STATISTICAL UNCERTAINTY IN DROUGHT FORECASTING USING MARKOV CHAINS AND THE STANDARD PRECIPITATION INDEX (SPI).

Resumo

Droughts affect basic human activities, and food and industry production. An adequate drought forecasting is crucial to guarantee the survival of population and promote societal development. The Standard Precipitation Index (SPI) is recommended by the World Meteorological Organization (WMO) to monitor meteorological drought. Using drought classification based on SPI to build Markov chains is a common tool for drought forecasting. However, Markov chains building process produce uncertainties inherent to the transition probabilities estimation. These uncertainties are often ignored by practitioners. In this study we analyze the statistical uncertainties of using Markov chains for drought annual forecasting. As a case study, the dry region of the State of Ceará (Northeastern Brazil) is analyzed, considering the precipitation records from 1911 to 2019. In addition to 100-year database (1911-2011) for Markov chain modeling and 8-year data (2012-2019) for forecasting validation, four fictional database extensions were considered in order to assess the effect of database size in the uncertainty. A likelihood ratio is used as index to assess model performance. The uncertainties assessment showed that an apparent performant Markov chain model for drought class forecasting may not be more informative than the historic proportion of drought class. Considering these uncertainties is crucial for an adequate forecasting with Markov chains.