LoginSign Up
← Back to all sparks
G

gofedf

INFRA · APIS
Velocity0.0

Goodness of Fit Tests Based on Empirical Distribution Functions

github.com/pnickchi/gofedf ↗

A two-test goodness-of-fit package opens itself up to any weight function

goodness-of-fitempirical-distributionhypothesis-testingextensibilitystatistics
◆Current state
gofedf runs goodness-of-fit tests built on the empirical distribution function. Its three releases trace a short, clean arc: existence in 2023, then p-values computed from an analytical solution of the integral equation in 2024, then in 2026 a user-supplied weight function that replaces the fixed menu. Cramer-von Mises and Anderson-Darling are now two points in a family rather than the two options.
◆Where it's heading
The package is generalising rather than accumulating. Each release removed a hard-coded decision: first how eigenvalues are computed, offering both the analytical route and a matrix approximation; then which weight function defines the statistic at all. The maintainer's own framing in 1.1.0 is a contrast against what earlier versions would not let you do, which is the shape of a package aiming to become a framework.
◆Prediction
An arbitrary weight function is the extensibility point that matters for EDF tests; what it lacks is calibration guidance, since Type I error behaviour was the argued benefit of the analytical eigenvalue route. Documented recommendations or diagnostics for user-chosen weights are the natural follow-up, though the entries do not announce one.

◆Recent moves

  1. 4mo ago

    User-supplied weight functions for Cramer-von Mises tests

    ⚡ SPARK

    The release that changes what gofedf is. Where earlier versions offered Cramer-von Mises or Anderson-Darling — two particular weight functions — a user can now supply their own, which turns the package from a menu into a generator of tests.

    View source ↗
  2. 1y ago

    p-values from an analytical eigenvalue solution

    The package gains p-value computation and two ways to get the eigenvalues behind it: the analytical solution of the integral equation, which the maintainer's simulations show controls Type I error better, or a matrix approximation. A statistic becomes a usable test.

    View source ↗
  3. 2y ago

    First release of the gofedf package

    ⚡ SPARK

    The starting point, and the whole note is that it is one. Everything the later releases generalise — the eigenvalue computation, the fixed choice of test — arrives here as a first working implementation.

    View source ↗