## 1. Share of World Trade Financed by SWIFT Letters of Credit

## Source details

**Canonical URL:** [1. Share of World Trade Financed by SWIFT Letters of Credit](https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021293-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2021/english/wpiea2021293-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2021/english/wpiea2021293-print-pdf.pdf.json)

---

### Introduction
- Paper complements an earlier IMF working paper published in November 2020 (CHMMR) by adding SWIFT data on documentary collections to improve short-term forecasts of international trade.
- Documentary collections are examined via SWIFT identification code MT 400 to assess their explanatory power as an early indicator of future trade.

### Documentary collections: definition and mechanics
- Documentary collections are a financial agreement between exporters, importers, and their financial institutions wherein the collecting bank releases importing documents to the importer only once the importer has paid for the imported goods.
- Process summary:
  - Exporter ships goods and submits shipping documents to the remitting bank.
  - Remitting bank forwards documents to the collecting bank.
  - Collecting bank releases documents to importer only after payment or agreed payment on due date.
- Key distinctions from letters of credit:
  - No financial guarantee of payment (unlike letters of credit).
  - Disputes typically handled within importing economy’s legal jurisdiction.
  - Lower cost to importer than letters of credit.
- Lead time:
  - SWIFT messages (MT 400 and MT 700) are typically sent before title documents are transferred and customs clearance; lead time varies by exporter production lags, geographic distance, and customs clearance time.

### SWIFT usage and geographic patterns
- SWIFT message shares in 2020:
  - Letters of credit (MT 700) accounted for about 12 percent of world merchandise trade in 2020.
  - Documentary collections (MT 400) were used to finance just over 1 percent of world trade in 2020.
- Main flows and users:
  - Letters of credit: within triangle between China, Hong Kong SAR, and Singapore.
  - Documentary collections: between Asia (notably China, India, Korea, and Singapore) and the United States.
  - Among top 10 importers in 2020, main users of SWIFT MT 400 messages: Hong Kong SAR, the United States, and India.
  - Among top 10 exporters in 2020, documentary collections used most by Korea, Hong Kong SAR, and the Netherlands.
- Caveat:
  - Results may reflect location of international banks associated with SWIFT MT 400 messages (e.g., Hong Kong SAR, the Netherlands, the United States) rather than direct trade flows of those economies.

### Regression framework used to assess explanatory power
- Reduced-form regression (Equation 1) of world trade (WT) in log differences:
  - Dependent variable: dlog(WT_t)
  - Regressors: up to four lags of dlog(WT), SWIFT MT 400 messages (SWIFT4) and SWIFT MT 700 messages (SWIFT7) up to four lags (including contemporaneous), Brent crude oil prices (Brent) up to four lags in log differences, and new export orders subcomponent of global manufacturing PMI (PMI) up to four lags (PMI re-centered around zero).
  - Lag indices: i = 1..4 for WT lags; j = 0..4 for SWIFT and other regressors.

### Regression results and interpretation (world trade aggregate)
- Specification with only SWIFT messages:
  - Contemporaneous coefficients on SWIFT4 and SWIFT7 and some of their lags are positive and highly significant.
  - R2 of just over 50 percent.
- Once Brent and PMI are included:
  - Contemporaneous coefficient on SWIFT4 remains highly significant.
  - SWIFT7 coefficients become insignificant.
  - Interpretation: SWIFT7 and Brent appear non-orthogonal; SWIFT7 contribution is diminished when Brent is included (possible link to hydrocarbon trade financing).
- Full specification diagnostics:
  - R2 of close to 80 percent.
  - Adjusted R2 of 75 percent.
- Key SWIFT coefficient estimates (selected, preserved exactly as reported):
  - SWIFT4: 0.3448 (0.110)***; 0.2371 (0.070)***; 0.1379 (0.045)***; 0.0770 (0.036)**.
  - SWIFT4 (-1): 0.2062 (0.071)***; 0.1586 (0.062)**; 0.0525 (0.040); 0.0489 (0.032).
  - SWIFT7: 0.1561 (0.055)***; 0.0186 (0.025); 0.0219 (0.025).
- Brent coefficients (selected):
  - Brent: 0.1010 (0.018)***; 0.0622 (0.016)***; 0.0804 (0.015)***; 0.1205 (0.020)***.
  - Brent (-1): 0.1030 (0.021)***; 0.0577 (0.019)***; 0.0638 (0.020)***; 0.1068 (0.022)***.

### Forecast of world trade during Covid-19 (SWIFT linear forecasts)
- Macroeconomic context:
  - World real GDP contracted by 3.1 percent in 2020.
  - World real GDP projected to rebound by 5.9 percent in 2021.
- World trade dynamics:
  - World trade contracted by more than 20 percent in the first four months of 2020, lowest in April/May 2020.
  - From June 2020, strong growth; surpassed pre-pandemic levels in October 2020.
- SWIFT-based forecast performance:
  - By end-March 2020, forecast indicated a significant decline in world trade while CPB data were only available for January 2020.
  - By July 2020, forecast pointed to a recovery (though not as rapid as observed).
  - Starting October 2020, forecast indicated a rapid recovery well above pre-pandemic levels; later forecasts indicated stabilization in second half of 2021, albeit still well above pre-pandemic levels.
  - Interpretation: Linear regression forecasts did relatively well at picking up turning points during Covid-19; stabilization in late 2021 likely reflects supply disruptions and limited shipping capacity.

### Dynamic Factor Model (DFM) comparison — methodology and findings
- DFM specification:
  - Uses 28 macroeconomic time series (global financial conditions, economic sentiment, manufacturing output, commodity prices, Baltic Dry and Container throughput indices).
  - Two-step estimation yields six common factors.
  - CPB world trade represented as linear combination of six factors.
  - Forecasts include 95 percent confidence intervals.
- DFM forecast performance (January 2020–December 2020):
  - February 2020 DFM forecast pointed to a significant decline; March and April forecasts reinforced decline.
  - DFM from May to August 2020 significantly underestimated the recovery.
  - Starting September 2020, DFM correctly anticipated the rapid recovery well above pre-pandemic level.
- Operational advantage:
  - DFM can be updated as soon as new data (e.g., PMI) are published, allowing more frequent forecasts than SWIFT linear regression.
- DFM empirical findings:
  - Decomposition: dlog(CPB_B_TRADE_t) = 0.8*f_t1 + 0.3*f_t2 -0.2*f_t3 + 0.3*f_t4 + 0.1*f_t5 + 0.1*f_t6 + ϑ_t.
  - Common component explains 86 percent of global trade.
  - Dataset balanced panel span: October 2010 to October 2020.
  - Hyperparameters chosen: Number of factors = 6; Lag length of underlying VAR = 4.

