El Niño–Southern Oscillation (ENSO) remains the dominant interannual signal of Earth's climate, affecting droughts, floods, and crops worldwide. Predicting it is difficult: fast noise, seasonal cycles, and slow decadal trends mix in a single time series. Existing approaches, from conceptual models to global climate models, work reliably only up to a year ahead. Machine learning has improved the situation, but nearly all successful models use filters that implicitly look into the future when smoothing the series. This makes them unsuitable for operational real-time forecasting. The work aligns with research on forecasting climate phenomena using reservoir computing and neural networks, but differs in that it is the first to build the entire pipeline, from raw data to prediction, strictly without future information.
The authors took monthly sea surface temperature data from the Niño-3.4 region (5°N–5°S, 170°W–120°W) from the HadISST dataset for 1870–2022. They subtracted the 1971–2000 climatology to obtain the anomaly without any moving average. Then they applied a new band-pass filter: a weighted moving average with a weight function that is zero for all future time points. Filter parameters (window width, cosine coefficients, decay rate) were tuned via Bayesian optimization to maximize the fraction of recurring patterns in the filtered series while maintaining high correlation with the original series. The filtered series was fed into an echo state network (reservoir computer) with delayed coordinates: the input vector consisted of nine values of the series with a delay step of 4 months. Network hyperparameters (reservoir size 244, spectral radius 0.712, leak rate 0.975, and others) were also tuned using Optuna on the training period 1986–1995, maximizing the correlation at the 24-month forecast lead.
The optimal filter passes oscillations with periods roughly from 4 to 8 years, matching the typical ENSO timescale. Due to the asymmetry of the weight function, the filtered series is shifted back by about 5 months, but the correlation with the original series reaches 0.837. On an independent test period from January 2001 to December 2015 (180 forecast starts), the average all-season correlation remains above 0.5 for 29 months. Accounting for the five-month filter shift, this means a real prediction horizon of 24 months, that is, two years. For comparison, random reservoir matrices with the same hyperparameters give noticeably worse results, confirming the importance of optimization. The model reproduces large positive and negative ENSO peaks in 2010–2014 well, but underestimates the record El Niño of 2015–2016.
The work differs from neighboring studies in that it is the first to propose a complete ENSO forecasting pipeline that does not use any future information at any stage, from filtering to model training. This removes a fundamental limitation for operational real-time forecasting. Furthermore, the authors show that a simple reservoir computer architecture with a properly chosen filter achieves a two-year horizon comparable to more complex deep neural networks. The method is not tied to ENSO: it can be applied to any time series with pronounced low-frequency oscillations where long-term forecasting without looking into the future is needed.
A natural direction is to test the method on other climate indices (e.g., Madden–Julian Oscillation, North Atlantic Oscillation) and on non-climatic series, from financial to biomedical signals. One could also investigate how the filter's passband tuning affects the forecast horizon in different systems. Increasing the length of training data, possibly through augmentation, would allow larger reservoirs and capture more complex dynamics. Combining the proposed filter with other machine learning architectures, such as deep reservoir computers or hybrid models, is promising.
The method could improve operational ENSO forecasts in meteorological agencies, which is important for agriculture, water resource management, and insurance. It also provides a universal tool for any field requiring long-term real-time time series prediction without future information.
The immediate next step is to validate the approach on other climate variables and independent datasets. It is also worth investigating the sensitivity of the forecast horizon to the choice of filter passband and to the length of the training window.
In climate science, the question of ENSO predictability beyond the 'spring barrier' and the role of decadal variability in limiting the forecast horizon remain open. This work does not solve these problems, but it offers a practical way to circumvent the limitation associated with the implicit use of future information in filters, which may increase confidence in long-term forecasts. The connection to fundamental questions about the chaotic nature of the climate system and the limits of predictability remains a topic for future research.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors built a bandpass filter that uses only past values of the series and combined it with a reservoir computer to forecast ENSO. On sea surface temperature data from the Niño-3.4 region, the method achieves a correlation above 0.5 at a 29-month horizon, which corresponds to two years after accounting for the filter shift. The filter and model hyperparameters were tuned via Bayesian optimization.
Where its strength is
A strong point is the strict adherence to causality at every stage, from filtering to forecasting, which makes the method suitable for operational use. Also noteworthy is the achieved prediction horizon of two years, which exceeds typical limits for ENSO.
Where it could go next
Worth trying
Why this may matter
The developed approach demonstrates that strictly causal methods can achieve a useful forecast horizon for ENSO, which is important for practical applications such as agricultural planning and water resource management. Moreover, the method is potentially transferable to other complex systems with similar time series properties.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.53, against a usual 0.64. It looks like few have drilled nearby.
The standard Niño-3.4 index used to declare El Niño is itself smoothed with a five-month moving average, meaning the official status of the event at any given time already contains information about future months. The authors even abandoned that to make the forecast honestly operational.
In Kim Stanley Robinson's novel 'Forty Signs of Rain', scientists try to predict abrupt climate changes, grappling with chaos and incomplete data. The idea of a filter that works only with the past resonates with the dream of a forecast that does not depend on knowledge of the future.