EVALUATION OF THE CLIMATE FORECAST SYSTEM (CFS) MODEL IN PREDICTING THE MADDEN JULIAN OSCILLATION (MJO) DURING BOREAL WINTER

• Main Author: Chania, Friska
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/84276
• Institution: Institut Teknologi Bandung

The Madden Julian Oscillation (MJO) has garnered attention among scientific practitioners as a key source of predictability for subseasonal to seasonal (S2S) forecasts, bridging the gap between weather and seasonal predictions. One of the S2S output models is the Climate Forecast System (CFSv2), which holds potential for predicting subseasonal conditions in the Indonesian region. Currently, predictions of the MJO index are available in the form of MJO phase diagrams. However, limitations in predicting the MJO index can constrain analysis of existing models. This study aims to evaluate the performance of the CFSv2 model in predicting the MJO index using the 'mjoindices' tool developed by Hoffman. Calculation of the MJO index was conducted using the 'mjoindices' tool from reanalysis data and operational CFSv2 model data for the boreal winter period (November, December, and January) of 2011-2021. Model performance was assessed deterministically using RMSE and CC metrics, and probabilistically using CRPS for evaluating the probabilistic distribution of weekly MJO index averages, and Brier Score for assessing weekly active MJO event occurrences. Active MJO criteria were applied when the MJO index exceeded 1 for five consecutive days. Based on this research, the 'mjoindices' tool proved effective in calculating the OMI index and establishing a predictive system for the MJO index using the ensemble prediction system of CFSv2 to forecast MJO phase and magnitude for the next 4 weeks. Evaluation of the CFSv2 model performance showed that although RMSE and CC did not indicate significant relationships between predictions and observations, probabilistic metrics such as CRPS and Brier Score provided a more comprehensive overview. CRPS indicated that the model was more accurate in short-term predictions with the best score of 0.368 at lead 1, but accuracy decreased with increasing lead time. The Brier score at lead 1 is 0.352, indicating low weekly prediction accuracy, but the model's predictability does not experience a significant decline across leads.