Bill of Lading Data in International Trade Research with an Application to the COVID-19 Pandemic

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Bill of Lading Data in International Trade Research with an Application to the COVID-19 Pandemic Aaron Flaaen, Flora Haberkorn, Logan Lewis, Anderson Monken, Justin Pierce, Rosemary Rhodes, and Madeleine Yi 2021-066 Please cite this paper as: Flaaen, Aaron, Flora Haberkorn, Logan Lewis, Anderson Monken, Justin Pierce, Rosemary Rhodes, and Madeleine Yi (2021). “Bill of Lading Data in International Trade Research with an Application to the COVID-19 Pandemic,” Finance and Economics Discussion Series 2021-066. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2021.066. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Bill of Lading Data in International Trade Research with an Application to the COVID-19 Pandemic* † Aaron Flaaen Flora Haberkorn Logan Lewis Anderson Monken Justin Pierce Rosemary Rhodes Madeleine Yi Federal Reserve Board September 2021 Abstract We evaluate high-frequency bill of lading data for its suitability in international trade research. These data offer many advantages over both other publicly accessible official trade data and confidential datasets, but they also have clear drawbacks. We provide a comprehensive overview for potential researchers to understand these strengths and weaknesses as these data become more widely available. Drawing on the strengths of the data, we analyze three aspects of trade during the COVID- 19 pandemic. First, we show how the high-frequency data capture features of the within-month collapse of trade between the United States and India that are not observable in official monthly data. Second, we demonstrate how U.S. buyers shifted their purchases across suppliers over time during the recovery. And third, we show howthedatacanbeusedtomeasurevesseldeliverybottlenecksinnearrealtime. JELclassifications: F14,F17,C81 Keywords: Billoflading,COVID-19trade,Firm-leveltrade *Corresponding author: aaron.b.flaaen@frb.gov. The views expressed here should not be interpreted as reflectingtheviewstheFederalReserveBoardofGovernorsoranyotherpersonassociatedwiththeFederal ReserveSystem. †LinktoGitHubrepository: https://github.com/maddieky/panjiva-code. 1

1 Introduction Researchers, policymakers, and firms increasingly turn to nontraditional, administrative, or other so-called “big data” to measure economic activity. These data are often available more quickly and offer a finer level of disaggregation than official statistics, but they can alsoposenewchallenges. Withoutaproperunderstandingoffeaturessuchasconceptual definitions, representativeness, and reporting details, such data can result in improper inference,biasedforecasts,ornon-replicableresults. Thispaperprovidesthefirstdetailed analysis of the utility of a major source of nontraditional administrative data related to international trade: the shipment-level bill of lading (BoL) data collected by U.S. ports. BoLdataprovideanumberofadvantagesanddisadvantagesrelativetootherpublicly accessible official data and confidential datasets. In this paper, we use S&P Panjiva as our source of BoL data, as they provide both the raw data and also a number of useful derivative variables, including identifiers that allow researchers to longitudinally track firms engaged in international trade. In Section 2, we describe the data in detail, while in Section 3, we explore the advantages and disadvantages of the data. Advantages of the data include detail, timeliness, and the data’s unrestricted nature. Data are available at the shipment level, often with company names for both the shipper (exporter or freight forwarder) and consignee (the importer, person, or firm taking final delivery of the merchandise). They are also available to researchers within weeks, rather than months or years in the case of some detailed confidential data. The ability to access the data outside of restricted environments allows easier merging with other datasets as well as diving into specific case studies that can help illuminate how these shipments work in practice for both researcher and reader. A list of the top 10 U.S. consignees and shippers in these data is simple and illustrative to show (see tables 4 and 5, below), but doing the same with public U.S. data would be impossible, and with confidential U.S. data, explicitly prohibited. As with all datasets, BoL data have disadvantages as well. U.S. law restricts public access to bill of lading records to only those shipped via vessel; some countries have broader access, but in this paper, we will largely focus on the strengths and weaknesses ofU.S. data. Inaddition, shipmentvaluesare missingfromBoLdata. Quantitymeasures and descriptions are included, but mapping these to commonly used product classificationsandestimatingvaluescouldintroducemeasurementerror. Companiesalsohavethe rightinU.S.lawtoredacttheirnamefromtherecords, whichcanhampereffortstotrack supply chains comprehensively. One of the most novel aspects of these data is information on the shippers (exporters) and consignees (importers) for each shipment. This is unique among publicly available 1

datasets, especially for the United States, where access to and disclosures from the U.S. Census Bureau’s confidential Longitudinal Firm Trade Transactions Database (LFTTD) are highly restricted. InSection4,wediveintothecharacteristicsoftheseshippersandconsigneestobetter understand global supply chains. We show how over 60 percent of U.S. consignees have only one foreign shipper, but that these consignees represent less than 20 percent of import volumes. We also find that most shipper-consignee pairs ship in only three or fewer months in a year, with a surprisingly small number of pairs shipping every month, thoughthese monthlyshippersrepresentover 50percentof trade. Wealso showhowthe number of shippers per consignee dropped dramatically in 2020 and remained volatile before recovering in mid-2021. Finally,weturntowaysinwhichthesedataarewell-positionedtoanalyzethestriking effects of the COVID-19 pandemic on U.S. trade. The daily frequency of the data show how exports from India to the United States fell within just a few weeks of the start of the pandemic, and given shipping lags, how that collapse took 5-10 weeks to show up in U.S. import data.1 Wecanalsoanalyzethemarginsonwhichimportscollapsedandsubsequentlyrecovered: theintensivemarginofchangesforagivenconsignee-shipperpair,thenetextensive marginofentryandexitofconsignees,andtheswitchingbyagivenconsigneetoadifferent shipper, a different country, or both. We begin to analyze these margins by focusing on an industry with particularly interesting trade patterns during this period: furniture. After initially plunging in the first half of 2020, demand for durable goods, such as furniture, skyrocketed, and furniture’s weight and size tends to preclude shipping by air, making it an ideal case to analyze with BoL data. We find that, during the initial collapse in trade, the extensive margin accounted for much of the plunge in trade volumes. Then, during the extraordinary rebound in U.S. imports, the intensive margin was most important for the first few quarters, with the extensive margin and switching margin slowly growing by the third quarter of the recovery. Wefindthattheintensivemarginissimilarlyimportantinthefirstfewquarters of the recovery in total U.S. imports. These results provide important lessons on the limitations of supply chain flexibility in the very short run and the time required to source products from new shippers, or for new importers to enter the market. Thefieldofinternationaltradeisparticularlywell-situatedtobenefitfromnewsources of nontraditional data, such as the port data we examine. Since the pioneering work of Bernard et al. (1995), trade economists have focused on firm-level participation in inter- 1We examine exports from India to the U.S. because India instituted a particularly stringent lockdown at the onset of the COVID-19 pandemic, and because China—another natural candidate country—stopped makingitsBoLdataavailableastensionswiththeU.S.rosein2018and2019. 2

national trade. Subsequent work by Monarch (2021) and Heise et al. (2019) has exploited information on the timing and frequency of trade transactions. This research has been conducted almost entirely using the confidential data available from the U.S. Census Bureau, which comes with strict access and disclosure restrictions. The administrative data collectedbyportsofferanalternativedatasourceforfirm-andtransaction-levelresearch, without these restrictions. Though little-used in international economics, this paper is not the first to make use of BoL data. The recent availability of this processed BoL data is enabling researchers to conduct more detailed studies of global trade flows, supply chains, and firm operations. Ganapati et al. (2021) use an extract of BoL data (also from Panjiva) and pair it with with vessel location data derived from transponders used for navigational safety purposes. They use these combined data to present new stylized facts on the shipping network of global trade flows, with corresponding implications for trade costs. Bonfiglioli et al. (2020) uses BoL data from Piers to show that richer countries have higher average sales per firm from two sources of heterogeneity. In related work, Bonfiglioli et al. (2021a) shows that market concentration in international trade has fallen overall.2 In addition, Feenstra and Weinstein (2017) use BoL data to estimate the concentration of exporters to the United States from markets outside of Canada and Mexico. Inadditiontothetradeliterature,BoLdatahavebeencombinedwithfinancialdatasets to yield new insights on the behavior of firms that operate internationally. Jain et al. (2014) construct a novel dataset by combining BoL data with publicly available countryyear-leveldataonbusinessregulationsandfirm-quarter-levelaccountingdatatoevaluate the participation of different firms and sectors in global trade. Jain and Wu (2020) use BoLdatatoexaminethesourcingofdifferentcategoriesofimportedgoodsbyfirmswith global supply chains, exploring the relationship between firms’ global sourcing strategy and future profitability. Bruno and Shin (2020) match BoL data from Mexico to financial data from Capital IQ (both available from S&P). They show that when the U.S. dollar appreciates, dollar wholesale-funded banks pare back credit to Mexican exporters, hampering their exports.3 2Bonfiglioli et al. (2021b) review the literature on heterogeneous firms in trade with additional results derivedfromBoLdata. 3They supplement the Panjiva data with estimates from PIERS to fill out the dollar value of imported goods,asthisvariableislargelymissingviaPanjiva. Thesepapersnotetheparticularchallengesthatcomes withworkingwithBoL,specificallywidespreadspellinginconsistencies,aswellasthevarioususeoftrade namesandsubsidiaries. 3

