【AP】Robust multi-period portfolio selection(4)

Pre Link

Robust multi-period portfolio selection(3)

Numerical results

Parameters estimation approach

令${\tilde{r}}_i^t=({\tilde{S}}_i^t-{\tilde{S}}_i^{t-1})/{\tilde{S}}_i^{t-1}$表示风险资产$i$在时刻$t$的平均收益率，计算对数收益率为
$$\{\begin{array}{rl}& \widetilde{{Y\text{Y}Y}^t}=\ln (1+{\widetilde{r\text{r}r}}_t)\\ & {\tilde{Y}}_i^t=\ln \frac{{\tilde{S}}_i^t}{{\tilde{S}}_i^{t-1}}=\ln (1+{\tilde{r}}_t^i)\\ & \widetilde{{Y\text{Y}Y}^t}=({\tilde{Y}}_1^t,\dots ,{\tilde{Y}}_n^t{)}^{\prime }\end{array}\text{\hspace{1em}(5.1)}$$
假设 { Y t ~ } \{\tilde{\pmb{Y}^t}\} {YYYt~}是一个VAR(1)过程
$$\begin{gathered} \widetilde{{Y\text{Y}Y}^t}=v\text{v}v+\Phi \text{Φ}\Phi {Y\text{Y}Y}^{t-1}+\widetilde{{\eta \text{η}\eta }^t} \\ \widetilde{{\eta \text{η}\eta }^t}\sim MVN[0\text{0}0,\Sigma \text{Σ}\Sigma ]\text{\hspace{1em}(5.2)} \end{gathered}$$
设在整个$t$期间，矩阵$\Sigma \text{Σ}\Sigma$为固定的，并且$\widetilde{{\eta \text{η}\eta }^t}=\widetilde{\eta \text{η}\eta }\sim MVN(0\text{0}0,\Sigma \text{Σ}\Sigma )$.
令
$$\widetilde{\xi \text{ξ}\xi }={\Sigma \text{Σ}\Sigma }^{-\frac{1}{2}}\widetilde{\eta \text{η}\eta }$$
可以知道$\widetilde{\xi \text{ξ}\xi }$满足标准条件，方程可以表示为
$$\widetilde{{Y\text{Y}Y}^t}=v\text{v}v+\Phi \text{Φ}\Phi {Y\text{Y}Y}^{t-1}+{\Sigma \text{Σ}\Sigma }^{-\frac{1}{2}}\widetilde{\xi \text{ξ}\xi }\text{\hspace{1em}(5.3)}$$
设当前时刻$t=0$，且$\widehat{v\text{v}v},\widehat{\Phi \text{Φ}\Phi },\widehat{\Sigma \text{Σ}\Sigma }$为参数$v\text{v}v,\Phi \text{Φ}\Phi$和$\Sigma \text{Σ}\Sigma$的估计值，在可观测样本序列 { Y } t ≥ − T 0 0 \{\pmb{Y}\}^0_{t\geq -T_0} {YYY}t≥−T0​0​，可以得到
$${\widehat{\mu \text{μ}\mu }}_Y^t=\mathbb{E}[\widetilde{Y\text{Y}Y}\mid {Y\text{Y}Y}^{t-1}]=\widehat{v\text{v}v}+\widehat{\Phi \text{Φ}\Phi }{Y\text{Y}Y}^{t-1}$$
令 { ξ t ^ } t ≥ − Y 0 0 \{\hat{\xi^t}\}^0_{t\geq -Y_0} {ξt^​}t≥−Y0​0​表示残差序列，可以通过定理3.3中的公式计算出$p_i$和$q_i$的值.

本文为了简化，设置$p_i$和$q_i$在所有阶段为固定值.

