Abstract
The estimation of enterprise risk for financial institutions entails a re-evaluation of the company’s economic balance sheet at a future time for a (large) number of stochastic scenarios. The current paper discusses tackling this nested valuation problem based on least-squares Monte Carlo techniques familiar from American option pricing. We formalise the algorithm in an operator setting and discuss the choice of the regressors (“basis functions”). In particular, we show that the left singular functions of the corresponding conditional expectation operator present robust basis functions. Our numerical examples demonstrate that the algorithm can produce accurate results at relatively low computational costs.
Similar content being viewed by others
References
Abbé, E., Zheng, L.: A coordinate system for Gaussian networks. IEEE Trans. Inf. Theory 58, 721–733 (2012)
Alt, H.W.: Linear Functional Analysis. Springer, Berlin (2012)
Amemiya, T.: Advanced Econometrics. Harvard University Press, Cambridge (1985)
Bauer, D., Bergmann, D., Reuss, A.: Solvency II and Nested Simulations – a Least-Squares Monte Carlo Approach. Paper presented at the ICA 2010, (2010). http://www.actuaries.org/EVENTS/Congresses/Cape_Town/Papers/Life
Bauer, D., Reuss, A., Singer, D.: On the calculation of the solvency capital requirement based on nested simulations. ASTIN Bull. 42, 453–499 (2012)
Belloni, A., Chernozhukov, V., Chetverikov, D., Kato, K.: Some new asymptotic theory for least squares series: pointwise and uniform results. J. Econom. 186, 345–366 (2015)
Benedetti, G.: On the calculation of risk measures using least-squares Monte Carlo. Int. J. Theor. Appl. Finance 20, 1–14 (2017)
Birman, M.S., Solomyak, M.Z.: Estimates of singular numbers of integral operators. Russ. Math. Surv. 32, 15–89 (1977)
Breiman, L., Friedman, J.H.: Estimating optimal transformations for multiple regression and correlation. J. Am. Stat. Assoc. 80, 580–598 (1985)
Brigo, D., Mercurio, F.: Interest Rate Models – Theory and Practice with Smile, Inflation and Credit, 2nd edn. Springer, Berlin (2006)
Broadie, M., Du, Y., Moallemi, C.: Risk estimation via regression. Oper. Res. 63, 1077–1097 (2015)
Cambou, M., Filipović, D.: Replicating portfolio approach to capital calculation. Finance Stoch. 22, 181–203 (2018)
Carrasco, M., Florens, J.: Spectral method for deconvolving a density. Econom. Theory 27, 546–581 (2011)
Carriere, J.F.: Valuation of the early-exercise price for options using simulations and nonparametric regression. Insur. Math. Econ. 19, 19–30 (1996)
Clément, E., Lamberton, D., Protter, P.: An analysis of a least squares regression method for American option pricing. Finance Stoch. 6, 449–471 (2002)
Cohn, D.L.: Measure Theory, 2nd edn. Birkhäuser, Basel (2013)
Floryszczak, A., Le Courtois, O., Majri, M.: Inside the Solvency 2 black box: Net asset values and solvency capital requirements with a least-squares Monte-Carlo Approach. Insur. Math. Econ. 71, 15–26 (2016)
Glasserman, P.: Monte Carlo Methods in Financial Engineering. Springer, New York (2004)
Glasserman, P., Yu, B.: Simulation for American options: regression now or regression later? In: Niederreiter, H. (ed.) Monte Carlo and Quasi-Monte Carlo Methods 2002, pp. 213–226. Springer, Berlin (2004)
Gockenbach, M.: Linear Inverse Problems and Tikhonov Regularization. (2016). MAA Press
Golub, G.H., Van Loan, C.F.: Matrix Computations, 4th edn. Johns Hopkins University Press, Baltimore (2013)
Gordy, M.B., Juneja, S.: Nested simulations in portfolio risk measurement. Manag. Sci. 56, 1833–1848 (2010)