2 Data Description BillofladingdatafromS&PPanjiva-—thedataproviderweuseforthisanalysis-—contain over one billion transaction-level records of goods traded across borders, with information including consignees and shippers, product descriptions, quantity, and, in limited cases, estimated values of shipment transactions (in USD). The data provide trade flows across 17-country-level datasets, including Bolivia, Brazil, Chile, China, Colombia, Costa Rica, Ecuador, India, Mexico, Panama, Pakistan, Paraguay, Peru, Sri Lanka, Uruguay, the United States, and Venezuela. For each of these countries, data users are able to observe both imports and exports of goods for all trading partners. We focus our analysis specifically on U.S. import data and, to a lesser extent, U.S. exportdata. Panjivaprovidestransactionssince2007forimportsandsince2009forexports. U.S. import data are updated several times per week, but U.S. export data updates are typically delayed by a 23-day lag for regulatory reasons (Panjiva). PanjivaacquiresthesedatabycollectingbillsofladingfromU.S.CustomsandBorder Protection (CBP), which are freely available under the Freedom of Information Act of 1966 (FOIA). A BoL is a legal document that serves as a record that a shipment has been transported from its origin to its final destination. It also details the contract between the shipper and consignee. Each BoL requires companies to fill out various fields, including shipper/consignee name and address, description of the goods, vessel name, transport companyname,portsoflading(loading)andunlading(unloading),weight,quantity,and containerinformation. (SeeAppendixFigures18and19fortheCBPinward(import)and outward (export) cargo declaration forms.) In addition to providing the raw information collected on Bills of Lading, Panjiva generates additional variables that may be of use to researchers. First, Panjiva imputes a standard measure of volume, twenty-foot equivalent units (TEUs), based on existing container information and other shipment characteristics. Second, while BoL forms require product descriptions, they do not collect dataon Harmonized System (HS) product codes. Panjiva attempts to assign HS codes to each shipment by searching product descriptionsforHScodesthatmayhavebeenoptionallyincludedbyshippersandbyusing a text processing algorithm to translate descriptions to HS codes. Third, Panjiva attempts to provide an estimate of the value of a transaction since this information is not required inaBoL.Asdiscussedbelow,thesevalues,whicharebasedonpublicly-availableaverage unit values, are only estimates and they are also currently unavailable for most transactions. Fourth, Panjiva also includes a unique company ID variable that can be used to linkthetradetransactionsofsomeshippersandconsigneestotheirassociatedcompanies inotherS&PGlobaldatasets,suchasS&PCapitalIQ.Onelimitationisthatthiscompany 4

IDlinkingvariableonlyexistsfor10-15percentofshippersandconsigneesinU.S.import dataatthistime,sonotalltransactionscanbelinkedtoS&P’sbroaderecosystemofdata. Table1listsvariablenamesanddescriptionsforsomeofthekeyvariablescontainedin thePanjivaBoLdata,withthetoppanelreportingrawBoLdatavariablesandthebottom panelreportingvariablesthatareimputedbyPanjiva. Next,wecompareaggregatedBoL data against official U.S. Census data. Table1: U.S.importdatadescriptionforselectvariables Rawvariable Description arrivaldate Arrivaldateofshipment shpname EntityResolvednameoftheshipper conname Thepartytotakefinaldeliveryofthemerchandise shpmtorigin LocationfromwhichshipmentleftfortheU.S portoflading Portoflading portofunlading Portofunlading weightkg Shipmentweightinkilograms vessel Nameofthevesselthattransportedthegoods Imputedvariable Description panjivarecordid UniquePanjivaIDforshipmentrecord shppanjivaid UniquePanjivaIDforpartyactingasshipper conpanjivaid UniquePanjivaIDforpartyactingasconsignee volumeteu VolumeofshipmentinTEU valueofgoodsUSD ValueofgoodsinUSD hscode HarmonizedItemDescriptionandCodingSystem(HS) companyid CapitalIQcompanyID The massive size of these datasets combined with continuous updating makes data management a particular challenge. In Appendix B, we describe some key features of thesystemwe’vecreatedattheFederalReserveBoardofGovernorstoupdate, store, and query the complete raw BoL data files. How Bill of Lading Data Compare to Census Data Here, we evaluate how well BoL aggregates align with official public trade aggregates. Figure 1 shows two measures of trade volume from BoL data: containers, measured by twenty-foot equivalent units, or TEUs (in blue) and shipments (in red), both normalized so that 2009 = 100. A shipment is the cargo, regardless of size, recorded in a single bill of lading.4 That TEUs and shipments track one another closely implicitly highlights the stabilityintheaveragenumberofTEUspershipment. Inordertoexcludetransshipments 4AsmallshareofshipmentsinthePanjivadatabaseshareaBoLnumberwithatleastoneothershipment. Thesemayrepresentduplicateobservations,thoughinatleastsomecasesthearrivaldateisdifferentwhile otherfieldsarethesame. 5

Figure1: ComparisonofbillofladingdataandCensuscontainerizedvesselvalueforU.S.imports Index, 2009 = 100 240 220 TEUs (Panjiva) Shipments (Panjiva) 200 Containerized Vessel Value (Census) 180 160 140 120 100 2009 2011 2013 2015 2017 2019 2021 Source: S&PGlobalMarketIntelligence,U.S.Census,andauthors’calculations. Notes: Seasonallyadjusted. that ultimately end up in a different country, we limit this analysis (as well as all our further analysis of U.S. import BoL data) to shipments where the consignee country is either listed as the United States or is missing. In addition, while BoL data contain noncontainerized vessel trade, in particular oil imports, these do not have corresponding TEU values and represent relatively few shipments of large value. Therefore, the most relevant publicly available measure of trade flows to compare to our BoL measures is the containerized vessel import value available from the Census Bureau. Importantly, Figure 1 indicates that this nominal measure aligns quite closely with the BoL volume measures over time.5 Figure 2 shows the comparison between BoL aggregates and total U.S. goods import value. Here, we see that BoL data still capture the broad pattern of trade growth, as well as the dramatic trade collapse and recovery during 2020. Relative to Figure 1, the Census Total Import Value in Figure 2 includes non-maritime trade as well as vessel trade not via containers, notably including oil imports, which leads to some modest differences with the Panjiva trade measures. Oil prices were elevated in 2011-2014, for example, contributing to the Census total value line being above the lines from BoL data. As we discuss in Section 3.2, the limitation to only maritime trade with U.S. BoL data should 5Importpriceinflationinthesegoodsisnearzero: fornonpetroleumgoods,BEAnationalaccountsdata showannualizedimportpriceinflationof−0.35%overthe11-yearspanof2010Q1to2021Q1. 6

Figure2: ComparisonofbillofladingdataandCensustotalvalueforU.S.goodsimports Index, 2009 = 100 240 220 TEUs (Panjiva) Shipments (Panjiva) 200 Total Import Value (Census) 180 160 140 120 100 2009 2011 2013 2015 2017 2019 2021 Source: S&PGlobalMarketIntelligence,U.S.Census,andauthors’calculations. Notes: Seasonallyadjusted. be carefully considered in the context of each research question. For example, as we examinethe2020tradecollapseandrecovery,wefocusoncategorieslikefurniturerather than medical equipment or semiconductors, as the latter two categories are more likely to use air shipping. 3 Advantages and Limitations of Bill of Lading Data 3.1 Advantages of Bill of Lading Data Bill of lading data have a variety of advantages relative to official trade statistics, making them a valuable resource for both researchers and policymakers. The first benefit of these data stems from the fact that shipments are associated with specific firms on both the shipper (exporter) and consignee (importer) side of the transaction. The combination of these data allows consideration of firm characteristics such as the frequency of shipments per consignee, which provides important information about the nature of firms’ procurement systems (Heise et al. 2019). We discuss interesting stylized facts based on exploiting the shipper and consignee identifiers in Section 4. A second benefit of the data is their high frequency. Official trade data are available at a monthly frequency, but BoL data track shipments arriving or departing the U.S. at a daily frequency. This higher frequency is important in many contexts, with one 7

