Probabilistic seasonal rainfall forecasts using semiparametric d-vine copula-based quantile regression

Abstract Skillful probabilistic seasonal rainfall forecasts play a vital role in supporting water resource users, developing agricultural risk-management plans, and improving decision-making processes. This chapter applies a novel statistical copula-based approach to develop a probabilistic seasonal rainfall forecast model using multiple large-scale oceanic and atmospheric climate indices. Here, a d -vine copula is used to forecast the seasonal cumulative rainfall in 16 weather stations across the Australia's Wheatbelt. These stations span different climate conditions recording historical data for the period 1889–2012. The seasonal rainfalls are forecast in different quantile levels using different climate predictor data sets. The corrected Akaike information criterion (AIC)–conditional log-likelihood is then used to screen the most influential covariates to be additively incorporated into the multivariate probabilistic forecast model, resulting in a parsimonious predictive model. The mutually inclusive correlations between El Nino–Southern Oscillation (ENSO) and Indian Ocean Dipole (IOD) indices and seasonal rainfall are found to be statistically significant. Therefore, using the climate information, skillful rainfall forecasts can be made three to 6 months ahead. The d -vine copula model is found to outperform the traditional quantile regression methods in forecasting rainfall in the median and the upper levels. The information from lagged, concurrent, and combined climate indices is therefore demonstrated to be a potentially useful predictor for forecasting seasonal rainfall in Australia's Wheatbelt region.