Prediction error minimization (PEM) method is used for modeling microelectromechanical (MEMS) system-based inertial sensor stochastic errors instead of autoregressive (AR) processes that are implemented in most of the recent studies. Since the sensor outputs are buried in a high-power white noise, the linear prediction methods used in the AR approach for bias instability modeling leads to biased estimates in parameter identification. Wavelet multiresolution techniques have been suggested so far for signal denoising to attenuate the effect of high-power white noise. However, the main sensor signal as well as the bias instability dynamics may be distorted during this procedure. The PEM method prepares the possibility to simultaneously model the bias instability and random walk as the most effective components of the MEMS inertial sensor's residual error. Therefore, the proposed method can be an alternative to the combination of the AR processes and wavelet decomposition in inertial sensor's stochastic error modeling. The performance of the proposed approach is illustrated in terms of prediction error indexes. Furthermore, the effectiveness of the modeling approach is shown in a static 2-D navigation application aided by zero velocity updates. The experimental results from real sensor's static data for both an inertial measurement unit (IMU) with conventional structure as well as a skew redundant IMU (SRIMU) show that the proposed PEM approach outperforms the AR method. The advantage of the method is more apparent for the SRIMU. The well-known $t$ -test statistical testing is exploited to guarantee the confidence of the results.