Nolds

Fixed

- Issue 13: corr_dim ignored the fit argument

Added

- documentation for `lyap_r_len`, `lyap_e_len` and the `hurst-nvals` example

Changed

- `hurst_compare_nvals` now also uses `np.asarray`

Fixed

- some formatting problems in the documentation

Added

- test function `hurst_compare_nvals` that compares different choices for the `nvals` parameter for `hurst_rs`
- example for `hurst_compare_nvals` (can be called using `python -m nolds.examples hurst-nvals`)
- helper functions `lyap_r_len` and `lyap_e_len` to calculate minimum data length required for `lyap_r` and `lyap_e`
- test cases `test_lyap_r_limits` and `test_lyap_e_limits` to ensure that `lyap_r_len` and `lyap_e_len` are calculated correctly
- description of parameter `min_nb` for `lyap_e`
- uses `np.asarray` wherever possible. The following functions should now also work with pandas objects and other "array-like" structures:
* `lyap_r`
* `lyap_e`
* `sampen`
* `hurst_rs`
* `corr_dim`
* `dfa `
- nolds documentation can now also be found on readthedocs.org: http://nolds.readthedocs.io/

Changed

- the previously internal helper function `expected_rs` is now available from the main module
- calculates minimum data length for lyap_r to provide better error messages
- uses `rcond=-1` in lstseq to keep behavior consistent between numpy versions
- mutes `ImportWarning`s from `sklearn` in unit tests
- disables an ugly hack when using `RANSAC` as fitting method and instead requires `sklearn>=0.19` that fixes the underlying issue
- makes test case for correlation dimension less strict
- added hint when `nolds.examples` is called with an unknown example name

Fixed

- note in the description of the parameter `tau` in `lyap_r` was misleading/wrong (probably a copy-pase error)

Removed

- distance values `"euler"`, `"chebychev"`, `rowwise_euler` and `rowwise_chebychev` for `sampen` and `corr_dim` (was deprecated)
- keyword parameter `min_vectors` for `lyap_r` (was deprecated)

Added

- function `logmid_n` that allows for a better choice of `nvals` parameter in `hurst_rs`

Changed

- adds more descriptions and instructions for comparing `hurst_rs` with other implementations

Added

- module `datasets`
+ dataset `brown72` that has a prescribed hurst exponent of 0.72
+ generators for the logistic and the tent map
+ true random numbers using the package `quantumrandom`
- test `test_hurst_pracma` that uses the same testing sequences for `hurst_rs` as the R-package `pracma`
- example function `plot_hurst_hist` that plots a histogram of hurst exponent values for random data
- example function `weron_2002_figure2`
- `fgn` for fractional gaussian noise in the `datasets` module
- documentation for unittests and examples
- parameter `unbiased` for `hurst_rs` that allows to choose between the new (fixed) behavior and the old one (using the wrong version of the standard deviation)
- parameter `corrected` for `hurst_rs` that applies the Anis-Lloyd correction factor to the example by default

Changed

- default choice for the parameter `nvals` in `hurst_rs` now favors higher n values and always uses 16 n values
- `fbm` is now moved to the `datasets` module

Fixed

- using fitting method `'ransac'` when sklearn was not installed resulted in an exception instead of a warning
- NaNs in `hurst_rs` where filtered from the set of (R/S)_n values, but the filtered values for n would remain in the calculation and fitting
- `hurst_rs` used the wrong standard deviation, since we estimate the mean of the samples from the data we need to set the parameter `ddof` to `1`

Added
- `lyap_r` now has a new parameter `fit_offset` that allows to ignore the first steps of the plot in the fitting process.

Changed
- The parameter `min_vectors` is now called `min_neighbors` in `lyap_r` and refers to the number of vectors that are candidates for the closest neighbor.

Fixed
- The algorithm for choosing the `lag` would always choose 0 in `lyap_r`.
- There was an error in the calculation of the number of vectors used for `min_vectors` in `lyap_r`.