Name fragmentation as measurement error

Abstract

Faculty who want a US importer panel from public ocean bills of lading currently buy entity identifiers from Panjiva / S&P. This paper shows, on a public file, why. Treating the raw AMS consignee_name as a firm ID invents volume collapses and supplier exits that are spelling. A researcher who types the 2025 Walmart string WALMART INC finds zero unique ocean bills in every year 2014–2017, then 26,057 in 2018. The 2016 modal string, WALMART STORES INC, is 1,849 bills in 2016 and one bill in 2025 — 0.0% of the 2025 Walmart prefix family. Home Depot’s 2025-legal string HOME DEPOT USA INC files 32,587 bills in 2019, four in 2020, 16,802 in 2021, and zero in 2022–2023, while HOME DEPOT USA without the suffix continues. Prefix families and a cheap INC/LLC strip do not replace an ID: they fail on Target (steel mills in the prefix), and they still split Walmart Stores from Walmart. The 2017 Walmart and 2020 Home Depot holes remain thin even after grouping, so coverage error is stacked on name error. The name error is the result. Replication: /research-name-fragmentation.

1. Introduction

A bill of lading is a string. The consignee field is whatever the filer typed that week: legal name, trade name, a 3PL, a store number, a Chilean subsidiary, a steel mill that happens to start with TARGET. Trade papers that use public AMS filings as a firm-year panel are then one exact-match away from a fake exit. Panjiva, ImportGenius, and the S&P Capital IQ linkage exist to stop that. They are not free. There is no public, citable demonstration of how badly the raw string lies, with the queries attached.

This paper makes that demonstration the result, not a data appendix. Two exhibits are already sitting in the Kirchner COVID and China Plus One papers as coverage caveats. Walmart’s locked-name series falls to 437 unique bills in 2017 against 1,849 in 2016 and 26,115 in 2018. Home Depot’s locked series falls to 1,139 in 2020 against 35,978 in 2019. Those sentences were written as “do not interpret the hole as de-globalization.” They are also the measurement-error paper. The 2025 legal name is not in the file before 2018. The 2016 legal name is not in the file in 2025. A naive panel built on either string is a different firm in the two halves of the decade.

Hypothesis

Conditional on a brand that is continuously importing on ocean bills, the unique-bill series of a single exact consignee_name is not a firm-year volume series. Holding the 2016 modal string fixed through 2025, and holding the 2025 modal string fixed back to 2016, both miss most of the prefix-family volume at the other end of the sample. A cheap punctuation-and-suffix cleaner does not close that gap for Walmart. Prefix matching is not an entity ID: it recovers the INC/USA flicker and also scoops TARGET STEEL.

2. Data and design

The source is Kirchner kirchner.bols on ClickHouse: US ocean AMS / CBP bills, 2014-01-01 through 2025-12-31 (173,043,175 rows; 158,548,137 unique bills). The unit is uniqExact(bill_of_lading). A US consignee filter drops names containing CANADA, MEXICO, DE MEXICO, or S DE R L. That filter is mechanical, not an address parser.

For each brand we keep four series. Naive is exact match on the string a 2025 researcher would type: WALMART INC, HOME DEPOT USA INC, falling back to the modal US prefix name if that string is absent. Locked is the three-name set used in the COVID and Plus One papers. Prefix (US) is startsWith(upperUTF8(consignee_name), …) on WALMART / WAL-MART, HOME DEPOT / THE HOME DEPOT, and the analog for IKEA, Costco, Mattel, and Target, with the US filter. Contains is the same tokens anywhere in the name. A fifth construction, legal-norm, uppercases, strips punctuation, and drops INC, CORP, LLC, LTD, CO, THE, USA, and similar tokens, then groups. Prefix is not Panjiva. It is the cheapest public substitute, and Target is in the sample to show that it is too cheap.

Concentration is the share of prefix-year bills sitting on the single most common string, and the HHI of strings inside that family. Held-fixed tests take the 2016 modal US-prefix string and follow it to 2025, and the 2025 modal string back to 2016. A false supplier exit is a shipper present on the naive string in year t, absent from that string in year t+k, still present on the US prefix family in t+k.

3. Walmart: the 2025 name is born in 2018

Table 1 is the paper. WALMART INC does not exist as a consignee string in this file in 2014, 2015, 2016, or 2017. It appears in 2018 with 26,057 unique bills and is 98.8% of the 2025 prefix family (29,195 of 29,549). The 2016 modal string is WALMART STORES INC: 1,849 bills that year, 437 in 2017, 58 in 2018, one in 2025. Hold 2016 fixed and Walmart exits. Hold 2025 fixed and Walmart is founded in 2018. The firm did not exit. The ticker did not change in a way that AMS spelling tracked. The filer stopped typing STORES.