### Forecast performance metrics: SWIFT vs DFM (January–December 2020; Table 2)
- Average Out-of-Sample Root Mean Squared Errors (Log difference)
  - Forecast Horizon (Months ahead): 1 2 3 4 5 6 Horizon Avg.
  - SWIFT Linear Regression Forecast RMSEs: 4.2 5.8 6.6 7.1 6.8 4.3 5.8
  - DFM RMSEs: 4.5 4.1 6.6 6.2 5.7 6.1 5.6
- Interpretation:
  - One-month ahead: SWIFT linear forecasts had better RMSEs.
  - Horizons up to five months: DFM outperformed SWIFT.
  - Over one- to six-month horizon: DFM forecast did somewhat better on average.

### Documentary collections, SWIFT messages, and national trade forecasting
- Main challenge:
  - SWIFT trade messages correlate with world trade but are blurred for national trade because large international banks may send/receive messages in countries different from trade origin/destination.
  - Blurring especially relevant for economies hosting international financial centers (Hong Kong SAR, Japan, Netherlands, United Kingdom, United States).
  - No data yet available to correct for this blurred relationship.
- MT 400 national correlations:
  - Simple contemporaneous correlations (MT 400 vs national merchandise exports/imports) computed for 60 economies.
  - Highest correlation coefficients: Vietnam and Egypt imports; Malaysia exports; Hong Kong SAR imports; Portugal, Spain, and Singapore exports.
  - MT 400 correlation pattern differs from MT 700.

### Regional patterns in SWIFT message relevance (selected results)
- Asia:
  - MT 700 positive and significant for Bangladesh imports/exports, China and Hong Kong SAR imports, Indonesia and India imports/exports, Japan and Korea imports, Philippines exports, Taiwan Province of China imports/exports, Vietnam imports.
  - MT 400 positive and significant for Bangladesh imports, China exports, Hong Kong SAR imports, Indonesia exports, Korea imports, Malaysia and Singapore exports, Thailand imports/exports, Vietnam imports.
- Oceania:
  - MT 700 positive and highly significant for Australia exports and New Zealand imports.
  - MT 400 positive and significant only for Australia exports.
- Europe:
  - MT 700 positive and significant for Belgium, Estonia, Finland, France, Greece, Latvia, Lithuania, Slovenia, Sweden, UK exports; Latvia, Portugal, Russia, Slovenia imports.
  - MT 400 positive and significant for France, Hungary, Lithuania, Netherlands, Portugal, Spain, Sweden exports; Russia imports.
  - Many European economies show insignificant coefficients for both messages.
- Western Hemisphere:
  - MT 700 positive and significant for Chile, Mexico, Peru exports; Argentina, Mexico, Peru imports.
  - MT 400 positive and significant only for Argentina and Peru exports; Peru imports.
  - Coefficients insignificant for Brazil, Canada, Colombia, Mexico, United States for both messages.
- Middle East:
  - MT 700 positive and significant for Israel, Pakistan, Turkey exports; Saudi Arabia, Turkey imports.
  - MT 400 positive and significant for Egypt, Saudi Arabia, Turkey imports; Israel and Saudi Arabia exports (Saudi Arabia non-oil).
- Africa:
  - MT 700 relevant for Ghana imports and Nigeria exports only; other regressions largely insignificant.
- Overall interpretation:
  - Letters of credit used mostly by emerging economies; advanced economies predominantly use open accounts; developing economies often use cash-in-advance systems.

### Machine learning algorithms (MLA) vs linear regression — forecasts and evaluation
- Forecast construction (linear):
  - One-step ahead forecasts recursively up to six-months ahead; SWIFT extended via AR(1) where needed; Brent prices extended using closing futures for 1-, 3-, and 6-month contracts with interpolation for other months; PMI extended via AR(1) when needed.
  - Example: November 30, 2020 linear regression forecast used SWIFT increases (20 percent letters of credit, 10 percent documentary collections from April trough), Brent recovery to about $47 per barrel and futures at $49 per barrel six months out, and export orders PMI in slight expansion.
- MLAs used:
  - Linear MLAs: Lasso and Ridge.
  - Single nonparametric MLAs: Decision Tree Regression and Support Vector Regression.
  - Ensemble nonparametric MLAs: Bagging, Gradient Boost, Random Forest.
- Evaluation design:
  - Training set up to August 2020; test set September 2020–August 2021; monthly rolling forecasts; RMSE computed over 1–6 month horizons.
- Findings:
  - Best-performing one-step ahead forecasts: about one third linear/parametric (Ridge, Lasso); two thirds non-linear MLAs.
  - Contrast with CHMMR: prior ratio was two-thirds linear and one-third non-linear — Covid-19 introduced significant non-linearities.
  - Diebold-Mariano tests: about one quarter of regressions have forecasts statistically better than a naïve constant-growth forecast.
- Example:
  - China imports August 31, 2021: linear regression predicted 8.3 percent rebound; Bagging predicted 3.5 percent rebound; actual data showed 4.1 percent increase — Bagging closer to outcome.

