From Measurement to Prediction: Supervised Topic Modelling and Ex-Ante Forecasting of Behavioral Customer Engagement

In the digital era, with the growing importance of social media platforms, this study advances customer engagement (CE) research by moving from ex-post measurement to ex-ante prediction. While recent social media CE literature has proposed sophisticated measures of CE, most indices are based on public metrics, leaving most of the work descriptive and retrospective in nature. However, marketing practitioners require predictive capabilities to optimise content and timing decisions before posting social media content. Responding to that challenge, this paper extends the CE measurement approach into a predictive modelling framework. The innovation lies in utilising supervised topic modelling (STM), specifically Anchored CorEx, to generate probabilistic indicators of content levers that are systematically integrated into predictive pipelines. The probabilistic lever features (Entertainment, Trendiness), combined with time and brand metadata, are incorporated into five predictive families: Random Forest, XGBoost, Keras MLP, OLS, and KNN. Results on an authentic luxury brand Twitter dataset (2017–2022) show that the proposed model yields satisfying predictive performance, improving both accuracy and rank validity. Notably, the predictive results validate the entropy-computed behavioral CE measure: the objective weighting and uncertainty-sensitive nature of Shannon entropy enable it to extract meaningful engagement signals from inherently noisy social media interactions. This, in turn, provides practitioners with a reliable basis for ex-ante content and timing decisions, strengthening data-driven strategic planning in dynamic online environments.