Is there a formula for a hit?

Every song that reached the Hot 100, one row each

The outcome is simple: did a song that reached the weekly Hot 100 make Billboard's Year-End Hot 100 that year? Each year about 100 songs did and about 360 more charted without making it, so every hit is compared with its own year.

Track record, genre and a danceable-but-restrained sound

Logistic regression with a fixed effect for each year and standard errors clustered by artist. Odds ratios above 1 raise the chance of making the year-end top 100. The artist's previous top-10 hits matter most; at the same number of top-10s, a long list of minor chart entries lowers the odds. Energy lowers the odds; loudness helps only once energy is held equal.

Feature (significant)	Odds ratio	95% CI	Unit
Artist's previous top-10 hits	1.56	1.43–1.69	per 1 SD
Pop artist	1.43	1.30–1.57	yes vs no
Hip hop artist	1.34	1.13–1.58	yes vs no
Dance / disco artist	1.34	1.21–1.49	yes vs no
Easy listening / soft rock	1.33	1.18–1.50	yes vs no
4/4 time	1.25	1.07–1.46	yes vs no
Loudness	1.19	1.12–1.26	per 1 SD
Danceability	1.18	1.12–1.25	per 1 SD
Rock artist	1.14	1.03–1.27	yes vs no
Song length	1.08	1.03–1.13	per 1 SD
Speechiness	0.94	0.90–0.98	per 1 SD
Sounds live	0.94	0.90–0.98	per 1 SD
Energy	0.77	0.71–0.82	per 1 SD
No genre tags	0.77	0.64–0.93	yes vs no
Country artist	0.71	0.62–0.81	yes vs no
Artist's previous chart entries	0.57	0.53–0.62	per 1 SD

Picks the bigger hit two times in three

Every song was scored by a model trained on the other 50 years. Within a year, the full model ranks a real year-end hit above a non-hit 65% of the time (logistic) or 67% (boosted trees); sound alone manages 57%.

What helped in one decade hurt in another

Danceability mattered from the 60s to the 80s

Odds ratios of 1.2–1.35 per step, then nothing in the 1990s.

Easy listening flipped

A plus in every decade to the 90s (about 1.3–1.5), then a minus in the 2000s (0.42).

Hip hop went from penalty to premium

0.50 in the 1980s, 1.46 in the 2000s. Energy lowers the odds in every decade.

The model's ten favourites: 4 hits, 6 misses

Its ten highest-scoring songs share the template: mid-tempo, danceable, loud, a title-phrase chorus, mostly R&B about attraction. The four hits each had a specific idea in the title; that hunch did not survive the 120-song test below.

Eighteen features, seven factors, two of them solid

Factor analysis on half the songs, confirmed on the other half. Factorability is mediocre (KMO 0.63) because Spotify designed its measures to be fairly independent. Rap/hip-hop and intensity replicate cleanly; the full seven-factor structure fits poorly (CFI 0.824, RMSEA 0.099). The seven factors still predict almost as well as all 18 features (AUC 0.615 vs 0.621).

Open each factor to see the ten year-end hits at the top, middle and bottom of it.

Ten song archetypes; one over-produces hits

Gaussian-mixture clusters on the factor scores. Lift compares each archetype's hit rate with songs from the same years, so decade trends don't masquerade as archetype effects.

Lyrics and hooks: a signature sound, a happy lyric, a love song

Every song in the study was summarised and coded (subject, mood, narrator, hook, signature sound, single order) from a model's general knowledge, blind to how it charted, with "don't know" allowed. The test below uses the 8,799 songs it knew well (3,754 year-end hits), adds the lyric codes to the sound-and-genre model with each year compared with itself, and clusters standard errors by artist. Only traits that survive a correction for testing many at once (Benjamini–Hochberg q < 0.05) are shown.

How far to trust it. The codes are the model's recollection, checked against sources on 135 songs (theme 94%, mood 97%, narrator 96% agreement; single order 76%). Famous riffs may be remembered better, which could flatter the signature-sound result, but among songs that all peaked in the top 10 it still holds (×1.52). Nothing else survives within that band, so these traits help a song into the year-end list more than they separate the biggest hits.

What changed from the pilot. The earlier 60-pair test flagged desire songs, playful and accusing narrators and standalone singles. None of these held up at scale; playful narrators now lean the other way (×0.78, not significant after correction). Signature sound and happy mood replicated.

The earlier test: 60 hits vs look-alike misses

60 year-end hits, each paired with a same-year song the model scored almost identically that missed. Research agents coded subject, mood, hook and structure from encyclopedic sources, blind to which songs were hits (no lyrics were reproduced). About 40 features were tested, so treat these as leads: none survives a formal correction.

53% vs 28%happy lyric mood, hits vs misses

57% vs 35%a signature riff, intro or sound

0 vs 10standalone singles (no parent album)

Feature	Hits	Matched misses	p (paired)
Standalone single (no album)	0%	17%	0.002
Happy lyric mood	53%	28%	0.014
Signature riff / intro / sound	57%	35%	0.024
Theme: desire	28%	10%	0.027
Playful narrator	17%	5%	0.016
Accusing narrator	2%	12%	0.031
Hook states a concrete situation	45%	43%	1.000
Has a bridge	58%	60%	1.000
Hook within 30 seconds	58%	50%	0.424
Title sung 10+ times	53%	45%	0.473
Lead single from album	48%	42%	0.572

The Choice Engine

Pick a song you love, and the engine ranks all 17,858 other songs by how much someone who loves it would value them. It uses a quadratic consumer-surplus score whose shadow prices come from the hit model, with plain cosine similarity alongside as a baseline. Switches add the artist's track record, the chart year's base rate and an era dial that favours songs from the reference's own years. Every card shows a synopsis, the hook and lyric-signal chips, and any ranking can be saved as a Spotify playlist.

Slow connection? Try the lighter version without synopses (4.8 MB instead of 9.3 MB).