The dynamic Allan variance (DAVAR) is a new method for analyzing stochastic error of gyroscope. However, it has difficulty in making a good tradeoff between tracking capabilities and variance reduction due to the window with fixed length. An improved algorithm based on kurtosis and time-variant window is proposed to quickly track variations in the signal and obtain a low variance of the estimate. The kurtosis is introduced into analysis of gyroscope output, and the window length function to truncate signal is built by taking kurtosis as variables,which can make window length change with the non-stationary of the signal automatically. Then the length of truncation window is estimated according to the function, the values of Allan variance is obtained in the windows, and the error coefficients can also be identified and extracted at the same time. The above data are expressed by three or two-dimensional to describe the dynamic characteristics of gyro. Simulation and experimental data analysis results show that the proposed algorithm can track the non-stationary of gyroscope more effectively and obtain a good confidence in stationary random process, and the method can be more effective for the extraction and identification of stochastic error coefficients.