There exists a fundamental as well as practical interest in being able to accurately forecast the future evolution of Earth's magnetic field at decadal to secular ranges. This work enables such forecasts by combining geomagnetic data with an Earth-like numerical model of a convection-driven fluid dynamo. The underlying data assimilation framework builds on recent progress in inverse geodynamo modelling, a method which estimates an internal dynamic structure for Earth's core from a snapshot of the magnetic field and its instantaneous rate of change at the surface, and takes advantage of linear relationships and long-range correlations between observed and hidden state variables. Here the method is further evolved into a single-epoch ensemble Kalman filter, in order to initialise at a given epoch an ensemble of states compatible with the observations and representative of the uncertainties in the estimation of hidden quantities. The ensemble dynamics, obtained by subsequent numerical integration of the prognostic model equations, are found to be governed by a thermal wind balance or equilibrium between buoyancy forces, the Coriolis force and the pressure gradient. The resulting core fluid flow pattern is a quasi-steady eccentric gyre organised in a column parallel to Earth's rotation axis, in equilibrium with a longitudinal hemispheric convective density anomaly pattern. The flow provides induction for the magnetic field, which also undergoes a realistic amount of diffusion. Predictions of the present magnetic field from data taken within the past century show that the ensemble has an average retaining good consistency with the true geomagnetic evolution and an acceptable spread well representative of prediction errors, up to at least a secular range. The predictability of the geodynamo thus appears to significantly exceed previous theoretical expectations based on the chaotic divergence of ensemble members. The assimilation generally outperforms the linear mathematical extrapolations from a 30-yr prediction range onwards, with a 40 per cent improvement in Earth-surface error at a secular range. The geomagnetic axial dipole decay observed over the past two centuries is predicted to continue at a similar pace in the next century, with a further loss of 1.1 ± 0.3 µT by year 2115. The focal (or minimum intensity) point of the South Atlantic geomagnetic anomaly is predicted to enter the South Pacific region in the next century, with the anomaly itself further deepening and widening. By year 2065, the minimum intensity is predicted to decrease by 1.46 ± 0.4 µT at the Earth surface and the focal point to move 12.8 ± 1.4 deg westwards with a slight northward component. This corresponds to a drift rate of 0.26 deg yr −1 , similar to the typical geomagnetic westward drift observed over the past four centuries. The same drift rate is also predicted until 2115 with a further (but more uncertain) intensity decrease