RDP 2024-04: Nowcasting Quarterly GDP Growth during the COVID-19 Crisis Using a Monthly Activity Indicator 1. Introduction

What is happening in the economy now? It is said that the future is uncertain, but so is the present. Policymakers operating in this environment need some way to understand what is happening now (i.e. current economic conditions). This need for timely information was most acute during the COVID-19 crisis when current conditions were rapidly evolving, requiring policymakers to make decisions under significant economic uncertainty. Policymakers are further hamstrung because the most comprehensive measure of economy activity, gross domestic product (GDP), is published with a substantial lag. Indeed, the full effect on economic activity of the first major lockdowns which occurred during June quarter 2020 were not realised until the release of National Accounts data in early September 2020; more than two months after the reference period.^[1] This delay limits its value to policymakers as a measure of the current state of the economy. Additionally, GDP is often revised in subsequent quarters which further limits its usefulness to policymakers for assessing current conditions.

In response, more and more higher-frequency partial indicators have become available in recent times; however, they are often not as comprehensive in their scope and coverage as traditional measures of economic activity such as GDP. And while these partial indicators do help fill the information gap, the signal they provide is often noisy. Further, one indicator might be useful in one context but not in another. For example, the unemployment rate is considered a key metric of economic activity, but during the COVID-19 crisis the Australian Government introduced the ‘JobKeeper’ program to keep workers employed, thereby limiting the rise in the unemployment rate caused by lockdowns.^[2] During this period, the underemployment rate was considered to provide a more accurate signal. Given this, it is not clear how policymakers should choose which indicator to focus on, and if there are multiple indicators available, how much weight they should give to each one. The answers to these decisions are subjective and will typically vary with time and across policymaker.

What is required is a method of combining the available partial indicators in a systematic manner to smooth out the noise and reveal the underlying signal. The most common tool for achieving this is via dynamic factor models (DFMs). DFMs are a dimension reduction technique that can summarise the common variation across a panel of time series data.^[3] In Australia, initial work exploring the usefulness of factor models for monitoring economic activity was undertaken by Gillitzer, Kearns and Richards (2005). They produced two coincident indicators, one summarising quarterly data and another summarising monthly data. Both indicators were estimated using the non-parametric methods developed by Stock and Watson (2002) and Forni et al (2000).^[4] This was followed by Sheen, Trück and Wang (2015), who introduced a daily business cycle indicator based on the work of Aruoba et al (2009). Their method uses a parametric estimation technique involving a state-space model estimated using the Kalman filter.

Besides being a successful tool for monitoring activity, another important use of DFMs is for prediction (Stock and Watson 2002), especially in relation to producing nowcasts.^[5] A significant amount of research effort has been devoted to this topic since the early works of Nunes (2005) and Giannone, Reichlin and Small (2008) (see Bańbura et al (2013)). However, there has been considerably less work done in Australia. For prediction, Gillitzer and Kearns (2007) had success, showing factor-based forecasts for key macroeconomic series can outperform standard time series benchmarks.^[6] The benefits of DFMs are less clear when focusing explicitly on nowcasting in Australia (Australian Treasury 2018; Panagiotelis et al 2019). Using different estimation techniques both suggest the sample mean is a difficult benchmark model to beat in relation to nowcasting quarterly GDP growth.^[7] But, while both works consider higher frequency data (monthly and higher), neither exploit this information in their nowcasting investigation. Instead, both convert all series in their respective datasets to a quarterly frequency before producing a nowcast.^[8],[9] This is a problem because it represents a loss of information. Further, there is extensive research highlighting the significant improvement in prediction accuracy that comes from working with mixed frequency data.^[10]

Our work bridges the gap that exists in the literature between employing factor models for monitoring and for nowcasting in Australia. Both issues are interrelated and are equally important for policymakers, so it is sensible to develop a framework that can achieve both objectives at once. In doing so, we build on previous work in Australia by incorporating more recent developments in factor modelling and nowcasting. We start by developing a monthly activity indicator (MAI) for Australia. The MAI aims to provide policymakers with a more immediate snapshot of prevailing economic conditions. We achieve this by using a ‘true’ DFM to summarise the information content from a dataset of 30 monthly targeted predictors selected for their ability to explain movements in first-release quarterly GDP growth.^[11] This is an important advance compared to previous studies as it links the variable of interest to the estimation of the DFM and has been shown to improve factor estimation and predictive ability (Bai and Ng 2008; Bulligan, Marcellino and Venditti 2015). We also extend the targeted predictor hard thresholding pre-selection step developed by Bai and Ng (2008) when estimating factor models to the mixed frequency setting. Further, the methodology we use to estimate the MAI allows us to use an unbalanced dataset which therefore means we can consider a broader collection of series over a longer time period than the competing indicators produced by Gillitzer et al (2005) and Sheen et al (2015).

