scikit-stats: statistics tools and utilities
scikit-stats like any other Python package:
pip install scikit-stats
or similar (use
virtualenv, etc. if you wish).
scikit-stats module includes modeling and hypothesis tests submodules. This a quick user guide to each submodule. The binder examples are also a good way to get started.
The modeling submodule includes the Bayesian Block algorithm that can be used to improve the binning of histograms. The visual improvement can be dramatic, and more importantly, this algorithm produces histograms that accurately represent the underlying distribution while being robust to statistical fluctuations. Here is a small example of the algorithm applied on Laplacian sampled data, compared to a histogram of this sample with a fine binning.
>>> import numpy as np >>> import matplotlib.pyplot as plt >>> from skstats.modeling import bayesian_blocks >>> data = np.random.laplace(size=10000) >>> blocks = bayesian_blocks(data) >>> plt.hist(data, bins=1000, label='Fine Binning', density=True, alpha=0.6) >>> plt.hist(data, bins=blocks, label='Bayesian Blocks', histtype='step', density=True, linewidth=2) >>> plt.legend(loc=2)
This submodule provides tools to do hypothesis tests such as discovery test and computations of upper limits or confidence intervals. scikit-stats needs a fitting backend to perform computations such as zfit. Any fitting library can be used if their API is compatible with scikit-stats (see api checks).
We give here a simple example of a discovery test, using zfit as backend, of gaussian signal with known mean and sigma over an exponential background.
>>> import zfit >>> from zfit.loss import ExtendedUnbinnedNLL >>> from zfit.minimize import Minuit >>> bounds = (0.1, 3.0) >>> obs = zfit.Space('x', limits=bounds) >>> bkg = np.random.exponential(0.5, 300) >>> peak = np.random.normal(1.2, 0.1, 25) >>> data = np.concatenate((bkg, peak)) >>> data = data[(data > bounds) & (data < bounds)] >>> N = data.size >>> data = zfit.Data.from_numpy(obs=obs, array=data) >>> lambda_ = zfit.Parameter("lambda", -2.0, -4.0, -1.0) >>> Nsig = zfit.Parameter("Ns", 20., -20., N) >>> Nbkg = zfit.Parameter("Nbkg", N, 0., N*1.1) >>> signal = Nsig * zfit.pdf.Gauss(obs=obs, mu=1.2, sigma=0.1) >>> background = Nbkg * zfit.pdf.Exponential(obs=obs, lambda_=lambda_) >>> loss = ExtendedUnbinnedNLL(model=signal + background, data=data) >>> from skstats.hypotests.calculators import AsymptoticCalculator >>> from skstats.hypotests import Discovery >>> from skstats.hypotests.parameters import POI >>> calculator = AsymptoticCalculator(loss, Minuit()) >>> poinull = POI(Nsig, 0) >>> discovery_test = Discovery(calculator, [poinull]) >>> discovery_test.result() p_value for the Null hypothesis = 0.0007571045424956679 Significance (in units of sigma) = 3.1719464825102244
The discovery test prints out the pvalue and the significance of the null hypothesis to be rejected.