Statistical Romberg extrapolation: a new variance reduction method and applications to option pricing (Q2496505): Difference between revisions

The main subject of the paper is studying the expectation \({\mathbb E}f(X_{T}),\) where \(f\) is a function with given properties and \(X_{T}\) is a solution to a stochastic differential equation. The Euler scheme with given steps is used for approximation of the solution to the considered differential equation. The classical Monte Carlo method (MCM) is used for approximation of the quantity \({\mathbb E}f(X_{T}).\) Theorem 3.1 shows that for obtaining a total error of order \({1 \over n^{\alpha}},\) for some \(\alpha \in [{1 \over 2}, 1],\) the minimal computation effort necessary to run the MCM is obtained for \(N = n^{2 \alpha}.\) The optimal complexity of MCM \({\mathcal O}( n^{2 \alpha + 1})\) is given. It is well-known that the rate of convergence of MCM depends on the variance of the function \(f.\) The author introduces a new variance reduction method, which can be viewed as a statistical analogue of Romberg extrapolation method. The computing algorithm of this method for approximation of \({\mathbb E}f(X_{T})\) is given. Theorem 3.2 shows that for obtaining a total error of order \({1 \over n^{\alpha}},\) for some \(\alpha \in [{1 \over 2}, 1],\) the minimal computational effort necessary to run the statistical Romberg method (SRM) applied to the Euler scheme with step \({T \over n^{\beta}},\) \(\beta \in (0,1),\) is obtained for \(N_{m} = n^{2 \alpha}\) and \(N_{n} = n^{2 \alpha - \beta}.\) The time complexity of SRM \({\mathcal O} \left( n^{\beta + 2 \alpha} + (n + n^{\beta}) n^{2 \alpha - \beta}\right)\) is obtained. The optimal choice of the parameter \(\beta\) \((\beta = {1 \over 2})\) minimizing the time complexity of SRM is shown. This permits the corresponding optimal parameters \(m = n^{1 \over 2},\) \(N_{m} = n^{2 \alpha}\) and \(N_{n}= n^{2 \alpha - {1 \over 2}}\) to be given. The optimal complexity of SRM \({\mathcal O}( n^{2 \alpha + {1 \over 2}})\) is obtained. The order of the complexity of SRM is smaller than the order of the complexity of MCM, which is the advantage of SRM. An application of SRM to the theory of Asian options is given. The trapezoidal scheme is used for approximating the integral of the asset price process. Theorem 4.1 studies the convergence of the process of trapezoidal discretization. Theorem 4.2 gives that for all \(\beta \in (0,1)\) for obtaining a total error of order \({1 \over n}\) the minimal computational effort necessary to run the SRM applied to the trapezoidal scheme with step numbers \(n\) and \(m = n^{\beta}\) is obtained for \(N_{m} = n^{2}\) and \(N_{n} = n^{2 - 2 \beta}.\) In this case the time complexity of SRM \({\mathcal O} \left( n^{\beta + 2} + (n + n^{\beta}) n^{2 - 2 \beta}\right)\) is given. The optimal choice of the parameter \(\beta\) which minimizes the time complexity of SRM applied to the trapezoidal scheme is shown. In this case the optimal complexity \({\mathcal O}(n^{7 \over 3})\) of SRM is obtained. For the same error of order \({1 \over n}\) the optimal complexity of MCM applied to the trapezoidal scheme with step numbers \(n\) is \({\mathcal O}(n^{3}).\) In some concrete examples concerning Asian options the efficiency of SRM is tested. The paper ends with a conclusion, where the priority of SRM over MCM is considered.