Skip to main content
Log in

On variance reduction of mean-CVaR Monte Carlo estimators

  • Original Paper
  • Published:
Computational Management Science Aims and scope Submit manuscript

Abstract

We formulate an objective as a convex combination of expectation and risk, measured by the \(\mathrm{CVaR }\) risk measure. The poor performance of standard Monte Carlo estimators applied on functions of this form is discussed and a variance reduction scheme based on importance sampling is proposed. We provide analytical solution for random variables based on normal distribution and outline the way for the other distributions, either by analytical computation or by sampling. Our results are applied in the framework of stochastic dual dynamic programming algorithm. Computational results which validate the previous analysis are given.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Fig. 1
Fig. 2
Fig. 3
Fig. 4

Similar content being viewed by others

References

  • Artzner P, Delbaen F, Eber J-M, Heath D (1999) Coherent measures of risk. Math Financ 9:203–228

    Article  Google Scholar 

  • Bayraksan G, Morton DP (2011) A sequential sampling procedure for stochastic programming. Oper Res 59:898–913

    Article  Google Scholar 

  • Blum M, Floyd RW, Pratt V, Rivest RL, Tarjan RE (1972) Linear time bounds for median computations. In: Proceedings of the fourth annual ACM symposium on theory of computing. pp 119–124

  • Denneberg D (1990) Premium calculation: why standard deviation should be replaced by absolute deviation. ASTIN Bull Int Actuar Assoc 20:181–190

    Article  Google Scholar 

  • Hesterberg TC (1995) Weighted average importance sampling and defensive mixture distributions. Technometrics 37:185–194

    Article  Google Scholar 

  • Knopp R (1966) Remark on algorithm 334 [G5]: normal random deviates. Commun ACM 12:281

    Article  Google Scholar 

  • Kozmík V, Morton D (2014) Evaluating policies in risk-averse multi-stage stochastic programming. Math Program. doi:10.1007/s10107-014-0787-8

  • Kozmík V (2012) Multistage risk-averse asset allocation with transaction costs. In: Proceedings of 30th international conference on mathematical methods in economics. pp 455–460

  • Krokhmal P, Zabarankin M, Uryasev S (2011) Modeling and optimization of risk. Surv Oper Res Manag Sci 16:49–66

    Google Scholar 

  • Markowitz HM (1952) Portfolio selection. J Financ 7:77–91

    Google Scholar 

  • Markowitz HM (1959) Portfolio selection: efficient diversification of investments. Wiley, New York

    Google Scholar 

  • Pereira MVF, Pinto LMVG (1991) Multi-stage stochastic optimization applied to energy planning. Math Program 52:359–375

    Article  Google Scholar 

  • Philpott AB, de Matos VL, Finardi EC (2013) On solving multistage stochastic programs with coherent risk measures. Oper Res 61:957–970

    Article  Google Scholar 

  • Rockafellar RT, Uryasev S (2002) Conditional value-at-risk for general loss distributions. J Bank Financ 26:1443–1471

    Article  Google Scholar 

  • Rudloff B, Street A, Valladao D (2014) Time consistency and risk averse dynamic decision models: definition, interpretation and practical consequences. Eur J Oper Res 234(3):743–750

  • Ruszczynski A (2010) Risk-averse dynamic programming for Markov decision processes. Math Program 125:235–261

    Article  Google Scholar 

  • Shapiro A (2009) On a time consistency concept in risk averse multistage stochastic programming. Oper Res Lett 37:143–147

    Article  Google Scholar 

  • Shapiro A (2011) Analysis of stochastic dual dynamic programming method. Eur J Oper Res 209:63–72

    Article  Google Scholar 

Download references

Acknowledgments

The research was partly supported by the project of the Czech Science Foundation P/402/12/G097 ‘DYME/Dynamic Models in Economics’.

Author information

Authors and Affiliations

Authors

Corresponding author

Correspondence to Václav Kozmík.

Appendix

Appendix

Proof of Proposition 3 is given below.

