Forecasting New Follower Growth Using ARIMAX, LSTM, and Stacking Fusion

Abstract

Against the backdrop of intensifying competition in social media content, new follower growth better reflects the conversion from content exposure to sustained engagement than metrics such as views, likes, and comments. This study uses daily operational data from influencers on a social media platform as the research sample. It incorporates metrics such as views, likes, comments, and new follower growth, and further constructs the interaction rate and day-of-week variables. ARIMAX, LSTM, a static weighted fusion model, and a Stacking fusion model are constructed to predict next-day new follower growth. Model performance is evaluated using mean squared error, root mean squared error, and mean absolute error. The results indicate that while ARIMAX can capture some linear relationships, its overall predictive performance is weak; LSTM significantly outperforms ARIMAX; the static weighted fusion model does not yield additional improvements; and the Stacking fusion model performs best. The study demonstrates that fusion methods can more effectively integrate information from different models to improve the accuracy of new follower growth forecasting, providing valuable insights for optimizing influencer operations and identifying creators on the platform.