【AP】Robust multi-period portfolio selection(3)

Pre Blog Link

Robust multi-period portfolio selection(2)

Robust conterparts of multi-period portfolio problems

本节主要分析在非对称不确定集合${\mathcal{F}}_{\Omega }$下，鲁棒投资组合模型$(2.14)$. 为此(To this end)，我们假设风险资产$i$的${\tilde{r}}_i^t,t=1,2,\dots ,T$由$m$个随机因子$\widetilde{{\xi \text{ξ}\xi }^t}=({\tilde{\xi }}_1^t,\dots ,{\tilde{\xi }}_m^t{)}^{\prime }\in {\mathbb{R}}^m$驱动，收益率向量$\widetilde{{r\text{r}r}^t}$可以表示为这些因子的线性组合.

从方程组$(2.7)$可以知道，由于${r\text{r}r}^0$是已知，可以得到收益率向量表达式如下
$${\widetilde{R\text{R}R}}^{t+1}={\mu \text{μ}\mu }^{t+1}+{A\text{A}A}^{t+1}\widetilde{{\xi \text{ξ}\xi }^t},t=1,\dots ,T\text{\hspace{1em}(4.1)}$$
其中${\mu \text{μ}\mu }^{t+1}$为常数，${A\text{A}A}^{t+1}\in {\mathbb{R}}^{n\times m}$是因子载荷矩阵，为了简化模型，设置${\widetilde{\xi \text{ξ}\xi }}^t$为序列无关，对于每阶段$t$满足标准条件$(3.6)$，且具有支撑集${\mathbb{S}}^t\in [-{l\text{l}l}^t,{u\text{u}u}^t]$. 可以得到随机向量的期望为
$$\mathbb{E}[{\widetilde{R\text{R}R}}^{t+1}]={\mu \text{μ}\mu }^{t+1},t=1,\dots ,T\text{\hspace{1em}(4.2)}$$
可以将非对称不确定集合${\mathcal{F}}_{\Omega }$扩展到多阶段的情况
F Ω t = { ξ t : ∃ u ‾ t , u ‾ t ∈ R + m , ξ t = u ‾ t − u ‾ t , ∥ ( P ) − 1 u ‾ + ( Q ) − 1 u ‾ ∥ 2 ≤ Ω t , ξ t ∈ [ − l t , u t ] } (4.3) \mathcal{F}_\Omega^t=\{\pmb{\xi}^t:\exist \underline{\pmb{u}}^t, \overline{\pmb{u}}^t\in\mathbb{R}_+^m, \pmb{\xi}^t=\overline{\pmb{u}}^t-\underline{\pmb{u}}^t, \lVert(\pmb{P})^{-1}\overline{\pmb{u}}+(\pmb{Q})^{-1}\underline{\pmb{u}}\rVert_2\leq\Omega^t, \pmb{\xi}^t\in[-\pmb{l}^t, \pmb{u}^t]\}\tag{4.3} FΩt​={ξ​ξ​​ξt:∃uuu​t,uuut∈R+m​,ξ​ξ​​ξt=uuut−uuu​t,∥(PPP)−1uuu+(Q​Q​​Q)−1uuu​∥2​≤Ωt,ξ​ξ​​ξt∈[−lllt,uuut]}(4.3)
其中，${P\text{P}P}^t=diag(p_1^t,\dots ,p_m^t)$且${Q\text{Q}Q}^t=diag(q_1^t,\dots ,q_m^t)$，矩阵元素满足
$$\{\begin{array}{l}{p}_i^t=p({\tilde{\xi }}_i^t)>0\\ q_i^t=q({\tilde{\xi }}_i^t)>0\end{array}$$
表示随机变量${\xi }_i^t$的forward和backward偏离值.

In the following, what we are interested in is the robust counterpart of $(2.14)$ when the random factors $\widetilde{{\xi \text{ξ}\xi }^t}$ vary in the uncertainty set ${\mathcal{F}}_{\Omega }^t$. To this end, we introduce some notation.

