Skip to content

Tutorial: Distribution Fitting

Mirrors 03_distribution_fitting.ipynb and the Distribution Fitting app page.

Fit one distribution

import numpy as np
from scipy import stats
from utils.distribution_utils import fit_distribution, perform_ks_test, compute_pdf, compute_cdf

data = np.random.default_rng(0).gamma(shape=2.0, scale=1.5, size=500)

params = fit_distribution(data, stats.gamma)   # (a, loc, scale) via MLE
print(params)
print(perform_ks_test(data, stats.gamma, params))

Compare several candidates

from utils.distribution_utils import fit_multiple_distributions

results = fit_multiple_distributions(data, [stats.norm, stats.gamma, stats.lognorm, stats.expon])
for r in sorted(results, key=lambda r: r.get("KS_stat", 1)):
    print(r["distribution"], round(r.get("KS_stat", float("nan")), 4), round(r.get("p_value", 0), 4))

The smallest KS statistic (largest p-value) is the best-fitting candidate.

Visualise the fit

from utils.viz_utils import plot_pdf_overlay, plot_cdf_overlay

plot_pdf_overlay(data, stats.gamma, params, title="Gamma PDF vs histogram")
plot_cdf_overlay(data, stats.gamma, params, title="Gamma CDF vs ECDF")

Why the frozen distribution matters

Goodness-of-fit is evaluated against the frozen distribution dist(*params).cdf. Passing the distribution name plus args= to kstest breaks on recent SciPy releases, which is why the utilities always freeze first.