Forecasting Pacific SSTs: Linear Inverse Model Predictions of the PDO

Abstract

A linear inverse model (LIM) is used to predict Pacific (30°S-60°N) sea surface temperature anomalies (SSTAs), including the Pacific decadal oscillation (PDO). The LIM is derived from the observed simultaneous and lagged covariance statistics of 3-month running mean Pacific SSTA for the years 1951-2000. The model forecasts exhibit significant skill over much of the Pacific for two to three seasons in advance and up to a year in some locations, particulary for forecasts initialized in winter. The predicted and observed PDO are significantly correlated at leads of up to four seasons, for example, the correlation exceeds 0.6 for 12-month forecasts initialized in January-March (JFM). The LIM-based PDO forecasts are more skillful than persistence or a first-order autoregressive model, and have comparable skill to LIM forecasts of El Niño SSTAs. Filtering the data indicates that much of the PDO forecast skill is due to ENSO teleconnections and the global trend.Within LIM, SST anomalies can grow due to constructive interference of the empirically determined modes, even though the individual modes are damped over time. For the Pacific domain, the basinwide SST variance can grow for ϳ14 months, consistent with the skill of the actual predictions. The optimal structure (OS), the initial SSTA pattern that LIM indicates should increase the most rapidly with time, is clearly relevant to the predictions, as the OS develops into a mature ENSO and PDO event 6-10 months later. The OS is also consistent with the seasonal footprinting mechanism (SFM) and the meridional mode (MM); the SFM and MM involve a set of atmosphere-ocean interactions that have been hypothesized to initiate ENSO events.

Systematic Estimates of Initial-Value Decadal Predictability for Six AOGCMs

“…Finally, note that the MLR method bears some similarities to the use of linear inverse modeling in predictability studies (Newman 2007;Alexander et al 2008), but one difference is that, in linear inverse modeling, a single propagation operator is assumed valid for all forecast ranges, whereas we construct a different MLR operator for each forecast range. Hence, unless the AOGCMs examined here are exactly first-order Markov processes, in which case the linear inverse modeling assumption is valid, the MLR operators we use are necessarily more accurate.…”

Pacific Decadal Oscillation and the Hydroclimatology of Western Canada—Review and Prospects

“…While the work of Keim et al (2002) was entirely retrospective, they suggested that PDO could be used to increase forecasting capability. Alexander et al (2008) developed a retrospective linear inverse model to predict Pacific sea surface temperature anomalies, including the PDO, which was useful in predicting PDO up to four seasons in advance. Their analysis indicated that much of the PDO variance was explained by ENSO teleconnections or a global trend.…”

The Impact of Extratropical Atmospheric Variability on ENSO: Testing the Seasonal Footprinting Mechanism Using Coupled Model Experiments

“…Several studies have used linear inverse models (LIMs) derived from simultaneous and lagged covariance statistics of observed SST anomalies to better understand and predict ENSO (e.g. Penland and Magorian 1993;Penland and Sardeshmukh 1995); the springtime SFM SST pattern closely resembles the "optimal structure", the pattern identified in LIMs as the most likely to grow into a large ENSO event (Penland and Sardeshmukh 1995;Xue et al 1997;Thompson and Battisti 2001;Alexander et al 2008). In addition, there is a close correspondence between the development of SST anomalies predicted by LIM and the evolution of the atmosphere-ocean system indicated by the SFM prior to ENSO events (Alexander et al 2008).…”