### Selected country regression highlights (exact coefficients and diagnostics preserved)
- Bangladesh Imports (Observations: 123)
  - SWIFT4: 0.0319 (0.018)*
  - SWIFT7: 0.168 (0.052)***
  - BRENT: 0.2376 (0.054)***
  - R2: 0.65; Adjusted R2: 0.59; F-Statistic: 17.35; Prob. (F-Statistic): 5.13e-24; Durbin-Watson: 1.922
- China Exports (Observations: 124)
  - SWIFT4: 0.3183 (0.113)***; SWIFT4 (-2): 0.3683 (0.099)***
  - SWIFT7: 0.0017 (0.112)
  - PMI: 0.0162 (0.01)
  - R2: 0.57; Adjusted R2: 0.46; F-Statistic: 5.115; Prob. (F-Statistic): 3.14e-9; Durbin-Watson: 2.085
- China Imports (Observations: 124)
  - SWIFT7: 0.1351 (0.056)**; SWIFT7 (-3): 0.2485 (0.076)***
  - PMI: 0.0058 (0.002)***
  - R2: 0.55; Adjusted R2: 0.44; F-Statistic: 11.11; Prob. (F-Statistic): 1.05e-18; Durbin-Watson: 2.035
- Egypt Imports (Observations: 119)
  - SWIFT4: 0.1979 (0.054)***; SWIFT4 (-1): 0.1566 (0.054)***
  - R2: 0.51; Adjusted R2: 0.38; F-Statistic: 7.344; Prob. (F-Statistic): 6.53e-13; Durbin-Watson: 2.039
- France Exports (Observations: 123)
  - SWIFT4 series positive across multiple lags (e.g., SWIFT4 (-1): 0.1212 (0.033)***)
  - BRENT: 0.145 (0.034)***
  - PMI: 0.007 (0.002)***
  - R2: 0.70; Adjusted R2: 0.62; F-Statistic: 6.182; Prob. (F-Statistic): 3.64e-11; Durbin-Watson: 2.025
- Korea Imports (Observations: 124)
  - SWIFT4: 0.048 (0.022)**; SWIFT4 (-2): 0.0799 (0.028)***; SWIFT7: 0.1861 (0.057)***
  - BRENT (-1): 0.1283 (0.029)***; BRENT (-2): 0.1227 (0.026)***; BRENT (-3): 0.101 (0.033)***
  - R2: 0.63; Adjusted R2: 0.54; F-Statistic: 26.44; Prob. (F-Statistic): 7.06e-33; Durbin-Watson: 2.108
- Pakistan Imports (Observations: 124)
  - SWIFT7 series large and positive across multiple lags (e.g., SWIFT7: 0.3022 (0.087)***; SWIFT7 (-1): 0.5523 (0.143)***)
  - R2: 0.56; Adjusted R2: 0.48; F-Statistic: 10.37; Prob. (F-Statistic): 1.97e-16; Durbin-Watson: 2.04
- Vietnam Imports (Observations: 121)
  - SWIFT4: 0.1191 (0.051)**; SWIFT4 (-3): 0.1415 (0.049)***
  - SWIFT7: 0.3617 (0.081)***; SWIFT7 (-4): 0.1726 (0.059)***
  - BRENT: -0.1633 (0.053)***
  - R2: 0.70; Adjusted R2: 0.63; F-Statistic: 13.42; Prob. (F-Statistic): 3.65e-21; Durbin-Watson: 2.061
- Cross-cutting patterns:
  - Lagged trade activity frequently negative and highly significant (examples: Bangladesh Imports (-1): -0.7829 (0.093)***; Vietnam Imports (-1): -0.8751 (0.099)***).
  - SWIFT4 and SWIFT7 often positive and sometimes significant across countries.
  - BRENT coefficients vary by country and lag; sign and significance differ across regressions.
  - PMI effects generally small; occasionally significant.

### Conclusions — key findings and methodological implications
- Key findings:
  - Documentary collections finance only about one percent of world trade but carry strong informational content for forecasting world trade and international trade in a selected number of economies, mostly in Asia.
  - The SWIFT linear regression forecast performed relatively well during the trough and rebound of the Covid-19 crisis in 2020-21.
  - The SWIFT linear regression performance was broadly equivalent to an alternative DFM forecast over the same period based on 27 different variables.
  - Covid-19 introduced significant non-linearities in the relationship between international trade and its regressors.
- Forecast performance and model comparison:
  - Linear SWIFT-based forecasts delivered performance comparable to a DFM using 27 variables during the 2020–21 trough and rebound.
  - Non-linear and machine learning algorithms can improve on linear regression forecasts, particularly during large shocks.
- Methodological implications:
  - It is useful to run MLA forecasts alongside linear regressions to capture regime changes and non-linear dynamics.
  - Documentary collections (SWIFT MT 400 and MT 700) provide timely explanatory variables that can appear with a lead relative to customs data, improving short-term forecast information content.
- Statistical evaluation:
  - Diebold-Mariano testing applied to one-step ahead forecasts; about one quarter of regressions show forecasts statistically better than a naïve constant-growth benchmark.

*Source: wpiea2021293-print-pdf - IMF working paper content provided in the source document.*

### 1. Share of World Trade Financed by SWIFT Letters of Credit ............................................................

### 1. Share of World Trade Financed by SWIFT Letters of Credit ............................................................................ 5

### Introduction
- Paper complements an earlier IMF working paper published in November 2020 (CHMMR) by adding SWIFT data on documentary collections to improve short-term forecasts of international trade.
- Documentary collections are examined via SWIFT identification code MT 400 to assess their explanatory power as an early indicator of future trade activity.

### Documentary Collections: definition and mechanics
- Documentary collections are a financial agreement between exporters, importers, and their financial institutions wherein the collecting bank releases importing documents to the importer only once the importer has paid for the imported goods.
- Process summary:
  - Exporter ships goods and submits shipping documents to the remitting bank.
  - Remitting bank forwards documents to the collecting bank.
  - Collecting bank releases documents to importer only after payment or agreed payment on due date.
- Key distinctions from letters of credit:
  - No financial guarantee of payment (unlike letters of credit).
  - Disputes typically handled within importing economy’s legal jurisdiction (less risk of lengthy international legal disputes).
  - Lower cost to importer than letters of credit; commonly used among established trading partners not yet on open-account terms.
- Lead time: SWIFT messages (MT 400 and MT 700) are typically sent before title documents are transferred and customs clearance, so they can serve as leading indicators of trade; lead time varies by exporter production lags, geographic distance, and customs clearance time.

### SWIFT usage and geographic patterns
- SWIFT message shares in 2020:
  - Letters of credit (MT 700) accounted for about 12 percent of world merchandise trade in 2020.
  - Documentary collections (MT 400) were used to finance just over 1 percent of world trade in 2020.
- Main export flows financed by letters of credit: within triangle between China, Hong Kong SAR, and Singapore.
- Main export flows associated with documentary collections: between Asia (notably China, India, Korea, and Singapore) and the United States.
- Among top 10 importers in 2020, main users of SWIFT MT 400 messages: Hong Kong SAR, the United States, and India.
- Among top 10 exporters in 2020, documentary collections were used most by Korea, Hong Kong SAR, and the Netherlands.
- Caveat: results may reflect location of international banks associated with SWIFT MT 400 messages (e.g., Hong Kong SAR, the Netherlands, the United States) rather than direct trade flows of those economies.

### Regression framework used to assess explanatory power
- Reduced-form regression (Equation 1) of world trade (WT) in log differences:
  - Dependent variable: dlog(WT_t)
  - Regressors: up to four lags of dlog(WT), SWIFT MT 400 messages (SWIFT4) and SWIFT MT 700 messages (SWIFT7) up to four lags (including contemporaneous), Brent crude oil prices (Brent) up to four lags in log differences, and new export orders subcomponent of global manufacturing PMI (PMI) up to four lags (PMI re-centered around zero).
  - Lag indices: i = 1..4 for WT lags; j = 0..4 for SWIFT and other regressors.

### Regression results and interpretation
- In the specification with only SWIFT messages (third column referenced):
  - Contemporaneous coefficients on SWIFT4 and SWIFT7 and some of their lags are positive and highly significant.
  - R2 of just over 50 percent.
- Once Brent crude oil prices and PMI are included:
  - Contemporaneous coefficient on SWIFT4 remains highly significant.
  - SWIFT7 coefficients become insignificant.
  - Interpretation: SWIFT7 and Brent appear non-orthogonal; SWIFT7 contribution is diminished when Brent is included. Possible explanation: SWIFT7 messages are partly used to finance a significant portion of hydrocarbon trade in Asia.
- Overall specification with all explanatory variables:
  - R2 of close to 80 percent.
  - Adjusted R2 of 75 percent.
- Conclusion: SWIFT MT 400 messages provide significant explanatory power for world trade in reduced-form regressions even after controlling for Brent and PMI.

