An essential step for improving climate change models' performance is to evaluate their ability to represent the current climate conditions, especially extreme events. On such background, this study aims at evaluating the performance of the Quantile Delta Mapping (QDM) as a bias correction method for annual maximum daily precipitation series (bmax) generated from downscaled climate change models under tropical–subtropical conditions of Brazil. We selected the QDM due to its ability to correct bias in extreme quantile of wet days. Climate projections obtained from 20 NASA Earth Exchange Daily Downscaled Projections models (NEX‐GDDP) from 1950 to 2005 were subjected to validation processes based on the QDM method. Two climate change scenarios (RCP 4.5 and RCP 8.5 W m⁻²) have also been considered. Several goodness‐of‐fit measures, such as root‐mean‐square‐error (RMSE),SD, percentual bias (pbias), mean absolute error (MAE), Pearson correlation test, modified Willmott test (d_m), have been calculated from the outcomes of the models and their corresponding observed data (obtained from rain gauges). These goodness‐of‐fit measures were calculated before and after applying the QDM method. The QDM was able to correct virtually all biases. More specifically, the QDM successfully adjusted the empirical cumulative distribution of climate change projections, removing the systematic error of raw data. The QDM also presented a suitable performance when applied to future projections (2020–2095). This statement holds for all NEX‐GDDP models, except for the ACCESS1‐0 model in RCP 8.5. In such a scenario, this latter model presented unrealistic rainfall values. Finally, with the improvement resulting from applying the bias correction method QDM, there was an increase in the number of climate projections suitable for end‐users in the study region.

The standardized precipitation index (SPI) was developed as a stationary method in which the parameters of the distribution used to estimate the cumulative probability of rainfall amounts do not vary over time. Considering that changing climatic conditions may violate the assumption of stationarity, a nonstationary version of the SPI (NSPI) has been proposed. However, since time‐varying parametric distributions tend to remove the effect of rainfall trends on NSPI estimates, this latter index may fail to quantify the severity of drought events. In this study, we developed a four‐step algorithm that uses information generated by the NSPI algorithm to improve the interpretation of SPI estimates in terms of their probability of occurrence under changing climate conditions. This statistical/computational algorithm uses a nonstationary approach to detect trends in rainfall quantities and to quantify the effect of such trends on the probability of a particular SPI value occurring. The suitability of this algorithm was evaluated through Monte Carlo experiments and through two case study applications. The first study case applied the algorithm to historical rainfall series for the UK. The second applied the algorithm to gridded rainfall data for South America. The results found in this study demonstrated that, in the presence of temporal changes in rainfall frequency distributions, the probabilities of drought events estimated by the four‐step algorithm were consistently closer to the observed frequency of SPI than the probabilities estimated by the original SPI algorithm. In conclusion, the four‐step algorithm improved the interpretation of SPI values in terms of their expected frequencies under climate changing conditions. This algorithm is able to detect changes in the frequency of SPI values, to quantify how these changes affect the probability of drought events according to the SPI classification system, isolating its effect on the central tendency and on the dispersion of SPI distributions.

The main objective of this study was to investigate the trends on average and extreme events in time series of daily precipitation from 1980 to 2010 in the Paraná River basin, Brazil. The nonparametric Mann–Kendall test was applied to detect monotonic trend in the precipitation series. The occurrence of extreme values was analysed based on three generalized extreme values (GEV) models: Model 1 (stationary), Model 2 (non‐stationary for location parameter), and Model 3 (non‐stationary for location and scale parameters). The GEV parameters were estimated by the Generalized Maximum Likelihood method (GMLE) and for the non‐stationary models, the parameters were estimated as linear functions of time. To choose the most suitable model, the maximum likelihood ratio test (D) was used. From the results observed at the monthly scale, it was possible to infer that the months with the highest probability of an extreme weather event occurrence are February (climates Aw and Cfa), July (Cfa and Cfb), and October (Aw, Cfa, and Cfb). Approximately 90% of the 1,112 stations presented no trend regarding the GEV parameters. The non‐stationarity showed by other stations (Models 2 and 3) might be associated with several factors, such as the alteration of land use due to the north expansion of the agricultural border of the Paraná River basin.

This study compared the performance of seven gap-filling methods in daily streamflow and precipitation data and assessed the maximum number of gaps’ days on which the methods perform well. Random (occur randomly in the data series) and sequential gaps (consecutive days with missing record) were considerate. The results indicated that the type of gaps affects the performance of the methods for gap-filling streamflow and precipitation data. Concerning random gaps, the best methods for streamflow were ARIMA (autoregressive integrate moving average) and Spline interpolation. For precipitation, the best methods were Inverse Distance Weight (IDW) and Linear Regression (LR). Regarding to sequential gaps, LR and MR (Multiple Regression) perform well in gap-filling up to 60 consecutive days in streamflow series. The other methods perform well up to 15 consecutive gaps’ days. For precipitation series, the methods performed well up to seven consecutive missing days. For longer gaps (15, 30, 45 and 60 days), the methods performed poorly.