Zeruya Yudea Siagian, Natasya Irene Poly, Halmar Halide
The El Niño–Southern Oscillation (ENSO) and the Indian Ocean Dipole (IOD) are major modes of interannual climate variability that significantly influence ocean–atmosphere dynamics and hydrometeorological conditions in Indonesia. Accurate prediction of these climate indices is essential for improving early warning systems and supporting climate risk mitigation. Due to the complex and nonlinear nature of climate variability, predictive models capable of capturing both linear and nonlinear relationships in time series data are required. Therefore, this study aims to develop and compare the performance of several linear and nonlinear models in predicting the Niño 3.4 and IOD indices at lead times ranging from 1 to 12 months.
Monthly Niño 3.4 anomaly data from September 1873 to July 2025 and IOD index data were obtained from the National Oceanic and Atmospheric Administration (NOAA). The Niño 3.4 dataset was chronologically divided into training data (1873–2009) and testing data (2010–2025). Modeling approaches included Multiple Linear Regression (MLR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Long Short-Term Memory (LSTM). Model performance was evaluated using the Anomaly Correlation Coefficient (ACC), Root Mean Square Error (RMSE), and Peirce Skill Score (PSS).
The results show that all models achieved excellent performance at a 1-month lead time for both climate indices. For the Niño 3.4 index, model performance decreased with increasing lead time; however, LSTM maintained the highest correlation at a 12-month lead time (ACC = 0.436; PSS = 0.136), indicating its capability to capture long-term temporal dependencies. Although Random Forest produced the lowest RMSE at longer lead times (0.731), its lower ACC and PSS suggest limitations in representing ENSO dynamics. For the IOD index, Multiple Linear Regression (MLR) achieved the best performance at a 1-month lead time with ACC of 0.97248, RMSE of 0.072261, and PSS of 0.94286. However, predictive skill decreased significantly at lead times beyond three months, with ACC and PSS approaching zero or becoming negative and RMSE increasing, indicating reduced predictability of the IOD at medium- to long-term horizons.
Overall, the results indicate that LSTM provides the most robust performance for long-term ENSO prediction, while the linear MLR model remains effective and consistent for short-term IOD prediction. These findings highlight the importance of selecting appropriate modeling approaches depending on the temporal characteristics of different climate variability phenomena.
Keywords: ENSO, Indian Ocean Dipole (IOD), LSTM, Multiple Linear Regression (MLR), time series prediction, ACC, PSS, RMSE.