记
$$\begin{gathered} h_0^t({y\text{y}y}^t,{z\text{z}z}^{t+1})=w_0^t-w_0^{t+1}+(\frac{1-\theta }{R_0^{t+1}}{z\text{z}z}^{s,t+1}-\frac{1+\theta }{R_0^{t+1}}{z\text{z}z}^{b,t+1}){\mu \text{μ}\mu }^{t+1} \\ t=1,\dots ,T-1\text{\hspace{1em}(4.4)} \end{gathered}$$
以及
$$\begin{gathered} h_j^t({y\text{y}y}^t,{z\text{z}z}^{t+1})={e\text{e}e}_j^{\prime }({A\text{A}A}^{t+1}{)}^{\prime }(\frac{1-\theta }{R_0^{t+1}}{z\text{z}z}^{s,t+1}-\frac{1-\theta }{R_0^{t+1}}{z\text{z}z}^{b,t+1}) \\ j=1,\dots ,mt=1,\dots ,T-1\text{\hspace{1em}(4.5)} \end{gathered}$$
其中${y\text{y}y}^t=(w_0^t;{w\text{w}w}^t)\in {\mathbb{R}}^{n+1},{z\text{z}z}^{t+1}=({z\text{z}z}^{b,t+1},{z\text{z}z}^{s,t+1})\in {\mathbb{R}}^{2m}$，并且${e\text{e}e}_j$是一个$m$维向量，其中位置$j$的元素为1，其他的位置为0，根据$(4.4)$和$(4.5)$重写随机不等式约束$(2.13)$

可以得到
$$w_0^t-w_0^{t+1}+\frac{1-\theta }{R_0^{t+1}}({\widetilde{R\text{R}R}}^{t+1}{)}^{\prime }{z\text{z}z}^{s,t+1}-\frac{1+\theta }{R_0^{t+1}}({\widetilde{R\text{R}R}}^{t+1}{)}^{\prime }{z\text{z}z}^{b,t+1}\ge 0\text{\hspace{1em}(4.6)}$$
将$(4.1)$代入到随机约束$(4.6)$中，由$(4.1)\sim (4.5)$随机约束等效于如下关于$({y\text{y}y}^t;{z\text{z}z}^{t+1})$和$({\widetilde{\xi \text{ξ}\xi }}^t)$的双线性方程(bilinear function)
$$\begin{gathered} G^t({y\text{y}y}^t,{z\text{z}z}^{t+1};\widetilde{{\xi \text{ξ}\xi }^t})=h_0^t({y\text{y}y}^t,{z\text{z}z}^{t+1})+\sum _{j=1}^m{\tilde{\xi }}_j^t{h}_j^t({y\text{y}y}^t,{z\text{z}z}^{t+1})\ge 0 \\ t=1,\dots ,T-1\text{\hspace{1em}(4.7)} \end{gathered}$$
进一步
$$g({y\text{y}y}^T,{z\text{z}z}^{T+1},{\widetilde{\xi \text{ξ}\xi }}^t)=w_0^T{R}_0^{T+1}+({\widetilde{R\text{R}R}}^{T+1}{)}^{\prime }{w\text{w}w}^T\text{\hspace{1em}(4.8)}$$
和
$$h_0^T({y\text{y}y}^T,{z\text{z}z}^{T+1})=w_0^T{R}_0^{T+1}+({\mu \text{μ}\mu }^{T+1}{)}^{\prime }{w\text{w}w}^T\text{\hspace{1em}(4.9)}$$
和
$$\begin{gathered} h_j^T({y\text{y}y}^T,{z\text{z}z}^{T+1})={e\text{e}e}_j^{\prime }({A\text{A}A}^{T+1}{)}^{\prime }{w\text{w}w}^T \\ j=1,\dots ,m\text{\hspace{1em}(4.10)} \end{gathered}$$
由$(4.1)$可以得到
$$\begin{array}{rl}g({y\text{y}y}^T,{z\text{z}z}^{T+1},\widetilde{\xi \text{ξ}\xi })& =w_0^T{R}_0^{T+1}+({\widetilde{R\text{R}R}}^{T+1}{)}^{\prime }{w\text{w}w}^T\\ & \stackrel{\text{分离出非随机部分}}{=}(w_0^T{R}_0^{T+1}+({\mu \text{μ}\mu }^{T+1}{)}^{\prime }{w\text{w}w}^T)+\sum _{i=1}^m{\tilde{\xi }}_j({e\text{e}e}_j^{\prime }({A\text{A}A}^{T+1}{)}^{\prime }{w\text{w}w}^T)\\ & \stackrel{\text{代入4.9和4.10}}{=}{h}_0^T({y\text{y}y}^T,{z\text{z}z}^{T+1})+\sum _{i=1}^m{\tilde{\xi }}_i{h}_j^T({y\text{y}y}^T,{z\text{z}z}^{T+1})\end{array}\text{\hspace{1em}(4.11)}$$
因此可以将原问题$(2.14)$改写如下MLPM
$$\begin{array}{rl}& \underset{(y^t,z^{t+1})\in \mathcal{X}}{\min }\rho (a,\lambda ;g({y\text{y}y}^T,{z\text{z}z}^{T+1},\widetilde{{\xi \text{ξ}\xi }^t}{)}_{+})\\ & =-\mathbb{E}[g({y\text{y}y}^T,{z\text{z}z}^{T+1},{\tilde{\xi }}^t)]+\lambda \mathbb{E}[(a-g({y\text{y}y}^T,{z\text{z}z}^{T+1},\widetilde{{\xi \text{ξ}\xi }^t}){)}_{+}]\\ & s.t.\{\begin{array}{l}{G}^t({y\text{y}y}^t,{z\text{z}z}^{t+1};\widetilde{{\xi \text{ξ}\xi }^t})=h_0^t({y\text{y}y}^t,{z\text{z}z}^{t+1})+\sum _{j=1}^m\widetilde{{\xi }_j^t}{h}_j^t({y\text{y}y}^t,{z\text{z}z}^{t+1})\ge 0\\ t=1,\dots ,T-1\end{array}\end{array}$$
其中${y\text{y}y}^t=(w_0^t;{w\text{w}w}^t)\in {\mathbb{R}}^{n+1},{z\text{z}z}^{t+1}=({z\text{z}z}^{b,t+1};{z\text{z}z}^{s,t+1})\in {\mathbb{R}}^{2n}$，并且在确定性约束条件的可行集$\mathcal{X}$为
$$\mathcal{X}=\left\{({y\text{y}y}^t,{z\text{z}z}^{t+1}):\begin{array}{rl}& R_0^1{w}_0^1+\sum _{i=1}^m{R}_i^1{w}_i^1=1,i=1,\dots ,n\\ & w_i^{t+1}=w_i^t+z_i^{b,t+1}-z_i^{s,t+1}\\ & w_i^t\ge 0\\ & z_i^{b,t}\ge 0\\ & z_i^{s,t}\ge 0\end{array}\right\}$$
定理4.1 模型$(2.14)$(MLPM)的目标函数具有紧的上界(tight upper bound)
$$\begin{array}{rl}\rho & =-\mathbb{E}[g({y\text{y}y}^T,{z\text{z}z}^{T+1},{\widetilde{\xi \text{ξ}\xi }}^T)]+\lambda \mathbb{E}[(a-g({y\text{y}y}^T,{z\text{z}z}^{T+1},\widetilde{{\xi \text{ξ}\xi }^T}){)}_{+}]\\ & \le -(1+\lambda )h_0^T({y\text{y}y}^T,{z\text{z}z}^{T+1})+\lambda a-\lambda (\gamma -1)\hat{\beta }+\lambda (-\hat{\beta }{)}_{+}\\ & +\lambda \gamma ((h_0^T({y\text{y}y}^T,{z\text{z}z}^{T+1})-a)+\underset{{\xi }^T\in {\mathcal{F}}_{\Omega }}{\max }\sum _{j=1}^m{h}_j^T({y\text{y}y}^T,{z\text{z}z}^{T+1}){\xi }_j^T)\end{array}\text{\hspace{1em}(4.12)}$$
其中方程参数定义如下

