Improving the Short-term Forecast of World Trade During the Covid-19 Pandemic Using Swift Data on Letters of Credit
  • 1 0000000404811396https://isni.org/isni/0000000404811396International Monetary Fund
  • | 2 0000000404811396https://isni.org/isni/0000000404811396International Monetary Fund
  • | 3 0000000404811396https://isni.org/isni/0000000404811396International Monetary Fund

An essential element of the work of the Fund is to monitor and forecast international trade. This paper uses SWIFT messages on letters of credit, together with crude oil prices and new export orders of manufacturing Purchasing Managers’ Index (PMI), to improve the short-term forecast of international trade. A horse race between linear regressions and machine-learning algorithms for the world and 40 large economies shows that forecasts based on linear regressions often outperform those based on machine-learning algorithms, confirming the linear relationship between trade and its financing through letters of credit.

Abstract

An essential element of the work of the Fund is to monitor and forecast international trade. This paper uses SWIFT messages on letters of credit, together with crude oil prices and new export orders of manufacturing Purchasing Managers’ Index (PMI), to improve the short-term forecast of international trade. A horse race between linear regressions and machine-learning algorithms for the world and 40 large economies shows that forecasts based on linear regressions often outperform those based on machine-learning algorithms, confirming the linear relationship between trade and its financing through letters of credit.

I. Introduction

An essential element of the work of the Fund is to monitor and forecast international trade. Trade is a leading indicator of global economic activity and a key channel of spillovers. An accurate and timely reading of trade developments can therefore provide useful insights into global economic activity and an economy’s short-term growth dynamics. This is particularly relevant during the current conjuncture as the decline in world trade in the first half of 2020 was an early indicator of the contraction in global economic activity due to the lockdowns and social distancing associated with the ongoing Covid-19 pandemic.

Trade financing can be an early indicator of international trade. Ahead of goods moving across countries, importers often arrange for trade financing, particularly in Asia. Banks provide letters of credit or trade advances that can be secured by the imported goods as collateral, and charge fees and interest until the importer has sold the goods and paid off the trade credit. Trade financing can therefore be a leading indicator of international trade by several weeks depending on the time needed to produce and transport the merchandise from the exporting country to the importing one.

SWIFT, the global provider of secure financial messaging services, provided the authors of this working paper with access to data on letters of credit.1 The monthly SWIFT data are available a few business days following the end of the reported month, aggregated by country and on an anonymous basis. The associated metadata provide the country where the message originates from (normally the importing country), the country where the beneficiary resides (usually, the exporting country), and the total amount of transactions during the month (in USD and in the currency of settlement).

This paper uses the SWIFT data to improve the short-term forecast of international trade, together with Brent crude oil prices and the new export orders subcomponent of manufacturing Purchasing Managers’ Index (PMI) where available. Both linear regressions and machine-learning algorithms are used to extract the lead information content of SWIFT trade messages to improve the short-term forecast of world and national trade for 40 large economies. In doing so, SWIFT trade messages are shown to have informational content to forecast world trade in the short run, and national trade in selected economies, particularly in Asia. In addition, a horse race between linear regression forecasts and machine-learning ones show that linear regression forecasts often outperform other forecasts. This likely reflects the linear relationship between trade data and its financing through SWIFT.

This paper is organized as follows. Section II provides a brief introduction to SWIFT messages and trade financing. Section III describes the forecast of world trade using SWIFT messages and the recent performance during the Covid-19 pandemic compared with other forecasts of world trade. The relationship between SWIFT messages and national trade is discussed in Section IV. Section V presents the horse race between forecasts based on linear regressions and those based on machine-learning algorithms. The main conclusions are summarized in Section VI.

II. Swift Messages and International Trade

SWIFT is a provider of secure financial messaging services. Its services are used by over 11,000 financial institutions in more than 200 countries and territories around the world. To this day, SWIFT represents the primary communication channel for corporates, financial institutions, and market infrastructures to settle international financial payments, securities, foreign exchange transactions, treasury operations, and trade flows.2

SWIFT messages and reference data are standardized, which ensures that data exchanged between financial institutions is unambiguous. SWIFT messages are classified into nine categories such as customer payments and checks, financial institution transfers, treasury markets, and documentary credits and guarantees. Each of these categories contains several types of messages. A SWIFT message starts with an identifier (MT) and is followed by a 3-digit number that represents category, group, and type.

The Fund acquired a proprietary database of SWIFT messages several years ago going back to 2010 and receives monthly updates a few business days after the end of the month. SWIFT data are aggregated by country at a monthly frequency and are anonymous. Each data point contains the message type code, message name, the code and name of the originating country, the code and name of the counterparty country, the currency used, the total number of SWIFT transactions sent and received during the month, and the total net amount sent and received in the original currency and US dollar equivalent.

SWIFT messages used in this paper are MT 700 messages and represent documentary credits, more commonly known as letters of credit.3 Importers usually contract a letter of credit from a bank to import merchandise goods against a fixed transaction fee and interest payments. When the letter of credit is obtained, the issuing bank sends an MT 700 message to the bank of the exporter (the advising bank) to indicate the terms and conditions of the letter of credit. Against this guarantee, the exporter then produces and ships the merchandise goods to the importer and produces the necessary shipping documentation (e.g., bill of laden) representing title to the goods. The issuing bank then checks the shipping documentation against the requirements under the letter of credit before making the payment to the advising bank. Letters of credit and associated obligations to handle documents under documentary credits are governed by the Uniform Customs and Practice for Documentary Credits rules established by the International Chamber of Commerce (ICC).4

SWIFT MT 700 messages can be a leading indicator of world trade. Since the letter of credit is sent before exporters ship their merchandise goods, the dollar amount of the letter of credit provides an early indicator of the nominal value of the trade that will take place. Figure 1 provides a graphical representation of the financial flows under a letter of credit (top row) before the actual movement of goods takes place (bottom row). It is worth noting that SWIFT MT 700 messages go in the opposite direction of the trade flow.

Figure 1.
Figure 1.

Financial Flow and Merchandise Trade

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Source: Authors’ representation.

