Predicting commercial performance of a thousand new tracks released in Q1 2019
To test the accuracy of our algorithm on a large collection of brand new music, we predicted commercial performance of a corpus of new songs one month after their release. Specifically, we took the popular global version of the New Music Friday playlist published by Spotify every week, and predicted the total number of streams each track would amass within one month. We chose New Music Friday as a test corpus because each week it contains a good number of tracks (about a hundred) representing a variety of music by a diverse range of artists operating at every level of the industry.
Here, we present an analysis of every track included on New Music Friday published between 11th January and 29th March 2019, i.e., Q1 2019. The dataset was comprised of a diverse range of new music broadly representative of the output of the global music industry:
The dataset contained 1128 tracks in total, and included songs performed in a broad range of genres and styles.
Songs were performed by 882 unique lead artists, from regional and emerging talent to global superstars like Ariana Grande and Drake.
Of the 1128 tracks analysed, 139 included one or more featured artists, and 16 tracks were performed by two or more joint-lead artists.
The most-streamed track after 1 month, with 196 million streams, was bad guy by Billie Eilish; the least-streamed track, with 39 thousand streams, was 2500 by Poundside Pop.
For each track, we predicted how many streams it would amass one month after release (or more precisely, since publication of the New Music Friday playlist upon which it appeared).
One month after the end of Q1, we took all our predictions, rank-ordered them from lowest to highest, and did the same for actual performance. When we compared predicted rank with actual rank, we found that over 70% of tracks fell within +/-10% of their actual rank, 60% fell within 5%, and over half fell within 1%.
We also more precisely measured comparative error between predicted number of streams and actual number of streams by standardising each artist’s predicted and actual values against their typical performance range. This revealed that 90% of tracks fell within 10% of their artists’ typical performance range, 70% fell within 5%, and over half fell within 2%.
Overall, average artist-standardised error was 6%, and industry-standardised error (calculated by standardising each artist’s predicted and actual values against performance range of the industry as a whole) was just 3%.
Taken together, our results suggest that our algorithm can predict commercial performance of a diverse range of new music on the day of it’s release with a high degree of accuracy. Over the coming weeks and months we’ll be digging deeper into this dataset to demonstrate how accurate our algorithm can really be when it comes to predicting new music on the day of its release (or even beforehand..).
If you have any questions or comments about these results, or would like to learn more about how we calculated them, we'd love to hear from you! For more general info about how Hyperlive quantifies listener engagement and forecasts future performance, check out our Home or About pages.
About Hyperlive. Music isn’t something we just listen to — music is something we experience with our heart and soul. Hyperlive captures this experience by modelling a range of neurobiobehavioural responses to music as well as the psychological processes that underpin them. This gives us a deep understanding of what drives musical engagement on a fundamental level, allowing us to quantify, model and predict that engagement — and the musical features that motivate it — with unmatched levels of precision.