The problem of optimally positioning sensors in lumped and distributed parameter dynamic systems for the purpose of system identification from time-domain input-output data is formulated and a methodology for its solution is presented. A linear relation between small perturbations in a finite-dimensional representation of the system parameters and a finite sample of observations of the system time response is used to determine approximately the covariance of the parameter estimates. The locations of a given number of sensors are then determined such that a suitable norm of the covariance matrix is minimized. The methodology is applied to the problem of optimally locating a single sensor in a building structure modeled by a shear beam, such that the estimates of the stiffness distributions, obtained from the records of strong ground shaking and the building response at the sensor location, are least uncertain.