Predicting food crises using news streams.

Area of Science:

• Computational Social Science
• Food Security Analysis
• Machine Learning Applications

Background:

• Effective food crisis prediction is essential for timely humanitarian aid allocation and mitigating suffering.
• Current predictive models often use delayed, incomplete, or outdated risk measures.
• There is a need for high-frequency, interpretable indicators for early food crisis detection.

Purpose of the Study:

• To develop and validate a novel approach for anticipating food crises using deep learning on news text.
• To extract interpretable, high-frequency precursors to food crises from a large corpus of news articles.
• To assess the improvement in food insecurity prediction accuracy using text-derived indicators.

Main Methods:

• Leveraged deep learning techniques on 11.2 million news articles (1980-2020) from food-insecure countries.
• Extracted interpretable, high-frequency precursors to food crises.
• Validated extracted indicators against traditional risk measures.
• Compared predictive performance of news-derived indicators against baseline models.

Main Results:

• News-derived indicators significantly improved district-level food insecurity predictions up to 12 months ahead (July 2009-July 2020) across 21 countries.
• The model demonstrated substantial improvements relative to baseline models lacking text information.
• Extracted precursors were interpretable and validated by traditional indicators.

Conclusions:

• News text analysis, powered by deep learning, offers a powerful tool for early food crisis anticipation.
• This approach enhances predictive accuracy for humanitarian aid allocation in data-scarce environments.
• Machine learning applications in analyzing unstructured data can significantly improve decision-making for global challenges.