Canaries in the coal mine: Boosted machine learning for the classification of recessions in the eurozone business cycle

Evans-Kaplan, Jordan

Permalink

https://hdl.handle.net/2142/109534 Copy

Description

Title

Canaries in the coal mine: Boosted machine learning for the classification of recessions in the eurozone business cycle

Author(s)

Evans-Kaplan, Jordan

Issue Date

2020-12-10

Director of Research (if dissertation) or Advisor (if thesis)

• Henry, Lucas A
• Parente, Stephen L

Department of Study

Liberal Arts & Sciences

Discipline

European Union Studies

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.A.

Degree Level

Thesis

Keyword(s)

• Euro
• Eurozone
• EU
• European Union
• Recession
• Great Recession
• 2008 Recession
• Machine Learning
• Boosting
• XGBoost
• Economic Contagion
• Classification
• Early Warning
• Canary
• Greece
• Cyprus
• Austria
• Shadow Economy
• Underground Economy
• Business Cycle
• Forecasting
• Double-Dip Recession
• Shapley Value
• Real Business Cycle Theory
• Periphery
• European Periphery

Abstract

This thesis considers the question of which eurozone nations are the most effective early warning predictors of general recession within the eurozone. This research challenges the idea within the literature that the largest economies (Germany, France, Italy) and their various industrial indices make for the most accurate predictive leading indicators. This assumption demonstrates a significant gap in the literature, which informs the policy of the European Central Bank as well as the EU’s member states. Research into the early warning signals surrounding major European recessions is not just critical for the euro’s stability; it helps identify and prevent economic contagion, keeping a recession from spreading at the very first symptoms. From the introduction of the euro to the 2008 recession, the euro enjoyed a period of near uninterrupted stability and convergence. Economists have had great difficulty analyzing what can be credited with this great moderation in bond rates and GDP. This paper seeks to analyze the relationship between the GDPs of individual eurozone nations and the eurozone business cycle using a gradient boosted decision tree-based machine learning framework. To facilitate this analysis, XGBoost was used to provide a gradient boosting framework. Using quarterly GDP values for all of the EA19 (with the exception of Ireland due to data availability), business cycle data was produced using the Bry-Boschan algorithm, coupled with seasonal and calendar smoothing via Eurostat. In order to identify the nations which serve as the best predictors of a recession, the quarterly GDP values for the eurozone (except Ireland due to data availability) were compared via an iterative boosting algorithm known as XGBoost. The result of this model is a hierarchical tree structure which was then used to produce a list of the most effective predictors, ranked by Gini importance and SHAP scores. From the top six nations, or the two most critical clusters out of three, I selected three nations for an individual case study due to their unique and highly effective interactions with the likelihood of recessions in the eurozone. These nations, Greece, Austria, and Cyprus represent unique and underutilized perspectives for the successful classification of recessions. This paper’s title, “Canaries in the Coal Mine,” is an apt comparison in the case of Greece. Further, when analyzing the countries belonging to the top two clusters, it was noted that only two of the six nations have shadow economies below 25% of their total GDP. As such, the potential classificatory relationship between the selection process and size of the informal economy was explored via a multi-disciplinary methodology informed by both European Union Studies and machine learning. To analyze the “black-box” nature of the machine learning model utilized, the marginal contributions of each nation were assessed as if they were players in a coalitional game, using a game theory technique known as Shapley Values. From this, each nation’s “optimal payout” is calculated, which corresponds to that nation’s marginal effect on the decision process. This combinatorial method serves to both provide a new perspective in the country variable-selection process for Eurozone recession prediction/classification, as well as to form statistical inferences that give new insight into the potential causes and signals of recessions. The result of this modeling process is the highest known accuracy in recession classification found in the literature from the period Q1 2000 to Q3 2019, as well as a set of nations well-suited to the accurate classification of recessions.

Graduation Semester

2020-12

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/109534

Copyright and License Information

Copyright 2020 Jordan Evans-Kaplan

Owning Collections