Spatial variation in long‐lead predictability of summer monsoon rainfall using a time‐varying model and global climatic indices.

Long‐lead prediction of summer monsoon rainfall in India is a challenging task, especially at finer spatial scale. The spatial variability in the long‐lead (one or two season in advance) prediction plays a vital role in planning of hydrological and agricultural aspects of the society. One of the major issues in this field is the climate‐induced time‐varying characteristics (non‐stationarity) that lead to a deteriorating model performance as the time passes by after the model calibration. This study proposes the use of time‐varying approaches in order to check such deteriorating performance over time. Considering the time‐varying association between Indian summer monsoon rainfall and large‐scale climatic indices (e.g., El Niño‐Southern Oscillation, Equatorial Indian Ocean Oscillation, North Atlantic Oscillation, Pacific Decadal Oscillation, and El Niño Modoki Index), a time‐varying approach based on hybrid graphical modelling (GM) and vine copula (GM‐Copula) is demonstrated for rainfall prediction over five homogeneous monsoon regions (HMRs) in India. The time‐varying characteristic is imparted in the GM‐Copula approach by recursively updating the model inputs and the corresponding model parameters at a regular time interval (τ) through recalibration. In the time‐varying framework, the parameter τ is referred to as optimum recurrence interval of model recalibration and it is identified as 5 years for the regions with moderate rainfall and 3 years for regions with above and below moderate rainfall. The developed time‐varying approach is able to yield reasonably good prediction performance (mean absolute percentage error being within 4–10% across HMRs) with a prediction lead time of 5 months. HMR‐wise seasonal rainfall predictions with such quality and lead time are expected to be highly useful. [ABSTRACT FROM AUTHOR]