This thesis proposes an optimization formulation to ensure accuracy and stability in the local volatility function calibration. The unknown local volatility function is represented by kernel splines. The proposed optimization formulation minimizes calibration error and an L1 norm of the vector of coefficients for the kernel splines. The L1 norm regularization forces some coefficients to be zero at the termination of optimization. The complexity of local volatility function model is determined by the number of nonzero coefficients. Thus by using a regularization parameter, the proposed formulation balances the calibration accuracy with the model complexity. In the context of the support vector regression for function based on finite observations, this corresponds to balance the generalization error with the number of support vectors. In this thesis we also propose a trust region method to determine the coefficient vector in the proposed optimization formulation. In this algorithm, the main computation of each iteration is reduced to solving a standard trust region subproblem. To deal with the non-differentiable L1 norm in the formulation, a line search technique which allows crossing nondifferentiable hyperplanes is introduced to find the minimum objective value along a direction within a trust region. With the trust region algorithm, we numerically illustrate the ability of proposed approach to reconstruct the local volatility in a synthetic local volatility market. Based on S&P 500 market index option data, we demonstrate that the calibrated local volatility surface is smooth and resembles in shape the observed implied volatility surface. Stability is illustrated by considering calibration using market option data from nearby dates.

This thesis proposes an optimization formulation to ensure accuracy and stability in the local volatility function calibration. The unknown local volatility function is represented by kernel splines. The proposed optimization formulation minimizes calibration error and an L1 norm of the vector of coefficients for the kernel splines. The L1 norm regularization forces some coefficients to be zero at the termination of optimization. The complexity of local volatility function model is determined by the number of nonzero coefficients. Thus by using a regularization parameter, the proposed formulation balances the calibration accuracy with the model complexity. In the context of the support vector regression for function based on finite observations, this corresponds to balance the generalization error with the number of support vectors. In this thesis we also propose a trust region method to determine the coefficient vector in the proposed optimization formulation. In this algorithm, the main computation of each iteration is reduced to solving a standard trust region subproblem. To deal with the non-differentiable L1 norm in the formulation, a line search technique which allows crossing nondifferentiable hyperplanes is introduced to find the minimum objective value along a direction within a trust region. With the trust region algorithm, we numerically illustrate the ability of proposed approach to reconstruct the local volatility in a synthetic local volatility market. Based on S&P 500 market index option data, we demonstrate that the calibrated local volatility surface is smooth and resembles in shape the observed implied volatility surface. Stability is illustrated by considering calibration using market option data from nearby dates.