This repository contains the data and code used for our master's thesis submission for Data Science Decision Making 2023. The project, conducted in partnership with the United Nations High Commissioner for Refugees (UNHCR), is a prediction task for predicting yearly refugee flows between dyads of countries.
data: The data folder contains all the data sources that were used for this project, with the exception of UNHCR data, which was shared under discretion. It contains aclean_data.ipynbfile, which is how therawdata is processed and subsequently moved to thecleanfolder.notebooks/models: This folder contains the three main notebooks that implement refugee outflows and flow predictions, along with helper functions files. It is the key part of the repository.notebooks/trends: Within this folder all of the Google Trends downloading and pretesting takes place. You will find different notebooks, helper functions files and helper csv files as well. As this was done in different stages, we recommend approaching the folder with caution. Feel free to consult the authors for specific queries if you need to.notebooks/descriptive_analysis: Some initial descriptive analysis notebooks can be found here.document: Here you will find the final document of our master's thesis.
This study presents a novel forecasting framework for global refugee flows, incorporating non-conventional data sources such as Google Trends, the GDELT project event dataset, conflict forecasts, among others. Our main objective is to generate accurate one-step ahead predictions for the number of new refugee arrivals per country pair. These predictions play a crucial role in facilitating effective humanitarian response and informed infrastructure planning. While existing literature focuses on developed countries, this research develops a comprehensive global model, considering the majority of refugees seeking refuge in neighboring countries. In addition, to overcome challenges with imbalanced and low-frequency data, two strategies are proposed using a rolling window framework: modeling refugee outflows from origin countries (outflow level), and modeling refugee flows between country pairs (dyad level). Our results reveal a significant improvement in prediction accuracy by augmenting traditional variables from UNHCR with high-frequency non-conventional data, with Random Forest and Gradient Boosting as effective regressors.
Daniela De los Santos - Barcelona School of Economics - daniela.de@bse.eu
Eric Frey - Barcelona School of Economics - eric.frey@bse.eu
Renato Vassallo - Barcelona School of Economics renato.vassallo@bse.eu