证明过程见论文的Online Appendix.
记
$$\psi (a,\lambda ;g({y\text{y}y}^T,{z\text{z}z}^{T+1},{\xi \text{ξ}\xi }^T))$$
表示方程$(4.12)$的右端，即$\rho$的紧的上界(tight upper bound). 因此可以转换原问题$(2.14)$的目标函数为$\psi$. 从定理$4.1$的证明过程中可以知道
max ⁡ ξ T ∈ F Ω ∑ j = 1 m h j T ( y T , z T + 1 ) ξ j T = min ⁡ d T { Ω ∥ d T ∥ ： P − 1 d T ≥ h T , Q − 1 d h ≥ − h T } \max_{\xi^T\in\mathcal{F}_\Omega}\sum_{j=1}^mh_j^T(\pmb{y}^T, \pmb{z}^{T+1})\xi_j^T=\min_{d^T}\{\Omega\lVert \pmb{d}^T\rVert：\pmb{P}^{-1}\pmb{d}^T\geq \pmb{h}^T, \pmb{Q}^{-1}\pmb{d}^h\geq-\pmb{h}^T\} ξT∈FΩ​max​j=1∑m​hjT​(y​y​​yT,zzzT+1)ξjT​=dTmin​{Ω∥dddT∥：PPP−1dddT≥hhhT,Q​Q​​Q−1dddh≥−hhhT}
由
$$(x{)}_{+}=x+(-x{)}_{+}$$
代入$\psi$表达式可以得到