### Forecast of World Trade during Covid-19 (context and performance)
- Macroeconomic context cited:
  - World real GDP is estimated to have contracted by 3.1 percent in 2020.
  - World real GDP is projected to rebound by 5.9 percent in 2021.
- World trade dynamics during Covid-19:
  - World trade contracted by more than 20 percent in the first four months of 2020, reaching its lowest level in April/May 2020.
  - Starting from June 2020, world trade showed strong growth and surpassed pre-pandemic levels in October 2020 (figure reference in source).
- Use of SWIFT-based regression:
  - Specification with SWIFT4, SWIFT7, Brent, and PMI provides a basis for short-term forecasting of world trade, including during the Covid-19 crisis period.
  - SWIFT4 (documentary collections) is useful as an early indicator because financing often precedes actual movement of goods.

### Key findings and implications
- Documentary collections (SWIFT MT 400) accounted for just over 1 percent of world trade in 2020 but have strong correlation with world trade and international trade in selected economies.
- SWIFT MT 400 messages are particularly relevant for financing imports from Asia, where parties prefer keeping legal arrangements within their jurisdiction rather than using letters of credit.
- SWIFT MT 400 messages retain significant contemporaneous explanatory power for world trade when Brent and PMI are included; SWIFT MT 700 messages lose significance once Brent is controlled for.
- The full specification achieves high explanatory power: R2 close to 80 percent and adjusted R2 of 75 percent, supporting the use of SWIFT-based reduced-form equations for short-term trade forecasting.

*INTERNATIONAL MONETARY FUND*

### 2020. By contrast, the Global Financial Crisis (GFC) of 2008-10 had a much sharper contraction over the first

### World Trade: Comparison of Different Regression Specifications

### Regression results (Table 1) — key coefficients and diagnostics
- Sample: April 2011-June 2021 (123 observations). Standard errors are heteroscedasticity and autocorrelation robust (HAC) using 1 lag and without small sample correction. Asterisks indicate significance at 10 percent (*), 5 percent (**), and 1 percent (***).
- Constant coefficients reported: 0.0012 (0.002); 0.0051 (0.002)***; 0.0043 (0.002)**; 0.0035 (0.001)***; 0.0023 (0.001)**; 0.0018 (0.001); 0.0028 (0.001)**.
- World Trade lag coefficients (selected):
  - World Trade (-1): 0.3205 (0.133)**; 0.1371 (0.088); -0.0858 (0.106); -0.1223 (0.099); -0.2323 (0.085)***; -0.2055 (0.078)***; -0.1034 (0.106).
  - World Trade (-2): -0.0582 (0.218); -0.1090 (0.153); -0.0865 (0.141); 0.0483 (0.105); 0.0452 (0.089); 0.0455 (0.089); -0.0141 (0.108).
  - World Trade (-3): -0.0359 (0.079); -0.1972 (0.090)**; -0.1498 (0.106); -0.0598 (0.119); 0.0648 (0.104); 0.1205 (0.097); -0.0252 (0.096).
  - World Trade (-4): 0.0437 (0.096); -0.1748 (0.103)*; -0.1976 (0.128); 0.0140 (0.108); 0.2005 (0.077)**; 0.2454 (0.067)***; 0.1166 (0.088).
- SWIFT coefficients (selected):
  - SWIFT4: 0.3448 (0.110)***; 0.2371 (0.070)***; 0.1379 (0.045)***; 0.0770 (0.036)**.
  - SWIFT4 (-1): 0.2062 (0.071)***; 0.1586 (0.062)**; 0.0525 (0.040); 0.0489 (0.032).
  - SWIFT4 (-3): 0.1330 (0.056)**; 0.0447 (0.050); 0.0384 (0.053); 0.0413 (0.044).
  - SWIFT4 (-4): 0.1873 (0.083)**; 0.1432 (0.066)**; 0.0675 (0.047); 0.0225 (0.038).
  - SWIFT7: 0.1561 (0.055)***; 0.0186 (0.025); 0.0219 (0.025).
  - SWIFT7 (-1): 0.1207 (0.033)***; -0.0135 (0.031); 0.0174 (0.027).
- Brent crude oil coefficients (selected):
  - Brent: 0.1010 (0.018)***; 0.0622 (0.016)***; 0.0804 (0.015)***; 0.1205 (0.020)***.
  - Brent (-1): 0.1030 (0.021)***; 0.0577 (0.019)***; 0.0638 (0.020)***; 0.1068 (0.022)***.
  - Brent (-2): 0.0194 (0.023); -0.0078 (0.019); -0.0029 (0.020); 0.0127 (0.023).
  - Brent (-3): 0.0258 (0.025); 0.0128 (0.024); 0.0152 (0.018)*; 0.0308 (0.019).
  - Brent (-4): -0.0003 (0.017); -0.0221 (0.014); -0.0259 (0.016)*; -0.0063 (0.019).
- PMI coefficients (selected): PMI 0.0047 (0.001)***; 0.0048 (0.001)***. PMI lags generally small and mostly insignificant.
- Diagnostics (selected models): R2 values reported as 0.097, 0.376, 0.507, 0.706, 0.801, 0.779, 0.661. Adjusted R2: 0.066, 0.326, 0.443, 0.652, 0.753, 0.750, 0.634. F-Statistic and Prob. (F-Statistic) vary by specification (examples: F-Statistic 2.807 with Prob. 0.0059; 29.334 with Prob. 29.334e-? reported formatting; Log-likelihood examples: 289.45, 312.15, 326.71, 358.46, 382.54, 375.90, 349.65). Durbin-Watson around 1.95–2.07.

### SWIFT linear forecasts: performance during Covid-19 (Figures 5–6 and text)
- SWIFT linear forecasts:
  - By end-March 2020, forecast indicated a significant decline in world trade while CPB data were only available for January 2020.
  - By July 2020, forecast pointed to a recovery in global trade, albeit not as rapid as observed.
  - Starting October 2020, forecast indicated a rapid recovery well above pre-pandemic levels; later forecasts indicated stabilization in second half of 2021, albeit still well above pre-pandemic levels.
- Interpretation:
  - Linear regression forecasts did relatively well at picking up turning points of world trade during the Covid-19 crisis.
  - Stabilization in late 2021 likely reflects supply disruptions and limited shipping capacity.

