Long-term prediction of El Niño-Southern Oscillation using reservoir computing with data-driven realtime filter.

The article focuses on developing a machine learning methodology for long-term, real-time prediction of the El Niño-Southern Oscillation (ENSO) using a novel bandpass filter combined with reservoir computing, a brain-inspired data-driven dynamical system modeling technique. The introduced realtime filter relies solely on past time-series data, avoiding the use of future information common in conventional smoothing methods, thus enabling operational forecasting without look-ahead bias. By applying Bayesian optimization to tune hyperparameters of both the filter and the reservoir computing model, the approach successfully extends ENSO prediction horizons up to 24 months, capturing multi-year dynamics in filtered sea surface temperature anomalies. While demonstrated on ENSO, the methodology is presented as broadly applicable to complex, high-dimensional dynamical phenomena requiring robust long-term forecasting.