0.4.1

0.4.1

New features

• The Distribution.fit method accepts a scipy_args dictionary which is
  passed to scipy's minimize function.
• The confidence interval computed by the profiler now uses root finding to
  find the bounds where the likelihood ratio test starts failing. This means
  confidence intervals can only be computed for the distribution's parameters.
• Upper bounds on dependencies were removed, improving compatibility with
  recent versions.

0.4.0

This is a long overdue release of pykelihood with many breaking changes that have accumulated over time.

Breaking changes

• Renamed stats_utils module to profiler
• Data must now be provided to kernels on creation, unbound kernels are
  no longer allowed
• Parameters are no longer subclasses of float, use .value to get
  their stored value
• ConditioningMethods were removed, their uses can be replaced with
  score functions
• The biv parameter to the Profiler was removed, confidence
  intervals are univariate only

Removed

Many distributions and utilities which were created with a specific use
case in mind and aren't generally useful have been removed:

• MixtureExponentialModel,
• ExtendedGPD,
• PointProcess,
• CompositionDistribution,
• DetrendedFluctuationAnalysis,
• pettitt_test,
• threshold_selection_GoF and threshold_selection_gpd_NorthorpColeman,
• extreme values visualisation routines,
• process samplers (Poisson and Hawkes).

New features

• Metrics: {pp,qq}_l{1,2}_distance, likelihood, expo_ratio
• Log-normal distribution
• Plotting functions now accept an ax argument to use instead of the
  global plt figure
• Constant kernel (most useful for testing)
• Kernels have a with_covariate method that returns a new kernel
  with the provided data as covariate, but all parameters are kept the
  same
• The random_state parameter to the Distribution.rvs method is now
  explicit and no longer hidden in the **kwargs

Bug fixes

• Fixed fit_instance for nested kernels with fixed values
• Fixed the TruncatedDistribution which forgot its bounds after fitting
• A parameter which shows up in several places in a distribution will
  keep the same value when fitting instead of returning independent
  parameters

Other

• Add section to README on fitting other score functions than the likelihood
• Add changelog with all version changes up to this one