### Dynamic Factor Model (DFM) comparison (methodology and findings; Figures 7–8)
- DFM specification:
  - Uses a wider set of 28 macroeconomic time series (global financial conditions, economic sentiment, manufacturing output, commodity prices, Baltic Dry and Container throughput indices).
  - Two-step estimation yields six common factors interpreted as major forces defining global economic developments.
  - CPB world trade (value) represented as linear combination of these six factors.
  - Forecasts include 95 percent confidence intervals.
- DFM forecast performance (January 2020–December 2020):
  - February 2020 DFM forecast pointed to a significant decline in world trade, reinforced by March and April forecasts.
  - DFM forecasts from May to August 2020 significantly underestimated the recovery in global trade.
  - Starting September 2020, DFM correctly anticipated the rapid recovery well above pre-pandemic level.
- Operational advantage:
  - DFM can be updated as soon as new data (e.g., PMI) are published, allowing more frequent forecasts than SWIFT linear regression.

### Forecast performance metrics: SWIFT vs DFM (Table 2)
- Table 2: Average Out-of-Sample Root Mean Squared Errors (Log difference; January to December 2020)
  - Forecast Horizon (Months ahead): 1 2 3 4 5 6 Horizon Avg.
  - SWIFT Linear Regression Forecast RMSEs: 4.2 5.8 6.6 7.1 6.8 4.3 5.8
  - DFM RMSEs: 4.5 4.1 6.6 6.2 5.7 6.1 5.6
- Interpretation:
  - One-month ahead: SWIFT linear forecasts had better RMSEs on average.
  - Longer horizons up to five months: DFM outperformed SWIFT.
  - Over one- to six-month horizon: DFM forecast did somewhat better on average, possibly due to inclusion of broader set of variables, including financial variables.

### Documentary collections, SWIFT messages, and national trade forecasting
- Main challenge:
  - SWIFT trade messages correlate with world trade but are blurred for national trade because large international banks may send/receive messages in countries different from where trade originates or is destined.
  - This blurring is especially relevant for economies hosting international financial centers (Hong Kong SAR, Japan, Netherlands, United Kingdom, United States).
  - No data yet available to correct for this blurred relationship.
- MT 400 (documentary collections):
  - Use varies across economies, mostly prevalent in trade with Asia.
  - Simple contemporaneous correlations (MT 400 vs national merchandise exports/imports) computed for 60 economies.
  - Highest correlation coefficients: Vietnam and Egypt imports; Malaysia exports; Hong Kong SAR imports; Portugal, Spain, and Singapore exports.
  - Correlation pattern for MT 400 differs from MT 700 (letters of credit) previously reported.
- Regression framework:
  - 120 regressions (exports and imports for 60 economies) estimated in log differences, using lagged customs data, SWIFT MT 700 and MT 400, Brent crude prices, and new export orders subcomponent of national manufacturing PMI where available. Ordinary least squares with HAC standard errors used.

### Regional patterns in SWIFT message relevance (findings from regressions and figures)
- Asia:
  - MT 700 positive and significant for Bangladesh imports/exports, China and Hong Kong SAR imports, Indonesia and India imports/exports, Japan and Korea imports, Philippines exports, Taiwan Province of China imports/exports, Vietnam imports.
  - MT 400 positive and significant for Bangladesh imports, China exports, Hong Kong SAR imports, Indonesia exports, Korea imports, Malaysia and Singapore exports, Thailand imports/exports, Vietnam imports.
  - Conclusion: both MT 700 and MT 400 used extensively in Asia, likely for different goods/origins/destinations.
- Oceania:
  - MT 700 positive and highly significant for Australia exports and New Zealand imports.
  - MT 400 positive and significant only for Australia exports.
- Europe:
  - MT 700 positive and significant for Belgium, Estonia, Finland, France, Greece, Latvia, Lithuania, Slovenia, Sweden, UK exports; Latvia, Portugal, Russia, Slovenia imports.
  - MT 400 positive and significant for France, Hungary, Lithuania, Netherlands, Portugal, Spain, Sweden exports; Russia imports.
  - Many European economies show insignificant coefficients for both messages, suggesting limited use of letters of credit and documentary collections.
- Western Hemisphere:
  - MT 700 positive and significant for Chile, Mexico, Peru exports; Argentina, Mexico, Peru imports.
  - MT 400 positive and significant only for Argentina and Peru exports; Peru imports.
  - Coefficients insignificant for Brazil, Canada, Colombia, Mexico, United States for both messages.
- Middle East:
  - MT 700 positive and significant for Israel, Pakistan, Turkey exports; Saudi Arabia, Turkey imports.
  - MT 400 positive and significant for Egypt, Saudi Arabia, Turkey imports; Israel and Saudi Arabia exports (Saudi Arabia non-oil).
- Africa:
  - MT 700 relevant for Ghana imports and Nigeria exports only; other regressions largely insignificant.
- Overall interpretation:
  - Results align with literature: letters of credit used mostly by emerging economies; advanced economies predominantly use open accounts; developing economies often use cash-in-advance systems.

### Horse race: Linear regression vs machine-learning forecasts (MLAs)
- Forecast construction (linear):
  - One-step ahead forecasts recursively up to six-months ahead; CPB and national customs extended using one-step forecast; SWIFT data extended via AR(1) where needed; Brent prices extended using closing futures for 1-, 3-, and 6-month contracts with interpolation for other months; PMI extended via AR(1) when needed.
  - Example: November 30, 2020 linear regression forecast indicated a significant rise in world trade going forward based on SWIFT increases (20 percent letters of credit, 10 percent documentary collections from April trough), Brent recovery to about $47 per barrel and futures at $49 per barrel six months out, and export orders PMI in slight expansion.
  - Linear forecast on November 30, 2020 broadly matched world trade up to December 2020; world trade was decidedly stronger in Q1 2021 than forecast.
- Machine-learning algorithms used:
  - Linear MLAs: Lasso and Ridge.
  - Single nonparametric MLAs: Decision Tree Regression and Support Vector Regression.
  - Ensemble nonparametric MLAs: Bagging, Gradient Boost, Random Forest (tree-based ensembles).
  - MLAs trained over full sample and used to produce 1–6 month ahead forecasts.
- Examples of MLA advantage:
  - China imports August 31, 2021 forecast: linear regression predicted 8.3 percent rebound in August; Bagging forecast predicted 3.5 percent rebound; actual data showed 4.1 percent increase — Bagging closer to outcome, suggesting non-linearities in China import relationships.

