LLM Sentiment Analysis-Integrated Sales Forecasting for E-Commerce—A Case Study on Amazon

In dynamic e-commerce scenarios, conventional sales forecasting models struggle to extract market sentiment embedded in online reviews. To address this issue, the paper proposes a novel forecasting framework that incorporates sentiment features extracted by large language models (LLMs) through few-shot prompting. The framework conducts both binary sentiment polarity classification and continuous sentiment quantification. Three feature settings are evaluated: a baseline model without sentiment features, a binary sentiment index, and a continuous sentiment index. These are fused with historical sales, price, and promotion data, and tested on five typical models (SARIMAX, Prophet, LSTM, XGBoost and LightGBM) using datasets from six Amazon products. Results indicate that 4-shot sentiment extractor based on DeepSeek-V3.2 yields an F1-score of 0.966. When integrated into sales forecasting pipelines, sentiment features consistently reduce forecasting errors. The continuous sentiment index further improves the performance of tree-based models, whereas the binary index exhibits universal effectiveness. Sentiment-enhanced XGBoost and LightGBM deliver the best overall performance. The proposed framework offers a scalable solution for e-commerce inventory and supply chain decisions.