In multi-period portfolio optimization, we need to estimated ${\mu \text{μ}\mu }^t$ in advance. Thus, we need to estimate the estimate ${\mu \text{μ}\mu }^t$ in advance. Thus, we need to estimate the possible realizations of { Y t } \{\pmb{Y}^t\} {YYYt} in advance using the following equation:

$$\widetilde{{Y\text{Y}Y}^t}=\widehat{v\text{v}v}+\widehat{\Phi \text{Φ}\Phi }{Y\text{Y}Y}^{t-1}+{\widehat{\Sigma \text{Σ}\Sigma }}^{1/2}\widetilde{\xi \text{ξ}\xi },t>0$$
设置$\widetilde{\xi \text{ξ}\xi }$服从如下分布

关于数值$J$的论述如下所示¹

Formally, we state the following general simulation procedure (GSP) to obtain a data sequence of { Y t } t = 0 T \{\pmb{Y^t}\}_{t=0}^T {YtYtYt}t=0T​.

GSP的过程如下

模拟产生的序列 { Y t } t = 0 T \{\pmb{Y}^t\}_{t=0}^T {YYYt}t=0T​可以用来估计模型参数${\mu \text{μ}\mu }^t,{A\text{A}A}^t$

如果是长期投资，可以使用滚动过程(M-stage)

When the investment horizon is long, we can use the following M-stage rolling procedure (M-SRP) to obtain the appropriate probability guarantee. Let the investment horizontal $T^{\prime }$ be long and the adjustment frequency be $\Delta t$ years. Consider a T-period portfolio model with $T^{\prime }=M\ast T$. The multi-stage rolling procedure for a long investment horizon, requres the T-period portfolio mode RLPM to be implemented M times.
Specifically, at the kth (k=1,2,..., M) stage, we run the RLPM to get the optimal portfolio sequence { w t k } t = 1 : T \{\pmb{w}_t^k\}_{t=1:T} {wwwtk​}t=1:T​.

策略运行过程描述如下

设计对比实验，主要对比模型有equally-weighted ,model(1/N), mean-WVaR², mean-CVaR³ and mean-WCVaR⁴

主要对比指标有expected return, volatility和Sharpe Ratio，引入风险回避因子$\lambda$，指标计算公式如下.

Numerical comparisons with real market data

数据选取

使用样本内数据(in-sample)估计VAR(1)的参数$v\text{v}v,\Phi \text{Φ}\Phi ,\Sigma \text{Σ}\Sigma$

根据方程$(5.7)$，使用GSP生成模拟序列，设置调仓频率为为1年，那么设置$K=252$，在得到模拟序列$\widehat{Y\text{Y}Y}$后，可以得到关于参数${\mu \text{μ}\mu }^t$和$\Sigma \text{Σ}\Sigma$的估计. 将历史数据得到的信息和模拟序列信息合并可以得到
$${\widehat{\Sigma \text{Σ}\Sigma }}^{sim}=Cov(\widehat{{Y\text{Y}Y}^t}\mid {\mathcal{F}}_{t-1})$$
更新$\Sigma \text{Σ}\Sigma$得到
$${\widehat{\Sigma \text{Σ}\Sigma }}^{new}=\delta \widehat{{\Sigma \text{Σ}\Sigma }^s}+(1-\delta ){\widehat{\Sigma \text{Σ}\Sigma }}^{sim}\text{\hspace{1em}(5.8)}$$

静态策略回测结果如Fig-2所示.
Fig-2: 给出了静态RLPM策略(static-RLPM)的表现.

Fig.2 indicates that the expected return, volatility and Sharpe ratio of the optimal portfolio of RLPM are decreasing with respect to the risk-aversion coeff. $\lambda$ for a fixed target $a$. For large $\lambda$, the lower partial moment is smaller, In particular
$$\lambda \times \mathbb{E}[(a-w_0^T{R}_0^{T+1}-({\widetilde{R\text{R}R}}^{T+1}{)}^{\prime }{w\text{w}w}^T{)}_{+}]\to 0$$
Fig.2 are approximately corresponding to asingle asset VaR with volatility $12.6\%$ per year and confidence level $0.975,0.90$ and $0.80$. Note that $a$ is the investment target at the terminal period and $1-a$ reflects the acceptable loss for the investor. Thus, the smaller the target $a$ is, the larger the acceptable loss is, which in return indicates the investor would like to find a big bet with possible large loss or return.

