Exploring Sub-Seasonal to Seasonal Prediction: Evaluating Deterministic and Probabilistic Forecasting Approaches Using Xcast on the S2S Scale

Sub-seasonal to seasonal (S2S) prediction has emerged as an important tool in anticipating climate variations over shorter timescales, from a few weeks to several months ahead. This research undertakes multiple evaluations of verification results derived from various deterministic and probabilistic forecasting approaches at the S2S scale, employing diverse techniques accessible within the Python tool named Xcast. Developed as a proficient utility, Xcast is a tool capable of utilizing statistical and machine learning methodologies to rapidly and effectively process various gridded climate data. The study conducts a comparative analysis of several methods, including multiple linear regression (MLR), extreme learning machine (ELM), and probabilistic output extreme learning machine (POELM). The assessment employs blended rain data from rain posts and Global Satellite Mapping of Precipitation (GSMaP), alongside S2S the European Center for Medium-Range Weather Forecasts (ECWMF) forecast data—both data are on a 10-day time scale with the period from 1996–2021 tailored for Indonesia region. The research domain employs a condition whereby the tercile probability is determined by data points that accumulate rainfall of over 50 mm per 10 days, with a 30% data point availability requirement. Evaluation metrics involve Pearson correlation and rank probability score (RPS). The evaluation of deterministic and probabilistic forecast models reveals that the POELM method outperforms in terms of deterministic forecast as indicated by Pearson correlation value, while MLR surpasses other methods in probabilistic forecasts when evaluated with RPS. Overall, the southwestern part of Kalimantan Island exhibits superior S2S forecast performance under the condition of rainfall accumulation exceeding 50 mm on a 10-day time scale.