AReMiC: Archive-driven probabilistic correction and statistical downscaling of weekly accumulated precipitation at subseasonal lead times

Coarse-grid ECMWF (EC) ensemble forecasts require probabilistic correction and statistical downscaling for local applications at subseasonal lead times. We propose AReMiC (Archive Retrieval, Error Migration, and Climate Mixing), which produces predictive distributions of weekly accumulated precipitation on a 0.1◦ grid for six forecast weeks (W1–W6). Current EC forecasts provide weather and precipitation signals, while historical fine-grid observations provide local rainfall magnitudes and occurrence frequencies. Paired historical EC forecasts and observations are retrieved and softly weighted using ensemble intensity, spread and meteorological conditions. Week-specific error migration adjusts positive historical rainfall using the difference between log1p-transformed current and historical EC ensemble means, while preserving original zeros. The resulting empirical distribution is mixed with local seasonal climatology using effective sample size and nearest-case feature distance. Other conditions being equal, the climatological contribution increases when case weights become more concentrated or the nearest-case distance increases. The retrieval, migration and mixing parameters are trained jointly using the continuous ranked probability score (CRPS) and Brier scores for local quantile events. An independent historical evaluation uses 345 initializations during 2021–2023 and 97,749 common land grid cells over mainland China. AReMiC achieves the best overall point estimates across all six reported scores among the compared methods. Its pooled six-week CRPS decreases from 6.521 mm for raw EC to 5.112 mm, a reduction of 21.60%. Relative to raw EC, the ranked probability skill score (RPSS10) increases by 0.7230, while the Brier skill scores (BSS) for Q90 and Q95 exceedance events increase by 0.7637 and 0.8510, respectively. CRPS is 1.97% lower than that of probabilistic bias correction. Q10–Q90 Brier skill remains positive in all six weeks. Evaluation on 688 operational initializations during 2024–2025 reduces CRPS by 12.96% relative to 101-member raw EC forecasts. Ablation experiments support the contributions of error migration, week-specific coefficients and adaptive climate mixing.