我们加总了每一期所有投资在风险资产上的权重，然后计算了平均值，如Fig-3所示

Fig-4对比了不同的策略（WVaR, CVaR, WCVaR,1/N, RLPM）的回测结果.

the WVaR model controls only the probability of extreme loss, and even though the CVaR and WCVaR can limit the expectation of loss-tail losses, it also only controls the expectation under the condition of left-tail losses, it also only controls the expectation under the condition of extreme loss. Extreme loss in general is a low-probability event whose contribution to controlling the whole variance and improving the sharpe ratio is possible small.
RLPM can also control the lower parity loss, but the left-tail event is not extreme and a target $a$ can be set by investors. Thus, the lower parity loss limited by RLPM has a greater probability than the extreme event of WVaR, WCVaR, CVaR models.

考虑非对称不确定集合的影响，作者提出

the asymmetry uncertainty set is significantly important to reduce the volatility of portfolio relative to the symmetry ellipsoidal uncertainty set $U_{\Omega }$.

Fig-5: 给出设置非对称风险不确定集时和对称风险不确定集时的策略表现对比

关于图中结果的解释：

One is that, as stated in Introduction, the asymmetry attitude of investors on the gain and loss can form the asymmetry distribution feature. The asymmetry uncertainty set model is clearly more appropriate than the symmetry ellipsoidal uncertainty set model.
The another is that, the asymmetry uncertainty ${\mathcal{F}}_{\Omega }$ is the generalization of the symmetry ellipsoidal uncertainty set $U_{\Omega }$ and can lead to the relative more conservative case than the ellipsoidal uncertainty set $U_{\Omega }$, which can lead to smaller variance of portfolio.

Fig-6: 样本外不同调仓频率的对比

In order to show that the Sharpe Ratio of RLPM is better than other models not only at the average level, but also at the statistical level, Table-1 reports a performance test of the Sharpe ratio using a significance test proposed by Jobson and Korkie (1981)⁵ and Memmel (2013)

Futher numerical comparisons

对于本文提出的RLPM模型和Rujeeraoaiboon et al(2016)提出的ROGP模型⁶(that can outperform equally weighted strategy).

RGOP model is essentially the worst-case VaR portfolio model and can be expressed as follows.

Appendix

---

1. Robust multiperiod portfolio management in the presence of transaction costs ↩︎

2. Mean-WVaR: min ⁡ x { − E [ x ′ r ~ ] + λ W V a R 1 − α : x ≥ 0 , ∑ i = 1 n x i = 1 } \min\limits_x\{-\mathbb{E}[x'\tilde{r}]+\lambda WVaR_{1-\alpha}: x\geq 0, \sum_{i=1}^n x_i=1\} xmin​{−E[x′r~]+λWVaR1−α​:x≥0,∑i=1n​xi​=1} with a normal distribution assumption. The $\alpha$ is the worst-case probability that the loss of portfolio is greater than worst-case VaR (WVaR), Ghaoui, 2003 ↩︎

3. Mean-CVaR: min ⁡ x { − E [ x ′ r ~ ] + λ C V a R 1 − α : x ≥ 0 , ∑ i = 1 n x i = 1 } \min\limits_x\{-\mathbb{E}[x'\tilde{r}]+\lambda CVaR_{1-\alpha}: x\geq 0, \sum_{i=1}^n x_i=1\} xmin​{−E[x′r~]+λCVaR1−α​:x≥0,∑i=1n​xi​=1} is solved by linear programming under discrete $\alpha$ are taken. Rokafellar and Uryasev, 2000 ↩︎

4. Mean-WCVaR: min ⁡ x { E [ x ′ r ~ ] + λ W C V a R 1 − α : x ≥ 0 , ∑ i = 1 n x i = 1 } \min\limits_x \{\mathbb{E}[x'\tilde{r}]+\lambda WCVaR_{1-\alpha}: x\geq 0, \sum_{i=1}^n x_i=1\} xmin​{E[x′r~]+λWCVaR1−α​:x≥0,∑i=1n​xi​=1} is solved by SOCP under the ellipsoidal uncertainty set $U_{\Omega }$. Zhu and Fukushima, 2009 ↩︎

5. Performance hypothesis testing with the Sharpe and Treynor measures ↩︎

6. Robust growth-optimal portfolios ↩︎