Tsfresh

==============
- new feature calculators:
- partial autocorrelation
- added list of calculated features to documentation
- added two ipython notebooks to
- illustrate PCA on features
- illustrate the Benjamini Yekutieli procedure
- fixed the following bugs
- improperly quotation of dickey fuller settings

=============
- new feature calculators:
- ratio_beyond_r_sigma
- energy_ratio_by_chunks
- number_crossing_m
- c3
- angle & abs for fft coefficients
- agg_autocorrelation
- p-Value and usedLag for augmented_dickey_fuller
- change_quantiles
- changed the calculation of the following features:
- fft_coefficients
- autocorrelation
- time_reversal_asymmetry_statistic
- removed the following feature calculators:
- large_number_of_peak
- mean_autocorrelation
- mean_abs_change_quantiles
- add support for multi classification in the feature selection
- improved description of the rolling mechanism
- added function make_forecasting_frame method for forecasting tasks
- internally ditched the pandas representation of the time series, yielding drastic speed improvements
- replaced feature calculator types from aggregate/aggregate with parameter/apply to simple/combiner
- add test for the ipython notebooks
- added notebook to inspect dft features
- make sure that RelevantFeatureAugmentor always imputes
- fixed the following bugs
- impute was replacing whole columns by mean
- fft coefficient were only calculated on truncated part
- allow to suppress warnings from impute function
- added missing lag in time_reversal_asymmetry_statistic

=============
- new features:
- linear trend
- agg trend
- new sklearn compatible transformers
- PerColumnImputer
- fixed bugs
- make mannwhitneyu method compatible with scipy > v0.18.0
- added caching to travis
- internally, added serial calculation of features

=============
- Breaking API changes:
- removing of feature extraction settings object, replaced by keyword arguments and a plain dictionary (fc_parameters)
- removing of feature selection settings object, replaced by keyword arguments
- added notebook with examples of new API
- added chapter in docs about the new API
- adjusted old notebooks and documentation to new API

=============

- added a maximum shift parameter to the rolling utility
- added a FAQ entry about how to use tsfresh on windows
- drastically decreased the runtime of the following features
- cwt_coefficient
- index_mass_quantile
- number_peaks
- large_standard_deviation
- symmetry_looking
- removed baseline unit tests
- bugfixes:
- per sample parallel imputing was done on chunks which gave non deterministic results
- imputing on dtypes other that float32 did not work properly
- several improvements to documentation

=============

- new rolling utility to use tsfresh for time series forecasting tasks
- bugfixes:
- index_mass_quantile was using global index of time series container
- an index with same name as id_column was breaking parallelization
- friedrich_coefficients and max_langevin_fixed_point were occasionally stalling