Exploring multi-annual regimes in total and extreme Argentinian precipitation using hidden 
Markov models
Authors: Jones, M.R., Katz, R.W., Rajagopalan, B.

The humid and semi-arid area of the Argentinian Pampas is influence by many external 
factors, of which ENSO is the major single source of seasonal to interannual climate 
variability. The region has also experienced marked inter-decadal climate variability and 
significant increases in annual precipitation until recently; however, it is unclear whether 
the variations form part of a longer term gradual trend, or arise from “regime shifts”. 
Recent increases in precipitation expanded the boundary of rainfed agriculture towards drier 
regions and have contributed to major changes in land use. However, these evolutions in land 
use may not be sustainable if the climate returns to a drier epoch, as suggested by recent 
drought.
Statistical analyses of annual to decadal climate variability are often modeled in terms of 
deterministic shifts in the mean or variance of a time series, using techniques such as 
change-point analysis. Instead, we use a fully probabilistic approach based on “hidden” 
mixtures of distributions, in which there is a probability of randomly shifting from hidden 
state to another during each year. We examine historical meteorological observations for 
evidence of trends and/or multiple climatic regimes to support the agricultural community in 
decision making over the next 10-30 years. Temperature statistics, such as annual daily 
maximum/minimum or maximum/minimum daily temperature range demonstrate clear trends 
consistent with both increases in global mean temperature and their associated atmospheric 
responses, and well documented urban heat island effects. While the seasonal temperature 
statistics tally well with seasonal measures of ENSO, there is little other evidence of 
multiple climatological states. In contrast there are few statistically significant trends 
in seasonal and annual precipitation statistics, and correlation with ENSO is less 
significant, but hidden states reflecting dry and wet years are more apparent in the >60 
year time series.
Closer examination of the annual and seasonal total wet day count and total precipitation 
reveal a significant improvement (tested using the AIC and BIC) in data representation when 
mixtures of two or more Gaussian or Poisson distributions are fitted. This supports the 
hypothesis that multiple states exist giving rise to wetter or drier years. “Regime-like” 
behavior can be introduced into the hidden mixture model through a Markov chain to allow for 
temporal persistence in the hidden states (i.e. a hidden Markov model; HMM), in addition to 
dependence on atmospheric covariates such as ENSO. The ultimate focus of this research is on 
the high impact weather phenomena which can have catastrophic consequences for agriculture. 
Therefore, we will extendborrow strength from the mixture models and HMMs for total 
precipitation and extend the technique to apply to annual/seasonal temperature and 
precipitation extremes using Extreme Value Theory.