Verification of Calibrated Multimodel Subseasonal Precipitation Predictions Cascaded From Global to Regional Scale Over Ceará State in Brazil

This paper assesses a multimodel subseasonal prediction system developed to produce calibrated global and regional precipitation subseasonal predictions for Northeast Brazil, including Ceará State, located within this region. This system was developed using four global models, based on a linear regression procedure of retrospective predictions on past observations for calibrating the multimodel ensemble mean of these models. The procedure used 1° by 1° spatial resolution retrospective predictions initialized on Wednesdays over the common 1999–2016 period for all models. For producing predictions over Ceará, the regression procedure was applied at a finer (0.15° by 0.15°) resolution using interpolated observations. This statistical downscaling procedure allows cascading predictions from global to regionally refined scales. The assessment was performed across different timescales—weekly, fortnightly, and extended accumulation periods (30 and 44‐days). Predictions expressed as accumulated precipitation anomalies and probability for the occurrence of positive anomalies were evaluated. The global assessment identified Northeast Brazil as a region with notably high subseasonal prediction performance, demonstrating reasonable quality while examining linear association (correlations exceeding 0.4) and discrimination (area under the relative operating characteristic [ROC] curve above 0.6). The benefit of using all available prediction information from the initialization day, taking advantage of initial conditions memory, to enhance longer accumulation periods (14, 30, and 44 days) performance was demonstrated by finding that predictions for these extended periods maintained performance levels comparable to those of Week 1. Longer lead predictions for Weeks 3 and 4, Fortnights 2 and 3 showed similar performance in the tropics, including Northeast Brazil, with correlations exceeding 0.4 and area under ROC curves above 0.6, with El Niño–Southern Oscillation (ENSO), Madden–Julian Oscillation (MJO), and tropical Atlantic variability suggested as potential predictability sources. Downscaled predictions over Ceará maintained similar performance levels to those obtained at coarser spatial resolutions, reassuring the adequacy of the cascading procedure used to generate regional scale predictions. The large uncertainty in the computed skill scores prevented demonstrating the benefit of calibrated multimodel predictions compared to individually calibrated model predictions. A limitation of the implemented approach is the need for high resolution historical precipitation observations to allow generating spatially refined calibrated predictions.