The fatigue life of coil springs is usually predicted with a stationary Gaussian vibration load and deterministic structural parameters. However, the obtained fatigue life is inconsistent with the observed fatigue life of fractured springs which varies within a wide range. The work aims to propose a method to predict the fatigue life of the coil spring by considering the time-varying vibration load, i.e., root mean square (rms) varies with time and the uncertainties of geometric parameters. First, a synthetic method for time-varying vibration loads is proposed. The time-varying load is decomposed into multiple stationary Gaussian short samples represented by their power spectral density (PSD). These PSDs are synthesized according to the distribution characteristics of spectral values, in which data that are non-Gaussian are processed with the Johnson system. Second, the influence of parameter uncertainties in the coil spring is studied by a Monte Carlo analysis of the stress frequency response function. Finally, the fatigue life is calculated and compared with the results predicted by using the measured stress. The results show that the synthetic spectrum has almost the same damage potential as the measured time-varying load. In comparison with results predicted from the measured stress, the synthetic spectrum gives much better estimates of the fatigue life of the coil spring than the average spectrum. Parameter uncertainties of coil springs significantly affect fatigue life and should be taken into account.