. More importantly, a formal statistical test is performed to compare the effects oftwo dimensions in GARCH models on accuracy of risk forecast, and the results suggestthe specification of distribution tail plays a more important role than the heteroscedasticprocess in forecasting tail risks in stock markets.Secondly, for the effect of tail dependence on risk management, this thesis proposesto calculate the risk contributions of individual assets in portfolios based on multivariate Copula simulation. Our research shows that this simulation approach not only providesa way to test whether the risk contributions of various assets are significantly different,but also generates results insusceptible to confidence level and risk measures. Moreover,this thesis further introduces a flexible tool of modeling multi-dimensional dependencestructure, Canonical Vine copula, to overcome the long-standing problem in literaturethat the multivariate Copula functions available now are rather few and imperfect either,each of them facing certain defects. The empirical study based on Shanghai, Hong Kongand Taiwan stock markets verifies the superiority of Canonical Vine Copula function indescribing the multi-dimensional dependence structure. Further, a finding more helpfulto practice is that, the risk forecast analysis based on both various strategies and varioussimulation samples indicates that modeling multi-dimensional dependence structure byCanonical Vine Copula function can make the VaR forecasts of portfolios become morerobust and accurate.Thirdly, for the effect of tail dependence on asset allocation, this thesis employs theMarkov Switching Copula model to capture both the non-linearity and the time-varietyof cross-asset relationship, and further constructs a procedure for investors to choose thetime of portfolio adjustment based on this model. The empirical evidence based on twostock portfolios (high-risk and low-risk portfolios) in Chinese market confirms that thedependence structures among financial assets are regime-switching indeed, and thus it isnot proper any longer for investors to build their portfolios only based on one invariantdistribution model during a relatively long investment horizon, like the static strategiesin previous studies. Therefore, this thesis asserts that it would be more proper to forecastthe possible changes of regimes (representing different dependence structures) using theMarkov Switching Copula model, and then adjust the portfolios weights in time only ifthe regimes have changed. The analysis on out-of-sample asset allocation performancefurther indicates that compared to those strategies used by previous studies, the timingstrategy constructed in this thesis could generate both higher realized returns and higherCertainty Equivalent Rate of Returns.Finally, for the effect of tail dependence on asset pricing, this thesis also focuses onthe tail dependence between individual stocks and the whole market, and then examinesits effect on stock returns. The existence of short-sale constraints in many stock marketsalways generates extreme downside market risk, and this extreme downside market risk is much stronger than upside market risk. However, linear beta cannot distinguish suchan asymmetry in market risk, so this thesis employs the tail dependence coefficients tomeasure this extreme downside market risk, i.e., the risk of individual stocks crashingdown together with the market during crisis. The empirical evidence based on ShanghaiA shares confirms that most stocks have non-negligible tail dependence with the market.More importantly, such tail dependence risk is proved to have explanatory power for themonthly returns on Shanghai A shares, and this explanatory power still exists even aftercontrolling the effects of beta and other factors. Therefore, such tail dependence mightprovide a new perspective for us to describe the market risk, and probably contain someinformation that linear beta and other asset pricing factors well-cited by previous studiescannot provide. Hence, tail dependence coefficients have the potential to become a newasset-pricing factor and should be taken into account in future study.