prominent example being an analysis of the timing of the collapse in trade associated with COVID-19. Our examination in Section 5.1 of U.S. imports from India during the initial days of the global pandemic lockdown reveals intra-month shifts in trade that are simply not observable at the monthly level. AthirdbenefitoftheBoLdataisthetimelinessoftheirreleaserelativetoofficialdata. Official monthly U.S. trade data typically lag the close of a month by more than 30 days. By contrast, and as summarized in Appendix Figure 15, BoL data are updated nearly continuously; data for a particular day are reasonably complete within 10-14 days. This timeliness allows for observation of supply chain disruptions, such as those arising from COVID-19 or the blockage of Suez Canal in essentially real-time. A final and intriguing benefit of BoL data is the potential of combining transactionlevel data from multiple trading partners. Combining data in this manner opens the possibility of linking shippers and consignees across multiple countries’ trade data, allowing for a level of detail on firms’ global supply chains that is not available elsewhere, even in confidential transaction-level trade data from the U.S. Census Bureau. Linking multiple countries’ data also holds the potential of observing trade networks (see e.g. BernardandMoxnes(2018)andDhyneetal.(2021))andthepropagationofsupplychain shocks across firms and borders (Boehm et al. 2019). 3.2 Limitations of Bill of Lading Data While the advantages for these data relative to publicly available sources can be substantial, there are also some limitations about which researchers should be aware. These limitations include missing or redacted data, as well as a general lack of non-imputed data on transaction values. 8

Limitation to Maritime Trade (United States) Figure3: U.S.tradesharesbymodeoftransport,2019 Share of total, percent Vessel Air Exports Other transport Imports 0 10 20 30 40 Source: U.S.Censusandauthors’calculations. Notes: Otherincludesrail,vehicle,pipeline,etc. Figure4: U.S.importsbymodeoftransport Billions of US dollars, monthly 100 80 60 40 Vessel Air Other 20 2009 2012 2015 2018 2021 Source: U.S.Censusandauthors’calculations. Notes: Otherincludesrail,vehicle,pipeline,etc. Seasonallyadjusted. 9

Oneofthekeylimitationsofbillofladingdataisitslackofinformationonnon-maritime trade for the United States. As indicated in Figure 3, maritime trade—i.e. trade transported by vessel—is the largest mode of transport by value, accounting for nearly 50 percentofthevalueofU.S.importsandnearly40percentofthevalueofU.S.exportsin2019. Nonetheless,theremainingvalueofU.S.trade,whichissplitbetweenairandland-based transport like trucks, railroads, and pipelines, is not available in U.S. bill of lading data. Moreover, asshownin Figure4, therelevance ofthis exclusionhasalso grownsomewhat over time, with land and air increasing in importance as modes of transportation. Table2: TradeSharesbyModeofTransport,2019 Imports Exports Total Value* Vessel Air Other Vessel Air Other Mexico 9.39 1.98 88.63 12.37 3.50 84.13 608.43 Canada 5.34 4.61 90.05 4.79 6.25 88.96 607.85 China 63.61 28.96 7.43 49.24 43.00 7.76 557.81 Japan 71.80 24.57 3.63 51.79 40.09 8.13 217.77 Germany 51.62 40.39 7.99 33.94 57.75 8.32 187.00 Source: U.S.Censusandauthors’calculations. Notes: Includestop5U.S.tradingpartnersbyvalue. *Inbillions. The exclusion of air and land-based trade also leads to substantial differences in coverage across major U.S. trading partners. As shown in Table 2, trade with Mexico and Canada—two of the largest trading partners of the United States—is conducted almost entirelyvialand-basedmodesoftransportation. BilateralU.S.tradewiththosecountries, therefore,islargelyexcludedfromBoLdata. However,thevesselshareoftradeis,unsurprisingly, much higher for other important U.S. trading partners outside North America. Trade by vessel accounts for 64 percent of the value of U.S. trade with China, 72 percent of the value of trade with Japan, and 52 percent of the value of trade with Germany. Missing Data Most big data sources suffer from missing information in some observations, and BoL data from Panjiva is no exception. There are two primary sources of missing data in Panjiva: fields for which a firm requests that the U.S. Customs and Border Protection (CBP) redact their identity in the shipper or consignee field, and fields like TEU, HS code, and value that Panjiva imputes from other information that is not always available. Generally speaking, fields that are directly filled in on CBP form 1302 (see appendix Figure 18) are available for the vast majority of observations. Table3reportstheshareofU.S.importobservationsforwhichparticularkeyvariables 10

Table3: MissingU.S.ImportDatabyVariable(Percent) Shipper ID Consignee ID HS Code TEU Value 2007 19.9 16.7 4.6 3.6 100.0 2008 22.6 22.6 4.9 3.8 95.2 2009 30.2 27.6 3.9 3.4 68.8 2010 33.8 31.5 5.0 3.1 69.9 2011 34.9 35.1 4.9 3.3 70.4 2012 33.5 31.0 4.4 3.4 69.8 2013 25.2 8.5 3.5 3.1 67.9 2014 23.7 8.0 4.2 3.0 65.7 2015 23.7 8.0 3.7 3.0 63.5 2016 28.4 10.6 3.8 2.7 63.8 2017 32.1 12.9 3.9 2.8 64.6 2018 32.6 14.2 3.7 2.7 68.3 2019 33.5 14.0 6.6 2.6 67.0 2020 31.4 12.3 7.3 1.8 67.5 2021 33.9 16.8 4.8 1.8 67.7 Source: S&PGlobalMarketIntelligenceandauthors’calculations. aremissing. Asshowninthetable,thevariablesfortheshipper/consigneeIDsandvalue have the highest probability of being missing, while the HS code and twenty-foot equivalent unit (TEU) fields are missing in a much lower share of observations.6 A few key variables, such as weight and shipment origin country, are not included in the table since they have nearly zero missing observations in U.S. import data. Importantly, the share of observations with missing data for particular variables can vary fairly substantially over years. For example, across the years 2007 to 2021, the share of observations with missing shipper (consignee) IDs ranges from 19.9 percent (8.0 percent) to 34.9 percent (35.1 percent). Firms’ requests for redactions of shipper and consignee information contribute to variation in the share of missing data over time. After a firm requests redaction, this requestisfulfilledfortwoyearsbeforerequiringrenewal. Whenarequestexpires,afirm’s transactions from that point forward are no longer redacted. These redaction requests must be made for a specific firm name, so firms that use multiple names on bills of lading must submit a request for each entity. Given that one feature Panjiva adds to the raw data is the matching of firm names (including likely typos) to a corporate entity in their overall data framework, this can lead to firms having some but not all of their shipments represented in the database. OneimportantillustrationofthisphenomenonisWalmart,whichappearstoredactits 6MissingTEUvaluescansimplyreflectshipmentsthatarenotcontainerized,suchasoilimports. 11

information imperfectly. As shown in Figure 5, Walmart’s monthly shipments generally hover around zero but spike briefly in 2007, 2012, and sporadically between 2017 and 2021, which suggests that some of Walmart’s redaction requests may have briefly expired before they were subsequently renewed. Figure5: Walmart,Inc.’smonthlyshipments Shipments, monthly 20000 Not fully redacted Not fully redacted 15000 10000 Not fully redacted 5000 Mostly redacted Mostly redacted 0 2009 2012 2015 2018 2021 Source: S&PGlobalMarketIntelligenceandauthors’calculations. Limited data on trade values While BoL data consistently report transaction weights, they typically lack data on the value of trade associated with each transaction. A small share of transactions include data on trade value pulled from the transaction description in the BoL. However, for the majority of other observations the shipment value is either missing or imputed by applying average unit values from public trade data to the BoL weights. As a result, over 60 percent of observations have missing data for shipment value. Product descriptions versus product codes As described above, CBP forms require shippers to report product descriptions, but not HS Product Codes. The HS codes provided in the data, therefore, are not official HS codes, but rather are assigned based on Panjiva’s proprietary algorithm. The assignment is actually quite comprehensive: as as indicated in Table 3, the imputed HS code variable is generally well-populated, with five percent or fewer of observations missing for this 12