Proof

Following are known integrals with normal pdfs:

$$\begin{aligned} \int z\phi (z) \, d z&= -\phi (z) + C , \\ \int z^2\phi (z) \, d z&= \varPhi (z) - z\phi (z) + C . \end{aligned}$$

We will write \(f\) instead od \(f_Z\) and \(g\) instead of \(g_Z\) for simplicity and expand the basic formula for variance:

$$\begin{aligned} \mathrm{var }_{g}\left[ \mathsf {Q}^i_Z\right]&= \mathbf{E}_{g}\left[ \left( \mathsf {Q}^i_Z\right) ^2\right] - \left( \mathbf{E}_{g}\left[ \mathsf {Q}^i_Z\right] \right) ^2 , \nonumber \\ \mathbf{E}_{g}\left[ \left( \mathsf {Q}^i_Z\right) ^2\right]&= \underbrace{\mathbf{E}_{g}\left[ \frac{f^2}{g^2} \left( 1 - \lambda \right) ^2 Z^2\right] }_{(1)} + \underbrace{\mathbf{E}_{g}\left[ \frac{f^2}{g^2} \frac{\lambda ^2}{\alpha ^2} \left[ Z - u_{Z}\right] _+^2 \right] }_{(2)} + \underbrace{\mathbf{E}_{g}\left[ \frac{f^2}{g^2} \lambda ^2 u_{Z}^2 \right] }_{(3)} \nonumber \\&\quad + \underbrace{2 \mathbf{E}_{g}\left[ \frac{f^2}{g^2} \left( 1 - \lambda \right) \frac{\lambda }{\alpha } Z \left[ Z - u_{Z}\right] _+\right] }_{(4)}\nonumber \\&\quad + \underbrace{2 \mathbf{E}_{g}\left[ \frac{f^2}{g^2} \left( 1 - \lambda \right) \lambda u_{Z}Z \right] }_{(5)} + \underbrace{2 \mathbf{E}_{g}\left[ \frac{f^2}{g^2} \frac{\lambda }{\alpha } \lambda u_{Z}\left[ Z - u_{Z}\right] _+\right] }_{(6)} . \end{aligned}$$
(25)

Next, we calculate all terms from the previous equations:

