Blind predictions of ground response at the Turkey Flat vertical array site during the 2004 Parkfield earthquake were performed using a series of nonlinear and equivalent-linear ground-response analysis codes. In this article, we report the outcome of the blind predictions, compare the predictions to measurements (released after the predictions were made), and provide discussion on residuals between the data and model predictions. The prediction exercise considered various sources of material variability and model variability. Specifically, these analyses utilized five different nonlinear analysis codes, each exercised with a suite of small-strain shear-wave velocity profiles and modulus reduction/damping curves that encompass the range of reported material properties. Below the site frequency of about 5 Hz, velocity profile variability dominates the computed response variability. Upon full release of strong-motion data from the site, the predictions were generally found to be biased towards underprediction at high frequencies. Sensitivity analyses and results of recent geophysical testing at the site suggest this underprediction bias is most likely a result of inaccuracies in prior estimates of near-surface shear-wave velocities. Analyses performed with the recent profile show no significant bias except at the site frequency, where the response is overpredicted.