The extent to which a SWIFT trade message is a leading indicator of bilateral trade varies across countries. The time it takes from the moment the MT 700 message is sent and the goods are recorded by the customs authorities of the importing country largely depends on the production time, the proximity between the trading partners, and the type of transportation method. For example, the lead time between the United States and Mexico may be very short (a week or two), while the lead time between China and Germany may be longer (one to two months), unless goods are already produced and shipped by plane. Thus, the lead time between SWIFT trade messages and bilateral trade will vary depending on the physical distance and method of transportation used for the merchandise to be shipped between countries. According to the ICC, the average maturity of a letter of credit is 60 days.5

The share of trade financed through SWIFT trade messages varies significantly across economies. Amongst the top 10 importers in the world, India, Hong Kong SAR, and China were the largest users of SWIFT MT 700 trade messages in 2019 with a share over 20 percent of their total merchandise trade (Figure 2). Amongst the largest 10 exporters, Hong Kong SAR and United Kingdom stand out as the largest user of SWIFT trade messages at over 25 percent, followed by Korea, Japan, and China. The relatively large shares for Hong Kong SAR, Japan, the United Kingdom, and the United States may, however, partly reflect the fact that banks located in these countries receive SWIFT trade messages on behalf of exporting clients in other parts of the world. Unfortunately, this cannot be verified through available data.

Figure 2.
Figure 2.

Share of Merchandise Trade Financed by SWIFT Trade Messages For the 10 Largest Importers and Exporters in the World, 2019

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Sources: National customs’ data, SWIFT, and authors’ calculations.

III. A Forecast of World Trade Using Swift Messages

SWIFT MT 700 messages financed about 15 percent of world trade in 2019. The relatively small share of trade financed through SWIFT is explained by the fact that letters of credit are mainly used by new customers, customers without payment record, in jurisdictions that are unfamiliar to the importer, or in countries where international contracts are difficult to enforce. Niepmann and Schmidt-Eisenlohr (2017) show that the use of letters of credit is mostly prevalent in bilateral trade with emerging markets, most notably in Asia, while open contracts are more prevalent in bilateral trade between advanced economies, and cash-in-advance amongst low-income and developing countries that do not have access to trade financing (Appendix I). This evidence is also confirmed by the forecasts below. Moreover, trade can be financed by a global bank or a multinational enterprise with subsidiaries or branches in both the exporting and importing country, which may require a simple crediting and debiting of the balance sheet within different units of the global bank or enterprise.

SWIFT trade messages have a strong correlation with world trade. Figure 3 shows the co-movement of the growth rate of SWIFT trade messages and world trade as measured by the Netherlands’ Bureau of Economic Policy Analysis (CPB).6 A visual inspection shows that SWIFT trade messages generally move broadly in line with world trade, while being more volatile and subject to unusual spikes and corrections that are unrelated to the underlying trade flows (see for example the large spikes and corrections in 2019). Most recently, SWIFT trade messages have closely mirrored the decline in world trade during the COVID pandemic, shown in the shaded blue area.

Figure 3.
Figure 3.

World Trade and SWIFT Trade Messages (Percentage Change in US dollars; 3-month moving average, seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Source: CPB, SWIFT, and authors’ calculations.

Another way to visualize this co-movement is through a scatterplot (Figure 4). The simple regression shown in the scatterplot has an R2 of 47 percent. The recent large declines in world trade associated with the COVID pandemic (the dots on the bottom left quadrant of the scatter plot) are broadly in line with the estimated regression, suggesting a stable linear relationship between the two variables, even during the current pandemic.

Figure 4.
Figure 4.

World Trade and SWIFT Trade Messages

(Percentage change; 3-month moving average; seasonally adjusted; December 2010-May 2020)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Sources: CPB, SWIFT, and authors’ regression.

Other variables, beyond SWIFT trade messages, are useful to explain the variation in world trade. In this paper, SWIFT trade messages are therefore complemented by Brent crude oil prices and a proxy for manufacturing exports to produce a more comprehensive forecast of world trade. Brent crude oil prices are generally a volatile determinant of the value of world trade in nominal terms as they capture both the value of traded hydrocarbons (as part of the world trade deflator) and the expectations of future global energy demand (and thus global economic activity) in volume terms. Global manufacturing activity also has a strong correlation with the volume of world trade as it is generally driven by global investment spending, which is highly trade intensive as evidenced during the Global Financial Crisis (see IMF 2016). Adding these two explanatory variables, in the form of Brent crude oil prices and the new export orders of the Global Manufacturing Purchasing Managers’ Index (PMI), significantly improves the fit in the scatterplot above and raises the R2 to 76 percent (Figure 5).

Figure 5.
Figure 5.

World Trade and SWIFT Trade Messages, Manufacturing PMI: New Export Orders, and Oil Prices

(Percentage change; 3-month moving average; seasonally adjusted; December 2010-May 2020)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Sources: CPB, Haver, IHS Markit, JP Morgan, SWIFT, and authors’ regression.

Based on these preliminary results, a complete regression is posited in Equation 1. World trade (WT) is regressed against its own lags, SWIFT trade messages (SWIFT), Brent crude oil prices (Brent), and the new export orders subcomponent of the global manufacturing PMI (PMI). All variables have up to four lags and are expressed in log differences except PMI, which is expressed in its original diffusion index form but re-centered around zero. The reduced-form equation is then as follows:

dlog(WTt)=α+βidlog(WTti)+γjdlog(SWIFTtj)+δjdlog(Brenttj)+ϑjPMItj+ϵt(1)

where the lag index i goes from 1 to 4 and the lag index j goes from 0 to 4.

The regression results of different specifications of equation (1) show that SWIFT trade messages, Brent crude oil prices, and manufacturing activity are all significant determinants of world trade (Table 1). The coefficients on lags of world trade are mostly insignificant, except under the two specifications in the center of Table 1 with Brent, and Brent and PMI as explanatory variables. The contemporaneous coefficients on SWIFT are positive as expected, (higher SWIFT financing should lead to higher trade activity), and highly significant in all specifications, while SWIFT lags are significant only in the specification without Brent and PMI. As expected, the coefficients on Brent and on PMI are positive (higher oil prices and more manufacturing activity lead to higher world trade) and highly significant, including in their first lags. When SWIFT is omitted from the regressions (the two right-hand columns of Table 1), coefficients on OIL and PMI (and some of their lags) continue to be significant and the R2 is high (0.673–0.768). This suggests that OIL explains most of the variation in world trade. Specifically, OIL and SWIFT seem to be non-orthogonal, and the contribution of SWIFT is smaller when OIL is included in the specification.

