Power Laws and Popularity

Power Laws and Popularity#

What 500 artists can and cannot tell you about streaming success

A synthetic Spotify catalog, treated as a small-n inference problem wearing an ML costume

Synthetic data. The Kaggle file is a 50,000-row track table with 500 invented artist names. The publisher states that no Spotify API data were used. Nothing below is a claim about the streaming industry.

Project Overview#

Kaggle will sell you 50,000 rows and 33 columns of “Spotify artist streaming analytics, 2020–2025”. That looks like a machine-learning problem. It is not. After a schema check the grain is track, not artist-year; the names are generated; audio features are orthogonal to streams; and popularity is almost log1p of the target. The interesting work is estimand definition, a leakage audit, validation that matches the clustering, and stopping when extra model capacity stops paying.

The honest conclusion is the one worth publishing: flexible models buy nothing at this generating process, leaked features buy (R^2 = 1), and the planted superstar tail is catalog size rather than quality.

The pages#

Code lives in src/spotify_powerlaws/ and numbered scripts; notebooks load artifacts.

1. Streams and leakage

Read the page →

Forensics, estimand, grouped CV, model ladder, power-law tests, conformal coverage

2. Genre from audio

Read the page →

The unused signal: recover planted genre from audio, with grouped-by-artist validation

Key questions#

  • Is this an artist panel, a track catalog, or a synthetic script with a Spotify costume?

  • Which columns are antecedent to streams, and which are the target in disguise?

  • Does grouped-by-artist CV change the honest model — or only fail to save a tautological one?

  • When does added capacity stop paying, and does partial pooling find artist effects the ANOVA already said were zero?

  • Did the generator couple audio to genre even though it did not couple audio to streams?

Preview figures#

Forensics dashboard

Leakage gap

Lorenz curves

Dataset#

  • Source: Kaggle: Spotify Artist Streaming Analytics 2020–2025 (CC BY 4.0)

  • File: 50,000 tracks × 33 columns; 500 invented artists

  • Grain: track (track_id unique). Not artist-year, not a listening panel

  • Rebuild: make download writes data/raw/ (gitignored). SHA-256 in artifacts/download.json

  • This site: notebooks read committed artifacts/ and figures/ only

Technical stack#

Python • pandas • scikit-learn • statsmodels MixedLM • matplotlib • scipy

Pipeline: projects/spotify-power-laws/Makefile (00_download.py … 09_genre.py)

Start with the forensics.

Open the analysis →

Contents