variable in all but two years. Nevertheless, it is important to emphasize that BoL records are based on shipments, and therefore an individual record (and hence unit of quantity) couldbecomprisedofmorethanone(andoftenmany)individualproducts. Thisfeature can make disaggregation by product an imperfect exercise. 4 Characteristics of Shippers and Consignees One of the most novel aspects of BoL data is the detailed, shipment-level information on shippersandconsignees. Subjecttothefirm-levelredactionsdescribedabove,researchers can track company-specific details over time, including a company’s trading partners, its frequencyandweightofshipments,itsportsofladingandunlading,andevenitscontact information. In addition, Panjiva assigns unique ID codes to all shippers and consignees after collecting and parsing firm names from bill of lading data, which makes it easier for users to identify and track specific companies as well as merge BoL data with other datasets. Table4: TopconsigneesbytotalTEU,2020 Consignee name Total TEU TEU (%) Shipments (%) Expeditors International 1,150,675 5.23 6.30 Ups Supply Chain Solutions 779,248 3.54 2.79 Dole Fresh Fruit Co. 236,312 1.07 0.52 Samsung Electronics 187,707 0.85 0.55 Chiquita Fresh North America Llc 171,205 0.78 0.10 Maersk Line 170,783 0.78 0.01 Fedex Trade Networks Transport 162,425 0.74 0.88 Seaboard Marine Ltd. 138,075 0.63 0.02 Geodis USA Inc. 123,801 0.56 0.36 Carmichael International Service 113,359 0.52 0.37 Source: S&PGlobalMarketIntelligenceandauthors’calculations. With these data, researchers can analyze certain industries or countries by reporting top suppliers and buyers. Tables 4 and 5, for example, report the top 10 U.S. consignees and foreign shippers, respectively, in U.S. import data. Table 4 reveals that eight of the top 10 consignees are freight and logistics companies, highlighting the importance of intermediaries in the actual execution of international trade. Table 5 shows that the top 10 foreign shippers to the United States are a mixture of these transportation companies, electronics and agricultural producers, and, improbably, Red Bull. As users of confidential Census Bureau data are well aware, revealing this type of information with those datasets is impossible. 13

Table5: TopshippersbytotalTEU,2020 Shipper name Country TEU TEU (%) Shipments (%) Thor Joergensen A S Denmark 170,351 1.04 0.01 Chiquita Brands International SARL Switzerland 153,933 0.94 0.10 Sm Line Corp. United States 75,207 0.46 0.00 Thai Samsung Electronics Co., Ltd. Thailand 60,748 0.37 0.20 Samsung Electronics Co., Ltd. South Korea 56,431 0.34 0.32 Lg Electronics Inc. South Korea 56,181 0.34 0.18 Samsung Electronics Digital Mexico 44,968 0.27 0.13 Red Bull GmbH Austria 43,309 0.26 0.03 Union De Bananeros Ecuatorianos S.A. Ubesa Ecuador 36,198 0.22 0.20 Seadom Units Dom. Republic 35,766 0.22 0.00 Source: S&PGlobalMarketIntelligenceandauthors’calculations. Figure6: ShippersandconsigneesbyTEU,2019 Percent Percent 80 80 Percent of total U.S. consignees Percent of total foreign shippers 60 60 Percent of total TEU Percent of total TEU 40 40 20 20 0 0 1 2−4 5−9 10−24 25+ 1 2−4 5−9 10−24 25+ Number of foreign shippers Number of U.S. consignees Source: S&PGlobalMarketIntelligenceandauthors’calculations. 14

Figure7: Frequencyoftransactionsbyshipper-consigneepair,2019 Percent 50 Percent of total long−term shipper−consignee pairs* 40 Percent of total TEU 30 20 10 0 0 2 4 6 8 10 12 Number of calendar months with at least one transaction Source: S&PGlobalMarketIntelligenceandauthors’calculations. *Includesshipper-consigneepairsthattradedatleastonceinthepreviousyear(2018). Bill of lading data offer further information on firm-level trade that are unobservable in public official data. As shown in the left panel of Figure 6, the majority of U.S. importers have a single foreign trading partner, but these firms account for a disproportionately small share of total U.S. imports by TEU. By contrast, only a small handful of U.S. importers have many trading partners (over 1000 partners, in some cases), but this small number of firms accounts for a disproportionately large share of imports by TEU. Moreover, the number of shippers and total TEU per consignee are positively correlated. These patterns are largely the same when we switch attention to the number of U.S. consignees per foreign shipper (left panel of Figure 6, and taken together, they highlight the significance of large firms in international trade.7 In addition, the majority of shipperconsignee pairs interact infrequently in a given year, which emphasizes the lumpiness of trade by pair. For example, in 2019, only 5% of all long-term shipper-consignee pairs traded at least once each month, while 47% of all pairs only traded in one or two months of the year (Figure 7). 7Forexample,thelargestconsignee,ExpeditorsInternational,accountedfor5.2%oftotalU.S.importsby TEUand6.3%oftotalshipmentsin2019. 15

Figure8: Changeinshippersperconsignee Percent change from same month in 2019 20 0 −20 −40 Jan. 2020 Jul. 2020 Jan. 2021 Jul. 2021 Source: S&PGlobalMarketIntelligenceandauthors’calculations. This shipper-consignee data can also be used to track how disruptions such as recent Covid-related lockdowns affect these relationships. In Figure 8, we plot monthly data on the percent change in the number of shippers per consignee relative to the previous year. As shown in the figure, the number of shippers per U.S. consignee dropped by over 35% in April 2020 relative to April 2019, as importers only managed to maintain a small share oftypicaltradingrelationships. Thispatternofreducedshippersperconsigneecontinued through much of 2020.. 5 Trade and the COVID-19 Pandemic Aswasjustmentioned,thetimelinessandgranularityofBoLdataareespeciallyvaluable in understanding the enormous changes to international trade patterns brought on by the COVID-19 pandemic. This section details several insights from these data about the collapse and resurgence of trade during 2020-2021. 5.1 The precise timing and effects of country-level lockdowns Unlike official statistics, the daily frequency of the BoL transaction-level data allow the observationofintra-monthpatternsoftrade. Thisfeatureisparticularlyusefulinevaluatingtheimpactofshockstotrade,withperhapsthelargestandmostabruptinthemodern era coming from the various country-level lockdowns associated with the early stages of 16

COVID-19. We leverage the multiple sources of information coming from BoL data to highlight the transmission of the trade shock from the March 2020 national lockdown in India to U.S. imports. Figure9: U.S.-IndiashipmentsduringtheCOVID-19lockdowns Index, Mar. 1, 2020 = 100 140 120 U.S. imports from India India exports to the U.S. 100 80 60 40 March 24, 2020 20 0 Mar Apr May Jun Jul Aug Source: S&PGlobalMarketIntelligenceandauthors’calculations. Notes: Thisfigureplotsthe7-daymovingaverageofshipmentsofU.S.importsfrom IndiaandIndiaexportstotheUnitedStates,witheachindexedtoequal100onMarch 1,2020. We focus on the specific case of India because that country instituted a particularly strict COVID-19 lockdown, because pandemic-era U.S.-India trade has been relatively unstudied, and because bill of lading data are available for Indian exports to the U.S.8 As showninFigure9,thenationallockdownannouncedbytheIndiangovernmentonMarch 24, 2020 is evident in the immediate decrease in India’s exports to the United States and thensubsequentlyinthedelayeddropinU.S.importsfromIndiaseveralweekslater. The high-frequency BoL data reveal a much sharper drop in Indian exports to the U.S. than would be visible with monthly-frequency publicly available data. Moreover, the patterns inFigure9revealimportantinformationonthetranslationofthisshockintoU.S.imports: The drop in U.S. imports from India is considerably less steep than the drop in Indian exports and lagged by 4 weeks. More broadly, Figure 9 indicates that BoL data can help researchers learn how the timing of such transmission of trade shocks varies across trading partner based on distance, shipping routes (such as the use of entrepôt trading hubs), and the particular characteristics of the shock. 8BoLdataonexportsfromChinatotheU.S.arenotavailableafterMarch2018. 17

5.2 Decomposing the collapse and subsequent surge in U.S. imports The enormous drop in trade in the first quarter of 2020 was followed by a remarkable recovery, such that U.S. import volumes surpassed typical levels by the middle of 2020. Giventhesurprisingspeedoftheresurgenceintrade,anaturalquestionishowimporters andexportersmanagedtoincreaseshipmentssodramatically. Foroneusefulperspective on both the collapse and subsequent surge in U.S. imports, we decompose the import changes based on the following margins at a quarterly frequency: • Entry/Exit of Consignees Margin: The changes in imports due to the net entry and exit of consignees across two quarters.9 • Add/DropShipperorCountryMargin: Thechangesinimportsacrosstwoquarters from a given consignee that changes either the shipper or the country associated with the import transaction. • IntensiveMargin: Thechangesinimportsfromagivenconsignee—shipper—country pair across two quarters. • Redacted: The changes in imports coming from changes in the pool of redacted consignees across two quarters. Apart from the complicating feature of redactions, the decomposition outlined above is similar in spirit to the work of Bernard et al. (2009), which uses confidential, firmlevel Census data. By contrast, with official public data, researchers are forced to define the extensive margin as something like an HS10 code coming from a particular country. That level of aggregation, however, would not capture the changes in relationships associated with entry/exit of consignees or switching among suppliers by continuing consignees. BoL data allow for the ability to track relationships defined at the consignee×shipper×country level. To focus attention on the dynamics introduced by COVID-19, we fix the baseline period to be the fourth quarter of 2019, and then track the change along each margin in subsequent quarters. We begin with a decomposition of furniture imports (Chapter 94 in the HS classification system) due to the dramatic changes in demand experienced by thisproductgroupduringourperiodofstudy. Inaddition, unlikesomeothercategories, furniture is unlikely to be moved by air. Finally, to account for the significant seasonal patterns and trends in the data, we calculate the identical decomposition for each of the previous three years (baseline quarters of 2016Q4, 2017Q4, and 2018Q4), and for each 9Aconsigneeisconsideredtohaveexitedinaparticularquarterifithasnoimportsduringthatquarter. Aconsigneeisconsideredanentrantinaparticularquarterifithadimportsduringthatquarterbuthadno importsin2019Q4. 18