The locked three-name set (INC, STORES INC, GLOBAL LOGISTICS) still shows the 2017 hole: 1,849 → 437 → 26,115. The US prefix family is 1,996 → 585 → 26,569. Contains-anywhere lifts 2017 only to 661. Notify-party names containing WALMART are 143, 843, and 10,872 in 2016–2018; they do not fill a 26,000-bill gap. Two errors are stacked. Name fragmentation is the zero on WALMART INC before 2018 and the zero on WALMART STORES INC after it. Coverage is the fact that even WALMART* is thin in 2017. The COVID paper was right to flag 2017 as unusable for turnover. This paper is about the zeros that grouping does not create and does not fix: the wrong string, followed across years, is a different company.

Two hundred six distinct US WALMART / WAL-MART consignee strings appear in 2014–2025. Mean top-1 share inside the family is 95.6%. The family is concentrated. The identity of the top string switches. That is enough to wreck an exact-name panel.

Table 1. Walmart unique ocean bills, 2014–2025

Naive = exact WALMART INC. Locked = INC + STORES INC + GLOBAL LOGISTICS. Prefix = US names starting WALMART or WAL-MART.

4. Home Depot: INC is a dummy for “the firm vanished”

Table 2 is the second exhibit. HOME DEPOT USA INC is 32,587 unique bills in 2019 and four in 2020. It returns to 16,802 in 2021, then files zero in 2022, zero in 2023, and one in 2024, while the prefix family files 3,542, 7,432, and 8,562. In those years the mass sits on HOME DEPOT USA without INC (3,343, 7,265, 8,410). In 2025 INC comes back (8,466) but is only 12.4% of the prefix family (68,059); HOME DEPOT USA is 66.7%. A researcher who believed INC was the company would record a near-total exit in 2020, a resurrection in 2021, another exit in 2022, and a modest 2025 return. That is not sourcing. That is a suffix.

Locked names recover 1,139 bills in 2020 against four on INC, and they recover the 2022–2024 volume that INC misses. They do not fill 2020 relative to 2019 (35,978 locked → 1,139). Prefix is 36,604 → 1,336. Contains is 1,361. As with Walmart 2017, 2020 is a coverage hole and a name hole at once. The name hole is the four-versus-1,135 split inside 2020, and the 2022–2024 zeros on INC while thousands of prefix bills continue. Ninety-five distinct US HOME DEPOT strings appear over the decade. Mean naive share of the prefix is 39.5%. Mean top-1 share is 83.4%, and the top string itself flips between INC and not-INC.

Table 2. Home Depot unique ocean bills, 2014–2025

Naive = exact HOME DEPOT USA INC. Locked = USA INC + USA + THE HOME DEPOT INC. Prefix = US names starting HOME DEPOT or THE HOME DEPOT.

2015–2017 is another INC collapse (32,488 → 64) that prefix only partly fills. The file has more than one Home Depot hole. 2020 and 2022–2024 are the clean suffix exhibits.

5. Held-fixed names, false exits, and a cleaner that is not an ID

Table 3 asks the panel question. Take the most common US-prefix string in 2016 and ask what share of 2025 prefix volume it still captures. Take the 2025 modal string and ask the same of 2016. For Walmart both answers are 0.0%. For Mattel the 2016 name MATTEL INC is 86.5% of 2016 prefix volume and 0.1% of 2025; the 2025 name MATTEL IMPORT SERVICES LLC is 0.0% of 2016. IKEA’s 2016 modal, IKEA DISTRIBUTION SERVICES INC, is 0.1% of 2025; IKEA SUPPLY AG took the family. Home Depot is the exception that proves the mechanism: the 2016 modal HOME DEPOT USA is still 66.7% of 2025, because that is the string that absorbed the INC flicker. Target’s 2016 modal is 0.6% of 2025, and the 2025 modal is TARGET LOS ANGELES.

Table 3. Share of prefix-family bills captured by a name held fixed

Table 4 takes shippers. Follow WALMART STORES INC from 2016 to 2018: 206 shipper strings in 2016, five still on that consignee in 2018 (naive exit 97.6%), 26 still on some US WALMART* consignee (false-exit share 12.6%). Follow MATTEL INC from 2016 to 2024: 145 shippers, zero remain on INC, 22 remain on the Mattel prefix (15.2%). Naive exit is 100%. Prefix-adjusted exit is 84.8%. The 15 percentage-point gap is suppliers a name-as-ID panel would count as gone who are still selling to Mattel under a different consignee spelling. Home Depot USA INC from 2021 to 2022: 581 shippers, zero stay on INC, 47 (8.1%) stay on the prefix. 2022 is a thin prefix year, so most of that “exit” is the coverage stack again. The Mattel window is the cleaner supplier-margin exhibit because 2024 prefix volume is in the same range as 2016.

Table 4. False supplier exits off a single consignee string

False exit = on the naive string in the from-year, not on that string in the to-year, still on the US prefix family in the to-year.