### Evaluation of linear vs machine-learning forecasts (RMSEs, DM tests, Table 3 summary)
- Evaluation design:
  - Training set up to August 2020; test set September 2020–August 2021; monthly rolling forecasts; RMSE computed over 1–6 month horizons.
- Findings:
  - Best-performing one-step ahead forecasts for world and 60 economies: about one third of best-performing forecasts are linear or parametric (Ridge, Lasso); two thirds are non-linear MLAs.
  - Contrast with CHMMR: prior ratio was two-thirds linear and one-third non-linear — Covid-19 introduced significant non-linearities.
  - Diebold-Mariano (DM) tests applied: about one quarter of regressions have forecasts statistically better than a naïve constant-growth forecast (based on DM P-statistics and RMSEs).
- Implication:
  - Non-linear MLAs often outperform linear approaches during periods with strong non-linearities (e.g., Covid-19); however, only a minority of forecasts are statistically superior to naïve benchmarks, indicating room for improvement (e.g., inclusion of additional explanatory variables, financial variables as in DFM).

*Source: IMF authors’ regression results and analysis contained in the cited document.*

### Conclusions

### Conclusions

### Key findings
- Documentary collections finance only about one percent of world trade but carry strong informational content for forecasting world trade and international trade in a selected number of economies, mostly in Asia.
- The SWIFT linear regression forecast performed relatively well during the trough and rebound of the Covid-19 crisis in 2020-21.
- The SWIFT linear regression performance was broadly equivalent to an alternative DFM forecast over the same period based on 27 different variables.
- The Covid-19 crisis introduced significant non-linearities in the relationship between international trade and its regressors, highlighting limitations of purely linear approaches.

### Forecast performance and model comparison
- During the Covid-19 trough and rebound (2020-21), linear SWIFT-based forecasts delivered performance comparable to a DFM using 27 variables.
- Non-linear and machine learning algorithms (MLA) can improve on linear regression forecasts, particularly during large shocks to the world economy.

### Statistical significance and evaluation
- Diebold-Mariano testing was used to compare one-step ahead forecasts; bolded forecasts in the accompanying table indicate tests significant at the 95th percentile.
- The Diebold and Mariano (1995) test framework was applied to assess forecast superiority.

### Methodological implications
- Given substantial non-linearities during major shocks, it is useful to run MLA forecasts alongside linear regressions to capture regime changes and non-linear dynamics.
- Documentary collections (SWIFT MT 400 and MT 700) provide timely explanatory variables that can appear with a lead relative to customs data, improving short-term forecast information content.

*Source: Conclusions (and accompanying tables/notes) from wpiea2021293-print-pdf - Conclusions*

### Appendix II. A Dynamic Factor Model of World

### Appendix II. A Dynamic Factor Model of World Trade

### Overview of the Dynamic Factor Model (DFM)
- DFMs estimate a limited number of common factors from a large dataset to capture the major forces shaping dynamics of macroeconomic variables (e.g., real GDP, inflation, international trade).
- Advantages of DFMs versus small-sample VARs:
  - Reduce risk of missing important information present in many series (example: the “price puzzle” in monetary policy).
  - Avoid loss of degrees of freedom and overfitting that bias parameter and impulse response estimates in large VARs.
- Conceptual common shocks summarized by factors can be interpreted as: economic activity, financial conditions, inflationary pressures, global trade momentum, etc.

### Model structure (equations and estimation approach)
- Observation equation (as given):
  - 푋푡′=(푋1,푡, ...푋N,푡) represented as 푋푡′=Λf S푡′+ξ푡 (Equation (1)), where Λf is N×P factor loadings and ξ푡 is N×1 white-noise disturbances.
- Factor dynamics (as given):
  - 푆푡′=A1(푆푡−1′)+⋯+Al(푆푡−l′)+ϑ푡 (Equation (2)), where A1...Al are coefficient matrices and ϑ푡 is mean zero common shocks with diagonal covariance matrix Θ.
- Estimation:
  - Two-step principal component procedure:
    1. Estimate factors using the first p principal components of 푋푡′ via Equation (1).
    2. Use estimated components (Ŝ푡′) to estimate Equation (2) with standard VAR techniques.
  - References to estimation approaches: Banbura and Modugno (2014), Doz et al (2005), Stock and Watson (2002a).

### Dataset used for CPB World Trade DFM
- Dataset design criteria:
  - Timely, high-frequency (monthly or higher), strongly correlated with global trade in value (CPB).
- Final dataset:
  - 28 global macroeconomic variables (monthly or higher frequency), covering global demand, production, transport availability, sentiment, financial indicators, prices, and hard indicators (industrial production, employment, unemployment rate, advanced economies’ retail sales, global car production).
- Data transformations and preprocessing:
  - Seasonal adjustment (X12) applied to all variables excluding CPB world trade.
  - Variables transformed in logarithmic differences.
  - Balanced panel constructed so all series start and end at the same exact date for estimation comparability.
  - Balanced panel span used: October 2010 to October 2020.
  - “Jagged edge” of the panel is used during forecasting to utilize real-time information.

### Variables and correlations with CPB World Trade (selected entries from Table II.1)
- CPB World Trade in Value (CPB Netherlands Bureau for Economic Policy Analysis; 2010M10–2020M10): correlation 1.0
- Industrial Production Volume Index (Haver Analytics; 2010M10–2020M10): correlation 0.76
- Employment Index (Haver Analytics; 2010M10–2020M09): correlation 0.73
- Car Production Volume (Haver Analytics; 2010M10–2020M11): correlation 0.69
- Primary Commodity Prices (IMF Commodity Data Portal; 2010M10–2020M12): correlation 0.65
- Available Passenger Capacity (International Air Transportation Association; 2010M10–2020M12): correlation 0.63
- Advanced Economies’ Retail Sales in Value (Haver Analytics; 2010M10–2020M11): correlation 0.60
- Producer Price Index (Haver Analytics; 2010M10–2020M11): correlation 0.52
- Consumer Price Index (Haver Analytics; 2010M10–2020M11): correlation 0.50
- PMI: Manufacturing New Export Orders (IHS Markit; 2010M10–2020M12): correlation 0.46
- TIGER Confidence Index (Brookings Institute; 2010M10–2020M08): correlation 0.39
- Available Cargo Capacity (IATA; 2010M10–2020M12): correlation 0.38
- PMI: Manufacturing Output Prices (IHS Markit; 2010M10–2020M12): correlation 0.36
- PMI: Manufacturing Employment (IHS Markit; 2010M10–2020M12): correlation 0.34
- PMI: Services Employment (IHS Markit; 2010M10–2020M12): correlation 0.32
- PMI: Manufacturing Backlogs of Work (IHS Markit; 2010M10–2020M12): correlation 0.30
- PMI: Manufacturing Stocks of Purchases (IHS Markit; 2010M10–2020M12): correlation 0.29
- Container Throughput Index (Institute of Shipping Economics and Logistics; 2010M10–2020M12): correlation 0.29
- Largest Containerized Shipping Companies’ Price (Bloomberg; 2010M10–2020M12): correlation 0.25
- Policy Related Interest Rate (Haver Analytics; 2010M10–2020M12): correlation 0.24
- Sentix Overall Economic Index (Sentix; 2010M10–2020M12): correlation 0.24
- Morgan Stanley’s Stock Price Index (Bloomberg; 2010M10–2020M12): correlation 0.16
- Baltic Exchange Dry Index (Baltic Exchange/ Haver Analytics; 2010M10–2020M12): correlation 0.15
- U.S. BB corporate bond spread over government securities (Bloomberg; 2010M10–2020M12): correlation -0.31
- Broad Money Index (Haver Analytics; 2010M10–2020M12): correlation -0.36
- J.P. Morgan Sovereign Emerging Market Bond Index (Bloomberg; 2010M10–2020M12): correlation -0.40
- J.P. Morgan Corporate Emerging Market Bond Index (Bloomberg; 2010M10–2020M12): correlation -0.44
- Unemployment Rate (Haver Analytics; 2010M10–2020M10): correlation -0.63

