Superstar collaborations don’t guarantee success: A technical analysis from beatBread data scientists

Earlier this month, Kim Petras’ collaboration with Sam Smith on “Unholy” led the emerging artist to a Grammy win. In light of the historic win, the beatBread data science team dug into the impact of collaborations on artists in the past year.

In this article, we add more detail to the analysis described by beatBread CEO Peter Sinclair in the latest edition of “Advancing Independence.” You can find his full analysis here.

Dataset

We started with a large dataset containing weekly streaming data since 2017 on over 14 million tracks. We then filtered it down to tracks associated with artists who had at least 10,000 Spotify monthly listeners. This is the dataset we used for the remainder of this analysis.

Track-level analysis

For each artist on each track, we calculate the release ratio as the track’s median streams in its first year over the artist’s median streams across all their tracks. This gives us an idea of how well the track performed relative to how a normal track performs for a given artist. If this release ratio is bigger than 1, the track performed better than its 50th percentile, and if it is smaller, the track performed worse than that.

We also calculated another feature which we call listener share. Listener shareis calculated by dividing an artist’s Spotify monthly listeners by the sum of the Spotify monthly listeners of all artists on the track. This feature indicates how big the artist is compared to the other artists on the track. If there is only one performer the listener share is exactly 1.

At this point, the tracks were separated into three groups based on the listener share of the main artist. The three groups were no collaborators (listener share = 1), small collaborators (listener share > 0.7), and no collaborators (listener share < 0.25). We then compare the distributions of release ratios of these groups to find the impact that collaboration status has on track performance. The empirical CDF for each of these three distributions is shown below. It appears that tracks with a small collaborator perform similarly to tracks with no collaborator, and that tracks with a big collaborator perform better than tracks with no collaborator.

Artist-level analysis

Our next question was whether collaboration can have positive effects that extend beyond the collaborated track to have an impact on the artist’s career. To do this, we went through each artist in our dataset and calculated a few statistics at any given point in time. First, we found out whether they had a new release in the last six months. If they had a new release, we also determined whether it was a collaboration. We also calculated the percent difference between the artist’s streams in the previous year, and the number of streams they had one or four years into the future. We also calculated the artist’s Spotify monthly listeners, the trend of the artist’s streams, and the average time since release of the artist’s catalog. These three metrics will let us control for the artist’s size, trend, and longevity.

Once we calculated the above statistics for each artist and date in our dataset, we performed a linear regression only using observations where an artist had any type of release in the previous six months. We had four features: whether they had collaborated in the last six months, Spotify monthly listeners, artist streams trend, and average catalog age. We had two targets: percent change in artist streams from last year to next year and percent change in artist streams from last year to four years in the future. The t-statistic for the collaboration feature was -3.817. This would mean that collaboration technically has a statistically significant negative correlation with the artist’s future streams. However, the fitted coefficient was 0.997, meaning that streams are expected to change by only 0.3%. The t-statistic for the year four model was -1.178, which is not statistically significant. We can essentially conclude that collaboration does not have much of a correlation with future streams.

Although it seems that collaboration does not help an artist’s career as a whole, we wonder if that changes based on the collaborator’s size. To do this, we took the same dataset that was used to answer the second question. We only took the rows corresponding with dates where collaborated tracks were released. For each of these tracks, we calculated the sum of monthly listeners of all artists collaborating on the track. Then, for each of those artists, we calculated the ratio of their monthly listeners and the total of all the artists’ listeners. We called this the artist’s listener share.

Next, we performed another linear regression. We used the same features and targets as those above, but replaced the collaboration indicator with listener share. The t-statistic for listener share was -1.936 for the one year model and was 2.602 for the four year model. Essentially, the results are very mixed for whether having a bigger collaborator is better for your future career than having a small one.

One could argue with these findings that being a featured artist or collaborating has a nominal negative impact on your near term career, however over the long term your career grows faster relative to other similar artists who do not collaborate. A couple hypotheses that could flow from this is that collaboration and being featured helps grow your long-term fan base and or give you experience in what clicks with fans so that you can incorporate it into your new work.

During this analysis, we learned more about the complexity surrounding musical collaborations. It can be difficult to try to get firm results. What we can say with the most confidence is that featuring a bigger artist on your track has a high correlation with the number of streams that track will receive. However, it does not seem to have much of an impact on your career as a whole.