$$\begin{aligned} \mathbf{E}_{g}\left[ \left( \mathsf {Q}^i_Z\right) \right]&= \int _{ u_{Z}}^{\infty } (z- u_{Z}) \frac{\alpha }{\beta } \frac{\lambda }{\alpha } \frac{\beta }{\alpha } \phi (z) \, d z+ \lambda u_{Z}\\&= \frac{\lambda }{\alpha } \left( \int _{ u_{Z}}^{\infty } z\phi (z) \, d z- u_{Z}\int _{ u_{Z}}^{\infty } \phi (z) \, d z\right) + \lambda u_{Z}\\&= \frac{\lambda }{\alpha } \left( \phi (u_{Z}) - u_{Z}(1 - \varPhi (u_{Z}) \right) + \lambda u_{Z}= \frac{\lambda }{\alpha } \phi (u_{Z}) . \end{aligned}$$
$$\begin{aligned} (1)\quad \mathbf{E}_{g}\left[ \frac{f^2}{g^2} \left( 1 - \lambda \right) ^2 Z^2\right]&= \int _{-\infty }^{ u_{Z}} z^2 \frac{(1-\alpha )^2}{(1-\beta )^2} \left( 1 - \lambda \right) ^2 \frac{1-\beta }{1-\alpha } \phi (z) \, d z\\&\quad + \int _{ u_{Z}}^{\infty } z^2 \frac{\alpha ^2}{\beta ^2}\left( 1 - \lambda \right) ^2 \frac{\beta }{\alpha }\phi (z) \, d z\\&= \frac{1-\alpha }{1-\beta } \left( 1 - \lambda \right) ^2 \left( \varPhi ( u_{Z}) - u_{Z}\phi (u_{Z})\right) \\&\quad + \frac{\alpha }{\beta }\left( 1 - \lambda \right) ^2 \left( 1 - \varPhi ( u_{Z}) +u_{Z}\phi (u_{Z})\right) \\&= \frac{1-\alpha }{1-\beta } \left( 1 - \lambda \right) ^2 \left( 1 - \alpha - u_{Z}\phi (u_{Z})\right) \\&\quad + \frac{\alpha }{\beta }\left( 1 - \lambda \right) ^2 \left( \alpha + u_{Z}\phi (u_{Z})\right) , \end{aligned}$$
$$\begin{aligned} (2)\quad \mathbf{E}_{g}\left[ \frac{f^2}{g^2} \frac{\lambda ^2}{\alpha ^2} \left[ Z - u_{Z}\right] _+^2 \right]&= \int _{ u_{Z}}^{\infty } (z- u_{Z})^2 \frac{\alpha ^2}{\beta ^2} \frac{\lambda ^2}{\alpha ^2}\phi (z) \frac{\beta }{\alpha }\, d z\\&= \frac{\alpha }{\beta } \frac{\lambda ^2}{\alpha ^2} \int _{ u_{Z}}^{\infty } (z^2 - 2 u_{Z}z+ u_{Z}^2) \phi (z) \, d z\\&= \frac{1}{\beta }\frac{\lambda ^2}{\alpha } \left( 1 - \varPhi (u_{Z}) + u_{Z}\phi (u_{Z})\right. \\&\quad \left. - 2 u_{Z}\phi (u_{Z}) + u_{Z}^2 - u_{Z}^2\varPhi (u_{Z}) \right) \\&= \frac{\lambda ^2}{\alpha \beta } \left( \alpha - u_{Z}\phi (u_{Z}) + u_{Z}^2 \alpha \right) , \end{aligned}$$
$$\begin{aligned} (3)\quad \mathbf{E}_{g}\left[ \frac{f^2}{g^2} \lambda ^2 u_{Z}^2 \right]&= \int _{-\infty }^{ u_{Z}} \frac{(1-\alpha )^2}{(1-\beta )^2} \lambda ^2 u_{Z}^2 \frac{1-\beta }{1-\alpha } \phi (z) \, d z+ \int _{ u_{Z}}^{\infty } \frac{\alpha ^2}{\beta ^2} \lambda ^2 u_{Z}^2 \frac{\beta }{\alpha }\phi (z) \, d z\\&= \frac{1-\alpha }{1-\beta } (1-\alpha ) \lambda ^2 u_{Z}^2 + \frac{\alpha }{\beta } \alpha \lambda ^2 u_{Z}^2\\&= \lambda ^2 u_{Z}^2 \left( \frac{(1-\alpha )^2}{1-\beta } +\frac{\alpha ^2}{\beta } \right) , \end{aligned}$$
$$\begin{aligned} (4)\quad 2 \mathbf{E}_{g}\left[ \frac{f^2}{g^2} \left( 1 - \lambda \right) \frac{\lambda }{\alpha } Z \left[ Z - u_{Z}\right] _+\right]&= 2 \int _{ u_{Z}}^{\infty } z(z- u_{Z}) \frac{\alpha ^2}{\beta ^2} \left( 1 - \lambda \right) \frac{\lambda }{\alpha } \frac{\beta }{\alpha } \phi (z) d z\\&= 2\frac{\alpha }{\beta } \left( 1 - \lambda \right) \frac{\lambda }{\alpha } \left( \int _{ u_{Z}}^{\infty } \!z^2 \phi (z) d z- u_{Z}\int _{ u_{Z}}^{\infty } z\phi (z) d z\!\right) \\&= 2 \frac{\lambda (1-\lambda )}{\beta } \left( 1 - \varPhi (u_{Z}) + u_{Z}\phi (u_{Z}) - u_{Z}\phi (u_{Z}) \right) \\&= 2 \frac{\lambda (1-\lambda )\alpha }{\beta } ,\end{aligned}$$
$$\begin{aligned} (5)\quad 2 \mathbf{E}_{g}\left[ \frac{f^2}{g^2} \left( 1 - \lambda \right) \lambda u_{Z}Z \right]&= 2\int _{-\infty }^{ u_{Z}} \frac{(1-\alpha )^2}{(1-\beta )^2} \lambda u_{Z}\left( 1 - \lambda \right) \frac{1-\beta }{1-\alpha } z\phi (z) d z\\&+ 2\int _{ u_{Z}}^{\infty } \frac{\alpha ^2}{\beta ^2} \lambda u_{Z}\left( 1 - \lambda \right) \frac{\beta }{\alpha } z\phi (z) d z\\&= 2\lambda u_{Z}\left( 1 - \lambda \right) \phi (u_{Z}) \left( \frac{\alpha }{\beta } - \frac{1-\alpha }{1-\beta } \right) , \end{aligned}$$
$$\begin{aligned} (6)\quad 2 \mathbf{E}_{g}\left[ \frac{f^2}{g^2} \frac{\lambda }{\alpha } \lambda u_{Z}\left[ Z - u_{Z}\right] _+\right]&= 2 \int _{ u_{Z}}^{\infty } \frac{\alpha ^2}{\beta ^2} \frac{\lambda }{\alpha } \lambda u_{Z}\frac{\beta }{\alpha } (z- u_{Z}) \phi (z) d z\\&= 2 \frac{\alpha }{\beta } \frac{\lambda }{\alpha } \lambda u_{Z}\left( \phi (u_{Z}) - \alpha u_{Z}\right) \\&= 2 \frac{\lambda ^2}{\beta } u_{Z}\left( \phi (u_{Z}) - \alpha u_{Z}\right) , \end{aligned}$$