### Estimation choices and forecasting procedure
- Hyperparameters chosen (based on out-of-sample forecasting performance analysis):
  - Number of factors: 6
  - Lag length of underlying VAR: 4
  - Justification: pair of hyperparameters minimizing RMSE over forecast horizon.
- Forecasting setup:
  - Produce 8 periods ahead out-of-sample forecasts.
  - Forecast evaluation composed of 13 cycles starting with dataset available at end-December 2019 and ending with dataset available at end-December 2020.
  - Balanced panel is used in each forecasting cycle; last 2 months of original dataset are dropped each time due to two-month data release lag for many indicators including world trade (example: balanced panel ends in October 2019 when using dataset available as of December 2019).
  - Conditional forecasting approach used to produce 2 months of newscast (nowcasts) and 6 months ahead forecasts, exploiting the “Jagged edge” to use all available real-time information.
  - Quasi-real time historical database constructed using data release calendar available as of December 2020 for each forecasting cycle.
- Model diagnostics:
  - The estimated underlying VAR meets all stability criteria: all eigenvalues of the coefficient matrix lie within the unit circle.

### Representation of CPB trade growth and empirical results
- Decomposition (as presented):
  - dlog(CPB_B_TRADE_t) = 0.8*f_t1 + 0.3*f_t2 -0.2*f_t3 + 0.3*f_t4 + 0.1*f_t5 + 0.1*f_t6 + ϑ_t  (Equation (3))
  - Interpretation: weighted sum of estimated common factors denotes common component (world trade growth momentum); ϑ_t denotes idiosyncratic shocks.
- Key empirical findings:
  - The common component commoves strongly with headline world trade series over the estimation sample.
  - Multivariate regression analysis result: the common component explains 86 percent of global trade.
  - Nowcast and forecast up to June 2021 confirm good out-of-sample fit of the forecast with actual data.
- Forecast specifics:
  - First 2 projection periods are nowcasts due to missing world trade data in real time.
  - Forecast horizon produced: up to 8 periods ahead (2 nowcast + 6 forecast).

*Source: Appendix II. A Dynamic Factor Model of World Trade (authors’ calculations; CPB and authors’ calculations and forecasts).*

### Appendix III.  Selected Linear Regression

### Appendix III. Selected Linear Regression Results

### Methodology and Diagnostics
- Data sources: Haver, national customs data, SWIFT, and authors’ regressions.
- Standard errors (in parenthesis) are heteroskedasticity and autocorrelation robust (HAC) using 1lag and without small sample correction.
- Asterisks indicate significance at the 10 percent (*), 5 percent (**), and 1 percent (***) levels.
- Typical diagnostics reported per regression: Observations, R2, Adjusted R2, F-Statistic, Prob. (F-Statistic), Log-likelihood, Durbin-Watson.

### Selected country regression highlights (coefficients and diagnostics reported exactly as in the source)

- Table III.1. Regression Results for Bangladesh Imports
  - Constant: 0.0119 (0.006)**
  - Imports (-1): -0.7829 (0.093)***
  - Imports (-2): -0.4466 (0.101)***
  - Imports (-3): -0.4094 (0.096)***
  - Imports (-4): -0.2007 (0.101)**
  - SWIFT4: 0.0319 (0.018)*
  - SWIFT4 (-1): 0.0375 (0.018)**
  - SWIFT4 (-2): 0.0348 (0.016)**
  - SWIFT4 (-3): 0.0442 (0.019)**
  - SWIFT4 (-4): 0.0373 (0.014)**
  - SWIFT7: 0.168 (0.052)***
  - BRENT: 0.2376 (0.054)***
  - BRENT (-1): 0.4172 (0.094)***
  - Observations: 123
  - R2: 0.65
  - Adjusted R2: 0.59
  - F-Statistic: 17.35
  - Prob. (F-Statistic): 5.13e-24
  - Log-likelihood: 172.6
  - Durbin-Watson: 1.922

- Table III.2. Regression Results for China Exports
  - Constant: 0.016 (0.006)**
  - Exports (-1): -0.6201 (0.084)***
  - Exports (-2): -0.5953 (0.157)***
  - Exports (-3): -0.1803 (0.096)*
  - Exports (-4): -0.1071 (0.085)
  - SWIFT4: 0.3183 (0.113)***
  - SWIFT4 (-1): 0.1784 (0.104)*
  - SWIFT4 (-2): 0.3683 (0.099)***
  - SWIFT4 (-3): 0.0795 (0.1)
  - SWIFT4 (-4): 0.2253 (0.097)**
  - SWIFT7: 0.0017 (0.112)
  - SWIFT7 (-3): 0.2117 (0.106)**
  - PMI: 0.0162 (0.01)
  - PMI (-1): -0.0101 (0.006)*
  - BRENT: -0.037 (0.08)
  - BRENT (-1): -0.2203 (0.132)*
  - BRENT (-4): -0.1109 (0.056)**
  - Observations: 124
  - R2: 0.57
  - Adjusted R2: 0.46
  - F-Statistic: 5.115
  - Prob. (F-Statistic): 3.14e-9
  - Log-likelihood: 154.64
  - Durbin-Watson: 2.085

- Table III.3. Regression Results for China Imports
  - Constant: 0.0117 (0.004)***
  - Imports (-1): -0.63 (0.084)***
  - Imports (-2): -0.446 (0.099)***
  - Imports (-3): -0.0861 (0.103)
  - Imports (-4): -0.0849 (0.077)
  - SWIFT4: -0.0533 (0.034)
  - SWIFT4 (-1): -0.1141 (0.056)**
  - SWIFT4 (-2): -0.0166 (0.041)
  - SWIFT4 (-3): 0.094 (0.063)
  - SWIFT7: 0.1351 (0.056)**
  - SWIFT7 (-3): 0.2485 (0.076)***
  - PMI: 0.0058 (0.002)***
  - PMI (-2): -0.0066 (0.002)***
  - BRENT coefficients (lags and levels) reported but many not statistically significant at the 1 percent level in this table.
  - Observations: 124
  - R2: 0.55
  - Adjusted R2: 0.44
  - F-Statistic: 11.11
  - Prob. (F-Statistic): 1.05e-18
  - Log-likelihood: 217.55
  - Durbin-Watson: 2.035