Next we return to the stochastic inequality constraints of problem (MLPM). With the uncertainty set ${\mathcal{F}}_{\Omega }$ and $3.10$，the robust counterpart of the stochastic inequality constraint is the determinate inequality.

$$G^t({y\text{y}y}^t,{z\text{z}z}^{t+1},{\xi \text{ξ}\xi }^t)\ge 0,{\xi \text{ξ}\xi }^t\in {\mathcal{F}}_{\Omega }^t\text{\hspace{1em}(4.13)}$$

In summary, we have the robust counterpart of original problem, denoted by RLPM:

定理4.2 对于任意$\lambda \ge 1,\epsilon >0$,${\epsilon }_t=\epsilon /(T-1)，\Omega ={\Omega }_t=\sqrt{-2\ln ({\epsilon }_t)}，\gamma =\exp (-{\Omega }^2/2)$，对于RLPM的任意可行解$({y\text{y}y}^t,{z\text{z}z}^{t+1})$，原问题的随机约束满足

作者对定理$(4.2)$的评论如下（主要分析了其内在的保守性）

The impact of Theorem 4.2 is double. On the one hand, one can always take a desired confidence level $\epsilon$ such as $\epsilon =0.05$，such that the probability guarantee (4.16) holds. However, on the other hand, this will lead to a very conservative individual constraint, in other words, if $\epsilon =0.05$，then we must take ε t ≤ ε / ( T − 1 ) = 0.0005 ; P { G t ( y t , z t + 1 ; ξ t ) ≥ 0 } ≥ 0.9995 \varepsilon_t\leq \varepsilon/(T-1)=0.0005; \mathbb{P}\{G^t(\pmb{y}^t, \pmb{z}^{t+1};\pmb{\xi}^t)\geq 0\}\geq 0.9995 εt​≤ε/(T−1)=0.0005;P{Gt(y​y​​yt,zzzt+1;ξ​ξ​​ξt)≥0}≥0.9995 for $T=101$. Indeed, this is too conservative and is unpractical.

The relationship between RLPM and CVaR

定义关于投资组合$x\text{x}x\in \mathcal{X}\subset {\mathbb{R}}^{n+1}$和随机向量$\widetilde{\xi \text{ξ}\xi }\in {\mathbb{R}}^m$的损失函数$f(x\text{x}x,\widetilde{\xi \text{ξ}\xi })$.

VaR is defined as the minimal level $\beta$, such that the probability that the portfolio loss $f(x\text{x}x,\widetilde{\xi \text{ξ}\xi })$ exceeds $\beta$ is below $\alpha$, where $\alpha \in (0,1)$ is the probability level specified by the user (typically, $\alpha =1\%$ or $5\%$).

V a R 1 − α ( f ( x , ξ ~ ) ) = min ⁡ { β ∣ P ( f ( x , ξ ~ ) ) ≤ α } VaR_{1-\alpha}(f(\pmb{x}, \tilde{\pmb{\xi}}))=\min\{\beta\mid \mathbb{P}(f(\pmb{x}, \pmb{\tilde{\xi}}))\leq \alpha\} VaR1−α​(f(xxx,ξ​ξ​​ξ~​))=min{β∣P(f(xxx,ξ~​​ξ~​​​ξ~​))≤α}
CVaR度量了在$f(x\text{x}x,\widetilde{\xi \text{ξ}\xi })\ge Va{R}_{1-\alpha }$下的条件期望损失，数学上的定义可以参考Rockafellar, R & Uryasev (2000)¹给出.
$$CVa{R}_{1-\alpha }(f(x\text{x}x,\widetilde{\xi \text{ξ}\xi }))=\mathbb{E}[f(x\text{x}x,\widetilde{\xi \text{ξ}\xi })\mid f(x\text{x}x,\widetilde{\xi \text{ξ}\xi })\ge Va{R}_{1-\alpha }]$$
对于给定的$\alpha$，如果设置LPM中的目标a的值为$a=1-Va{R}_{1-\alpha }$，并且$g(x\text{x}x,\widetilde{\xi \text{ξ}\xi })=1-f(x\text{x}x,\widetilde{\xi \text{ξ}\xi })$，那么可以得到
$$\begin{array}{rl}LPM(a;g(x,\tilde{\xi }))& =\mathbb{E}[a-g(x\text{x}x,\widetilde{\xi \text{ξ}\xi }){]}_{+}=\mathbb{E}[a-g(x\text{x}x,\widetilde{\xi \text{ξ}\xi })\mid a-g(x\text{x}x,\widetilde{\xi \text{ξ}\xi })\ge 0]\\ & =\mathbb{E}[f(x\text{x}x,\widetilde{\xi \text{ξ}\xi })-Va{R}_{1-\alpha }(f(x\text{x}x,\widetilde{\xi \text{ξ}\xi }))\mid f(x\text{x}x,\widetilde{\xi \text{ξ}\xi })\ge Va{R}_{1-\alpha }]\end{array}$$
如果$\alpha$的取值使得$Va{R}_{1-\alpha }$为非负值，那么可以得到
$$LPM(a;g(x\text{x}x,\widetilde{\xi \text{ξ}\xi }))\le CVa{R}_{1-\alpha }(f(x\text{x}x,\widetilde{\xi \text{ξ}\xi }))$$
当$\alpha$和$\tilde{\xi }$的分布为给定时，即$Va{R}_{1-\alpha }(f(x\text{x}x,\widetilde{\xi \text{ξ}\xi }))$为确定值，可以得到
$$LPM(a;g(x\text{x}x,\widetilde{\xi \text{ξ}\xi }))=CVa{R}_{1-\alpha }(f(x\text{x}x,\widetilde{\xi \text{ξ}\xi }))-Va{R}_{1-\alpha }(f(x\text{x}x,\widetilde{\xi \text{ξ}\xi }))$$

