Seasonal Precipitation Predictability Based on Sea Surface Temperature Forecasts From a Cyclostationary Linear Inverse Model

Abstract. Developing skillful seasonal precipitation forecasts remains a key challenge over many areas of the world. In this study, we test how much seasonal precipitation forecast skill can be derived from sea surface temperature (SST) forecasts from a cyclostationary linear inverse model (CSLIM) trained on observed estimates of seasonally varying predictable dynamics. Here we introduce a novel method that learns lead- and season-specific relationships between CSLIM SST forecasts and observed precipitation empirical orthogonal functions to generate precipitation forecasts up to one year in advance. Furthermore, we assess the precipitation forecast skill related to modes of SST variability, including the El Niño / Southern Oscillation (ENSO) and a long-term warming trend. Precipitation skill generated using CSLIM SST forecasts is better than climatology for many key rainy seasons worldwide, in many cases 6-12 months in advance. A single ENSO mode, which captures most but not all ENSO variability, is the leading source of precipitation skill for several seasons such as the October-December ‘short’ rains in East Africa, the main Central America rainy season, and winter precipitation over the southeast and southwest United States. Conversely, a long-term trend accounts for most of the skill over the Sahel. West Pacific SST variability distinct from the ENSO mode contributes to skill in some cases such as the March–May ‘long’ rains in East Africa and the Middle East / Southwest Asia region. Some precipitation seasons, such as warm season southeast US precipitation, are not skillful from CSLIM SST forecasts, suggesting a limited or nonlinear relationship between SSTs and precipitation, and/or one that is hindered by seasonally unpredictable dynamics.