Substituting all above terms into (25):

$$\begin{aligned} \mathrm{var }_{g}\left[ \mathsf {Q}^i_Z\right]&= \frac{1-\alpha }{1-\beta } \left( 1 - \lambda \right) ^2 \left( 1 - \alpha - u_{Z}\phi (u_{Z})\right) + \frac{\alpha }{\beta }\left( 1 - \lambda \right) ^2 \left( \alpha + u_{Z}\phi (u_{Z})\right) \nonumber \\&\quad + \frac{\lambda ^2}{\alpha \beta } \left( \alpha - u_{Z}\phi (u_{Z}) + u_{Z}^2 \alpha \right) + \lambda ^2 u_{Z}^2 \left( \frac{(1-\alpha )^2}{1-\beta } +\frac{\alpha ^2}{\beta } \right) \nonumber \\&\quad + 2 \frac{\lambda (1-\lambda )\alpha }{\beta }\nonumber \\&\quad + 2\lambda u_{Z}\left( 1 - \lambda \right) \phi (u_{Z}) \left( \frac{\alpha }{\beta } - \frac{1-\alpha }{1-\beta } \right) \nonumber \\&\quad + 2 \frac{\lambda ^2}{\beta } u_{Z}\left( \phi (u_{Z}) - \alpha u_{Z}\right) - \left( \frac{\lambda }{\alpha } \phi (u_{Z}) \right) ^2 . \end{aligned}$$
(26)

Consider a random variable \(X = Z+ \frac{\mu }{\sigma }\). Then by using Proposition 1:

$$\begin{aligned} \mathrm{var }_{g_X}\left[ \mathsf {Q}^i_X\right] = \mathrm{var }_{g_Z}\left[ \mathsf {Q}^i_Z\right] + 2\frac{\mu }{\sigma } \mathbf{E}_{}\left[ \left( \frac{f_Z}{g_Z} - 1 \right) \mathsf {Q}^s_Z\right] + \frac{\mu ^2}{\sigma ^2} \frac{(\alpha -\beta )^2}{\beta (1-\beta )} \end{aligned}$$
(27)

