KeyTrend: Automated Keyword Synthesis via LLMs for Demand Forecasting with Google Trends

Google Trends (GT) can provide useful external signals for demand forecasting, but choosing the right keywords is usually done manually and depends heavily on domain knowledge. In this work, we introduce KeyTrend, a framework where an LLM generates GT keyword candidates from e-commerce websites, and a correlation-based filter selects the most relevant ones. We evaluate on four real-world daily logistics shipment datasets from a major Canadian freight forwarder and find that a fixed set of 25 filtered GT keywords consistently improves XGBoost SMAPE across all datasets, even without any hyperparameter tuning. We also benchmark against foundation models (Chronos, TimesFM) and classical methods, and observe that no single approach dominates across all domains. The performance of the foundation models is similar to the statistical baseline. We also discuss practical lessons about prompt strategies and keyword selection that can guide future work in LLM-augmented forecasting.