El Niño/Southern Oscillation (ENSO) is traditionally described using monthly averages, which smooth out fast processes. However, in recent years it has become clear that high-frequency components may carry important information about the system's dynamics. This work stands alongside studies that search for hidden periodicities in climate series and examine long-range correlations, but it differs by using daily data instead of monthly and focusing on intraseasonal modes. This allows a finer structure of the signal to be seen and tests how predictable short oscillations are.
The authors took daily SOI values from January 1999 to March 2007 (3006 points). To extract periodicities, they used the multitaper method for power spectrum estimation with five tapers, which reduces estimate variance. Peak significance was tested against the null hypothesis of red noise and locally white noise. The signal was then modeled as a sum of harmonics with the found periods plus noise, with harmonic amplitudes considered linear functions of time. Parameters were fitted using least squares. After removing the reconstructed signal, the residual was analyzed using detrended fluctuation analysis (DFA) and first return time distributions.
The power spectrum revealed peaks at periods of 24, 27, 37, 76, 100, and 365 days, exceeding the 99% confidence interval. The 24-day period is associated with the Bransstator-Kushnir wave, 27 days with the lunar cycle, 37 days with the Madden-Julian oscillation, 76 days with intraseasonal tropical atmospheric oscillations, and 100 days may be a beat between 37 and 27 or 76 and 365 days. Signal reconstruction with these periods allowed a nine-month forecast. Forecast quality was assessed by the area under the ROC curve: AUC = 0.73, which is considered good for binary classification of the SOI sign. After removing the periodicities, DFA did not show linear behavior across the entire range of scales, indicating the absence of simple self-similarity. The return time distribution for the cleaned signal is better described by a stretched exponential rather than a Poisson distribution, pointing to correlations among extreme events.
This work shows that daily SOI data contain significant high-frequency components that can be used for medium-term forecasting. This distinguishes it from neighboring studies that rely on monthly averages and miss fast oscillations. Furthermore, analysis of the residual after removing periodicities demonstrates that the stochastic component is not simple white noise but has a complex correlation structure. This challenges models that assume uncorrelated fluctuations and underscores the need to account for memory in climate series.
In the future, periods could be refined by abandoning integer values and using longer series to test the robustness of the found modes. It is also promising to apply a similar approach to other climate indices, such as the North Atlantic Oscillation index or the Pacific Decadal Oscillation index. Developing nonlinear forecasting methods based on the identified periodicities could improve prediction quality.
The results may be useful for meteorologists and climatologists working on El Niño forecasts and associated precipitation and temperature anomalies. They are also important for understanding the statistics of extreme events in the climate system.
The next step could be to validate the found periods on independent data and compare with numerical model results. It is also worth investigating how these high-frequency modes interact with the low-frequency component of ENSO.
This work touches on a fundamental problem in climatology: how to separate deterministic periodicities from stochastic noise in climate series. It also contributes to understanding the nature of extreme events, such as strong El Niño, and their predictability.
The 100-day period may arise as a beat between 37 and 27 days: 1/(1/27 - 1/37) ≈ 100 days. This is an example of how the interaction of two oscillations generates a third, slower one.