margin of adjustment and then subtract out the average change across each time horizon from the COVID-19 period. The results are displayed in panel (a) of Figure 10. The black line shows the overall change in U.S. furniture imports, relative to 2019Q4. Importsfellmodestlyinthefirstquarterof2020andthenmoresignificantlyinthesecond quarter.10 The surge in imports for product categories such as furniture is evident in subsequent quarters, with imports up over 65 percent (seasonally adjusted) by 2021Q1 from pre-pandemic levels. We derive several useful lessons from decomposing these overall changes into the marginsofadjustmentoutlinedabove, whichareillustratedbythecoloredbarsinFigure 10a. First, the drop in U.S. imports during the initial lockdowns of COVID-19 in 2020Q1 were driven largely by net consignee exit (the red bars), a feature that continued into the second quarter of 2020. Second, although the intensive margin (light blue) accounts for the largest individual share of the increase at the end of our sample period, when we combine the two extensive margins—i.e. the net consignee exit (in red) and add/drop shipper or country margin (in dark blue)—their contribution is slightly larger than the intensivemargin.11 Hence,by2021-Q1,roughlyhalfofthegrowthinfurnitureimports(a nearly 30 percentage point increase relative to 2019-Q4) came from trading relationships that did not exist in 2019-Q4. Finally, increases in consignee redactions (in gray) are also an important component in the overall increase in imports; without the consignee redaction, we would have been able to allocate these transactions into one of the other margins of adjustment. Panel (b) of Figure 10 provides a contrasting perspective by decomposing the growth in overall BoL imports during this time period. The most obvious difference relative to the decomposition for furniture imports in Figure 10a is the smaller and more gradual increase following the 2020-Q2 nadir: overall imports were up 15 percent in early 2021 relative to the baseline compared with the roughly 60 percent increase for furniture imports. A second and more surprising result comes from the component margins of this increase revealed by the decomposition. The contributions from the two extensive margins—theentry/exitofconsigneesmarginalongwiththeadd/dropshipper/country margin—provide a notably smaller contribution to the overall increase in imports than is the case for furniture imports shown in Figure 10a. The extensive margin also accounted for a larger fraction of the decrease in 2020-Q2. The clear differences in the role of the extensive margin suggests there are likely important variations across products and in- 10Importsfellconsiderablymoreinthefirstquarterof2020onanon-seasonallyadjustedbasis. However, furnitureimportstendtopeakinthefourthquartereachyear,andthenfallsubstantiallyinthefirstquarter. 11AsfirstdiscussedinBernardetal.(2009)theextensivemarginbecomesmoreimportantasthehorizon lengthens. Inourcase,theswitchingproduct/countrymarginofadjustmentispredominantlycomposedof caseswheretheconsigneeswitchessuppliersbutmaintainsthesamesourcecountry. 19

Figure10: Decomposingpercentchangeinimportsof(byTEU)relativeto2019Q4 (a)FurnitureImports(HS94) Relative to Average Change in 2017−2019 Total change Redacted Intensive Margin Add/Drop Shipper or Country Margin 60 Entry/Exit of Consignees Margin 40 20 0 −20 2020−Q1 2020−Q2 2020−Q3 2020−Q4 2021−Q1 2021−Q2 (b)TotalImports Relative to Average Change in 2017−2019 Total change Redacted Intensive Margin Add/Drop Shipper or Country Margin 25 Entry/Exit of Consignees Margin 15 5 −5 −15 2020−Q1 2020−Q2 2020−Q3 2020−Q4 2021−Q1 2021−Q2 Source: S&PGlobalMarketIntelligenceandauthors’calculations. Notes: ThisfigureplotsthequarterlychangeinU.S.imports(byTEU)relativeto2019- Q4alongfourmarginsdescribedinthetext. Thequarterlychangeforeachmarginis netoftheaveragechangeduringtheequivalentquarterduring2017-2019toaccount forseasonalvariationandtrendgrowth. Panel(a)restrictstoimportsoffurniture(HS Chapter94)whereasPanel(b)reportsthedecompositionfortotalimports. 20

dustry that warrant further study. Some of these differences could result from 1) the mix of supply and demand factors during the heart of COVID-19 lockdowns, 2) idiosyncratic factors affecting the ease with which trading relationships form and break, and 3) capacity constraints in existing foreign suppliers. In summary, the BoL data allow researchers to understand the mechanisms underlying the extraordinary growth in imports during the onset of the COVID-19 pandemic. These decompositions would have been invisible using traditional, publicly available datasets. 5.3 Real-time measures of shipping bottlenecks during the COVID-19 trade recovery The dramatic resurgence of trade in the second half of 2020 led to some much-discussed bottlenecksacrossmanytransportationmodes. Inthissection,weshowhowtheBoLdata canbeusedtoexaminecharacteristicsofvesselshippingthatshedlightontheprevalence and effects of bottlenecks in oceanic vessel shipping in nearly real time. The use of BoL data to study the shipping network is highlighted by Ganapati et al. (2021) when used in conjunction with newly available vessel transponder data (otherwise known as Automatic Identification System (AIS) data) that tracks vessel ship movements.12 While BoL data alone can identify the presence of indirect shipping—the primary topic of interest in Ganapati et al. (2021)—based on shipments coming from many different ports of lading on a given vessel-port of unlading combination, the key drawback is a lack of date associated with a shipment’s foreign departure. The time stamp on AIS vessel movements enable researchers to track the precise route of a vessel through multiple ports of call. However, the key limitation of AIS data is a lack of any easily quantifiable measure of trade volume associated with each vessel. The analysis below leveragesthevesselandportsofunladingvariablesthataretypicallyreportedintheBoL data, and focuses attention on the vessel congestion centered in the ports of Los Angeles / Long Beach in late 2020 and into 2021. We take several steps to convert the raw BoL data into a dataset useful for tracking vessel arrivals at U.S. ports. First, we clean and standardize vessel name and a corresponding vessel identifier to account for inconsistencies in these variables.13 Second, for many analyses at a vessel-port level it is helpful to restrict attention to container vessels. While external lists can identify vessels based on vessel type, for our purposes, we classified container vessels based on a measure of observed capacity: whether the maximum 12SeeHeilandetal.(2021),Cerdeiroetal.(2020)andCerdeiroandKomaromi(2020)forexamplesofrecent papersusingAIStransponderdata. 13WeprovidedetailforthisprocessinAppendixC. 21

observedTEUsunloadedataparticularpointoftimeforavesselsurpassedathreshold.14 Third,wemustidentifyaspecificdateforavesselunloadingcargoataU.S.port. The difficulty here lies in the fact that the “arrival date” associated with BoL records typically reflect when individual shipments clear customs. Generally speaking, a large majority of BoL import shipment records from a container vessel at a port of unlading are listed as arrivingwithinaoneortwodayperiod. However,therearefrequentexceptionsinwhich a vessel’s shipments are reported as arriving over more extended periods of time, which could lead to an incorrect inference for a vessel arrival date. These records could reflect delays in clearing customs, typos in arrival date, or differences in identifying arrival date by exporters or importers. To account for these concerns, we take our baseline dataset of daily vessel-port observations and then eliminate a daily record if that day’s shipments fromaparticularvesselwereaverylowshareofthevessel’s(observed)maximumcapacity. Finally,weconsolidateavessel’sarrivaldateintoasingledayifsubstantialshipments occur over a period of less than five days. Figure11: Mediannumberofdaysbetweenvesselvisitsatport (a)PortofLosAngeles/LongBeach (b)MajorEastCoastPorts Number of days Number of days 60 80 76 56 72 52 68 48 64 44 60 40 56 2018 2019 2020 2021 2018 2019 2020 2021 Source: S&PGlobalMarketIntelligenceandauthors’calculations. Notes: This figure plots, for a given month, the median number of days since a vessel last visited the port. Theblacklineineachfigurerepresentstheaveragenumberofdaysduringtheperiod2013-2017. MajorEast CoastportsincludetheportsofCharleston,Newark/NY,Norfolk,andSavannah. For a first look at the insights from this new dataset, we quantify the delays in vessel movementsbroughtonbytheshippingcongestionexperiencedin2020and2021. Tomea- 14Forthediscussionbelow,wesetthisthresholdatarelativelylowvalueof200,thoughforotherpurposes researchersmaywanttofocusonvesselswithlargercapacity. 22