Overall, the regressions confirm the economic theory of positive signs on the coefficients of the explanatory variables and their statistical significance. The specification with all explanatory variables has an R2 of close to 80 percent (adjusted R2 of 75 percent), which provides a good basis to use this reduced-form equation to forecast world trade in the section below.

Forecast of World Trade

The predictive power of the regressions above is evident by forecasts of world trade. An initial forecast was undertaken on January 15, 2019, based on data available at the time, namely CPB world trade data up to October 2018 and SWIFT messages up to December 2018. A simple autoregressive process at the time predicted world trade to continue growing at a constant rate in late 2018/early 2019, as shown by the green line in Figure 6. The forecast with only SWIFT messages up to December 2018 instead predicted a gradual decline in world trade in the latter part of 2018 and early 2019 as shown by the blue line. Adding Brent crude oil prices extended by oil futures available on January 15, 2019, produced a forecast that was even more negative as shown by the purple line. In fact, as shown by the red line, world merchandise trade declined rapidly in November and December, and stabilized in the first four months of 2019, broadly in line with the forecast with oil prices.

Repeating the same forecast exercise above on a regular basis for 2019 and the first eight months of 2020 shows that these forecasts are close to the realized data (Figure 7). During the current pandemic, the April 16, 2020 forecast correctly pointed to a significant decline in world trade, albeit not as sharp as it finally occurred. In contrast, CPB data was available only up to January 2020 and indicated only a stabilization in world trade. The latest SWIFT forecast as of August 17, 2020, predicts a moderate rebound in world trade in the third and fourth quarter of 2020. The CPB data for June 2020, which became available in late August, confirmed the rebound, albeit stronger than predicted by the August 17 forecast.

Table 1.

World Trade: Comparison of Different Regression Specifications

Dependent Variable: World Trade Sample: April 2011-May 2020 (110 observations) (Coefficients and Standard Errors in Parenthesis)*

article image
* Standard errors are heteroscedasticity and autocorrelation robust (HAC) using 1 lags and without small sample correction. Asterisks indicate significance at 10 percent (*), 5 percent (**), and 1 percent level (***).Source: Authors’ regression results.
Figure 6.
Figure 6.

Forecast of World Trade, January 15, 2019 (Billions of US dollars; seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Sources: CPB, Haver, SWIFT, and authors’ regression and forecast.
Figure 7.
Figure 7.

Rolling Forecasts of World Trade (Billions of US dollars; seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Sources: CPB, Haver, IHS Markit, JP Morgan, SWIFT, and authors’ regressions and forecasts. All forecasts are based on the regression specification with lagged world trade, SWIFT trade data, and Brent crude oil prices, except for the August 17, 2020 forecast which includes the new export orders subcomponent of the global manufacturing PMI.

Comparison with Other Forecasts of World Trade

It is useful to compare the forecasts above with other potential forecasts of world trade. Specifically, this section compares the SWIFT forecast with a forecast based on the Baltic Dry Index and a statistical forecast based on an autoregressive model.

The Baltic Dry Index provided a good early indicator of developments in world trade in the past. During the Global Financial Crisis, it foretold the collapse in world trade that occurred in late 2008-early 2009. A formal regression of CPB world trade, based on a similar specification in log differences as in Equation 1 above, confirms that the coefficients on the Baltic Dry Index (BDIt) and its lags are statistically significant and are able to explain about a quarter of the variation in world trade (Table 2).

A comparison of the SWIFT and the BDI forecasts provides interesting insights. For the comparison purposes, both forecasts were built based on the information available on April 16, 2020, namely world trade data up to January 2020, SWIFT data up to March 2020, and the BDI data available until mid-April. Specifically, the BDI showed a significant decline in February/March followed by a rebound in early April. Accordingly, the forecast, based on the regression in Table 2, predicated a significant decline in world trade until May 2020, followed by a rapid rebound in June and July (blue line in Figure 8). The SWIFT forecast instead predicted a slower decline in world trade until March 2020, followed by a more rapid decline until July 2020 (red line in Figure 8). The actual world trade data through May shows a rapid decline in February/March, closely matching the BDI forecast, followed by a much sharper decline in April than either the BDI or the SWIFT forecast predicted (green line in Figure 8). Overall, the BDI forecasts performed better one- to two-months ahead, while the SWIFT forecast performed better over the longer forecast horizon. The better short-term performance of the BDI forecast may, in part, reflect the fact that the BDI forecast includes additional information up to mid-April 2020, while the SWIFT data only covered trade developments up to March 2020.

A similar comparison can be undertaken with the statistical model. The statistical model is based on a simple autoregressive process of order 2 (AR(2)) with a slow-moving long-term mean. As such, the AR(2) forecast relies only on past lags of world trade data and does not consider other explanatory variables, like Brent crude oil prices, SWIFT trade messages, or new export orders for the SWIFT forecast. The AR(2) forecast does not cover all countries in the world and excludes about 9 percent ($1.5 trillion annually on the import side) of world trade compared with the CPB measure.

Table 2.

World Trade: Regression Results, Baltic Dry Index (BDI) Dependent Variable: World Trade Sample: May 2006-January 2020 (165 observations) (Coefficients and Standard Errors in Parenthesis)*

Dependent Variable: World Trade Sample: April 2011-May 2020 (110 observations) (Coefficients and Standard Errors in Parenthesis)*

article image
* Asterisks indicate significance at 10 percent (*), 5 percent (**), and 1 percent level (***).Sources: Bloomberg, CPB, and authors’ regression results.
Figure 8.
Figure 8.

BDI and SWIFT Forecasts of World Trade

(Billions of US dollars; seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Sources: Bloomberg , CPB, Haver, IHS Markit, JP Morgan, SWIFT, and authors’ regression results and forecasts.

