Poverty Prediction Challenge: Advancing Real-Time Monitoring of Poverty through Data Science

Overview

Timely and reliable poverty data are essential for effective policymaking. Yet in many countries, household consumption surveys, the backbone of monetary poverty measurement, are conducted infrequently due to cost and logistical constraints. As a result, policymakers often lack up-to-date information on household welfare, particularly during periods of rapid economic change.

To help address this gap, the World Bank launched the Poverty Prediction Challenge in November 2025, a global data science competition designed to advance methods for predicting household consumption and poverty across surveys. Hosted on the DrivenData platform, the Challenge invited participants from diverse analytical backgrounds to develop models that could estimate household expenditure and predict poverty rates using survey data.

The competition closed on February 4, 2026, with strong global engagement: 1,322 participants registered and more than 500 valid solutions were submitted.

The Challenge

The Poverty Prediction Challenge was designed as a supervised learning task. Participants were provided with test survey datasets and asked to:

• Predict household-level expenditure using available survey information (for example, household size, education, employment, housing conditions, and access to services).
• Use those predictions to estimate poverty rates, defined as the share of the population living below a given expenditure threshold.

Participants were free to use any modeling approach, including machine learning, classical econometric techniques, ensemble methods, or hybrid approaches. The competition was open to statisticians, data scientists, researchers, students, and practitioners from around the world.

The objective was not only to identify high-performing models, but also to explore how different approaches perform when using older survey data to estimate poverty in newer surveys, with particular attention to how well these methods work across different contexts.

Why This Matters

Traditional household surveys remain the gold standard for measuring poverty. However, because of their periodic nature, they are less useful for governments seeking to respond quickly to shocks, economic downturns, or emerging vulnerabilities.

Improving methods to predict consumption and poverty between survey rounds can:

• Support more frequent updates of poverty estimates.
• Inform adaptive policy design during crises.
• Strengthen monitoring of distributional impacts.
• Enhance the cost-effectiveness of statistical systems.

The Challenge contributed to the World Bank's broader Real-Time Monitoring agenda, which aims to combine traditional data sources with innovative methods to provide more timely and policy-relevant welfare indicators.

Winning Solutions

Following verification, the top-performing teams were:

Top-performing solutions relied on advanced data-driven methods and careful testing approaches to ensure their predictions worked well across different survey contexts.

The code for all winning solutions, including full model implementations and documentation, is publicly available in the official repository.

Key Lessons (Preliminary)

Several key insights emerge from the Challenge:

• Predicting poverty across surveys requires approaches that can perform well even when the data changes from one survey to another.
• Hybrid approaches can improve performance in complex, heterogeneous datasets.
• Adjusting model outputs (known as "calibration") plays a critical role in ensuring that predicted poverty rates accurately reflect real-world population patterns.
• Transparency and reproducibility are essential when translating predictive models into policy-relevant tools.
• Competition-based approaches can accelerate methodological innovation by engaging a broad analytical community.

A forthcoming working paper will provide a detailed technical analysis. These lessons will inform future work on survey-to-survey imputation and the integration of predictive methods into official poverty monitoring systems.

Participation and Submissions

The Challenge attracted a highly diverse pool of participants across countries and disciplines and featured a wide range of approaches.

All models were evaluated using predefined metrics that combined household-level prediction accuracy with performance in estimating poverty rates across multiple thresholds. Final rankings were determined through a rigorous verification process to ensure reproducibility and methodological soundness.