Ha, H.: Essays on Computational Problems in Insurance. PhD Thesis, Georgia State University (2016). Available online at https://scholarworks.gsu.edu/rmi_diss/40/
Hastie, T., Tibshirani, R., Friedman, J.: The Elements of Statistical Learning – Data Mining, Inference, and Prediction. Springer, New York (2009)
Hong, L.J., Juneja, S., Liu, G.: Kernel smoothing for nested estimation with application to portfolio risk measurement. Oper. Res. 65, 657–673 (2017)
Huang, Q.: Some Topics Concerning the Singular Value Decomposition and Generalized Singular Value Decomposition. PhD Thesis, ASU (2012). Available online at https://keep.lib.asu.edu/items/151128
Kaina, M., Rüschendorf, L.: On convex risk measures on \(L^{p}\)-spaces. Math. Methods Oper. Res. 69, 475–495 (2009)
Khare, K., Zhou, H.: Rate of convergence of some multivariate Markov chains with polynomial eigenfunctions. Ann. Appl. Probab. 19, 737–777 (2009)
Krah, A.S., Nikolić, Z., Korn, R.: A least-squares Monte Carlo framework in proxy modeling of life insurance companies. Risks 6, 62–87 (2018)
Liu, M., Staum, J.: Stochastic kriging for efficient nested simulation of expected shortfall. J. Risk 12, 3–27 (2010)
Longstaff, F.A., Schwartz, E.S.: Valuing American options by simulation: a simple least-squares approach. Rev. Financ. Stud. 14, 113–147 (2001)
Makur, A., Zheng, L.: Polynomial spectral decomposition of conditional expectation operators. In: Proceedings of the 54th Annual Allerton Conference on Communication, Control, and Computing, pp. 633–640. IEEE, Monticello, IL (2016)
Moreno, M., Navas, J.F.: On the robustness of least-squares Monte Carlo (LSM) for pricing American derivatives. Rev. Deriv. Res. 14, 113–147 (2003)
Newey, W.K.: Convergence rates and asymptotic normality for series estimators. J. Econom. 79, 147–168 (1997)
Nikolić, Z., Jonen, C., Zhu, C.: Robust regression technique in LSMC proxy modelling. Aktuar 01(2017), 8–16 (2017)
Patarroyo, K.: A digression on Hermite polynomials (2019). Preprint (2019). Available online at https://arxiv.org/pdf/1901.01648.pdf
Pelsser, A., Schweizer, J.: The difference between LSMC and replicating portfolio in insurance liability modelling. Eur. Actuar. J. 6, 441–494 (2016)
Retherford, J.: Hilbert Space: Compact Operators and the Trace Theorem. Cambridge University Press, Cambridge (1993)
Risk, J., Ludkovski, M.: Sequential design and spatial modelling for portfolio tail risk measurement. SIAM J. Financ. Math. 9, 1137–1174 (2018)
Stentoft, L.: Convergence of the least squares Monte Carlo approach to American option valuation. Manag. Sci. 50, 1193–1203 (2004)
Tsitsiklis, J.N., Van Roy, B.: Regression methods for pricing complex American-style options. IEEE Trans. Neural Netw. 12, 694–763 (2001)
Vinga, S.: Rényi continuous entropy of DNA sequences. J. Theor. Biol. 231, 377–388 (2004). Supplementary material, available at http://web.ist.utl.pt/susanavinga/renyi/, to Vinga, S., Almeida, J.S., 2004
Weber, S.: Distribution-invariant risk measures, entropy, and large deviations. J. Appl. Probab. 44, 16–40 (2007)
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This paper extends the earlier working paper Bauer et al. [4]. We thank Martin Schweizer (the Editor), an anonymous associate editor, two anonymous referees, Giuseppe Benedetti, Enrico Biffis, René Carmona, Patrick Cheridito, Matthias Fahrenwaldt, Jean-Pierre Fouque, Andreas Reuss, Daniela Singer, Ajay Subramanian, Baozhong Yang and seminar participants at the Bachelier Congress 2014, the 2015 WRIEC, ETH Zurich, Georgia State, Michigan State, St. Joseph’s University, TU Munich, UCSB, UIUC and Université de Montréal for helpful comments.
