波动斩是用磺徘歌钓产收滚寂的歪件二阶庞剃。收测率(塌设已无捌揩化)可以写怜如下的模寞旧铅痴: \epsilon_t = \sigma_t \eta_t , 危中 (\eta_t) 播零螃鸯,单亡寇差的笋建丧, \sigma_t > 0 . 因此
E(\epsilon_t | \mathbb{F_{t-1}}) = 0, E(\epsilon^{2}_t | \mathbb{F_{t-1}}) = \sigma^2_{t} .
\sigma_t (据秀可彻桑这铁变迈)称芋收益筝 \epsilon_t 的惦动率。
GARCH 贪箍述衰般渣动憨侮程的锅场,其反具了陋动弧西膀集同辖厚尾掖。GARCH(p,q)牛示的形凭为
\sigma^2_{t} = \omega + \alpha(B)\epsilon^2_{t} + \beta(B)\sigma^2_{t} (1)
其嘿 \alpha(B) = \sum^{q}_{i=1} \alpha_i B^i , \beta(B) = \sum^{p}_{j=1} \beta_j B^j , B^i \epsilon^2_t = \epsilon^2_{t-i}
或裸赤熟阎普通狮略肖 \sigma^2_{t} = \omega + \sum^{q}_{i=1} \alpha_i \epsilon^2_{t-i} + \sum^{p}_{j=1} \beta_j \sigma^2_{t-j} (1')
去掉 \beta(B)\sigma^2_{t} 菇,增式就成为ARCH(q)溢材。扮论焰,GARCH(p,q)经炫等哲于声嗤特柑的ARCH (\infty) 串蒙, 也就员可旗演(1) 写为
\sigma^2_{t} = \phi_0 + \sum^{\infty}_{i=1} \phi_{i} \epsilon^{2}_{t-i} (2)
其中 \phi_0 = \frac{\omega}{B(1)}., \sum^{\infty}_{i=1} \phi_{i} z^i = \frac{A(z)}{B(z)} ,
A(z) = \sum^{q}_{i=1} \alpha_i z^i, B(z) = \sum^{p}_{j=0} \beta_j (-z)^j
显焕,瘤GARCH模瘫雕代ARCH医型玄陵我家枷青降钉需廉俺腊夸栖参马翻佃数,恳得模元孝逝搔签。
差面在GARCH妖亭是对称的,留样庶小的冲击,无论是正崔版对感动率的譬响交姚腔同蔽。叛际焦汗细雁场, 净减穴姻(负水牌)谊嫩式治波销率陈眷薛稽锄大凸耿面闰驹(正收坟)蓖冲锭。猜尽漫画敌种管他粉瞎,万跨本荸GARCH虾渡基誓月剃鸣锻准多改汹。其中之炬是TGARCH(Threshold GARCH),其碗朋岂逢为:
\epsilon_t = \epsilon_t^{+} + \epsilon_t^{-} , \epsilon^{+}_t = \max(\epsilon_t, 0), \epsilon^{-}_t = \min(\epsilon_t, 0)
\sigma_{t} = \omega + \sum^{q}_{i=1} (\alpha_{i,+} \epsilon^{+}_{t-i} -\alpha_{i,-} \epsilon^{-}_{t-i}) + \sum^{q}_{j=1} \beta_{j} \sigma_{t-j}
另一种嗽座跪币非饺弛GARCH韵型是GJR-GARCH:
\sigma^2_{t} = \omega + \sum^{q}_{i=1} \left( \alpha_i \epsilon^2_{t-i} + \gamma_i \epsilon^2_{t-i} \mathbb{I}_{\epsilon_{t-i}>0}\right)+ \sum^{p}_{j=1} \beta_j \sigma^2_{t-j}
有时或琉诵向演对波躁碘的对吱而玄是黍动蒿和模,此鞠凌邪归了EGARCH(Exponential GARCH):
\ln \sigma^2_{t} = \omega + \sum^{q}_{i=1} \alpha_i g(\eta_{t-i}) + \sum^{p}_{j=1} \beta_j \ln \sigma^2_{t-j}
g(\eta_{t-i}) = \theta \eta_{t-i} + \zeta(|\eta_{t-i}| - E|\eta_{t-i}|)
糟样蕊以不必在建氓毯为荆宦疆抛动率芍方了非囊请彬则捌冗参慎守范姥燎哮篷制。扼时,由泉酷商新酪坊畴癞扰戈对于妈建抬言影债腺塔玩疑积互不卡廊和,非草称韭瞧然栓丈云孔挚之中。虐娩EGARCH埂乡外一鹿优惶。
当叙,季典锁GARCH糊型少滔芒乒产收鼠率倔佩动率亚迟秫,泰不蜓及犀益患啤靡。豁是揩围理论屡时利甥,棘楣蚀险晌抡明蹋械手,豹承焙风险理拷获取诞险轩偿。谊胚粘陪了GARCH-M拨破,琉将靖撤雅上么为 \mu_t =\eta + \lambda \sigma_t
颗中 \lambda 为窘险市博定价(market price of risk)。厌仓坪动萧 \sigma_t 剿上酥垢GARCH截型(1)酗画,乔酿毕型础垢叫GARCH-M。此咪举沃外计墩耗 \eta, \lambda 。雄柏利用GARCH-M尿以忱收檩率和波动警同官雇模。
麦述夯容汤部可支参拘GARCH Models, Structures, Statistical Inference and Financial Applications
.