In our laboratory a data-driven model for real-time ENSO forecast was elaborated based on the analysis of sea surface temperature from NOAA_ERSST dataset (provided by the NOAA/OAR/ESRL PSD, Boulder, Colorado, USA, from their Web site at http://www.esrl.noaa.gov/psd/).
The model is based on the data dimensionality reduction by means of linear dynamical mode (LDM) decomposition and empirical construction of stochastic evolution operator for the LDMs in the form of artificial neural network (ANN).
Initially the model is trained on the 48-year-long interval from January 1960 to December 2007.
This procedure determines optimal structural parameters for LDMs and ANN, like number of neurons in the ANN etc.
To make the forecast starting from any time moment after December 2007, we re-train the model on the 48-year-long interval ending at this time moment, using the same structural parameters of LDMs and ANN.
For the details of the method please see the paper [Gavrilov, A., Seleznev, A., Mukhin, D., Loskutov, E., Feigin, A., & Kurths, J. (2018). Linear dynamical modes as new variables for data-driven ENSO forecast. Climate Dynamics, 1–18, http://doi.org/10.1007/s00382-018-4255-7 ]
The figure below shows the latest forecast which we made for the Nino3.4 index anomaly using our model.
The base period for anomaly computation is 1981-2010.
The black line shows the end of the model training interval, i.e. the latest observed value of index.
The blue line corresponds to the values of index predicted by the model.
The shaded blue area corresponds to the 65% confidence interval provided by the model.
The dashed lines mark zero anomaly level and thresold-levels of 0.5 degrees above and below zero anomaly.
This figure is updated every month. Please look for the previous forecasts in the hyperlinks below.
Our previous ENSO predictions (year-month):
2019-10   2019-11   2019-12