At timescales shorter than about 2 yr, non-tidal length-of-day (LOD) variations are mainly excited by angular momentum exchanges between the atmospheric, oceanic and continental hydrological fluid envelopes and the underlying solid Earth. But, neither agreement among different geophysical models for the fluid dynamics nor consistency with geodetic observations of LOD has reached satisfactory levels. This is mainly ascribed to significant discrepancies and uncertainties in the theories and assumptions adopted by different modelling groups, in their numerical methods, and in the accuracy and coverage of global input data fields. Based on careful comparisons with more accurate geodetic measurements and satellite gravimetry products (from satellite laser ranging, SLR), observed LOD and C 20 geopotential time-series can provide strong constraints to evaluate or form combined geophysical models. In this study, wavelet decomposition is used to extract several narrow-band components to compare in addition to considering the total signals. We then make refinements to the least difference combination (LDC) method proposed by Chen et al., to form multimodel geophysical excitations. Two combination variants, called the weighted mean combination (WMC2 and WMC4), are also evaluated. All the multimodel methods attempt to extract the best-modelled frequency components from each geophysical model by relying on geodetic excitation and the C 20 series as references. The comparative performances of the three combinations LDC, WMC2 and WMC4 and the original single models are determined. We find that (1) Estimating the Circulation and Climate of the Ocean and Max-Planck-Institute for Meteorology Ocean Model give a more reliable view of the ocean redistributions than the Ocean Model for Circulation and Tides used by European Centre for Medium-Range Weather Forecasts, especially for the annual component; (2) C 20 series from SLR can provide a rigorous constraint for the total matter excitation of the geophysical fluids, especially for broad-band parts; (3) the Sea-Level Angular Momentum functions term, correcting for sea-level effects (global mass balance) put forward by the Earth System Modelling group at GFZ German Research Centre for Geosciences, can significantly improve the Hydrospheric Effective Angular Momentum functions matter terms; (4) the LDC/WMC combinations are much better than the original individual geophysical model excitations, reducing the magnitude of unexplained LOD excitations to roughly the 10 μs level; (5) the level of residual LOD variations after removing models or model combinations is remarkably invariant with respect to LOD periods between ∼2 months and ∼3 yr, being 12-14 μs for the best original models and 7-12 μs for our combinations; (6) while differences between the IERS 14C04 and the JPL SPACE2015 geodetic LOD time-series are not negligible, errors in both series are still not large compared to the geophysical models (for periods >2 months) so the impact on excitation studies is minimal except at semiann...