<italic>Background:</italic> Optically pumped magnetometers (OPMs) have made moving, wearable magnetoencephalography (MEG) possible. The OPMs typically used for MEG require a low background magnetic field to operate, which is achieved using both passive and active magnetic shielding. However, the background magnetic field is never truly zero Tesla, and so the field at each of the OPMs changes as the participant moves. This leads to position and orientation dependent changes in the measurements, which manifest as low frequency artefacts in MEG data. <italic>Objective:</italic> We model the spatial variation in the magnetic field and use the model to predict the movement artefact found in a dataset. <italic>Methods:</italic> We demonstrate a method for modelling this field with a triaxial magnetometer, then show that we can use the same technique to predict the movement artefact in a real OPM-based MEG (OP-MEG) dataset. <italic>Results:</italic> Using an 86-channel OP-MEG system, we found that this modelling method maximally reduced the power spectral density of the data by 27.8 <inline-formula><tex-math notation="LaTeX">$ \pm $</tex-math></inline-formula> 0.6 dB at 0 Hz, when applied over 5 s non-overlapping windows. <italic>Conclusion:</italic> The magnetic field inside our state-of-the art magnetically shielded room can be well described by low-order spherical harmonic functions. We achieved a large reduction in movement noise when we applied this model to OP-MEG data. <italic>Significance:</italic> Real-time implementation of this method could reduce passive shielding requirements for OP-MEG recording and allow the measurement of low-frequency brain activity during natural participant movement.