Empirically derived Markov models and prediction of tropical Pacific sea surface temperature anomalies

Empirical dynamical modeling (EDM) is employed to determine if ENSO forecas ting skill using monthly mean SST data can be enhanced by including subsurf ace temperature anomaly data. The Nino 3.4 index is forecast first using an EDM constructed from the principal component time series corresponding to EOFs of SST anomaly maps of the central and eastern tropical Pacific (32 de grees N-32 degrees S, 120 degrees E70 degrees W) for the period 1965-93. Cr oss validation is applied to minimize the artificial skill of the forecasts , which are made over the same 29-yr period. The forecasting is then repeat ed with the inclusion of principal components of hear content of the upper 300 m over the northern tropical Pacific (30 degrees N-0 degrees, 120 degre es E-72 degrees W). The forecast skill using SST alone and SST plus subsurface temperature is c ompared for lead times ranging between 3 and 12 months. The EDM, which incl udes the subsurface information, forecasts with grater skill at all lead ti mes: particularly important is the second principal component of the heal c ontent, which appears to contribute information on the transition phase bet ween warm and cold ENSO events. The apparent improvement by including subsu rface information, although robust, does not appear to be statistically sig nificant. However, the temporal and spatial coverage of the subsurface data is limited, so this study probably underestimates the usefulness of includ ing subsurface temperature data in efforts to predict ENSO. Finally, cross- validated forecasts using a Markov model that includes an annual cycle are shown to be less skillful than forecasts using a seasonally invariant Marko v model. The reason for this appears to be that dividing the data yields an insufficient database to derive an accurate Markov model.