A visual comparison of the SWIFT and AR(2) forecasts during the current COVID pandemic shows that the SWIFT forecast generally picks up turning points earlier than the AR(2) forecasts (Figure 9). In January and February 2020, both forecasts were showing a relatively flat projection of world trade. In March, April, and May 2020, however, the SWIFT forecast was predicting a significant decline in world trade, while the AR(2) forecast was still signaling a relatively flat world trade forecast, notwithstanding an initial decline. In June 2020, on the contrary, the AR(2) forecast was significantly more pessimistic than the SWIFT forecast, and correctly predicted the steep decline in global trade that happened in April 2020. In July and August, the SWIFT forecast signals a moderate rebound in global trade, while the AR(2) forecast predicts a continued decline, despite a correction in June. The CPB data for June, instead, showed a larger rebound in world trade than predicted by the SWIFT forecast. The AR(2) forecast also generally provides a forecast over two to three months, while the SWIFT forecast provides a longer forecast up to six months.

Overall, the AR(2) forecast generally provides a relatively good forecast during normal times. It picks well the trend growth in world trade. However, given that the AR(2) forecast is purely backward looking, it is unable to pick up turning points the way the BDI or SWIFT forecast can.

It is unfortunately not possible at this stage make a statistical comparison of the accuracy of each forecast, given the short overlap of SWIFT and AR(2) forecasts. An accurate statistical comparison would require at least 40 overlapping forecasts, while for now there are only 12 overlapping forecasts. A formal statistical comparison will therefore need to be undertaken as part of future research work.

IV. Swift and National Customs Trade

The ability to forecast world trade with SWIFT trade messages derives from its use to finance national trade, particularly in Asia. However, the relationship between SWIFT trade messages and national trade flows is blurred by the fact that SWIFT trade messages are often handled by international banks located in a different country from the one where the merchandise trade originates from or is destined to. While this is not a problem at a world level, it is an issue at a national level because it reduces the correlation between SWIFT trade messages and the underlying merchandise trade flows. As indicated before, this is particularly relevant for economies that host financial centers, like Hong Kong SAR, Japan, the United Kingdom, and the United States, where most of the international banks that handle SWIFT trade messages are located. Unfortunately, no data are available to verify or correct this blurred relationship.7

Figure 9.
Figure 9.

Comparison of SWIFT and AR(2) Forecasts of World Trade (Billions of US dollars; seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Sources: Bloomberg, CPB, IHS Markit, JP Morgan, SWIFT, and authors’ regression results and forecasts.

Figure 10 shows the simple contemporaneous correlation between SWIFT messages and total merchandise exports and imports, as reported by the national customs authorities of each of the 40 countries in the sample.8 It shows that the highest correlation is driven by Turkey and Asian countries (Korea, China, Vietnam, Bangladesh, Indonesia, Malaysia, and Thailand), while the lowest correlation is for several advanced economies (Ireland, Sweden, Belgium, Norway, Japan, and the US) and several emerging markets (Colombia, Philippines, South Africa, and Poland). Moreover, the contemporaneous correlation is likely to understate the overall correlation, given the lagged correlation between SWIFT trade messages and national trade mentioned above. The significance of the contemporaneous coefficient on the SWIFT variable in the regression results below is shown by the number of asterisks next to the name of the national customs data in Figure 10.

To build a linear forecast of national trade based on SWIFT trade messages, a simple regression is estimated for both exports and imports for the 40 countries in the sample, based on Equation 1 in the previous section. The equation is estimated in log differences for both exports and imports (cus_x and cus_m) as reported by the national customs authorities for each country (80 regressions) using lagged customs data, Brent crude oil prices and the new export orders subcomponent of the national manufacturing PMI where available. The equation is estimated using ordinary least squares with heteroskedastic-consistent and autocorrelated robust (HAC) standard errors.9

As expected, the regression results vary broadly in line with the correlation structure shown in Figure 8 (see Appendix II for selected regression results). The overall fit of the regression results for the countries in the sample broadly match the contemporaneous correlation structure, except for countries where the coefficients on SWIFT are insignificant while the ones on Brent crude oil prices and PMIs are highly significant. For those regressions where SWIFT coefficients are significant, they are also positive, indicating that higher trade financing is associated with higher trade activity. The magnitude of the coefficients on SWIFT vary in line with the contribution of SWIFT trade messages in explaining the variation of national trade. The contemporaneous coefficients on Brent crude oil prices and PMIs are also generally positive as expected.

In Asia, the regression results for Bangladesh, India and Korea, Thailand, and Vietnam exports; and China, Indonesia, Korea, Malaysia, and Vietnam imports show a relatively high R2 (0.60–0.79) and positive and significant coefficients on SWIFT trade messages. On the other hand, the coefficient on SWIFT trade messages for Hong Kong SAR, Japan, the Philippines, Taiwan Province of China, (both exports and imports) are insignificant or of the wrong sign. In the case of Hong Kong SAR and Japan, this probably reflects the fact that banks domiciled in Hong Kong SAR and Japan intermediate significant SWIFT trade messages unrelated to their trade flows as discussed above. Overall, the regression results confirm the relevance of letters of credit in financing selected Asian trade.

Figure 10.
Figure 10.

Correlation Between SWIFT Messages and National Customs Data (Contemporaneous Correlation in log difference, after adjusting for outliers)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Sources: CPB, national customs data, SWIFT, and authors’ calculations. Asterisks next to the name of the country indicate the significance level of the contemporaneous coefficient on SWIFT in the regression results at the 10 percent (*), 5 percent (**), and 1 percent (***) levels when the coefficient is positive.

In Europe, results are mixed. The coefficients on SWIFT trade messages are significant for France, Italy, Latvia, Lithuania, Norway, and Sweden exports; and for Latvia, Lithuania, and Russia imports, with a relatively high R2 (0.49–0.78). On the other hand, the coefficients on SWIFT trade messages are insignificant for the regressions for Belgium, Denmark, Estonia, Germany, Ireland, Netherlands, Poland, Portugal, Spain, and the United Kingdom (both exports and imports). This suggests that SWIFT trade messages mostly finance imports to Eastern Europe (Latvia, Lithuania, and Russia) and significant portion of exports for selected European countries. In the case of the United Kingdom, the London financial center acts as a significant financial intermediary for SWIFT trade messages and as such blurs the relationship between SWIFT trade messages and United Kingdom trade.