Since \(\mathbf{E}_{}\left[ Z\right] = 0\):

$$\begin{aligned}&\!\!\!\mathbf{E}_{}\left[ \left( \frac{f_Z}{g_Z} - 1 \right) \mathsf {Q}^s_Z\right] = \mathbf{E}_{}\left[ \left( \frac{f_Z}{g_Z} - 1 \right) \left( \left( 1 - \lambda \right) Z+ \lambda \left( u_{Z}+ \frac{1}{\alpha } \left[ Z- u_{Z}\right] _+ \right) \right) \right] \nonumber \\&\quad = \left( 1 - \lambda \right) \underbrace{\mathbf{E}_{}\left[ \frac{f_Z}{g_Z}Z\right] }_{(1)} + \lambda u_{Z}\underbrace{\mathbf{E}_{}\left[ \left( \frac{f_Z}{g_Z} -1\right) \right] }_{(2)} + \frac{\lambda }{\alpha } \underbrace{\mathbf{E}_{}\left[ \left( \frac{f_Z}{g_Z} -1\right) \left[ Z- u_{Z}\right] _+ \right] }_{(3)}\qquad \\&\quad (1) \qquad \mathbf{E}_{}\left[ \frac{f_Z}{g_Z}Z\right] = \int _{-\infty }^{ u_Z} \frac{1-\alpha }{1-\beta } z\phi (z) d z+ \int _{ u_Z}^{\infty } \frac{\alpha }{\beta } z\phi (z) d z= \phi (u_Z)\left( \frac{\alpha }{\beta } - \frac{1-\alpha }{1-\beta } \right) ,\nonumber \\&\quad (2) \qquad \mathbf{E}_{}\left[ \left( \frac{f_Z}{g_Z} -1\right) \right] = \frac{(1-\alpha )^2}{1-\beta } + \frac{\alpha ^2}{\beta } - 1 = \frac{(\alpha -\beta )^2}{\beta (1-\beta )} ,\nonumber \\&\quad (3) \qquad \mathbf{E}_{}\left[ \left( \frac{f_Z}{g_Z} -1\right) \left[ Z- u_{Z}\right] _+ \right] =\int _{ u_Z}^{\infty } \left( \frac{\alpha }{\beta } -1 \right) \left( z- u_Z\right) \phi (z) d z\nonumber \\&\quad \qquad \qquad \qquad \quad \qquad \qquad \qquad \quad \qquad \qquad \, =\left( \frac{\alpha }{\beta } -1 \right) \left( \phi (u_Z) - \alpha u_Z\right) ,\nonumber \end{aligned}$$
(28)

Substituting all above terms into (28):

$$\begin{aligned} \mathbf{E}_{}\left[ \left( \frac{f_Z}{g_Z} - 1 \right) \mathsf {Q}^s_Z\right]&= \left( 1 - \lambda \right) \phi (u_Z)\left( \frac{\alpha }{\beta } - \frac{1-\alpha }{1-\beta } \right) \nonumber \\&\quad + \lambda u_{Z}\frac{(\alpha -\beta )^2}{\beta (1-\beta )} + \frac{\lambda }{\alpha } \left( \frac{\alpha }{\beta } -1 \right) \left( \phi (u_Z) - \alpha u_Z\right) \end{aligned}$$
(29)

Combining (26), (27) and (29):

$$\begin{aligned} \mathrm{var }_{g_X}\left[ \mathsf {Q}^i_X\right]&= \frac{1-\alpha }{1-\beta } \left( 1 - \lambda \right) ^2 \left( 1 - \alpha - u_{Z}\phi (u_{Z})\right) + \frac{\alpha }{\beta }\left( 1 - \lambda \right) ^2 \left( \alpha + u_{Z}\phi (u_{Z})\right) \nonumber \\&\quad + \frac{\lambda ^2}{\alpha \beta } \left( \alpha - u_{Z}\phi (u_{Z}) + u_{Z}^2 \alpha \right) \nonumber \\&\quad + \lambda ^2 u_{Z}^2 \left( \frac{(1-\alpha )^2}{1-\beta } +\frac{\alpha ^2}{\beta } \right) + 2 \frac{\lambda (1-\lambda )\alpha }{\beta }\nonumber \\&\quad + 2\lambda u_{Z}\left( 1 - \lambda \right) \phi (u_{Z}) \left( \frac{\alpha }{\beta } - \frac{1-\alpha }{1-\beta } \right) \nonumber \\&\quad + 2 \frac{\lambda ^2}{\beta } u_{Z}\left( \phi (u_{Z}) - \alpha u_{Z}\right) - \left( \frac{\lambda }{\alpha } \phi (u_{Z}) \right) ^2 \nonumber \\&\quad + 2\frac{\mu }{\sigma } \left( 1 - \lambda \right) \phi (u_Z)\left( \frac{\alpha }{\beta } - \frac{1-\alpha }{1-\beta } \right) \nonumber \\&\quad + 2\frac{\mu }{\sigma } \frac{\lambda }{\alpha } \left( \frac{\alpha }{\beta } -1 \right) \left( \phi (u_Z) - \alpha u_Z\right) \nonumber \\&\quad + \frac{(\alpha -\beta )^2}{\beta (1-\beta )} \left( 2\frac{\mu }{\sigma } \lambda u_{Z}+ \frac{\mu ^2}{\sigma ^2} \right) . \end{aligned}$$
(30)

