Sales Prediction Based on Textual Features of Online Product Reviews

Abstract

Accurate sales forecasting is essential for effective enterprise decision-making. Reliable predictions not only optimize inventory management and reduce resource waste but also enhance customer experience. While prior research has primarily focused on improving traditional time series models, relatively limited attention has been given to forecasting sales based on textual features extracted from online product reviews. This study addresses this gap by incorporating consumer-generated reviews into a sales prediction framework. Using Amazon.com as a case study, we analyze the open-source Amazon food review dataset. Through feature engineering, three categories of textual variables are constructed: review topics (extracted via the Gensim library and Latent Dirichlet Allocation, LDA), star ratings, and review usefulness. These features, together with lagged sales, are used as inputs into a ridge regression model. Experimental results show that the proposed model achieves an R² of 0.88 on the training set and 0.78 on the test set, indicating both feasibility and strong predictive accuracy. Compared with traditional time series methods, the review-based approach leverages text mining to capture consumers’ genuine perceptions and market responses. This innovation enhances forecasting accuracy and offers theoretical as well as practical implications for enterprises, including improved sales planning, more adaptive market strategies, and more efficient warehouse and inventory management.