In the Western Hemisphere, the results are also mixed. The coefficients on SWIFT trade messages are significant for Mexico exports, and Brazil, Chile, and Mexico imports, with a relatively high R2 (0.54–0.88). However, the coefficients are insignificant for Canada, Colombia and the United States (both exports and imports). The results for the United States again may reflect the significant activity of U.S. banks in intermediating SWIFT trade messages on behalf of exporters and importers outside of the United States.

In the Middle East, the results confirm the significant use of SWIFT trade messages to finance trade in Israel and Turkey. For Israel, the coefficients on SWIFT trade messages are significant both on the export and the import side and the R2 is relatively high (0.62 on the export side and 0.72 on the import side). For Turkey, the coefficients on SWIFT trade messages are highly significant both in the export and import regression and the R2 is among the highest in the sample (0.70 on the export side and 0.72 on the import side).

In Africa, the results are significant for Ghana and Nigeria, but not for South Africa. For Ghana, the coefficients on SWIFT trade messages are significant for the import equation, namely the contemporaneous coefficient and the ones on 1 and 3 lags, while the R2 is 0.50. For Nigeria, the coefficients on SWIFT trade messages are significant on the export side, namely on both the contemporaneous coefficient and the ones on 1 and 2 lags, with an R2 of 0.41. On the other hand, the coefficients on SWIFT trade messages for South Africa are insignificant both in the export and import regressions, suggesting that banks in South Africa may also be intermediating SWIFT trade messages for other countries, thus blurring the relationship with national trade.

V. Horse Race Between Linear Regression and Machine-learning Forecasts

It is useful to compare the linear regressions forecasts with machine-learning forecasts in order to rule out significant non-linearities in the relationship between SWIFT trade messages and the underlying trade flows. For this purpose, this section presents a horse race between linear regression and machine-learning forecasts for world trade and the 40 countries in the sample. It provides an overview of the linear regression and machine-learning forecasts, and an overall assessment of the forecasts. For a detailed description of the machine-learning algorithms (MLAs) used in this paper, together with the advantages and disadvantages of linear regressions vs. machine–learning algorithms, please refer to Appendix III. Appendix IV provides a description of the forecast methodology used.

Linear Regression Forecasts

This section discusses linear regression forecasts. Linear regression forecasts are built using the estimated regressions of equation (1) to forecast one-step ahead and then recursively forecasting longer horizons.10 Customs data are extended using the one-step ahead forecast. SWIFT data are generally available one to three months ahead of the customs data. For the remainder, SWIFT data are extended through an AR(1) process. Brent crude oil prices are extended through the closing futures prices on the date of the forecast for the one-, three-, and six-month contracts, while interpolating the other months. The new export orders subcomponent of the PMI is usually available one to three months ahead of customs data. For the remainder of the forecast horizon, it is extended through an AR(1) process.

A good example of a linear regression forecast is the one for world trade as of April 16, 2020. At the time, the CPB world trade data was only available up to January 2020, while SWIFT data up to March 2020 already indicated a significant decline in global trade activity. Brent crude oil prices had plummeted from $68 per barrel in early January to $27 per barrel on April 16, and Brent futures indicated a further decline in the months ahead. The new export orders subcomponent of the global manufacturing PMI for March had fallen from 49.5 in January to 43.0 in March, signaling a significant contraction in manufacturing exports going forward.

Based on these explanatory variables, the linear regression forecast on April 16, 2020 indicated a significant decline in world trade going forward (black line in Figure 11). Other machine-learning forecasts indicated a smaller decline and possibly a stabilization in world trade (e.g., Decision Tree). World trade (red line) turned out close to the linear regression forecast for February and March 2020, before collapsing in April 2020.

Figure 11.
Figure 11.

World Trade: Forecasts Based on Linear Regression and MLAs, April 16, 2020

(Billions of US dollars, Seasonally Adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Sources: CPB, SWIFT, Brent crude oil prices and futures, JP Morgan Global Manufacturing PMI and authors’ regressions and forecast results.

Machine-learning forecasts11

Three categories of MLAs are used in this paper, namely linear, single nonparametric, and ensemble nonparametric MLAs. Linear MLAs comprise Lasso and Ridge regressions, which are variations of a linear regression that weigh regressors based on their significance. Single nonparametric MLAs comprise Decision Tree Regression and Support Vector Regression. Ensemble MLAs are built based on regression trees to identify non-linearities in subsamples of the dataset. The tree-based ensemble MLAs comprise Bagging, Gradient Boost, and Random Forest. Each MLA is trained over the full sample period. Based on this training, each MLA is then used to produce a forecast one to six-month ahead based on the same one-step ahead forecast as for linear regression forecast.

A good example of an MLA forecast that is superior to a linear regression one is in the case of Korea imports (Figure 12). As of April 16, 2020, Korea merchandise import data for March 2020 were already available and signaled a rebound from the decline experienced the previous month. However, with SWIFT, Brent crude oil prices and futures, and Korea’s new export orders of the manufacturing PMI all signaling a decline in imports in the near future, the linear regression forecast as of April 16, 2020 pointed to a steep decline up to May 2020, before a rebound in June and July. Customs data released subsequently turned out to follow a similar pattern with the exception that the trough was in April instead of May. The Lasso forecast in this respect did marginally better at capturing this pattern than the linear regression forecast, but the difference in forecast errors between the two is marginal.

Figure 12.
Figure 12.

Korea Imports: Forecasts Based on Linear Regression and MLAs, April 16, 2020

(Billions of US dollars, Seasonally Adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Sources: Brent crude oil prices and futures, IHS Markit, national customs data, SWIFT, and authors’ regressions and forecast results.