surethetypicaltransittimesforcontainervesselsatagivenport,wecalculatethenumber of days between return arrivals of a given vessel and calculate the monthly median value for a given port. Panel (a) of Figure 11 indicates that a typical vessel would unload cargo at the Ports of Los Angeles/Long Beach (LA/LB) about every 43 days during normal times (2013-2017). This value was relatively stable in 2018 and 2019, but spiked in early 2020 following country-level lockdowns and the more general slowdown in trade during theearlyperiodofCOVID-19. Round-triptransittimesnormalizedinthethirdquarterof 2020 but subsequently increased in late 2020 and early 2021 due to the congestion at the Port of LA/LB. The median number of days in between port visits of 52 during 2021-Q1 and 2021-Q2 reflects an increase of roughly 8 days from typical levels. Panel (b) of Figure 11 shows that there has been no such systematic delays in ship processing at an average of major U.S. East Coast ports during this period. Panel (b) also shows the longer average round-trip transit time of East Coast ports, a fact which reflects the increased prevalence of multi-stage trips common for vessels servicing these ports.15 Given reports that the congestion at the Ports of LA/LB resulted in vessels being rerouted to unload at other ports on the U.S. West Coast, we next attempt to quantify this degree of rerouting from our BoL-based dataset of vessel-port traffic. We first identify the sample of vessels that visited the Port of LA/LB on a consistent basis in a pre-Covid period,i.e. inbothQ3andQ4of2019. Inthesubsequentsixquarters(2020-Q1to2021-Q2) we identify the potential set of vessel re-routings as those vessels that are not observed visitingthePortofLA/LBbutareobservedvisitingadifferentU.S.port. Wemeasurethe magnitude of these re-routings as the number of TEUs unloaded at alternative ports in a given quarter, which are then displayed as a fraction of total inbound TEUs at the port of LA/LB in that quarter. Finally, because what we define as re-routing may occur even in normal times, we calculate identical statistics from baseline periods in each of 2016-2018 and subtract the average of these “normal” vessel re-routings from the period of study. 15MediantimebetweenportvisitsalsotendstobenoisierforEastCoastportsbecauseWestCoastports havemorededicatedport-to-portvesselroutes,whichtendtorunonmorepredictableschedules. 23

Figure12: PercentofinboundLosAngeles/LongBeachactivityre-routedtootherports Percent of inbound LA−LB TEUs 8 Ports of Seattle−Tacoma Port of Oakland 6 Avg. of Major East Coast Ports 4 2 0 −2 22002200−−QQ11 2020−Q2 2020−Q3 2020−Q4 22002211−−QQ11 2021−Q2 Source: S&PGlobalMarketIntelligenceandauthors’calculations. Notes: This figure plots the percent of quarterly inbound LA/LB TEU imports that are identified as being re-routed to other ports. These values are net of the average observed percent re-routed to these ports during the period 2017-2019. Major East CoastportsincludetheportsofCharleston,Newark/NY,Norfolk,andSavannah. The result is plotted in Figure 12 for three likely destinations of re-routings from the Ports of LA/LB: Seattle-Tacoma, Oakland, and an aggregate of four major East Coast ports. Figure 12 reveals that vessel re-routings from LA/LB to Seattle-Tacoma spiked in the first quarter of 2021 (following the onset of port congestion in late 2020) to an amount equal to roughly 8 percent of inbound TEUs at the ports of LA/LB. This rerouting declines somewhat in the second quarter of 2021 but remains elevated relative to normal levels. While some re-routings were documented in press reports to the Port of Oakland, our data indicate that these did not constitute a significant fraction of inbound TEUsfromLA/LB.Similarly,thedataalsoconfirmthatfew,ifany,vesselswerere-routed on net from LA/LB to the East Coast of the United States during this period. In summary, the unique features of BoL data, together with timely access, provide bothresearchersandpolicymakerswithausefultooltoanalyzedisruptionstotradesuch as those accompanying COVID-19. 24

6 Conclusion This paper provides the first detailed analysis of the utility of data from bills of lading for international trade research, specifically the information available on U.S. imports via Panjiva. These data provide a near real-time, firm-level dataset useful for addressing a variety of economic questions that cannot be addressed with other data. Furthermore, some of the limitations of U.S. import data—including a general lack of trade values, redaction of some firm names, and being restricted to vessel shipping—do not apply to the same data available for other countries. We use the unique elements of the data to analyze international trade relationships. Over 60 percent of consignees (importers) have only one foreign shipper (exporter), but these consignees represent less than 20 percent of import volumes. Most shipperconsignee pairs ship in three or fewer months per year, though the surprisingly small number of pairs that ship every month account for over 30 percent of U.S. imports by TEU.IntheCOVID-inducedtradecollapsein2020,thenumberofshippersperconsignee dropped dramatically and remained volatile before recovering in mid-2021. Finally, we explore other aspects of international trade during the COVID-19 crisis. The daily frequency shows how quickly exports from India to the United States fell following lockdowns in March 2020. Furthermore, the resulting drop in U.S. imports weeks later demonstrates clearly how international shipping lags transmit these shocks with a delay. Following the collapse, U.S. goods demand recovered briskly, and these data demonstrate the margins on which imports can rise. In the very short run, within a few months, higher imports were mostly achieved within existing shipper-consignee pairs. Over subsequent quarters, however, imports rose by consignees switching shippers or source countries, and also by the entry of new consignees. Our work and the recent literature demonstrate that bill of lading data remains underutilized in international trade. With some caveats, these data provide a useful complementary dataset to disaggregated official public data and confidential datasets. Moreover, the ability to see most firm names of shippers and consignees opens the possibility of merging BoL data with other firm-level datasets. Researchers in international trade should consider the possibility of using BoL data in their future work. 25

References Bernard, Andrew B. and Andreas Moxnes, “Networks and Trade,” Annual Review of Economics,2018,10(1),65–85. _eprint: https://doi.org/10.1146/annurev-economics- 080217-053506. , J. Bradford Jensen, and Robert Z. Lawrence, “Exporters, Jobs, and Wages in U.S. Manufacturing: 1976-1987,” Brookings Papers on Economic Activity. Microeconomics, 1995, 1995, 67–119. Publisher: Brookings Institution Press. , , Stephen J. Redding, and Peter K. Schott, “The Margins of U.S. Trade (Long Version),” Working Paper 14662, National Bureau of Economic Research January 2009. Boehm,ChristophE.,AaronFlaaen,andNityaPandalai-Nayar,“InputLinkagesandthe Transmission of Shocks: Firm-Level Evidence from the 2011 To¯hoku Earthquake,” The Review of Economics and Statistics, March 2019, 101 (1), 60–75. Bonfiglioli, Alessandra, Rosario Crinò, and Gino Gancia, “Firms and Economic Performance: A View from Trade,” September 2020. , , and , “Concentration in international markets: Evidence from US imports,” Journal of Monetary Economics, July 2021, 121, 19–39. , , and , “International Trade with Heterogeneous Firms: Theory and Evidence,” CEPR Working Paper, June 2021, (16249). Bruno, Valentina and Hyun Song Shin, “Dollar and Exports,” April 2020. Cerdeiro,DiegoA.andAndrasKomaromi,“SupplySpilloversDuringthePandemic: Evidence from High-Frequency Shipping Data,” IMF Working Papers, December 2020, (WP/20/284). , ,YangLiu,andMamoonSaeed,“WorldSeaborneTradeinRealTime: AProof ofConceptforBuildingAIS-basedNowcastsfromScratch,”IMFWorkingPapers,May 2020, (WP/20/57). Dhyne,Emmanuel,AyumuKenKikkawa,MagneMogstad,andFelixTintelnot,“Trade and Domestic Production Networks,” TheReviewofEconomicStudies, March 2021, 88 (2), 643–668. Feenstra, Robert C and David E Weinstein, “Globalization, Markups, and US Welfare,” Journal of Political Economy, August 2017, 125 (4), 35. 26

