A Review of Sales Forecasting Research Based on the Fusion of Online Consumer Sentiment and Time-Series Data

Abstract

Sales forecasting in digital markets is no longer driven only by historical sales, price and promotion records. Online reviews, social media posts and other user-generated texts often appear before sales changes become visible, so they may provide early clues about attention, satisfaction and purchase intention. In this paper, online consumer sentiment is used as an umbrella term for affective signals extracted from these different forms of user-generated content. The paper reviews sales forecasting research based on the fusion of online consumer sentiment and time-series data. The discussion is organized around three connected layers: sentiment modeling, time-series forecasting and multimodal fusion. The reviewed studies suggest that sentiment features are most useful when demand is affected by reputation, product experience or online discussion, whereas time-series models remain essential for trend, seasonality and autocorrelation. Feature-level fusion is easy to implement, decision-level fusion is robust to noisy inputs, and deep or attention-based fusion can capture richer interactions when enough data are available. The paper also points out several practical limits, including noisy sentiment extraction, temporal misalignment, shallow fusion design, weak interpretability and deployment difficulty. Future work should pay more attention to aspect-level sentiment, lag structure, explainable forecasting and lightweight systems that can be transferred across product categories.