Evaluation of linear and machine-learning forecasts

The evaluation of linear and machine-learning forecasts is based on the root mean squared error (RMSE) of the two forecast methods. To calculate the RMSEs, the forecasts are computed based on a subsample up to May 2019 (the training set). The forecasts thus computed are then evaluated over the data between June 2019 and May 2020 (the test set), based on monthly rolling forecasts, and then calculating the RMSE over the one-, two-, three-, four-, five-, and six-month ahead forecasts.

The RMSEs for Latvia imports provide a useful example of this evaluation (Figure 13). They are relatively small, ranging between 5 and 14 percent, with longer forecast horizons having larger RMSEs as expected. Most notably, there is a clear difference in performance in the shorter forecast horizons, with the linear regression, ridge, support vector regression, and lasso forecasts having much smaller RMSEs than the other forecasts. This advantage then disappears over the longer forecast horizons. This suggests that linear regression methods have a clear advantage in forecasting Latvia’s imports over the short run, while over five to six months this advantage disappears.

Figure 13.
Figure 13.

Latvia Exports: RMSEs for Different Forecasts

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Sources: Authors’ forecasts and calculations.

The best-performing forecasts based on the lowest RMSE for the world and the 40 countries are presented in Table 3. It is worth noting that the linear regression forecasts perform the best over a 1-month ahead forecast in nearly half the sample. The ratio increases to two thirds when Lasso and Ridge forecasts (which are more parsimonious version of linear regression forecasts) are added. This means that linear forecasts generally outperform other non-linear MLAs, which confirms a linear relationship between SWIFT trade messages and the underlying merchandise trade as expected by economic theory.

All forecasts are also evaluated formally using the Diebold-Mariano (DM) test.12 The DM test assesses the statistical significance of the differences in RMSEs of each forecast against a so-called naïve forecast, which is based on the simple average growth rate of exports/imports over the sample period. The results of the DM tests for the best one-step ahead forecast are also shown in Table 3, together with the corresponding P-statistics and RMSEs. The results show that a few countries have statistically significant forecasts that are different than a naïve forecast. This suggests that more needs to be done to improve the quality of the forecasts, possibly by considering additional explanatory variables.

VI. Conclusions

This paper has documented the usefulness of using SWIFT trade messages, Brent crude oil prices, and the new export orders subcomponent of manufacturing PMI (where available) to forecast short-term merchandise trade. While SWIFT trade messages only finance 15 percent of world trade, they have predictive power to forecast world trade. They can help forecast turning points in world trade and are particularly relevant to forecast trade in Asia. However, the relationship at a national level between SWIFT trade messages and the underlying merchandise trade becomes blurred, particularly in economies that host financial centers, like Hong Kong SAR, Japan, the United Kingdom, and the United States. SWIFT trade messages are therefore less reliable in forecasting national trade particularly in those countries.

A horse race between linear regression forecasts and other forecasts based on machine-learning algorithms has also shown that linear regression forecasts, together with its Lasso and Ridge variants, are superior in about two thirds of the sample, which suggests that the underlying relationship between SWIFT messages and merchandise trade for most countries is likely to be linear. Future research will need to assess whether SWIFT forecasts perform statistically better than forecasts based on the Baltic Dry Index or on a simple autoregressive process.

Table 3.

Diebold-Mariano Tests of One-Step Ahead Forecast13

article image
article image
Source: Authors’ forecasts and calculations.

References

  • Basak, D., Srimanta Pal and Dipak C. Patranabis, D. C., 2007, “Support Vector Regression,” Neural Information Processing-Letters and Reviews 11(10): 203224.

    • Search Google Scholar
    • Export Citation
  • Baum, C., 2011, “DMARIANO: Stata Module to Calculate Diebold-Marian Comparison of Forecast Accuracy,” EconPapers, available at https://econpapers.repec.org/software/bocbocode/s433001.htm

    • Search Google Scholar
    • Export Citation
  • Boehmke, B., & Brandon M. Greenwell, 2019, Hands-On Machine Learning with R. CRC Press.

  • Breiman, L., J.H. Friedman, R.A. Ohlsen, and C.J. Stone, C. J., 1984, “Classification and Regression Trees.” Wadsworth & Brooks/Cole.

  • Breiman, L., 2001, “Random Forests.” Machine Learning 45(1):532.

  • CPB Netherlands Bureau for Economic Policy Analysis, 2019, “CPB World Trade Monitor”, January 2019, available at https://www.cpb.nl/en/worldtrademonitor.

    • Search Google Scholar
    • Export Citation
  • Danske Bank (2019). “Documentary Credits – Protecting Payments in International Trade,” available at https://danskebank.com/en-gb/Corporate-Banking/Trade-and-Export-Finance/Documentary-Credits/Pages/The-Documentary-Credit-process.aspx.

    • Search Google Scholar
    • Export Citation
  • Diebold, F.X. and R.S. Mariano, 1995, “Comparing Predictive Accuracy”. Journal of Business and Economic Statistics, 13: 25363.

  • Dietterich, T. G., 2000, “Ensemble Methods in Machine Learning.” in International Workshop on Multiple Classifier Systems, 115. Berlin and Heidelberg: Springer.

    • Search Google Scholar
    • Export Citation
  • Hayashi, F., 2000, Econometrics, Princeton University Press.

  • ICC, 2018, Global Trade–Securing Future Growth, 2018, Technical Report, International Chamber of Commerce.

  • IMF, 2016, “Global Trade: What’s Behind the Slowdown?Chapter 2 of “Subdued Demand: Symptoms and Remedies,” World Economic Outlook, October 2016, International Monetary Fund.

    • Search Google Scholar
    • Export Citation
  • Jones, S., David Johnstone, and Roy Wilson, 2015, “An Empirical Evaluation of the Performance of Binary Classifiers in the Prediction of Credit Ratings Changes.” Journal of Banking & Finance. 56: 7285.

    • Search Google Scholar
    • Export Citation
  • Lantz, B., 2013, Machine Learning with R. Birmingham: Packt Publishing Ltd. (Second edition).

  • LeCun, Y., Yoshua Bengio, and Geoffrey Hinton, 2015, “Deep Learning.” Nature 521(7553): 436.