Unlike previous nowcasting studies in Australia (e.g. Australian Treasury 2018; Panagiotelis et al 2019), which have focused exclusively on quarterly frequency data, we undertake the first investigation of nowcasting in Australia using a mixed frequency modelling framework. We exploit the MAI's high-frequency information content within a factor augmented unrestricted MIDAS (MIxed Data Sampling) model (FA-U-MIDAS).^[12] We assess the model's ability to nowcast first-release quarterly GDP growth using a recursive out-of-sample evaluation exercise covering a 34-year period (1988:Q2–2022:Q2), longer than previous works including Gillitzer and Kearns (2007), Australian Treasury (2018), and Panagiotelis et al (2019). Additionally, since we use monthly data, we can generate four nowcasts for each quarterly GDP growth observation as new monthly data becomes available across the quarter. Finally, as in previous evaluations, we use the standard benchmark forecasting models of the sample mean and an AR(1) process for comparison.

Our results show that incorporating monthly information provides more accurate predictions compared to the benchmark models based on smaller estimated root mean squared error. The improvement over the benchmark models (sample mean and AR(1) models) is also found to be statistically significant as well.^[13] Crucially, predictive accuracy of the models with monthly data is largest during the COVID-19 crisis compared to the benchmark models relying solely on quarterly data, highlighting the benefit to policymakers from having timely information. Our results also support previous findings which suggest that model predictive performance can change depending on the state of the economy (see Chauvet and Potter (2013), Siliverstovs (2020) and Jardet and Meunier (2022)).

We begin by describing in detail the methods we follow to construct the MAI in Section 2. In Section 3 we discuss how we use the MAI to predict quarterly GDP growth as well as the steps we follow to implement the out-of-sample evaluation exercise before concluding in Section 4. Some additional results are provided in the appendices.

Footnotes

The actual publication date was 2 September 2020 (see <https://www.abs.gov.au/statistics/economy/national-accounts/australian-national-accounts-national-income-expenditure-and-product/jun-2020>). [1]

The JobKeeper scheme was a wage subsidy for businesses introduced in March 2020 by the Australian Government to support the economy during the COVID-19 crisis. [2]

For a comprehensive review of DFMs, see Stock and Watson (2016) and references therein. Examples where DFMs have been used in policy institutions include: Matheson (2006); Aruoba, Diebold and Scotti (2009); Cunningham et al (2012); Bańbura and Modugno (2014); Higgins (2014); Bok et al (2017); and more recently, Lewis et al (2021). [3]

The method proposed by Stock and Watson (2002) uses principal components analysis (PCA) while the method developed by Forni et al (2000) uses dynamic PCA. [4]

That is, using higher frequency information to predict the current value of an (unpublished) lower frequency variable. [5]

They focused on quarterly data covering the period 1960 to 2005 and produced forecasts using a recursive scheme for growth in GDP, non-farm GDP, private final demand, household final consumption expenditure with horizons from 2, 4 and 8 quarters ahead. The benchmark model was an AR(1) process. [6]

Australian Treasury (2018) follows the Federal Reserve Bank of Atlanta's ‘GDPNow’ methodology (Higgins 2014) which uses a parametric model based on a state-space model to estimate the DFM. In contrast, Panagiotelis et al (2019) estimate a static factor model by PCA. To predict GDP, they use the factor-augmented (i.e. diffusion index) model approach of Stock and Watson (2002). [7]

Australian Treasury (2008) estimates a monthly factor initially but converts this to a quarterly frequency for use in bridging equations. Panagiotelis et al (2019) converts all monthly series in their dataset to quarterly before extracting any factors. Both implement temporal aggregation by taking the average of the three months in each quarter. [8]

Anthonisz (2021) is an exception and considers mixed frequency data in his analysis. [9]

For a non-exhaustive list, see Clements and Galvão (2008, 2009), Galvão (2013), Foroni and Marcellino (2014), Leboeuf and Morel (2014), Schorfheide and Song (2015), Ferrara and Marsilli (2019), Galvão and Lopresto (2020), Siliverstovs (2020), Baumeister and Guérin (2020), and Jardet and Meunier (2022). [10]

Bai and Wang (2015) define a true DFM as one that incorporates dynamics between the observed series and the factors. See Section 2 for more details. [11]

For MIDAS, see Ghysels, Santa-Clara and Valkanov (2004) and Ghysels, Sinko and Valkanov (2007). For unrestricted MIDAS see Foroni, Marcellino and Schumacher (2015) and for factor augmented MIDAS see Marcellino and Schumacher (2010), Ferrara and Marsilli (2019) and Jardet and Meunier (2022). [12]

Curiously, except for Anthonisz (2021), none of the previous studies in Australia conducted examinations comparing predictive accuracy using formal statistical tests. [13]