A data-based procedure is presented to develop spectra-compatible time-history records that are based on the dominant probabilistic features of an ensemble of records corresponding to the non-stationary stochastic phenomena of interest (e.g., earthquakes, wind loads, etc). The method requires a statistically significant collection of time-history records that are used to construct the associated covariance kernel of the random process. Subsequently, orthogonal decomposition approaches are used to determine the dominant eigenvectors of the covariance matrix, and these vectors are then linearly combined, with an adjustable amplitude-scale and phase-shift, to determine, via a nonlinear optimization scheme (employing a combination of stochastic and deterministic approaches), a time-history record that matches the target spectrum within a specified error bound. The utility of this approach is demonstrated with several collections of earthquake records from different regions of the world (Japan, Los Angeles, and San Francisco) that are then used to match various spectra widely used in seismic design applications. Issues that impact the selection of the bases vectors to construct the optimum spectra-matching record are discussed, and guidelines are provided for successful implementation of the proposed methodology. [ABSTRACT FROM AUTHOR]