A standard goal of many neurophysiological investigations is to obtain enough insight into a neuron's behavior that it becomes possible to predict responses to arbitrary stimuli. Techniques have been developed to solve this system identification problem, and the numerical method presented here adds to this toolbox. Stimuli and responses, beginning as functions of time, are transformed to complex-valued functions of both time and temporal frequency, giving amplitude and phase at each frequency and time point. The transformation is implemented by wavelets. The kernel describing the system is then derived by simply dividing the response wavelet by the stimulus wavelet. The results are averaged over time, incorporating median filtering to remove artifacts. Estimated kernels match well to the actual kernels, with little data needed. Noise tolerance is excellent, and the method works on a wide range of kernels and stimulus types. The algorithm is easy to implement and understand, but can be applied broadly.