  • Mahdevari, S., Kourosh Shahriar, Saffet Yagiz, and Mohsen A. Shirazi, (2014), “A Support Vector Regression Model for Predicting Tunnel Boring Machine Penetration Rates.” International Journal of Rock Mechanics and Mining Sciences 72: 214229.

    • Search Google Scholar
    • Export Citation
  • McCulloch, W.S. and Pitts, W., 1943, “A Logical Calculus of the Ideas Immanent in Nervous Activity,” The Bulletin of Mathematical Biophysics 5(4):,115133.

    • Search Google Scholar
    • Export Citation
  • Moustapha, M., Jean Marc Bourinet, J., Benoît Guillaume, and Sudret, Bruno, 2018, Comparative study of Kriging and support vector regression for structural engineering applications. ASCE-ASME Journal of Risk and Uncertainty in Engineering Systems, Part A: Civil Engineering, 4(2), 04018005.

    • Search Google Scholar
    • Export Citation
  • Mullainathan, S. and Jann Spiess, 2017), “Machine Learning: An Applied Econometric Approach.” Journal of Economic Perspectives, 31(2): 87106.

    • Search Google Scholar
    • Export Citation
  • Niepmann, F. and Tim Schmidt-Eisenlohr, 2017, “International Trade, Risk and the Role of Banks,” in Journal of International Economics 107: 11126.

    • Search Google Scholar
    • Export Citation
  • Pedregosa, F., Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, and Jake Vanderplas, J., 2011, “Scikit-learn: Machine Learning in Python.” Journal of Machine Learning Research, 12(10): 28252830.

    • Search Google Scholar
    • Export Citation
  • Russell, S.J. and Peter Norvig, 2016, “Artificial Intelligence: A Modern Approach.” Malaysia: Pearson Education Limited.

  • Seabod, S. and Josef Perktold, 2010, “Statsmodels:Econometric and Statistical Modeling with Python,” Proceedings of the 9th Python in Science Conference.

    • Search Google Scholar
    • Export Citation
  • SWIFT (2019a), “SWIFT History”, available at https://www.swift.com/about-us/history.

  • SWIFT (2019b), “Messaging Standards,” available at https://www.swift.com/about-us/discover-swift/messaging-standards.

  • Verikas, A., Evaldas Vaiciukynas, Adas Gelzinis, James Parker, and M. Charlotte Olsson, 2016, “Electromyographic patterns during golf swing: Activation sequence profiling and prediction of shot effectiveness,” Sensors, 16(4), 592.

    • Search Google Scholar
    • Export Citation

Appendix I—International Trade Financing

This appendix summarizes the four main ways in which international trade can be financed. It draws heavily from earlier work by Niepmann and Schmdt-Eisenlohr (2017).

International trade is usually financed in four ways. These are: cash-in-advance, open account, letter of credit, and documentary collection. Each of these methods has tradeoffs in terms of risks and financial costs for the exporter and importer. Table 1 summarizes these tradeoffs, which are explained below.

Table I.1.

Tradeoffs in Trade Financing

article image
Source: Niepmann and Schimdt-Eisenlohr (2017)

A. Cash-in-advance

Cash-in-advance financing requires the importer to prepay the exporter for the delivery of imported goods. As such, the importer will normally draw on a credit from a domestic bank to finance the imported goods in advance of their delivery. Once sold, the importer will repay the credit line with interest.

Risks and financial costs of cash-in-advance financing fall exclusively on the importer. The importer carries the risk that the imported goods will not be delivered on time or will be of a substandard quality. S/he also bears the financial costs of paying interest on its credit line. The exporter will not bear risks or financial costs given that it receives the payment in advance of exporting the goods. S/he also does not incur any risks, unless the importer sues in court for breach of contract, which is likely to be an expensive proposition across international borders.

Cash-in-advance financing is most prevalent in low-income countries, where importers do not have access to other forms of trade financing or where the cost of this financing is prohibitive.

B. Open account

An open account is a legal contract requiring the importer to pay the exporter once goods are received. Accordingly, the importer can verify the quality of the goods received and charge penalty for delays ahead of making a payment.

With open accounts, the risk and financial costs fall exclusively on the exporter. It is the exporter that assumes all the risks of producing and shipping the goods ahead of being paid by the importer. In addition, the financial costs also fall on the exporter as s/he will need working capital to produce the goods and ship them until payment is received.

Open accounts are very common in international trade amongst advanced economies. The reason is that it is easier for contracts to be legally enforced and relations between exporters are importers are long-standing and well-established, leading to a higher degree of trust that open accounts will be honored.

C. Letters of credit

A letter of credit (also called documentary credit) is a financial instrument that allows exporters and importers to mitigate the risk and share the financial cost of international trade. A letter of credit is a guarantee by the bank servicing the importer to make a payment to the exporter once the exporter produces documents (normally a bill of laden) that prove that the goods have been shipped. The letter of credit is usually presented to the exporter once the goods are ordered and extinguished once the goods are shipped. For the exporter, the letter of credit can be used as collateral to request a credit line from its own bank to produce the goods and ship them to the importer. For the importer, the letter of credit can turn into a credit line from its own bank once the goods are shipped until they are sold. The importing bank usually charges the importer a fixed fee and interest on a letter of credit.

With letters of credit, risks and financial costs are shared between the importer and the exporter. The risk is borne by the exporter until the goods are shipped, and they fall to the importer after that. The financial costs are shared. The exporter pays interest for the working capital until the goods are shipped. The importer pays for the credit line after the goods have been shipped. The fees for the letter of credit are often split between the importer and the exporter, but this varies from case to case.

Letters of credit are most frequently used for trade with emerging markets, particularly in Asia. They mitigate the risks between exporters and importers and share the financial burden, while generating fees and interest income for banks producing the letters of credit. They also mitigate the likelihood of legal disputes as letters of credit are usually governed by rules and regulations established by the International Chamber of Commerce (ICC).

D. Documentary collection

A documentary collection involves the transfer of the shipping/ownership documents from the exporter (or his/her bank) to the importer’s bank. The importer’s bank will then transfer the shipping/ownership documents to the importer only after payment is made to the exporter. There is no guarantee of payment involved (like in the case of a letter of credit). Rather, a documentary collection is more of a form of insurance for the exporter that the importer will pay for the goods upon receipt.

