Nowcasting Economic Activity with Electronic Payments Data: A Predictive Modeling Approach *
Predicao imediata da atividade economica com dados de pagamentos eletronicos. Um enfoque de modelado preditivo
The lack of timely information about the current state of the economy is a well-recognized problem among policy makers (Evans, 2005). That explains why 'nowcasting', defined as "current-period estimates" (Galbraith & Tkacz, 2017) or "the prediction of the present, the very near future and the very recent past" (Banbura et al., 2013), has become a standard activity for central banks (Tiffin, 2016, Hinds et al., 2017).
Economic activity is one among many lagged key macroeconomic variables. Gross domestic product (GDP) estimates are released with a four-week lag in the United States and the United Kingdom; six in Japan; six or seven in the Euro area; eight in Canada; and up to one or two years in Lebanon (see Banbura et al., 2013; Tiffin, 2016; Bragoli, 2017; Galbraith & Tkacz, 2017). Short-term economic activity indicators attempt to reduce this lag. However, as most short-term indicators tend to rely on a subset of GDP inputs or other non-high-frequency and lagged variables, they usually have a non-negligible delay.
In the Colombian case, the national bureau of statistics (Departamento Administrativo Nacional de Estadistica, DANE) releases GDP quarterly figures with an approximate two-month lag (DANE, 2017). The same bureau releases the Economic Monitoring Index (Indicador de Seguimiento a la Economia, ISE), which is a monthly indicator with an approximate two-month lag that provides the market with a dynamic index of economic activity (DANE, 2016). Consequently, an accurate current-period estimate of ISE could provide the central bank, financial market participants, and other economic agents (e. g., government, real sector) with better tools for decision-making and modeling.
Our aim in this article is to answer a single question: is it possible to nowcast changes in the ISE with a dataset of electronic payment instruments? (1) Unlike most literature on economic activity nowcasting, we rely on electronic payments among individuals, firms, and the central government as high-frequency indicators (i. e., the inputs) from which reliable signals of economic activity are to be extracted. Using electronic payments as inputs is not fortuitous: they convey information about economic activity as they are representative of agents' expenditure.
As nowcasting is better understood as a prediction task, we undertake a predictive modeling approach (Shmueli, 2010) that relies on the well-documented performance and flexibility of artificial neural networks (ANNS) for prediction. Hence, unlike a traditional econometric approach, we do not attempt to test any causal theory or to explain which inputs excel as explanatory determinants of economic activity, but to attain an accurate out-of-sample prediction model by discovering complex structures in data that are not specified in advance. In this vein, following Shmueli (2010), Varian (2014) and Mullainathan and Spiess (2017), our approach departs from traditional econometric methods (i. e, explanatory modeling), and pertains to machine learning methods.
Evans (2005) and Giannonne et al. (2008) are among the first and most influential works on nowcasting economic activity. Our work is closely related to that of Galbraith and Tkacz (2017), who claim to be the first to make economic activity nowcasting based on payments data. However, there are several differences with Galbraith and Tkacz (2017) as we i) use a broader and larger spectrum of payment instruments that incorporates expenditures from individuals, firms, and the central government; ii) implement a Non-linear AutoRegressive eXogenous Artificial Neural Network model (NARX-ANN) that is documented for its predictive high-performance and flexibility; and iii) attempt to nowcast economic activity with a model that does not include traditional inputs (e. g., macroeconomic data, financial variables, surveys), but relies on electronic payment instrument data and lags of the corresponding economic activity indicator only.
Results suggest that a dataset containing ISE's lagged data and contemporary electronic payments data allows to make a fair current-period estimate of ISE's change circa two months before its release. That is, nowcasting economic activity with electronic payment instruments data and a NARX-ANN model is feasible. Also, results validate that electronic payments data significantly reduces the nowcast error of a benchmark autoregressive model. Therefore, we contribute to decision-making and modeling by providing a suitable nowcasting alternative for economic activity indicators. Furthermore, our work may be a starting point for other nowcasting attempts based on payments data.
