The orientation of the earth in space changes unpredictably in a rapid and irregular manner, in addition to the uniform rotation of the earth. Observations of extra-terrestrial objects from the surface of the earth are affected by these variations, and knowledge of these changes is required for a variety of geodetic and astrometric purposes as well as being of interest in its own right. The orientation of the earth (specified by a three dimensional rotation vector) is measured by a variety of techniques; combination of these data sets is complicated by irregular changes in the spacing and accuracy of the various time series, and also by the existence of lower dimensional measurements of different linear combinations of the rotation vector components. A Kalman filter has been developed at JPL to smooth and predict earth orientation changes for application to spacecraft navigation by the NASA Deep Space Network. The filter, which provides estimates of the earth orientation changes (and of the excitation of these changes) based on whatever measurements are available, has been used for a number of research applications, both in the reduction of geodetic and astrometric data, and in research into the geophysical causes of earth orientation changes. The JPL Kalman filter uses stochastic models to account statistically for otherwise unpredictable changes in earth orientation; these models make it possible to provide reasonable estimates of the error in the smoothed time series, and to automatically vary the amount of smoothing according to the accuracy and density of the data. The derivation of the stochastic models used by the filter, the implementation of the models into the filter, a statistical description of what the filter does, and the results of filtering specific data sets will be discussed.