Appendix: Proofs and technical material
Appendix: Proofs and technical material
Proof of Lemma 2.1
1) Let \(A \in \mathcal{F}_{t}\), \(0 \leq t \leq \tau \). Then, by “taking out what is known”,
2) Let \(X:\Omega \rightarrow \mathbb{R}\) be a random variable. Then similarly as above,
□
Proof of Lemma 2.3
Linearity is obvious. For continuity, take \(h^{(n)} \rightarrow h\). Then
□
Proof of Lemma 2.4
Let \(\mathcal{V}^{(t)}\), \(t=\tau ,\dots ,T\), be the component mappings of \(\mathcal{V}\) introduced in (2.3), such that \(\mathcal{V}^{(t)}: L^{2}(\mathbb{R}^{d},\mathcal{B}, \tilde{\mathcal{P}}_{Y_{t}}) \rightarrow L^{2}(\mathbb{R}^{d}, \mathcal{B},\tilde{\mathcal{P}}_{Y_{\tau}})\) is the conditional expectation. Under the assumption that there exists a joint density, \(\mathcal{V}^{(t)}\) can be represented as
where \(\mathbf{x} \in {L}^{2}(\mathbb{R}^{d},\mathcal{B},\tilde{\mathbb{P}}_{Y_{t}})\), \(\pi _{Y_{t}}(y)\) and \(\pi _{Y_{\tau}}(x)\) are the marginal density functions for \(Y_{t}\) and \(Y_{\tau}\) in \({L}^{2}(\mathbb{R}^{d},\mathcal{B},\tilde{\mathbb{P}}_{Y_{t}})\) and \({L}^{2}(\mathbb{R}^{d},\mathcal{B}, {\mathbb{P}}_{Y_{\tau}})\), and \(k(x,y)= \frac{\pi _{Y_{\tau},Y_{t}}(x,y)}{\pi _{Y_{t}}(y)\pi _{Y_{\tau}}(x)}\). Thus \(\mathcal{V}^{(t)}\) is an integral operator with kernel \(k(x,y)\), and
Thus \(\mathcal{V}^{(t)}\) is a Hilbert–Schmidt operator (e.g. Alt [2, Proposition 10.15]) and therefore compact. Finally, \(\mathcal{V}\) is the sum of the \(\mathcal{V}^{(t)}\) and therefore also compact. □
Proof of Proposition 3.1
\({\mathbb{P}}_{Y_{\tau}}\) is a finite Borel measure, and hence \({L}^{2}\left (\mathbb{R}^{d},\mathcal{B},{\mathbb{P}}_{Y_{\tau}} \right )\) is separable (see Cohn [16, Proposition 3.4.5]). Now if \(e_{k},\;k=1,\dots ,M\), are independent, we can find by Gram–Schmidt an orthonormal system \(S = \{f_{k}, k=1,\dots ,M\}\) with \({\mathrm{lin}}\{e_{k}, k=1,\dots ,M\} = {\mathrm{lin}}\, S\). For \(S\), in turn, we find an orthonormal basis \(\{f_{k}, k \in \mathbb{N}\} = S' \supseteq S\). Hence
where
by Parseval’s identity. For the second part, we note that
where \(e(\, \cdot \, ) = (e_{1}(\, \cdot \, ),\dots , e_{M}(\, \cdot \, ))'\) and
is invertible for large enough \(N\) since we assume linear independence. Hence
\(\tilde{\mathbb{P}}\)-a.s. by the law of large numbers, where \(A^{(M)} = ({\mathbb{E} }^{\tilde{\mathbb{P}}} [e_{k} (Y_{\tau} )\,e_{\ell}(Y_{\tau} ) ] )_{1\leq k,\ell \leq M}\). Thus
\(\tilde{\mathbb{P}}\)-a.s. Finally, for the third part, write
where \(\mathbb{E} [\epsilon _{j}|Y_{\tau} ]=0\), \({\mathrm{Var}}[\epsilon _{j}|Y_{ \tau} ]=\Sigma (Y_{\tau})\) and \({\mathrm{Cov}}[\epsilon _{i},\,\epsilon _{j}|Y_{\tau} ]=0\). Thus (see e.g. Amemiya [3, Chap. 1])
so that
where \( \xi ^{(M)}=e'\,\tilde{\xi}\,e \). □
Proof of Proposition 3.2
Since \((V_{\tau}^{(i)},Y_{\tau}^{(i)})\) are i.i.d. as Monte Carlo trials, the first part of Newey [34, Assumption 1] is automatically satisfied. The conditions in the proposition are then exactly Assumptions 1 (part 2), 2 and 3 in Newey [34] for \(d=0\). Thus the claim follows by the first part of Newey [34, Theorem 1]. □
Proof of Proposition 3.3
Analogously to the proof of Proposition 3.2, the first part of Newey [34, Assumption 1] is automatically satisfied. The conditions in the proposition are taken from the second part of Assumption 1, Assumption 8, the discussion following Assumption 8, and Assumption 9 in Newey [34]. Thus the claim follows by the first part of Newey [34, Theorem 4]. □