We want to find a minimum with respect to the parameter \(\beta \) and therefore we differentiate:

$$\begin{aligned} \frac{\partial }{\partial \beta } \left( \mathrm{var }_{g_X}\left[ \mathsf {Q}^i_X\right] \right)&= \frac{1-\alpha }{(1-\beta )^2} \left( 1 - \lambda \right) ^2 \left( 1 - \alpha - u_{Z}\phi (u_{Z})\right) \nonumber \\&\quad - \frac{\alpha }{\beta ^2}\left( 1 - \lambda \right) ^2 \left( \alpha + u_{Z}\phi (u_{Z})\right) \\&\quad - \frac{\lambda ^2}{\alpha \beta ^2} \left( \alpha - u_{Z}\phi (u_{Z}) + u_{Z}^2 \alpha \right) + \lambda ^2 u_{Z}^2 \left( \frac{(1-\alpha )^2}{(1-\beta )^2} -\frac{\alpha ^2}{\beta ^2} \right) \\&\quad - 2 \frac{\lambda (1-\lambda )\alpha }{\beta ^2} + 2\lambda u_{Z}\left( 1 - \lambda \right) \phi (u_{Z}) \left( - \frac{\alpha }{\beta ^2} - \frac{1-\alpha }{(1-\beta )^2} \right) \\&\quad - 2 \frac{\lambda ^2}{\beta ^2} u_{Z}\left( \phi (u_{Z}) - \alpha u_{Z}\right) \\&\quad + 2\frac{\mu }{\sigma } \left( 1 - \lambda \right) \phi (u_Z) \left( - \frac{\alpha }{\beta ^2} - \frac{1-\alpha }{(1-\beta )^2} \right) \nonumber \\&\quad - 2\frac{\mu }{\sigma } \frac{\lambda }{\beta ^2} \left( \phi (u_Z) - \alpha u_Z\right) \\&\quad + \frac{2\alpha ^2\beta -2\alpha \beta ^2-\alpha ^2+\beta ^2}{\beta ^2(1-\beta )^2} \left( 2\frac{\mu }{\sigma } \lambda u_{Z}+ \frac{\mu ^2}{\sigma ^2} \right) . \end{aligned}$$

By Proposition 2 we have that for \(Y = \sigma X\): \(Y \sim \mathbf{N}(\mu ,\sigma ^2)\) & \(\mathrm{var }_{g_Y}\left[ \mathsf {Q}^i_Y\right] = \sigma ^2 \mathrm{var }_{g_X}\left[ \mathsf {Q}^i_X\right] \). Therefore by solving the quadratic equation

$$\begin{aligned} \frac{\partial }{\partial \beta } \left( \mathrm{var }_{g_X}\left[ \mathsf {Q}^i_X\right] \right) = 0 , \end{aligned}$$

we obtain the optimal selection of \(\beta \) with respect to the choice of parameters \(\alpha \) and \(\lambda \) under the assumption of normal distribution. \(\square \)

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Kozmík, V. On variance reduction of mean-CVaR Monte Carlo estimators. Comput Manag Sci 12, 221–242 (2015). https://doi.org/10.1007/s10287-014-0225-7

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1007/s10287-014-0225-7

Keywords

Mathematics Subject Classification

Navigation