Assessing Box Office Performance Using Movie Scripts: A Kernel-Based Approach

Abstract

We develop a methodology to predict box office performance of a movie at the point of green-lighting, when only its script and estimated production budget are available. We extract three levels of textual features (genre and content, semantics, and bag-of-words) from scripts using screenwriting domain knowledge, human input, and natural language processing techniques. These textual variables define a distance metric across scripts, which is then used as an input for a kernel-based approach to assess box office performance. We show that our proposed methodology predicts box office revenues more accurately (29 percent lower mean squared error (MSE)) compared to benchmark methods.