Leveraging News Data for AI-powered Food Insecurity Forecasts
October 28 | Berkeley, CA
Ananth Balashankar (Google DeepMind, work done while at NYU)
Coauthors: Lakshminarayanan Subramanian (NYU), Samuel Fraiberger (World Bank)
Panelist
Challenge: crises preparedness and response in fragile contexts
Sources: McDougal & Patterson (2021), Global Network Against Food Crises (2024), The World Bank (2024)
2 billion people live in Fragility, Conflict, and Violence (FCV) settings
Estimated $367 billion spent on domestic and international humanitarian aid in 2023
Anticipatory transfers can increase impact of aid and lower humanitarian cost by up to 30%
282 million people in 59 countries experienced high acute hunger in 2023
2
Status quo: existing early warning tools are limited by data gaps
Expensive
Delayed
Incomplete
Survey-based
3
Solution: scalable AI system using news for food crisis predictions
Anticipatory signals of food crises appear in on-the-ground news before traditional data
Our forecasting model can identify +46% food crisis outbreaks up to 12 months ahead relative to existing approaches such as Lentz et al. (2019)
Our AI models can convert news data into high-frequency food crises risk indicators at the district level globally
+46% outbreaks identified
4
Methodology: Extracting risk factors from news
Source: Swayamdipta et al., (2017)
Cause
Keyword Effect
Semantic frame parsing
Causation frames
Peer-reviewed articles on famine
News at ADMIN-2 District Level from 2009-20
Expansion around word2vec clusters
Granger-causality with FEWS levels
5
Methodology: Forecasting model
Methodology: Extracting news indicator factors
6
Methodology: a new perspective to the study of food insecurity
Once we identify these news indicators (“News”), we combine them with traditional food insecurity indicators (“Traditional”) in forecasting models (Random Forest Regression)
7
Rainfall anomaly
News mentions of climate shocks
Interpretability: why and how does our approach work?
Percentile of value
IPC Phase
Legend
Extreme drought struck Ethiopia in 2009. IPC in the Majang zone raised from a normal to a crisis phase in July 2010 when the precipitation levels were normal
News mentions of climate shocks (“drought”, “failed rain”, “abnormally low rainfall”, “prolonged dry spell”, etc.) peaked 6 months before the crisis and enabled our model to correctly predict the outbreak 4 months ahead
Machine learning predictions of food crises are interpretable by enabling to trace back to the underlying cause of the outbreak mentioned in the news
Rainfall anomaly
News mentions of climate shocks
Food crisis outbreak
IPC (Ground Truth)
Traditional
Trad. + News
Trad. + News
(excl. climate terms)
8
AI-powered public good to better anticipate and respond to food crises; scalable, real-time, interpretable and cost-effective
Inform World Bank’s food systems resilience programs across Africa ($3 billion in 16 countries over 10 years) and the Crisis Response Window ($2.5 billion)
Prioritize allocation of emergency food assistance to improve preparedness and reduce human suffering
Complement existing data products for food security incl. Hunger Map and the Global Food and Nutrition Security Dashboard
Takeaway: better forecasting, better policies
9
Potential extensions: More local, and anticipatory actions
Radio mining as localized mirror of public sentiments
AI-based anticipatory cash transfers for resilience
SMS-based alerts on emerging food crises to beneficiaries
Additional languages to enhance linguistic representation
Credit: Philipp Zimmer & Samuel Fraiberger, World Bank
10
11
Appendix
12
13
Adoption: how the World Bank is now thinking about bringing the insights to policymakers
* Frontend demo with dummy data
14
15
16
17
18
19
20
21
22