Assessing the utility of statistical downscaling for subseasonal temperature forecasts

Subseasonal temperature forecasts can guide timely action against heat risks, but their coarse resolution limits regional usefulness. We apply a subseasonal downscaling framework to benchmark 27 statistical methods, including configurations of bias correction, linear and logistic regression, and analogs, to assess how they transfer forecast skill from coarse to local resolution at the weekly scale. As a test case, retrospective forecasts from CFSv2 (Climate Forecast System version 2; ~100 km) are downscaled to ~ 5 km for initializations issued one to four weeks before each target week of the Paris 2024 Olympics. Skill is assessed with the Brier Skill Score at the 10th and 90th percentiles for extremes. We also test whether incorporating atmospheric patterns adds value to downscaling. Methods are implemented with both daily and weekly data to examine the role of temporal resolution. Results show that, although most methods successfully transfer CFSv2 skill to higher resolution, method choice remains critical, as some degrade skill while others enhance it. Methods incorporating atmospheric patterns show promise at longer lead times when the relevant pattern, climatologically a key driver of heat in the target week, is well predicted. At the subseasonal scale, downscaling with weekly predictors outperforms that based on daily predictors.