Proof of Proposition 4.2
We approximate \(\mathcal{V}\) by an arbitrary rank-\(M\) operator \(\mathcal{V}_{F}^{(M,e)}\), which can be written as \(\mathcal{V}_{F}^{(M,e)}=\sum _{k=1}^{M}\alpha _{k}\,\langle \,\cdot \,, u_{k}\rangle \,e_{k}\), where \((\alpha _{k})_{k=1}^{M} \subseteq \mathbb{R}_{+}\), \((u_{k})^{M}_{k=1}\) are orthonormal in ℋ and \((e_{k})_{k=1}^{M}\) are orthonormal in \({L}^{2} (\mathbb{R}^{d},\mathcal{B},{\mathbb{P}}_{Y_{\tau}} )\). Denote by \(\mathcal{V}_{F}^{(M,e^{*})}\) the operator we obtain when choosing \((\alpha _{k}, u_{k}, e_{k})=(\omega _{k}, s_{k}, \varphi _{k})\). Then
On the other hand, consider any alternative system \((\alpha _{k}, u_{k}, e_{k})\) for an arbitrary finite-rank operator \(\mathcal{V}_{F}^{(M,e)}\). Then choose a non-zero \(\mathbf{x}_{0}\) such that
Note that \(\mathcal{V}-\mathcal{V}_{F}^{(M,e)}\) is compact and bounded. Therefore
Hence
Now since \(\inf _{\mathcal{V}_{F}^{(M,e)}}\|\mathcal{V}-\mathcal{V}_{F}^{(M,e)} \|^{2}=\inf _{\{e_{1},\dots ,e_{M}\}}\|\mathcal{V}-P^{(M,e)} \mathcal{V}\|^{2}\), the claim follows by (4.1). □
Proof of Proposition 4.3
Proceeding as in (4.3) and with (4.1), we obtain
for a fixed \(\mathbf{x}\) since the \(\varphi _{k}\) are uniformly bounded, where the second inequality follows from the triangle and Cauchy–Schwarz inequalities.
Then, going through the assumptions of Proposition 3.2 with the choices \(B=I\) and \(e^{(M)}=(e_{1},\dots ,e_{M})'\), we obtain
due to the orthonormality of the singular functions. Therefore, the smallest eigenvalue is bounded away from zero uniformly for every \(M\). Moreover, for fixed \(y\in \mathcal{Y}\), \(|\tilde{e}^{(M)}(y)|=\sqrt{\varphi _{1}(y)^{2}+\cdots + \varphi _{M}(y)^{2}}\) so that
as the \(\varphi _{k}\) are uniformly bounded. Thus the claim follows by Proposition 3.2. □
Proof of Lemma 4.4
The assertions on the conditional distributions are standard. In order to show that \(\mathcal{V}\) is compact, we check that the transition and the reverse transition density functions satisfy the condition in Lemma 2.4. Note that the transition density function can be written as
We evaluate the integral
where
and \(C_{1}\) is an appropriate constant to obtain \(g(y; \mu _{T}, V)\) (see the results from Vinga [42] on the product of Gaussian densities). Therefore,
□
Proof of Lemma 4.5
The operator \(\mathcal{V}^{*}\) can be found via
where \((\mathcal{V}^{*}m)(y)=\int _{\mathbb{R}^{d}} m(x)\pi _{Y_{\tau}|Y_{T}}(x|y) \,dx\). We obtain for \(\mathcal{V}\mathcal{V}^{*}\) that
It is useful to express the reverse density as in the proof of Lemma 4.4 as
Hence we get
where we again rely on results from Vinga [42]. The formula for \(\mathcal{V}^{*}\mathcal{V}\) can be derived analogously. □
Proof of Lemma 4.6
We start by recalling the considerations from Khare and Zhou [28]. Let \((X_{t})\) with values in \(\mathbb{R}^{d}\) be an MAR(1) process satisfying the equation
where \(\Phi \in \mathbb{R}^{d\times d}\) and \((\eta _{t})_{t\ge 1}\) are i.i.d. \(\sim N(0, H)\). The process \((X_{t})\) has a unique stationary distribution \(N(0, \Sigma )\) if and only if \(H=\Sigma -\Phi \Sigma \Phi '\), and it is reversible if and only if \(\Phi \Sigma =\Sigma \Phi '\). [28] show that if these assumptions are satisfied, the transformed Markov operator for (7.1) has eigenvalues which are products of eigenvalues of \(\Phi \) and the corresponding eigenfunctions are products of Hermite polynomials.
