Sudden stratospheric warmings (SSWs) have a long-lasting effect within the stratosphere as well as impacts on the underlying troposphere. However, sub-seasonal forecasts of the winter polar stratosphere fail to use their full potential for predictability as they tend to underestimate the magnitude and persistence of these events already within the stratosphere. The origin of this underestimation is unknown. Here, we demonstrate that the associated polar stratospheric cold bias following SSW events in sub-seasonal hindcasts can be halved by increasing vertical model resolution, suggesting a potential sensitivity to gravity wave forcing. While the predictability of the planetary Rossby wave flux into the stratosphere at lead times longer than a week is limited, the existence of a critical layer for gravity waves with a low zonal phase speed caused by the disturbed polar vortex provides predictability to the upper stratosphere. Gravity wave breaking near that critical layer can, therefore, decelerate the zonal flow consistently with anomalous subsidence over the polar cap leading to warmer temperatures in the middle polar stratosphere. Since the spectrum of gravity waves involves vertical wavelengths of less than 4000 m, as estimated by wavelet analysis, a high vertical model resolution is needed to resolve the positive feedback between gravity wave forcing and the state of the polar vortex. Specifically, we find that at a spectral resolution of TCo639 (approximate horizontal grid spacing of 18 km) at least 198 levels are needed to correctly resolve the spectrum of gravity waves in the ECMWF Integrated Forecasting System. Increasing vertical resolution in operational forecasts will help to mitigate stratospheric temperature biases and improve sub-seasonal predictions of the stratospheric polar vortex.