IMF Decides, Machine Learns: An AI Approach to IOs

• In Event: International Collaboration on Economic Issues

Abstract

Why does the International Monetary Fund (IMF) favor some countries over others? Although the Fund is often criticized for its “one size fits all approach” as its programs to a large extent consist of a list of the conditions set, it is usually acknowledged that the IMF’s lending policies vary across time and space. The IMF is frequently argued to be an agent of its most powerful shareholders. Challenging the common belief that the strategic allies of the United States and/or G5 countries always get better deals from the IMF, whereas it is the IMF staff who has the main leverage over the design of conditionality when low-income countries are borrowing from the Fund, this paper will present a novel framework, drawing upon a rich body of literature on the IMF. Building a comprehensive model that takes into account all actors having an impact on IMF program design, the following questions are at the core of the research project this paper will present: What factors influence the terms (loan size, number of conditions, and conditionality waivers granted) of an IMF program? And how do those factors play into shaping the design of the programs? In this context, the project focuses on three particular aspects of IMF lending: (1) the size of IMF loans; (2) the number of conditions attached to the loans; and (3) the number of condition waivers granted to borrowing countries during program implementation.

This research creates a natural language processing (NLP) tool for automated, fast analysis of the IMF’s Executive Board meeting minutes, which is able to capture elements including individual board member sentiments, alliance between representatives of different countries and G5 stance with high accuracy. For this purpose, we have annotated over 50 thousand pages of IMF documentation matching our dataset covering IMF programs between 1978-2014. The contributions of this project to the literature on the IMF and study of international organizations (IOs) is four-fold:

1. This project develops a comprehensive theoretical model for understanding IMF program design and implementation that integrates all explanatory variables from the existing literature as well as novel ones to be introduced through the analysis of Executive Board meeting minutes.
2. The project develops machine learning-based models of the IMF lending process to provide high predictive power for the outcomes of program design under various macroeconomic and political conditions, providing a new methodological contribution to the field, which has traditionally used statistical prediction models.
3. The research takes on board an eclectic approach, using mixed methods involving Machine Learning, Natural Language Processing, statistical analyses, as well as process tracing of in-depth case studies to account for the variation in the terms of IMF programs. Most of the scholarly work done so far in this literature are either based on statistical analyses or case studies. To the best of our knowledge, this will be the first research using the said methods together.
4. By developing a tool for automated analysis of Executive Board meeting minutes to extract data for Executive Board views on IMF program design, the project will greatly facilitate future research on international organization document analysis. The machine learning (ML) framework developed is also an important tool for political scientists to apply ML-based modeling to different problems in IOs and other domains alike.

Author

• Merih Angin, Koç University