Numerical examples

Table of Links

Abstract and 1. Introduction

2. Bermudan option pricing and hedging
3. Sparse Hermite polynomial expansion and gradient
4. Algorithm and complexity
5. Convergence analysis
6. Numerical examples
7. Conclusions and outlook, Acknowledgments, and References

We consider the example of Bermudan geometric basket put from [13, 25]. The exact prices are computed by solving the reduced one-dimensional problem via a quadrature and interpolation-based method [25] for Bermudan options. We present in Table 2 the computed option prices and their relative errors using G-LSM and LSM, with the same ansatz space for the CVF. The results show that G-LSM achieves higher accuracy than LSM in high-dimensions: G-LSM has a relative error 0.55% for d = 15, which is almost ten times smaller than that by LSM. That is, by incorporating the gradient information, the accuracy of LSM can be substantially improved.

Table 3 gives the computed option prices and their relative errors by G-LSM with different maximum polynomial orders p for the 20-dimensional geometric basket put. The relative error decays steadily as the order p of Hermite polynomials increases, which agrees with Theorem 5.6.

6.2. Example 2: American geometric basket call. Now we consider the example of American geometric basket call option from [21, 6] to demonstrate that the proposed G-LSM can achieve the same level of accuracy as the DNN-based method [6], and take M = 720, 000 samples as in [6]. The prices and deltas are given in Table 4 and Table 5, respectively, where the results of [6] are the average of 9 independent runs. From Table 4, both methods have similar accuracy for the price. From Table 5, the relative error of delta using G-LSM and DNN varies slightly with the dimension d. This is probably because the DNN-based method computes the delta via a sample average, while G-LSM uses the derivatives of the value function directly. The relative error of the delta computed by G-LSM increases slightly for larger dimensions, possibly due to the small magnitude of the exact delta values.

Next, Figure 2 shows the classification results of continued and exercised data using GLSM and LSM with the number of simulated paths M = 100, 000 in d = 7 or 20. Compared with the exact exercise boundary, G-LSM achieves better accuracy in determining the exercise boundary than LSM despite of using the same ansatz and number of paths. Thus, even with the right ansatz space, LSM might fail the task of finding exercise boundary in high dimensions using only a limited number of samples. Compared with [6, Figure 5] and [18, Figure 6], Figure 2 demonstrates that G-LSM can detect accurate exercise boundary with fewer number of paths than the DNN-based method.

6.3. Example 3: Bermudan max-call with symmetric assets. To benchmark G-LSM on high-dimensional problems without exact solutions and to validate the complexity analysis in section 4, we test Bermudan max-call option and report the computing time. The computing time is calculated as follows. For a fixed time step, Tbas is the time for generating basis matrix Φ, Tmat is the time for assembling matrix A, Tlin is the time for solving linear system, and Tup is the time for updating values. The overall computing time is Ttot ≈ (N − 1)(Tbas + Tmat + Tlin + Tup).

Table 6 presents the prices and computing time (in seconds) for Bermudan max-call options with d symmetric assets. The reference 95% confidence interval (CI) is taken from [3]. The reference CI is computed with more than 3000 training steps and a batch of 8192 paths in each step, which in total utilizes more than 107 paths. The last five columns of the table report the computing time for the step 6, 7, 8, 9 in Algorithm 4.1, and the total time, respectively. All the computation for this example was performed on an Intel Core i9-10900 CPU 2.8 GHz desktop with 64GB DDR4 memory using MATLAB R2023b. It is observed that the prices computed by G-LSM fall into or stay very close to the reference 95% CI, confirming the high accuracy of G-LSM. Furthermore, the time for generating basis matrix, Tbas, dominates the

overall computing time. Hence, the cost mainly arises from evaluating Hermite polynomials on sampling paths, which is also required by LSM. In comparison with LSM , Tmat is the extra cost to incorporate the gradient information and takes only a small fraction of the total time. Therefore, G-LSM has nearly identical cost with LSM.

Figure 4 shows the classification of continued and exercised sample points computed by G-LSM and LSM in the example of two-dimensional max-call. G-LSM yields a smoother exercise boundary than LSM. Compared with the exercise boundary computed in literature [19, Figure 3], G-LSM exhibits higher accuracy, albeit that the same ansatz space for the CVF is employed.