In this case, the minimization of CVaR is equivalent to the minimization of the sum of LPM and VaR. Thus, the differences between mean-CVaR and mean-LPM are clear.
First, in the minimization of CVaR problem, we can understand that the target $Va{R}_{1-\alpha }(f(x\text{x}x,\widetilde{\xi \text{ξ}\xi }))$ is in fact endogenous. But in LPM, the target $a$ is exogenous.
Second, the advantage of exogenous $a$ can avoid the effect of uncertainty set in robust portfolio choice which can reduce a chance constraint in multi-period case.

根据CVaR的一致性，可以得到
$$\underset{x}{\min }CVa{R}_{1-\alpha }(f(x\text{x}x,\widetilde{\xi \text{ξ}\xi }))=\underset{x,{\gamma }_0}{\min }\{{\gamma }_0\mid CVa{R}_{1-\alpha }f(x\text{x}x,\widetilde{\xi \text{ξ}\xi })+{\gamma }_0\le 0\}\text{\hspace{1em}(4.17)}$$

The right hand side of problem $(4.17)$ includes a CVaR constraint which is still difficult to be expressed explicitly when the distribution information $\widetilde{\xi \text{ξ}\xi }$ is only given by $(3.7)$

可以将CVaR约束看做机会约束的近似
$$CVa{R}_{1-\epsilon }(f(x\text{x}x,\widetilde{\xi \text{ξ}\xi })+{\gamma }_0)\le 0\to \mathbb{P}\{f(x\text{x}x,\widetilde{\xi \text{ξ}\xi })+{\gamma }_0\ge 0\}\ge 1-\epsilon \text{\hspace{1em}(4.18)}$$
但是这并不是充分条件.

As stated in Natarajan et al. (2008) to solve the minimization of CVaR under uncertainty set ${\mathcal{F}}_{\Omega }$, it needs to add a new constraint (the right hand side of (4.18)) , which will lead to a more conservative solution from $(4.2)$ than LPM.
Third, when the exogenous target a is taken not too extreme value, it can help LPM reduce the variance of portfolio. For example, if a is close to the mean return of portfolio, then LPM in fact controls the down-side semi-variance which clearly is important to limit the variance. (如果目标值a与投资组合的收益均值接近，那么LPM实际上控制了下行方差（半方差）的风险).

可以由VaR给出外生变量a的参考值. 令投资组合的收益率${\tilde{r}}_p(\widetilde{\xi \text{ξ}\xi })$服从正态分布. 令$a=1-VaR$²，可以得到
$$a=1-{\mathcal{Z}}_{1-\alpha }{\sigma }_{{\tilde{r}}_p}\text{\hspace{1em}(4.19)}$$

---

1. Optimization of conditional value-at-risk ↩︎

2. 在normal distribution下可以计算VaR值为$VaR={\mathcal{Z}}_{1-\alpha }{\sigma }_{{\tilde{r}}_p}+{\mu }_{{\tilde{r}}_p}$，本文的数值实验中，使用Hull (2015)的结果，他在计算过程中忽略了期望收益${\mu }_{{\tilde{r}}_p}$，因为和波动率相比，期望收益的值太小了. ↩︎