The SuperDARN radar observations were processed to remove ground scatter, and to eliminate measurements with too low power (lower than 3dB), or which had a poor-quality flag (identified in RSTv4.0). When binning the data, range gates below 11 and above 150 (where those values correspond to a multiple of the 45 km range distance from the radar array location) were not used, since these gave inaccurate locational estimates.

The Shore et al. (2022) eigenanalyses of the radar data were performed on a spatial grid with 559 locations, so we keep the same grid for consistency. We recombined all 10 spatial and temporal modes for the 12-year interval. The resulting velocities have NaN (not a number) values for co-latitudes greater than 30 degrees, and for the single point exactly at the pole. For this reason, our forecast model regression coefficients also have NaN values for co-latitudes greater than 30 degrees and the pole.