Abstract: A class of time-varying nonlinear stochastic systems subject to unmodeled dynamics and disturbances is considered. Through online filter parameter identification, a robust adaptive filter (RAF) is proposed. The optimal filtering parameters, such as covariance of state errors and filtering residuals, are determined by minimizing the covariance of state errors and ensuring the orthogonality of the filtering residuals at two adjacent times. The simulation example is a nonlinear time-delay stochastic system, in which mean and covariance of measurement errors are changed randomly and abruptly to simulate sensor faults. Even in such severe scenario, the RAF has strong robustness against measurement errors and shows satisfactory adaptive ability to track changes of time-delay and parameters no matter whether such changes are abrupt or slow.