Ganapati, Sharat, Woan Foong Wong, and Oren Ziv, “Entrepôt: Hubs, Scale, and Trade Costs,” July 2021. Heiland,Inga,AndreasMoxnes,KarenHeleneUlltveit-Moe,andYuanZi,“TradeFrom Space: Shipping Networks and The Global Implications of Local Shocks,” March 2021. Heise, Sebastian, Georg Schaur, Justin R Pierce, and Peter K Schott, “Tariff Rate Uncertainty and the Structure of Supply Chains,” January 2019. Jain, Nitish and Di (Andrew) Wu, “Can Global Sourcing Strategy Predict Stock Returns?,” SSRN Scholarly Paper ID 3606884, Social Science Research Network, Rochester, NY May 2020. , Karan Girotra, and Serguei Netessine, “Managing Global Sourcing: Inventory Performance,” Management Science, May 2014, 60 (5), 1202–1222. Monarch, Ryan, ““It’s Not You, It’s Me”: Prices, Quality, and Switching in U.S.-China Trade Relationships,” The Review of Economics and Statistics, February 2021, pp. 1–49. Panjiva, “S&P Global Market Intelligence, Panjiva Supply Chain Intelligence Platform & Data Feed.” Wasi, Nada and Aaron Flaaen, “Record Linkage Using Stata: Preprocessing, Linking, and Reviewing Utilities,” The Stata Journal, October 2015, 15 (3), 672–697. Publisher: SAGE Publications. 27

A Comparing Panjiva, U.S. Census, and Port Data Inthissectionoftheappendix,wecompareaggregatedshippingvolumeBoLdatatodata directly reported by ports themselves and also to official U.S. Census Bureau data. These checks provide information to researchers considering the representativeness of the BoL data. We focus on comparing volume measures in BoL (weight and TEU), as they are more comprehensively available than imputed values, which suffer from both missing observations and extensive measurement error. Generally speaking, weight and TEU are very similar volume measures over time and could substitute for one another given most questions. In short, we find that BoL data closely track official port data in TEUs. A.1 Comparing statistics by port Table 6 compares measures of trade weight and number of TEUs by port, as reported in BoL data and by Census. BoL weight measures tend to exceed Census measures somewhat. Still, as the columns labeled “share” demonstrate, the proportion of imports going to each port is similar between Census and BoL, with the notable exception of Los Angeles and Long Beach: Here, the sum of the two ports is more comparable than their individual identification. This adds to the list of reasons why it is best practice to treat LA/LB as a single economic entity for most questions with these data. The right two columns of Table 6 provide the total TEU count in 2019 by port for BoL data and data provided by the ports themselves. In most cases, these correspond remarkably closely. Table6: ComparisonofPanjivaandOfficialStatistics,2019 Panjiva Census Official Panjiva Census Panjiva Weight Weight Port Weight* Weight* TEU* Share Share TEU* Houston, TX 50,497 55,630 0.08 0.09 1.23 1.24 Los Angeles, CA 43,092 25,190 0.07 0.04 4.63 4.71 Long Beach, CA 35,524 54,085 0.06 0.09 3.72 3.76 Newark, NJ 51,425 56,963 0.08 0.09 3.65 3.77 Savannah, GA 20,803 20,049 0.03 0.03 2.20 2.22 Seattle/Tacoma, WA 14,726 14,398 0.02 0.02 1.49 1.37 Source: S&PGlobalMarketIntelligence,HaverAnalytics,U.S.Census,andauthors’calculations. Notes: PanjivaaggregatesforSeattle/TacomaandHoustonexcludeshipmentswheretheconsignee countryismissing. *Inmillions. 28

A.2 Comparing containers by port over time Next, we compare the number of imported TEUs reported by Panjiva to the volumes reportedbyports. Inparticular, Figure13displaysmonthlyPanjivaandofficialimported TEU volumes for the top six U.S. ports. Importantly, both sources tend to give similar signals for the level and changes in trade from month to month. In terms of timeliness of data reporting, the official data on container volumes by port are available from Haver with a lag of about 3 weeks on average, while data are available from Panjiva with a lag of only about 7-14 days. While this improvement in timeliness from Panjiva data is relatively modest, it may nonetheless be valuable during times when shipping is being interrupted, such as during the COVID-19-related plunge and the backups at West Coast ports during the subsequent recovery. Figure13: ComparisonofPanjivadataandofficialportstatisticsbyport Los Angeles Long Beach Seattle and Tacoma Thousands Thousands Thousands 550 450 150 500 400 140 450 130 400 350 120 350 300 110 300 100 250 250 90 200 2018 2019 2020 2021 2018 2019 2020 2021 2018 2019 2020 2021 Newark Savannah Houston Thousands Thousands Thousands 140 400 240 130 380 360 220 120 340 200 110 320 180 100 300 90 280 160 80 260 2018 2019 2020 2021 2018 2019 2020 2021 2018 2019 2020 2021 Official port statistics Panjiva Source: S&P Global Market Intelligence, individual ports via Haver Analytics, and authors’calculations. Notes: PanjivaaggregatesforSeattle/TacomaandHoustonexcludeshipmentswhere theconsigneecountryismissing. 29

Figure14: ComparisonofPanjivadataandofficialportstatistics TEUs, thousands 3500 Panjiva: All ports Panjiva: Major ports* 3000 Official port statistics: Major ports* 2500 2000 1500 1000 500 0 2009 2011 2013 2015 2017 2019 2021 Source: S&P Global Market Intelligence, individual ports via Haver Analytics, and authors’calculations. Notes: *Major ports include Houston, Long Beach, Los Angeles, Savannah, Seattle/Tacoma,NewYork,andNewark. A.3 Comparing lags in data reporting Figure 15 illustrates the timeliness of when data are available for a given month. As shown in the figure, data are updated continuously with roughly three-quarters of a given month’s final TEU value in by the end of the month. The data then reach close to 100 percent of the final monthly value by around 7 to 14 days after the end of the month. This reporting is sooner than the port-level reporting and significantly sooner than the Advance Economic Indicators trade report released by the U.S. Census Bureau. 30

Figure15: PanjivaBoLdatacompleteness Percent of final TEU estimate for the month 100 90 CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeennnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuussssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssss aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaannnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnncccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccceeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee ttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrraaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee rrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrreeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeelllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllleeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaasssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssseeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee 80 70 AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvgggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggg................................................................................................................................................................................................................................................................ ppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppoooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooorrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrtttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttt−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllleeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeevvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvveeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeellllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll ddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaattttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa rrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrreeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeelllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllleeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaasssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssseeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee**************************************************************************************************************************************************************************************************************************************************************** 60 50 Georgia Ports Authority, Savannah, Georgia New York/Newark Area, Newark, New Jersey Port of Long Beach, Long Beach, California 40 The Port of Los Angeles, Los Angeles, California 30 −14−12−10 −8 −6 −4 −2 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 Days following the end of the month Source: S&PGlobalMarketIntelligence,individualportsviaHaverAnalytics,CensusBureau,andauthors’ calculations. Notes: 100percentreflectsthe“final”levelofTEUsestimatedforagivenmonth. A.4 Comparing firm-level trading information Asdiscussedabove,oneofthekeybenefitsoftheBoLdata,relativetopublicdatasources, istheavailabilityoffirmidentifiersformosttransactions. Comparingfirm-levelinformation from BoL data to similar information in other datasets, such as the Census Bureau’s Longitudinal Foreign Trade Transaction Database (LFTTD), is difficult given the confidentiality associated with official statistical datasets. Nonetheless, the Census Bureau does publish some information on characteristics of firms engaged in international trade, which can be compared to BoL sources. One piece of information about trading firms that the Census Bureau reports is a histogram of the value of trade by the number of destination countries for each exporting firm(Seetopchartonpage3,CensusBureau2020). InFigures16and17,wedisplaysimilar figures based on Panjiva data for both exporters and importers, respectively, though our histograms are in terms of the number of TEUs and shipments. Our figures include all firms and are therefore most comparable to the blue bars in the histogram provided by the Census Bureau. Figure 16, for exports, shows a rightward skew of the distributions for TEUs and shipments based on BoL data, indicating the importance of firms that export to many 31

countries in overall trade volumes. This rightward skew is consistent with, but actually somewhat less pronounced than that reported for the value of exports in Census Bureau (2020), which is reproduced in the gray bars of the Figure. Figure17indicatesthat,incontrasttoexports,firmsthatimportfromasmallnumber ofdestinationsaccountforarelativelylargershareofU.S.importvolumes. Thisdifference may be indicative of smaller fixed costs associated with importing, relative to exporting. Figure16: PercentofTEUs,shipments,andvaluebynumberofpartnercountriesforU.S.exports, 2018 Percent 50 TEUs 40 Shipments Known export value (Census) 30 20 10 0 1 2−4 5−9 10−24 25−49 50−200 Number of partner countries Source: S&PGlobalMarketIntelligence,U.S.Census,andauthors’calculations. 32

