LLM-Grounded Visual Alerts for Humanitarian Food-Price Crisis Dashboards: Reproducible Anomaly Detection on WFP VAM Market Data

Abstract

Humanitarian non-governmental organizations need food-price dashboards that turn volatile market data into alerts that field teams, donors, and accountability officers can read quickly. This paper presents and evaluates a reproducible food-price anomaly detection workflow for an NGO dashboard built on World Food Programme Vulnerability Analysis and Mapping market data. The experiment uses a pre-2022 public WFP food-price CSV subset covering Afghanistan, Benin, Burkina Faso, Burundi, Chad, the Democratic Republic of the Congo, Ethiopia, Kenya, and Lebanon from 2000 to 2021. After deterministic filtering, the dataset contains 76,822 valid retail price records, 526 eligible market-commodity time series, and 12,813 held-out observations for 2019-2021. A controlled anomaly benchmark injects 1,239 reproducible price shocks into the test period while preserving the original clean series for operational alert review. Seven detectors are evaluated: percent-change, exponentially weighted moving average residuals, six- and twelve-month rolling median absolute deviation, seasonal median absolute deviation, trend-seasonal residuals, and a rank ensemble. The exponentially weighted moving average detector achieves the strongest F1 score, 0.5146, with precision 0.8510, recall 0.3688, AUROC 0.9303, and AUPRC 0.7116. The paper also implements a bounded LLM-style warning-copy generator that verbalizes only measured variables, producing map labels, severity tags, and donor-facing summaries without adding unsupported claims. The resulting dashboard design prioritizes auditability: every alert links to country, market, commodity, month, price, anomaly score, month-to-month change, threshold, and verification text. The results show that a transparent, high-precision visual alert pipeline can support NGO food-security monitoring while limiting false escalation.