We propose a fixed-lag particle smoother (FLPS) for time-varying frequency-selective and nonlinear channel estimation. Compared to the standard particle filter (PF) which tracks the filtering distribution of the state variable, the FLPS tracks the fixed-lag smoothing distribution of the state variable. The choice of the proposal distribution and the computation of the importance weights are derived. Simulations are provided to illustrate the performance of the FLPS under different system settings. Performance comparison of the FLPS with the standard PF is provided. Simulation results show that the FLPS outperforms the PF with the similar computational costs.