To solve the calibration problem numerically, we propose a multigrid technique that couples the calibration process with the model solver. For well-behaved problems, we find that the calibration problem can be solved for about three times the cost of solving the control problem with a fixed set of parameters. Computational evidence suggests that this holds independent of the number of parameters to be calibrated. In short: If you can solve a model numerically, it is within your computational budget to calibrate as well, provided it can be calibrated.
We illustrate this technique and its limitations with a series of examples. One example is a recent model from financial economics. This has only one parameter to calibrate but requires careful formulation to realize good numerical results. Another example we consider is a classical optimal stopping problem in two-dimensions. In this problem, the location of the stopping boundary is viewed as an infinite-dimensional parameter that must be chosen to satisfy the smooth pasting and value matching conditions. Fourier analysis of the pseudo-differential operators implicit in this problem show that the natural formulation is ill-posed for numerical purposes. This analysis suggests a reformulation to regularize the problem without introducing distortions and a process for smoothing errors. This is used to construct a fast multigrid solver.