Can Machine Learning Catch the Covid-19 Recession?

Abstract

Based on evidence gathered from a newly built large macroeconomic dataset (MD) for the UK, labelled UK-MD and comparable to similar datasets for the United States and Canada, it seems the most promising avenue for forecasting during the pandemic is to allow for general forms of nonlinearity by using machine learning (ML) methods. But not all nonlinear ML methods are alike. For instance, some do not allow to extrapolate (like regular trees and forests) and some do (when complemented with linear dynamic components). This and other crucial aspects of ML-based forecasting in unprecedented times are studied in an extensive pseudo-out-of-sample exercise.

A large Canadian database for macroeconomic analysis

“…Figure 3 plots the number of factors selected recursively by Bai and Ng (2002) and Hallin and Liska (2007) methods. 8 We observe that the number of static and dynamic factors is generally increasing since 1990, a similar pattern found with other large macroeconomic datasets (McCracken and Ng, 2016;Goulet Coulombe et al, 2021). Many explanations on the time-varying nature of the number of factors are plausible: structural changes in terms of the correlation structure, presence of group-specific factors, finite-sample sensitivity of selection procedures, and so on.…”

Forecasting in the absence of precedent

“…2021) find that machine learning methods improved forecasting performance over certain benchmark linear models similar to Equation (1) during the pandemic, it is not clear how these methods compare to parametric extensions to Equation (1), such as those discussed in Section 3.1. In addition, machine learning methods and nonlinear models more generally have not been found to have systematic benefits over linear models before the pandemic (see Ferrara et al, 2015 andCoulombe et al, 2021 for details). 2 The pandemic triggered an explosion of papers connecting macroeconomic and epidemiological outcomes (e.g., Acemoglu et al, 2020;Alvarez et al, 2021;Eichenbaum et al, 2020).…”

“…The time‐varying parameter (TVP) VAR (e.g., Primiceri, 2005; Lubik & Matthes, 2015) takes the form () but allows B and C to vary over time. More involved methods, especially from the field of machine learning, have been used for macroeconomic forecasting before and during the COVID‐19 recession (e.g., Coulombe et al., 2021; Huber et al., 2020). While we do not provide a detailed description of these methods here, we note that these methods introduce a high degree of flexibility for the variables $$X_{t}$$ to evolve differently during contrasting episodes, which can be thought of as less parametric approaches to the ideas in Section 3.1 1…”

Big data forecasting of South African inflation

“…To address some of the challenges that may arise when looking to forecast macroeconomic variables, following a significant structural change or large departure from steady-state values, Galvao (2021) summarises a number of developments from the international literature, while Castle et al ( 2021) and Coulombe et al (2021) note that statistical learning models that are able to adapt to various changes may perform better than well-specified structural models. 4 In addition, the use of statistical learning models that incorporate nonparametric nonlinear features have gained significant attention over recent periods of time, partially due to the fact that they may be applied to large datasets to yield impressive results.…”

“…In addition, for the period that incorporates the financial crisis,Stock and Watson (2010) suggest that this model should incorporate a stochastic trend that reacts to the unemployment recession gap, where the short-term response of inflation is consistent with an increase in this gap, while the long-term response is dependent upon the persistence in trend inflation. As is the case with most low-and middle-income countries, South Africa does not have a reliable measure for the unemployment recession gap that could be applied in an investigation that makes use of monthly observations of time-series variables.4 In particular,Coulombe et al (2021) advocate for the use of nonlinear statistical learning models, while in a similar investigation,Koop et al (2021) suggest that modelling specifications that accommodate time variation in forecasting uncertainty may also provide improved results.…”