Projects / Which new tracks a music label must promote first

Which new tracks a music label must promote first

Course project, Michigan Ross MBA. Hit prediction models in R on 289,941 public Spotify tracks, tested on later release years. Not client or employer work.

R
R scriptModels and the business result
SituationA music label releases many tracks each year and has promotion money for a small part of them.
ComplicationAbout 1 track in 4 becomes a hit. The label must choose which tracks to promote at release, before there are streaming numbers.
QuestionCan a model find the hits at release, and what does it use?
AnswerA model picks a top 10% of new tracks in which 77% are hits, against 26% for all tracks. Genre and the record of the artist decide; the sound adds little.

The return on promotion depends on the share of hits among the picks; three kinds of information can raise it

  • Hits among the tracks promoted77% in the answer
    • Tracks picked
      • Promotion budgetThe limit: the top 10% of new tracks
    • Information at release
      • Sound of the trackEleven audio measures

        Does the sound predict a hit?

        A little. 48% hits among the picks, against 26% at random.

      • Genre

        Does genre add to the sound?

        Yes. 73% hits among the picks.

      • Record of the artistThe lever: earlier tracks and earlier hits

        Does the record add more?

        Yes. 77% with all inputs. Artists with more than 50% earlier hits: 62% of new tracks are hits.

Lever what the decision changes Limit what the answer must respect Fixed outside the decision

Hits among the top 10% of picks

77%

26% for all new tracks: 3.0 times better than chance

With the sound of the track only

48%

audio measures such as energy, tempo and loudness

Hits that the picks contain

30%

1,778 of the hits, from 10% of the tracks

The sound says little; genre and the record of the artist give most of the gain

A hit is a track in the top quarter of popularity among the tracks of the same release year. Models were fitted on releases from 2000 to 2021 and tested on 23,180 tracks released in 2022 and 2023.

An artist with a record of hits is the safest pick, and a new artist is a better pick than an artist with no hit

Record of the artist in the years before the track: share of earlier tracks that were hits. Test tracks, 2022 and 2023.

The model

Decision

Which new tracks get a place in the promotion plan.

Objective

The highest share of hits among the tracks picked.

Method

Logistic regressions and a random forest in R on 289,941 tracks by 45,228 artists (a 25% sample of a public file).

  • In the random forest, genre and the record of the artist carry 43% of the importance. The eleven audio measures share the other 57%.
  • The popularity score favors recent tracks. With one cutoff for all years, the model learns which tracks are new. The hit label is set inside each release year to remove this.
  • The record of the artist uses earlier years only, so the model knows nothing that a label does not know at release.
  • Genre is strong in part because of how the public file was collected, by genre lists. A label must test the model on its own catalog.