Did Streaming Break How Songs Are Built?

A Multi-Method Statistical Analysis of 1M+ Spotify Tracks (2000 -- 2023)

Welch's t-test  ·  Pearson's Correlation  ·  Levene's Variance Test  ·  Cohen's d  ·  Fisher Z-transform

---

Overview

The shift from physical media to algorithmic streaming reshaped the economics of music. Artists now earn revenue per play, not per minute. Playlist algorithms curate discovery. Short-form platforms reward instant hooks.

Did these incentives measurably change how songs are composed?

This project tests that question using a dataset of 1,048,575 Spotify tracks spanning 82 genres and 22 years, applying inferential statistics covered in an introductory data analytics course.

---

Key Findings

Hypothesis	Method	Verdict	Result
Songs are shorter post-2018	Welch's t-test	Supported	-28 sec (-11.2%), d = 0.315
Songs are faster post-2018	Welch's t-test	Negligible	+1 BPM, d = 0.034
Louder tracks are less happy	Pearson's r	Rejected (reversed)	r = +0.277, louder = happier
Tempo is converging	Levene's test	Rejected	U-shaped variance, highest in 2020-23

Bottom line: Streaming shortened songs by ~28 seconds, but it hasn't made them faster, sadder, or more uniform in tempo. The industry trimmed the container -- not the music inside it.

---

Research Architecture

                     Main Question
          Did streaming change song structure?
               (Welch's t-test on duration & tempo)
                    /              \
                   /                \
          Branch 1                  Branch 2
   Loudness-Valence             Tempo Convergence
   Trade-off                    Across Eras
   (Pearson's r)                (Levene's test)
                    \              /
                     \            /
                  Unified Conclusion
           Has streaming homogenized music?

---

Visualizations

Duration & tempo: pre-streaming vs streaming era	Loudness vs valence: positive, not negative
Tempo variance: U-shaped, not converging	Loudness-valence coupling strengthening over time

---

Dataset

Property	Value
Source	Spotify Audio Features API
Tracks	1,048,575
Years	2000--2009, 2012--2023
Genres	82
Variables	19 (popularity, danceability, energy, loudness, valence, tempo, duration, etc.)
Missing values	0
Duplicates	0

Era groupings used in analysis:

Group	Years	n
Pre-streaming	2000--2009	425,924
Transition	2012--2017	309,313
Streaming era	2018--2023	313,338

---

Methods & Statistical Pipeline

Preprocessing

1. Missing value check -- none found, no imputation needed
2. Outlier detection -- IQR method with winsorization (capping, not removal)
3. Z-score standardization -- for correlation analysis variables

Inferential Analysis

• Shapiro-Wilk and Kolmogorov-Smirnov tests for normality (on n=5,000 subsamples)
• Levene's test for homoscedasticity
• Welch's t-test (independent, unequal variances) for era comparison
• Cohen's d for practical effect size
• Pearson's r for loudness-valence relationship
• Fisher Z-transformation to compare correlations across eras
• Spearman's rho for tempo variance trend over time

---

Repository Structure

.
├── README.md                        # This file
├── research_findings.md             # Full research report (Introduction, Method, Results, Discussion)
├── final_analysis.ipynb             # Jupyter notebook with all runnable Python code
├── final_presentation.pdf           # Slide deck (26 slides, dark theme)
├── spotify_data_edited.csv.gz       # Dataset (1M+ tracks, gzip compressed)
└── figures/
    ├── phase1_initial_distributions.png
    ├── phase2_before_after_boxplots.png
    ├── phase3_era_comparison.png
    ├── phase3_yearly_trends.png
    ├── phase4_loudness_valence_scatter.png
    ├── phase4_yearly_correlation_trend.png
    ├── phase5_tempo_convergence.png
    └── phase5_variance_bars.png

---

How to Reproduce

# Clone the repo
git clone https://github.com/yarorazum/Did-Streaming-Break-How-Songs-Are-Built-.git
cd Did-Streaming-Break-How-Songs-Are-Built-

# Decompress the dataset
gunzip -k spotify_data_edited.csv.gz

# Install dependencies
pip install pandas numpy scipy matplotlib seaborn

# Run the analysis
jupyter notebook final_analysis.ipynb

The dataset is provided as a gzip archive (spotify_data_edited.csv.gz, ~69 MB). Decompress it before running the notebook. The notebook loads spotify_data_edited.csv and reproduces every figure and statistical test from scratch.

---

Tech Stack

Tool	Purpose
Python 3.10	Core language
pandas	Data loading, manipulation, grouping
NumPy	Numerical operations
SciPy	Statistical tests (t-test, Pearson, Shapiro-Wilk, Levene)
Matplotlib	Plotting (histograms, scatter, line charts)
seaborn	Styled visualizations
Jupyter	Interactive analysis notebook

---

What I Learned

• Statistical significance is not practical significance. With 1M data points, a 1 BPM difference yields p < 0.001 -- but Cohen's d = 0.034 reveals it's meaningless.
• Hypotheses can be wrong -- and that's the point. The loudness-valence correlation ran opposite to our prediction, leading to a richer discussion about confounding variables (energy as a common cause).
• Preprocessing choices matter. Winsorization vs. removal produces different downstream results; documenting before/after stats is essential for transparency.
• Effect sizes belong next to every p-value. They anchor statistical results in the real world.

---

_{Built as the final project for DS108: Introduction to Data Analysis with Python}