Examining Whether Commonly Used Acoustic-Based Hit Prediction Methods Can Be Applied to the Dutch Top 40

Abstract

Hit prediction for global top charts is frequently studied within the research area referred to as Hit Song Science. Hit Song Science is concerned with predicting the hit potential of a song. As of now, research to hit prediction on local charts remains relatively unexplored. This thesis examines whether methods that are commonly used for hit prediction on global charts can be applied to the Dutch Top 40. The problem is scoped to acoustic based features. To address this issue a novel approach on non-hit song selection has been deployed; only non-hits that appear on a release album with at least one hit are selected. Additionally, three experiments are designed by each following an influential hit prediction study. A Wide and Deep Network (E1), Random Forest (E2), and Support Vector Machine (E3) were tuned, trained, and assessed. Showing that the predictive performance of acoustic-based audio features is similar between the Dutch Top 40 and Billboard Top 100. Each experiment has a min-f1 score of approximately 0.58 for both charts. This is more accurate than a random prediction, and still provides substantial potential for improvement. Future research could dive into what additional features complement the features used in this thesis and improve the predictive performance of the experiments.