The term 'nowcasting' is a contraction for 'now' and 'forecasting', and has been used for a long time in meteorology and recently also in economics (Banbura et al., 2013). It consists of exploiting the information that is available early and at higher frequencies to make a current-period or early estimate of a lagged or low-frequency variable.
Nowcasting economic activity entails two main challenges. First, finding a suitable set of high-frequency indicators (i. e, inputs) from which reliable signals of economic activity (i. e, outputs) are to be extracted. Second, selecting an appropriate prediction method for extracting reliable signals of economic activity that serve as fair current-period or early estimates of economic activity. This section presents how we tackle the first challenge.
Literature on economic activity nowcasting is fairly abundant (Evans, 2005; Bell et al., 2014, Bec & Mogliani; 2015, Bragoli, 2017). Most literature sets the GDP as the output variable. Our aim is to nowcast the ISE (Economic Monitoring Index). ISE is a monthly indicator of economic activity that combines information about the production of goods and services pertaining to the most important economic activities in Colombia. (2) It is estimated and released by the Colombian bureau of statistics with an approximate two-month lag.
By choosing the ISE instead of GDP we attain a greater number of observed outputs to work with. The ISE is available monthly from January 2000 onward. As payment inputs are available from January 2001 onward (see inputs section below), working on GDP quarterly figures would restrict our sample to about 48 observations, whereas working on ISE provides 192 observations. Moreover, as it is harmonized with quarterly GDP figures, working on ISE provides a fair and coherent estimate of changes in economic activity (DANE, 2016).
Figure 1 exhibits ISE from January 2001 to December 2016; descriptive statistics and the logarithmic returns of ISE are reported in Table 6 and Figure 6 (in the Appendix). As electronic payments data is non-seasonally adjusted (i. e, seasonal patterns are not removed), we use the non-seasonally adjusted ISE, at constant prices. We work with the revised index instead of preliminary estimates; as highlighted by Bell et al. (2014), ultimately, it is the official data that is most important for policy makers.
Most economic activity nowcasting literature relies on macroeconomic data, financial variables, surveys, or a mixture of them (Evans, 2005; Banbura et al., 2013; Bell et al., 2014; Bec & Mogliani, 2015; Bragoli, 2017). These inputs tend to convey a non-negligible lag (e. g., days, weeks or even months). Lately, research on how to nowcast economic activity from online data has surfaced. In that case, news articles, social media, web search data and other related sources are used to measure economic activity; for instance, McLaren (2011) and Choi and Varian (2012) illustrate how Google query indexes may be used for short-term economic activity prediction.
We choose electronic payment instrument data as our inputs for nowcasting economic activity. Our choice is unusual but has been already put into practice by Galbraith and Tkacz (2017). Galbraith and Tkacz use a set of electronic payments data comprising the value and number of operations from debit card transactions (i. e., point-of-sale payments) and 'small' cheques (i. e., under $50,000) that clear through the Canadian banking system. They argue that both sources of payment information are representative of consumption expenditure, which is a major component of GDP, meaning that their set of inputs provide an incomplete but direct source of information on GDP changes. Also, they argue that electronic payments are available quickly and virtually free of sampling error.
In our case we use a dataset comprising the value and number of operations from two different electronic payment instruments. First, from electronic transfers ordered by individuals, firms, and the central government, which are processed and cleared in both existing automated clearing houses (ACH), namely ACH Colombia and ACH Cenit; second, from cheques that are processed and cleared in the local cheque clearing house, Cedec. Unlike Galbraith and Tkacz (2017), we do not consider debit and credit card transactions because their contribution to the total value of electronic payments is rather low (i. e, below 5 percent), and because data is available with a lag that is incompatible with a nowcasting task, i.e. two or three months. Also, we do not limit cheques to being 'small'. A notable shortcoming common to both works is the absence of information about cash payments--the prevalent payment instrument in Colombia--. (3)
ACH Colombia is owned and managed by local banks. It fulfills the nowadays prevalent purpose of ACH s (McAndrews, 1994), namely to clear transfers between different banks regarding small repetitious payments among economic agents, such as payrolls, supplies, social security, mortgage installments, insurance premiums, dividends, and...