We study statistical properties of stochastic variations in pulse arrival times, timing noise, in radio pulsars using a new analysis method applied in the time domain. The method proceeds in two steps. First, we subtract low‐frequency wander using a high‐pass filter. Secondly, we calculate the discrete correlation function of the filtered data. As a complementary method for measuring correlations, we introduce a statistic that measures the dispersion of the data with respect to the data translated in time. The analysis methods presented here are robust and of general usefulness for studying arrival time variations over time‐scales approaching the average sampling interval. We apply these methods to timing data for 32 pulsars. In two radio pulsars, PSRs B1133+16 and B1933+16, we find that fluctuations in arrival times are correlated over time‐scales of 10–20 d with the distinct signature of a relaxation process. Though this relaxation response could be magnetospheric in origin, we argue that damping between the neutron star crust and interior liquid is a more likely explanation. Under this interpretation, our results provide the first evidence independent from pulsar spin glitches of differential rotation in neutron stars. PSR B0950+08 shows evidence for quasi‐periodic oscillations that could be related to mode switching.