We revisit the empirical moving window filtering method of Swenson and Wahr (Geophys Res Lett 33:L08402, 2006) and its variants, Chambers (Geophys Res Lett 33:L17603, 2006) and Chen et al. (Geophys Res Lett 34: L13302, 2007), for reducing the correlated errors in the Stokes coefficients (SCs) of the spherical harmonic expansion of the GRACE determined monthly geopotential solutions. Based on a comparison of the three published approaches mentioned, we propose a refined approach for choosing parameters of the decorrelation filter. Our approach is based on the error pattern of the SCs in the monthly GRACE geopotential solutions. We keep a portion of the lower degree-order SCs with the smallest errors unchanged, and high-pass filter the rest using a moving window technique, with window width decreasing as the error of the SCs increases. Both the unchanged portion of SCs and the window width conform with the error pattern, and are adjustable with a parameter. Compared to the three published approaches mentioned, our unchanged portion of SCs and window width depend on both degree and order in a more complex way. We have used the trend of mass change to test various parameters toward a preferred choice for a global compromise between the removal of the correlated errors and the minimization of signal distortion.