Essays on recessions through machine learning techniques

Resumo

This thesis consists of two essays on recession forecasting using machine learning techniques. In the first essay, we seek to evaluate the applicability of boosting models, not merely as tools to predict economic recessions but also as mechanisms for ranking the predictive efficacy of different variables. Building on the pioneering work introduced by Ng (2014), we have expanded the pool of potential predictors. Our enhanced model integrates widely used indicators, alternative measures of labor market slack, and a selection of aggregated series to assess their ability to enhance forecasts. Our findings reveal that the yield curve slope typically offers a more precise prediction of recessions than yield curve spreads. In another discovery, the firm-side equivalent unemployment rate — an underexplored variable that factors in job openings and quit rates — emerged as a potent recession predictor. The second essay of this thesis provides a comparative analysis between traditional logit models and more contemporary LASSO, Ridge and Elastic Net regressions. The study utilizes an extensive data set, and its dimension was reduced applying both principal component analysis and factor analysis. In addition, we also employed the regularized techniques to disaggregated series, a process not feasible with standard logits. Although we did not find a definitive advantage concerning forecasting ability by using regularized techniques, they proved to be significantly more efficient for handling data-rich environments. The research further suggested that implementing two types of dimension reduction is key to better interpret the relationship between variables. Our findings indicate a two-year lag where a strong housing market and higher short-term interest rate spreads are usually indicators of a forthcoming downturn. Moreover, one year prior, the earliest indicators suggestive of a recession within four to six quarters include average hours worked, unfilled durable orders, and declining stock markets. Overall, our findings reinforce that machine learning techniques - now increasingly being used in macroeconomics - present a robust alternative to conventional models. Given their design, these techniques facilitate the use of extensive data sets and a more streamlined selection of variables. They can also yield credible estimates of the likelihood of a recession within a specified timeframe. This allows a comprehensive and dynamic understanding of economic movements, which is essential given that recession are infrequent but consequential events, and economists currently have to deal with a vast array of indicators.

Ficha do documento

Tipo

Tese

Ano

2023

Instituição

Fundação Getulio Vargas

Fonte

Repositório da FGV

Idioma

Inglês

Acesso

Acesso aberto

Identificador

oai:repositorio.fgv.br:10438/34128

Temas

EconomiaModelagem EconômicaDados

Palavras-chave

Machine learningBoostingRecessionLASSORecessãoEconomiaAprendizado do computadorRecessão (Economia)Modelagem de dadosAnálise de regressão