A framework to quantify bias for improved drought forecasting 

Abstract

Listen

Translate article

<p>The Standardized Precipitation Index (SPI) is one of the most popular indices for characterizing the meteorological drought on a range of time scales.  To date, SPI has been thoroughly used to monitor and predict drought in the precipitation signal and to further support early warning and climate services. While many studies focus on the performance improvement of drought models, there is to our knowledge no reference around the correct computation of SPI in a drought forecasting setting. As SPI is typically computed on the entire data set, prior to model-validation, bias is introduced to both the training and validation sets. This stems from the fact that the distribution parameters of the index are estimated using observations from the validation and test sets leading to information leakage. Here, we propose a modified calculation of SPI oriented for forecasting applications by measuring the bias introduced to the SPI values in the training set. Moreover, we propose the best practice for calculating the SPI during model-validation and encapsulate these in a drought forecasting framework. The proposed framework is further demonstrated using a 50-year data set from Sweden. Our findings suggest that the amount of bias introduced to the training sets increases with increased SPI scale, significantly affecting in some cases more than 80% of the available basins.</p>