Figure17: PercentofTEUs,shipments,andvaluebynumberofpartnercountriesforU.S.imports, 2018 Percent TEUs 30 Shipments Known import value (Census) 20 10 0 1 2−4 5−9 10−24 25−49 50−200 Number of partner countries Source: S&PGlobalMarketIntelligence,U.S.Census,andauthors’calculations. 33

A.5 Bill of Lading Forms for U.S. Imports and Exports Figure18: CustomsandBorderProtectionBillofLadingFormforU.S.Imports 34

Figure19: CustomsandBorderProtectionBillofLadingFormforU.S.Exports B Handling the Panjiva data feed While the storage and analysis of data is typically a relatively minor concern in economics research, these issues require extensive consideration when handling massive datasets, such as Panjiva’s BoL data. In this section, we provide detailed discussion of the computing solutions we used to effectively make use of these data. Our hope is that this information will assist other researchers as the use of BoL data becomes more widespread. Scalable programming tools are critical to effectively ingesting and analyzing this dataset. While there are a variety of potential big data solutions to handle 100 GB of data, our research solution balanced performance and usability of the data. Harnessing primarilyopensourcetoolsfromApacheandthePythonSoftwareFoundation,weloaded the Panjiva data into a cluster Hadoop environment to provide scalable data storage and processing. PanjivaprovidesresearchersaccesstotheunderlyingdatathroughanFTPserverthat hoststherawfilesinazipformat. Oncerawfilesaredownloaded,theyaredecompressed and converted out of their “phrase-separated” values file into a more useful format for querying. Panjiva’s file format uses a non-standard characters to separate records and fields, which can cause performance bottlenecks. The files are large enough to warrant 35

Table7: U.S.importdatadescriptionforremainingvariables Variablename Description billofladingnumber Billofladingforshipment billofladingtype Typesofbillsoflading:House,SimpleorMasterdesignation carrier Nameofthecompanythattransportsthegoods concity Cityoftheconsignee’sdomesticlocation concountry Countryoftheconsignee’sdomesticlocation confulladdress Fulladdressoftheconsignee’slocation conoriginalformat Thepartytotakefinaldeliveryofthemerchandise(originalformat) conpostalcode Postalcodeoftheconsignee’sdomesticlocation conroute Streetaddressoftheconsignee’sdomesticlocation constateregion State/regionoftheconsignee’sdomesticlocation containermarks Symbolsprintedonboxes/cratestodeterminehowtohandleshipment containermarksid Symbolsprintedonboxes/cratestodeterminehowtohandleshipment containernumbers Containeridentificationnumbers containernumbersid Containeridentificationnumbers containertypeofservice Indicatesthetypeofserviceprovidedforthecustomer containertypeofserviceid Indicatesthetypeofserviceprovidedforthecustomer containertypes Indicatesthetypeofcontainerusedintheshipment containertypesid Indicatesthetypeofcontainerusedintheshipment dangerousgoods Substancesormaterialsthatposeunreasonablerisktohealthandsafety dangerousgoodsid DangerousgoodsID dividedLCL Indicateswhethershipmentsarecombinedwithothershipments dividedLCLid Indicateswhethershipmentsarecombinedwithothershipments filedate Representsthedatethatthedatawaspubliclyavailable FROB Foreigncargoremainingonboard goodsshipped Freetextdescriptionoftheproduct goodsshippedid UniqueIDforrecordswithingoodsshippedtables hasLCL Denoteswhethertheshipmenthasconsolidatedcargo hscodeid HarmonizedItemDescriptionandCodingSystem(HS) inbondcode IndicateswhethertheshipmentisIn-BondornotIn-Bond iscontainerized Indicateswhetherashipmentwascontainerized(Panjivaderived) manifestnumber Identificationnumberofmanifestonwhichgoodswerelisted masterbillofladingnumber Identificationnumberofthemasterbilloflading measurement Additionaldescriptionofmeasurementusedontheshipment notifyparty Nameofnotifyparty notifypartySCAC StandardCarrierAlphaCode(SCAC)fornotifyparty numberofcontainers Totalnumberofcontainersintheshipment placeofreceipt Locationwherethegoodswerereceivedfortransporttothevessel portofladingcountry Countryofportoflading portofladingregion Regionofportoflading portofunladingregion Regionofportofunlading quantity Quantityofitemsintheshipment shpcity Cityinwhichtheexporterislocated shpcountry Countryofshipper shpfulladdress Fulladdressofshipper shpmtdestination Countryofshipmentdestination shpmtdestinationregion USgeographicregionofthefinaldestinationofthegoods shporiginalformat Nameoftheshipper(originalformat) shppostalcode Entityresolvedpostalcodeoftheshipper’sdomesticlocation shproute Streetaddressoftheshipper’sdomesticlocation(Panjivaderived) shpstateregion State/regionoftheshipper’sdomesticlocation(Panjivaderived) transportmethod Modeoftransportation vesselvoyageid VoyageIDforvesselcarryingshipment volumecontainerTEU VolumeofcontainerinTEU volumecontainerTEUid VolumeofcontainerinTEU weightoriginalformat Shipmentweightasoriginallyreportedonshipmentrecord weightT Shipmentweightinmetrictons parallelization though their format of the files forces an initial single core processing bottleneck. Parallelizationisrequiredtoeffectivelyingestthesefilessinceasinglefilecan 36

contain tens to hundreds of millions of records. Using the python packages Dask and Pandas, the data are saved into a column-oriented storage object called Apache Parquet. This file type is popular among big data experts due to its effective compression and querying performance. The structure of Panjiva updates involves large snapshot files and smaller modification files that could include updates, additions, or deletions to the primary data. TheseoptimizeddatafilesarepartitionedbasedonthestructureprovidedbyPanjiva. For U.S. Imports data, records are separated into four blocks by arrival date: 2007-2009, 2010-2014, 2015-2019, 2020-2024. Partitioning the data improves query performance on the cluster by minimizing the number of files being scanned when requesting time-based subsets of the data. The data files are structured into a partitioned Apache Hive tables with the added capabilities of Apache Impala. We then utilize either PySpark or SQL protocols to query specific data subsets from the hundreds of millions of records and produce basic summary statistics of the data. C Data Cleaning and Vessel Standardization Fortheportanalysisexercise,wepullallshipment-leveldatafrom2012topresent,resulting in approximately six million observations. As with most “big data” datasets, spelling errors and name variations are widespread. Over two-and-a-half million observations are missing a vessel International Maritime Organization identifier (IMO). We remove obvious duplicates, standardize the vessel names, and impute the missing vessel IMOs. First,wegenerateacrosswalkusingasubsetofthedatathatcontainsavesselIMO.We clean the vessel names by removing any non-alphanumeric characters and removing any trailingorleadingspaces. Next,wecollapseourdatabyvesselnameandvesselIMO–of the 15,000 unique vessel IMOs in our known IMO dataset, over 2,000 are associated with more than one vessel name. Because of this, we create a “primary” vessel name based on the name which is most commonly associated with each individual vessel IMO. We then merge back the standardized crosswalk of known IMOs on the full dataset, resulting in almost 3.5 million observations to be standardized. Next, we aim to add in the missing vessel IMOs of over 2.5 million observations. This process is similar to our standardization process, except we keep all observations with an unknown vessel IMO. We then clean the vessel names in the same manner (remove trailing and leading spaces and non-alphanumeric characters). Next, we merge these onto the de-duplicated known vessel IMO and standardized vessel name crosswalk such that only unique names in the unknown IMO dataset will merge to unique names in the 37

knownIMOcrosswalk. Over2millionobservationsmergedatthispoint,leavingroughly half a million remaining. The rest of the data cleaning is an iterative process. After skimming through the unmatched observations, we find that almost 8,000 observations actually have the vessel IMO in the vessel name column. We also find that of the unknown vessel IMO observations, some observations will likely remain unidentified as they have meaningless vessel names; we choose to drop names that are a random mix of numbers and letters and are also too short (less than seven characters) to accurately identify. At this stage, we have less than half-a-million observations without a vessel IMO to identify. Using reclink2, a natural language processing package ((Wasi and Flaaen 2015)), we attempt to merge the remaining unknowns to our known dataset. We manually go through 2,000 unique name matches resulting from the fuzzy merge. The fuzzy merge providesa“score”ofhowaccuratetheunidentifiedstringistoaknownIMOstringvessel name. We tag all matches over 99 percent accuracy and generate a crosswalk from the fuzzy merge which yields us an additional 50,000 observations. Finally,wemanuallysearchforthevesselIMOsbasedonthevesselnameandroute(if there are multiple IMOs associated with a given vessel) of the largest number of unidentified vessels. Merging these imputed vessel IMOs results in almost 100,000 identified vessels. After these iterative processes, our data cleaning and standardization results in over 5.6 million of the 6 million observations. 38