We present a comparative study of the performance of different basis functions for the nonparametric modeling of neural activity in response to natural stimuli. Based on naturalistic video sequences, a generative model of neural activity was created using a stochastic linear-nonlinear-spiking cascade. The temporal dynamics of the spiking response is well captured with cubic splines with equidistant knot spacings. Whereas a sym4-wavelet decomposition performs competitively or only slightly worse than the spline basis, Haar wavelets (or histogram-based models) seem unsuitable for faithfully describing the temporal dynamics of the sensory neurons. This tendency was confirmed with an application to a real data set of spike trains recorded from visual cortex of the awake monkey.