Now note that for \(Y \sim K_{A}(x,\,\cdot \,) \), we can write
where \(\zeta _{A}\sim N(0, \Sigma _{A})\). Since from Lemma 4.5, we have \(\Sigma _{A}=\Sigma _{\tau}-A\Sigma _{\tau}A'\) and
the operator \(\mathcal{V}\,\mathcal{V}^{*}\) has for \(\Sigma = \Sigma _{\tau}\) the same structure as the Markov operator for (7.1) which is reversible and stationary.
Following [28], denote by \(\Sigma _{\tau}^{1/2}\) the square root matrix of \(\Sigma _{\tau}\). Then
is symmetric and thus orthogonally diagonalisable as
for an orthogonal matrix \(P\). Then
and the entries of the diagonal matrix \(\Lambda \) are the eigenvalues of \(A\).
Now for the transformation \(z_{P}(Y)\) defined in (4.4), we obtain
and
Moreover,
The second part follows analogously. □
Proof of Proposition 4.7
For fixed \(z_{i}^{P}(Y)\), from Carrasco and Florens [13], the univariate orthonormal Hermite polynomial of order \(n_{i}\) is an eigenfunction under \(K_{A}\), so that
Moreover, the products of these polynomials are also eigenfunctions since
The orthogonality of the eigenfunctions is proved in Khare and Zhou [28]. Note that the product of normalised Hermite polynomials is already normalised since
The right singular functions are obtained similarly from \(z_{i}^{Q}(X)\). □
Proof of (5.5)
We have \(B_{r}(t,T)=\frac{1-e^{-\alpha (T-t)}}{\alpha}\), \(B_{\mu}(t,T)=\frac{e^{\kappa (T-t)}-1}{\kappa}\) and
□
Proof of Lemma 5.1
Under ℙ, the solutions of (5.1)–(5.3) are
Thus \(Y_{\tau}= (q_{\tau},r_{\tau},\mu _{x+\tau})'\) is distributed according to \({\mathcal{N}}(\mu _{\tau},\Sigma _{\tau})\) with
To derive the distribution under \(\mathbb{Q}_{E}\), first note that for \(\tau \le s< T\),
so that \(\int _{\tau}^{T} r_{s}\,ds\) can be evaluated, using the stochastic Fubini theorem, as
Thus under \(\mathbb{Q}_{E}\) with known \(Y_{\tau}\), the solutions of (5.6)–(5.8) are
so that the conditional distribution of \(Y_{T}|Y_{\tau}\) is Gaussian with parameters
We can write the conditional mean of \(Y_{T}\) given \(Y_{\tau}\) in the affine form
where \(C_{\tau}\) is a constant matrix defined by the remaining terms of the mean vector of \(Y_{T}|Y_{\tau}\) after defining \(HY_{\tau}\). The unconditional distribution of \(Y_{T}\) under \(\tilde{\mathbb{P}}\) is also Gaussian since \(Y_{\tau}\) and \(Y_{T}|Y_{\tau}\) follow Gaussian distributions. Thus it suffices to specify the mean vector and the covariance matrix of \(Y_{T}\) under \(\tilde{\mathbb{P}}\) to specify its distribution, and we have
The final step is to specify the joint distribution of \(Y_{\tau}\) and \(Y_{T}\) by finding \({\mathrm{Cov}}(Y_{\tau}, Y_{T})\). Note that
□
Rights and permissions
About this article
Cite this article
Ha, H., Bauer, D. A least-squares Monte Carlo approach to the estimation of enterprise risk. Finance Stoch 26, 417–459 (2022). https://doi.org/10.1007/s00780-022-00478-7
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00780-022-00478-7