Prediction of Jazz Tracks' Success Using Supervised Machine Learning Techniques

Hit song prediction is a widely studied concept that is still considered to be lacking in methodological diversity. Over the last 20 years, the trends in the music industry have significantly changed, from the production and perception to the appearance of new artists. Music streaming revenue on platforms like Apple Music, Spotify, Deezer, Pandora, etc. accounted for 83% of the total music revenue in the US while only 9% came from physical sales. This seems to be a promising trend for independent artists as it lowered barriers to market entry. Although many predictive models have been proposed to predict music success, only few did the predictions based on the pre-release information. This paper strives to fill the gap by exploring the predictability of jazz tracks' success defined as making it to the top charts on Billboard relying solely on the features that are available prior to the track release. A total number of 17,511 observations have been collected. Several dimensions are examined, i.e., audio features, the popularity of the song’s artist, label the song is released under, as well as time-specific and artists-specific controls. To make the predictions, different classification machine learning techniques are applied, i.e., Logistic Regression, Random Forests, Support Vector Machine, and K-Nearest Neighbours. Based on the results, The Support Vector Machine and the Random Forest are the best performing models with the accuracy level of 98% and 95.3%, and AUC of 87.6% and 98.5% respectively. It is also found that that there is a non-linear relationship between the audio characteristics and the success of jazz tracks. In addition, despite the historical dominance of major record labels, jazz tracks released with independent labels tend to be more successful.