An Atmospheric Signal Lowering the Spring Predictability Barrier in Statistical ENSO Forecasts

Abstract

Statistical models are known to be simple and effective tools for interseasonal predictions of ENSO dynamics (Barnston et al., 2012; Jan van Oldenborgh et al., 2005). The IRI/CPC ENSO Predictions Plume (Barnston et al., 2012)-an ensemble forecast of the Niño 3.4 index defined as the average sea surface temperatures (SST) in the region (5°N-5°S, 170°W-120°W)-demonstrates that both statistical and dynamical models yield close prediction skills at lead times up to 12 months. This similarity likely reflects the near-linearity of the seasonal tropical Indo-Pacific SST predictability studied by Newman and Sardeshmukh (2017). The main factor limiting statistical forecasts is the spring predictability barrier (SPB), also called the spring persistence barrier, that is, the empirically observed loss of autocorrelations in the tropical Pacific climate dynamics in May-June (Barnston et al., 2012; Torrence & Webster, 1998). Since many statistical models rely on SST anomalies (SSTAs) in the tropics, the SPB impacts statistical models more than dynamical models during forecasts beginning in spring (Barnston et al., 2012). Basically, the SPB phenomenon can be explained as a manifestation of ENSO seasonality related to the phase locking of ENSO dynamics with a seasonal cycle (Liu et al., 2018). In the tropical SSTA variability, there is a distinct one-year temporal pattern, hereinafter referred to as the ENSO cycle, that lasts from June to May of the following year, with persistent SST anomalies developing in the middle of the cycle (autumn-winter), whereas smaller and noisier anomalies appear at the beginning and end of the cycle (summer and spring, respectively). In particular, Tippett and L'Heureux (2020) recently showed that approximately 90% of the Niño 3.4 index variability can be explained by a one-dimensional deterministic signal defined on the June-May interval multiplied by different amplitudes in different years, with extrema in December and the lowest absolute values in May and June.

Improving statistical prediction and revealing nonlinearity of ENSO using observations of ocean heat content in the tropical Pacific

“…In this work we use two different parameterizations of deterministic part f of the model (2) which account phase locking of the ENSO dynamics to the annual cycle (Chen and Jin, 2020). The first one is a linear parameterization, suggested by Mukhin et al (2021):…”

“…To avoid overfitting of the model, the choice of the hyperparameters should be statistically justified, or optimal. According to (Gavrilov et al, 2017(Gavrilov et al, , 2019Mukhin et al, 2021), we use the Bayesian optimality criterion for estimating them, which relies on assessing the probability density function of data given the particular model; see details in Appendix A.…”

“…It is often noted (Martinez-Villalobos et al, 2019;Deser et al, 2009;Bamston et al, 1997) that the first EOF of SSTA in the tropical Pacific is associated with ENSO and the corresponding the forecasts they produce are random sequences of states, which means that different model runs yield different forecasts. According with (Gavrilov et al, 2019;Mukhin et al, 2021), the future value of a quantity x is assessed from a model forecast as the ensemble median x over a large number of the model runs.…”

“…Relying on theoretical understanding of ENSO, many studies are focused on finding additional atmospheric and oceanic predictors which can help to lower the SPB. Various predictors based on ocean heat content (OHC) (Clarke and Van Gorder, 2003), warm water volume (Meinen and McPhaden, 2000;, as well as atmospheric fields (Clarke and Van Gorder, 2003;Byshev et al, 2016;Mukhin et al, 2021) has been suggested. Nevertheless, there is still no conventional way to derive statistically justified predictors from data and to include them into prognostic models.…”

“…Nevertheless, there is still no conventional way to derive statistically justified predictors from data and to include them into prognostic models. Often Mukhin et al, 2021) such predictors are determined by finding significant lagged correlations between time series of SST-based ENSO index which needs to be predicted and corresponding time series of another ENSO-related climate variables. Typically, the obtained predictors are passed to the model as a fixed forcing (e.g.…”