With a documentary collection, it is still the exporter who carries the risk and the financial cost. The only advantage compared with an open account is that the documentary collection ensures prompt payment by the importer once the goods are received. However, the risk and financial cost are borne by the exporter.

Documentary collections are a relatively new means of financing global trade. According to the ICC, documentary collections accounted for about one tenth the volume of letters of credit in 2017 and their use declined significantly that year.

Appendix II—Charts and Linear Regression Results14

Figure II.1.
Figure II.1.

Bangladesh Exports and SWIFT Trade Messages

(Percentage change; 3-month moving average, seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Table II.1.

Regression Results for Bangladesh Exports

article image
Standard errrors (in parenthesis) are heteroskedasticity and autocorrelation robust (HAC) using 1 lag and without small sample correction. Asterisks indicate significance at the 10 percent (*), 5 percent (**), and 1 percent (***) levels.Sources: Haver, national customs data, SWIFT, and authors’ regressions.
Figure II.2.
Figure II.2.

Brazil Imports and SWIFT Trade Messages

(Percentage change; 3-month moving average, seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Table II.2.

Regression Results for Brazil Imports

article image
Standard errrors (in parenthesis) are heteroskedasticity and autocorrelation robust (HAC) using 1 lag and without small sample correction. Asterisks indicate significance at the 10 percent (*), 5 percent (**), and 1 percent (***) levels.Sources: Haver, IHS Markit, national customs data, SWIFT, and authors’ regressions.
Figure II.3.
Figure II.3.

Chile Imports and SWIFT Trade Messages

(Percentage change; 3-month moving average, seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Table II.3.

Regression for Results Chile Imports

article image
Standard errrors (in parenthesis) are heteroskedasticity and autocorrelation robust (HAC) using 1 lag and without small sample correction. Asterisks indicate significance at the 10 percent (*), 5 percent (**), and 1 percent (***) levels.Sources: Haver, IHS Markit, national customs data, SWIFT, and authors’ regressions.
Figure II.4.
Figure II.4.

China Imports and SWIFT Trade Messages

(Percentage change; 3-month moving average, seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Table II.4.

Regression Results for China Imports

article image
Standard errrors (in parenthesis) are heteroskedasticity and autocorrelation robust (HAC) using 1 lag and without small sample correction. Asterisks indicate significance at the 10 percent (*), 5 percent (**), and 1 percent (***) levels.Sources: Haver, IHS Markit, national customs data, SWIFT, and authors’ regressions.
Figure II.5.
Figure II.5.

France Exports and SWIFT Trade Messages

(Percentage change; 3-month moving average, seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Table II.5.

Regression Results for France Exports

article image
Standard errrors (in parenthesis) are heteroskedasticity and autocorrelation robust (HAC) using 1 lag and without small sample correction. Asterisks indicate significance at the 10 percent (*), 5 percent (**), and 1 percent (***) levels.Sources: Haver, IHS Markit, national customs data, SWIFT, and authors’ regressions.
Figure II.6.
Figure II.6.

Ghana Imports and SWIFT Trade Messages

(Percentage change; 3-month moving average, seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Table II.6.

Regression Results for Ghana Imports

article image
Standard errrors (in parenthesis) are heteroskedasticity and autocorrelation robust (HAC) using 1 lag and without small sample correction. Asterisks indicate significance at the 10 percent (*), 5 percent (**), and 1 percent (***) levels.Sources: Haver, national customs data, SWIFT, and authors’ regressions.
Figure II.7.
Figure II.7.

India Exports and SWIFT Trade Messages

(Percentage change; 3-month moving average, seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Table II.7.

Regression Results for India Exports

article image
Standard errrors (in parenthesis) are heteroskedasticity and autocorrelation robust (HAC) using 1 lag and without small sample correction. Asterisks indicate significance at the 10 percent (*), 5 percent (**), and 1 percent (***) levels.Sources: Haver, IHS Markit, national customs data, SWIFT, and authors’ regressions.
Figure II.8.
Figure II.8.

Indonesia Imports and SWIFT Trade Messages

(Percentage change; 3-month moving average, seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Table II.8.

Regression Results for Indonesia Imports

article image
Standard errrors (in parenthesis) are heteroskedasticity and autocorrelation robust (HAC) using 1 lag and without small sample correction. Asterisks indicate significance at the 10 percent (*), 5 percent (**), and 1 percent (***) levels.Sources: Haver, IHS Markit, national customs data, SWIFT, and authors’ regressions.
Figure II.9.
Figure II.9.

Israel Exports and SWIFT Trade Messages

(Percentage change; 3-month moving average, seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Table II.9.

Regression Results for Israel Exports

article image
Standard errrors (in parenthesis) are heteroskedasticity and autocorrelation robust (HAC) using 1 lag and without small sample correction. Asterisks indicate significance at the 10 percent (*), 5 percent (**), and 1 percent (***) levels.Sources: Haver, IHS Markit, national customs data, SWIFT, and authors’ regressions.
Figure II.10.
Figure II.10.

Israel Imports and SWIFT Trade Messages

(Percentage change; 3-month moving average, seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Table II.10.

Regression Results for Israel Imports

article image
Standard errrors (in parenthesis) are heteroskedasticity and autocorrelation robust (HAC) using 1 lag and without small sample correction. Asterisks indicate significance at the 10 percent (*), 5 percent (**), and 1 percent (***) levels.Sources: Haver, IHS Markit, national customs data, SWIFT, and authors’ regressions.
Figure II.11.
Figure II.11.

Italy Exports and SWIFT Trade Messages

(Percentage change; 3-month moving average, seasonally adjusted)

Citation: IMF Working Papers 2020, 247; 10.5089/9781513561196.001.A001

Table II.11.

Regression Results for Italy Exports

article image
Standard errrors (in parenthesis) are heteroskedasticity and autocorrelation robust (HAC) using 1 lag and without small sample correction. Asterisks indicate significance at the 10 percent (*), 5 percent (**), and 1 percent (***) levels.Sources: Haver, IHS Markit, national customs data, SWIFT, and authors’ regressions.