[2.5.0] - 2021-04-14
--------------------
- Added option to not use (i.e. overwrite) existing local bias adjustment or 
  statistical downscaling results. This is the new default.
- Added option to do bias adjustment in running-window mode with adjustable
  step size. Compared to a month-by-month bias adjustment the new mode reduces
  discontinuities in statistics such multi-year daily mean values at each turn
  of the month.
- Added options to ignore trends in frequencies of values beyond threshold and
  not limit the climate change signal transfer to values within threshold, both
  for a better bias adjustment of hurs in the case of too many supersaturated
  hurs values in the simulated input data.
- Simplified transfer of climate change signal in frequencies of values beyond
  threshold for a better bias adjustment under climate change of variables such
  as prsnratio.



[2.4.1] - 2020-06-15
--------------------
- Fixed bug that occurred when there are no values within threshold in the data
  to be quantile-mapped but there should be such values after bias adjustment.
  In this case, non-parametric quantile mapping is now applied.



[2.4.0] - 2020-06-10
--------------------
- Changed how the pseudo future observations are generated by limiting the
  climate change signal transfer to values within threshold for a better bias
  adjustment of variables such as hurs and pr.
- Introduced brute force quantile mapping of the values to be quantile mapped
  to the distribution of all simulated future values within threshold prior to
  the actual quantile mapping for a better bias adjustment of variables such as
  hurs and pr in the case of biased frequencies of values beyond threshold.



[2.3.1] - 2020-05-27
--------------------
- Added rough goodness-of-fit test to fit function to notice bad distribution
  fits. The test is based on the Kolmogorov-Smirnov test statistic. In the case
  of a noticed bad fit, quantile mapping is done non-parametricly.



[2.3] - 2020-02-06
------------------
- Added non-parametric quantile mapping option for a more robust bias adjustment
  of bounded variables. Non-parametric quantile mapping is applied if no
  distribution type is specified for parametric quantile mapping.
- Changed climate change signal transfer to empirical percentiles of bounded
  variables and frequencies of values beyond threshold (equations (8) and (9) in
  Lange (2019) <https://doi.org/10.5194/gmd-12-3055-2019>) for a better bias
  adjustment under climate change.
- Changed sampling of invalid values for months without valid values. In these
  cases the average of the valid values from all months is used. Only if there
  are no valid values at all if_all_invalid_use is used.



[2.2] - 2020-01-29
------------------
- Changed randomization of values beyond threshold for a better bias adjustment
  of variables such as hurs: randomization is now applied to all values beyond
  threshold, ranks of values beyond threshold are preserved, random numbers are
  no longer raised to higher power.
- Fixed numerical instability of climate change signal transfer to upper bound
  climatology to improve bias adjustment of variables such as rsds.



[2.1] - 2020-01-18
------------------
- Changed sampling of invalid values to mimic trend in valid values to better
  preserve within-period trends for variables such as prsnratio.
- Removed the automatic calculation of the number of iterations of the
  (modified) MBCn algorithm. Made univariate bias adjustment and statistical
  downscaling with 20 iterations the new defaults.
- Made detrending conditional on trend being significantly (at the 5 % level)
  different from 0.
- Fixed numerical instability of upper bound scaling to improve bias adjustment
  of variables such as rsds.
- Fixed minor bugs related to missing output directories, inconsistent spatial
  shapes, and running the code with different numpy versions.
- Facilitated resumption of canceled jobs by using already existing local
  results.



[2.0] - 2019-08-18
------------------
- Added multivariate bias adjustment option.
- Added support of input NetCDF files with missing values.
- Added support of different downscaling factors in different spatial
  dimensions.
- Increased execution speed by eliminating various I/O bottlenecks.
- Reduced memory usage by sharing resources between processes and saving local
  results in numpy stack.
- Fixed minor bugs related to bilinear regridding, detrending, and applications
  to fewer than 12 calendar months.



[1.0] - 2019-03-07
------------------
- Reference version for Lange (2019) <https://doi.org/10.5194/gmd-12-3055-2019>.