- Table III.4. Regression Results for Egypt Imports
  - Constant: 0.018 (0.009)*
  - Imports (-1): -0.5434 (0.089)***
  - Imports (-2): -0.0062 (0.091)
  - Imports (-3): -0.0816 (0.095)
  - Imports (-4): -0.1416 (0.086)
  - SWIFT4: 0.1979 (0.054)***
  - SWIFT4 (-1): 0.1566 (0.054)***
  - SWIFT7: 0.0491 (0.035)
  - PMI coefficients small and not strongly significant
  - BRENT series not significant at conventional levels in this table
  - Observations: 119
  - R2: 0.51
  - Adjusted R2: 0.38
  - F-Statistic: 7.344
  - Prob. (F-Statistic): 6.53e-13
  - Log-likelihood: 143.64
  - Durbin-Watson: 2.039

- Table III.6. Regression Results for France Exports
  - Constant: 0.0034 (0.003)
  - Exports (-1): -0.4354 (0.094)***
  - Exports (-2): -0.3051 (0.09)***
  - SWIFT4: 0.0511 (0.023)**
  - SWIFT4 (-1): 0.1212 (0.033)***
  - SWIFT4 (-2): 0.1169 (0.037)***
  - SWIFT4 (-3): 0.1362 (0.033)***
  - SWIFT7: 0.0681 (0.035)*
  - PMI: 0.007 (0.002)***
  - PMI (-2): -0.0036 (0.002)**
  - BRENT: 0.145 (0.034)***
  - Observations: 123
  - R2: 0.70
  - Adjusted R2: 0.62
  - F-Statistic: 6.182
  - Prob. (F-Statistic): 3.64e-11
  - Log-likelihood: 251.85
  - Durbin-Watson: 2.025

- Table III.12. Regression Results for Korea Imports
  - Constant: 0.0082 (0.003)***
  - Imports (-1): -0.7003 (0.077)***
  - Imports (-2): -0.5061 (0.103)***
  - Imports (-4): -0.1465 (0.086)*
  - SWIFT4: 0.048 (0.022)**
  - SWIFT4 (-2): 0.0799 (0.028)***
  - SWIFT4 (-3): 0.058 (0.027)**
  - SWIFT7: 0.1861 (0.057)***
  - SWIFT7 (-1): 0.2492 (0.057)***
  - PMI (-3): 0.0042 (0.002)**
  - BRENT (-1): 0.1283 (0.029)***
  - BRENT (-2): 0.1227 (0.026)***
  - BRENT (-3): 0.101 (0.033)***
  - Observations: 124
  - R2: 0.63
  - Adjusted R2: 0.54
  - F-Statistic: 26.44
  - Prob. (F-Statistic): 7.06e-33
  - Log-likelihood: 265.35
  - Durbin-Watson: 2.108

- Table III.16. Regression Results for Pakistan Imports
  - Constant: 0.0006 (0.007)
  - Imports (-1): -0.8052 (0.107)***
  - Imports (-2): -0.596 (0.157)***
  - Imports (-3): -0.3676 (0.138)***
  - SWIFT4: -0.165 (0.134)
  - SWIFT4 (-1): -0.2764 (0.147)*
  - SWIFT4 (-2): -0.2443 (0.139)*
  - SWIFT7: 0.3022 (0.087)***
  - SWIFT7 (-1): 0.5523 (0.143)***
  - SWIFT7 (-2): 0.4842 (0.127)***
  - SWIFT7 (-3): 0.3456 (0.114)***
  - SWIFT7 (-4): 0.2519 (0.094)***
  - BRENT (-3): -0.203 (0.094)**
  - Observations: 124
  - R2: 0.56
  - Adjusted R2: 0.48
  - F-Statistic: 10.37
  - Prob. (F-Statistic): 1.97e-16
  - Log-likelihood: 131.43
  - Durbin-Watson: 2.04

- Table III.27. Regression Results for Vietnam Imports
  - Constant: 0.0241 (0.006)***
  - Imports (-1): -0.8751 (0.099)***
  - Imports (-2): -0.5682 (0.115)***
  - Imports (-3): -0.3755 (0.089)***
  - Imports (-4): -0.1961 (0.065)***
  - SWIFT4: 0.1191 (0.051)**
  - SWIFT4 (-3): 0.1415 (0.049)***
  - SWIFT7: 0.3617 (0.081)***
  - SWIFT7 (-1): 0.1847 (0.08)**
  - SWIFT7 (-4): 0.1726 (0.059)***
  - BRENT: -0.1633 (0.053)***
  - BRENT (-2): 0.1378 (0.065)**
  - Observations: 121
  - R2: 0.70
  - Adjusted R2: 0.63
  - F-Statistic: 13.42
  - Prob. (F-Statistic): 3.65e-21
  - Log-likelihood: 183.14
  - Durbin-Watson: 2.061

### Cross-cutting patterns in the reported regressions
- Lagged trade activity (Imports (-1) or Exports (-1)) is frequently negative and highly significant across countries (examples: Bangladesh -0.7829 (0.093)***; China Exports -0.6201 (0.084)***; Korea Imports -0.7003 (0.077)***; Vietnam Imports -0.8751 (0.099)***), indicating strong autoregressive dynamics in the monthly trade series estimated.
- SWIFT indicators (SWIFT4 and SWIFT7) often enter with positive and sometimes significant coefficients (examples: Bangladesh SWIFT7 0.168 (0.052)***; China Exports SWIFT4 0.3183 (0.113)***; Pakistan SWIFT7 series with multiple significant positive lags).
- BRENT (oil price) coefficients vary by country and lag; several regressions show significant positive associations (examples: Bangladesh BRENT 0.2376 (0.054)***; France BRENT 0.145 (0.034)***; Peru Exports BRENT 0.25 (0.062)***), while others show negative significant coefficients (example: Vietnam BRENT -0.1633 (0.053)***).
- PMI effects are generally small in magnitude; in some country regressions PMI or its lags are significant (examples: France PMI 0.007 (0.002)***; India PMI 0.0071 (0.002)***; Korea PMI (-3) 0.0042 (0.002)**).

*Appendix III. Selected Linear Regression Results, "Another Piece of the Puzzle: Adding SWIFT Data on Documentary Collections", Working Paper No. WP/21/293*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021293-print-pdf.pdf_