Consumer Review Sentiment Analysis and Market Trend Prediction Based on GPT-3

Abstract

Existing market-share forecasting approaches often rely on simplified sentiment proxies that fail to capture nuanced semantics and long-range contextual dependencies in consumer reviews, limiting their predictive power in dynamic ecommerce environments. To address this gap, this paper proposes an integrated framework that uses GPT-3 for highfidelity sentiment time-series extraction and embeds these signals into a Temporal Fusion Transformer for forwardlooking market-share prediction in the textile sector. GPT-3 generates confidence-weighted monthly positivity and net sentiment scores, which are fused with 10 additional features covering static product attributes, historical review dynamics, and future-known inputs such as public holidays and major promotional events. Experimental results show that GPT-3 achieves a Macro-F1 of 0.842 in sentiment classification, outperforming RoBERTa and VADER, especially in neutral sentiment recognition, where F1 reaches 0.789. The TFT+GPT-3 model achieves MAE of 0.024, RMSE of 0.032, MAPE of 4.29%, and directional accuracy of 83.7% in market-share prediction. Ablation analysis shows that removing GPT-3-derived sentiment features increases MAE to 0.035, confirming their critical contribution to forecast accuracy and interpretability.