A graduate student will say: strip INC and LLC. The legal-norm grouping does that. For Home Depot it mostly works, because USA is also dropped and INC/not-INC collapse to HOME DEPOT (top legal-norm share 99.8% in 2023–2024, 99.4% in 2025). For Walmart it fails the actual rename: WALMART STORES versus WALMART remain different keys (93.3% on STORES in 2016, 98.1% on WALMART in 2018). For Mattel it splits MATTEL from MATTEL IMPORT SERVICES. Mean legal-norm groups per year are 27.5 at Walmart and 20 at Mattel — still a panel of names, not a firm. Shipper strings in 2024 are not saved by the same cleaner: Walmart’s top 300 shippers collapse only 300 → 280 groups (residual ratio 1.07).

Target is why prefix is not for sale as an ID. One hundred distinct US TARGET* strings; mean top-1 share 42.6%; mean HHI 0.326. The 2016 name list includes TARGET STEEL INC. The 2025 modal is TARGET LOS ANGELES. Naive TARGET STORES DIVISION OF TARGET C is 19,862 bills in 2014, 4,097 in 2016, and three in 2025. Costco’s typed string COSTCO WHOLESALE is 3.8% of the prefix on average; the real mass is COSTCO WHOLESALE CORP, plus CORPORTION.

6. What faculty are paying for

Panjiva, PIERS, and S&P Capital IQ company identifiers are a product because Tables 1–3 are the default. A WRDS pull that already merged Panjiva company IDs is not magically true — proprietary matchers have their own Type I and Type II — but it is not the raw AMS string. This paper does not reverse-engineer Panjiva. It measures the bias of not buying the match: fake births, fake deaths, fake supplier churn, and, if you use prefix instead, a steel mill in the Target panel.

The public substitute that is defensible in a methods paragraph is the one used in the COVID and Plus One papers: inspect the top strings, lock a short US list, drop the hole years, and print the SQL. That is not an entity resolver. It is a way to keep Walmart 2017 and Home Depot 2020 from becoming a de-globalization coefficient. Researchers who need a true firm panel still need IDs. The citation for that sentence can now be this replication rather than a vendor slide.

7. Conclusion

Raw consignee names are measurement error. Walmart INC is missing for four years and then is almost the entire family. Home Depot USA INC has four bills in 2020 and none in 2022–2023. The 2016 names are not the 2025 names. A suffix strip helps Home Depot and does not help Walmart Stores. Prefix helps and then includes Target Steel. 2017 and 2020 remain thin even after grouping; coverage error is real, and it is not an excuse to treat the name zeros as economic. Faculty pay Panjiva for IDs because this is how badly the strings lie. The queries are public.

Appendix. Core queries

Every statement in the snapshot, including name discovery, is on the replication page. The four designs are below.

A. Naive string by year

SELECT toYear(actual_arrival_date) AS year,
       uniqExact(bill_of_lading) AS bols
FROM bols
WHERE actual_arrival_date >= toDate('2014-01-01')
  AND actual_arrival_date <= toDate('2025-12-31')
  AND consignee_name = 'WALMART INC'
GROUP BY year
ORDER BY year

B. US prefix family

SELECT toYear(actual_arrival_date) AS year,
       uniqExact(bill_of_lading) AS bols,
       uniqExact(consignee_name) AS names
FROM bols
WHERE actual_arrival_date >= toDate('2014-01-01')
  AND actual_arrival_date <= toDate('2025-12-31')
  AND (startsWith(upperUTF8(consignee_name), 'WALMART')
    OR startsWith(upperUTF8(consignee_name), 'WAL-MART'))
  AND positionCaseInsensitive(consignee_name, 'CANADA') = 0
  AND positionCaseInsensitive(consignee_name, 'MEXICO') = 0
GROUP BY year
ORDER BY year

C. Locked names (COVID / Plus One set)

SELECT toYear(actual_arrival_date) AS year,
       uniqExact(bill_of_lading) AS bols
FROM bols
WHERE consignee_name IN (
    'WALMART INC',
    'WALMART STORES INC',
    'WALMART GLOBAL LOGISTICS')
  AND actual_arrival_date >= toDate('2014-01-01')
  AND actual_arrival_date <= toDate('2025-12-31')
GROUP BY year
ORDER BY year

D. False supplier exit

WITH
pre AS (
  SELECT DISTINCT upperUTF8(trimBoth(shipper_name)) AS shipper
  FROM bols
  WHERE consignee_name = 'MATTEL INC'
    AND toYear(actual_arrival_date) = 2016
),
exact_post AS (
  SELECT DISTINCT upperUTF8(trimBoth(shipper_name)) AS shipper
  FROM bols
  WHERE consignee_name = 'MATTEL INC'
    AND toYear(actual_arrival_date) = 2024
),
prefix_post AS (
  SELECT DISTINCT upperUTF8(trimBoth(shipper_name)) AS shipper
  FROM bols
  WHERE startsWith(upperUTF8(consignee_name), 'MATTEL')
    AND toYear(actual_arrival_date) = 2024
)
SELECT uniqExact(pre.shipper) AS suppliers_from,
       uniqExactIf(pre.shipper, exact_post.shipper != '') AS stayed_on_naive,
       uniqExactIf(pre.shipper, exact_post.shipper = '' AND prefix_post.shipper != '') AS false_exit
FROM pre
LEFT JOIN exact_post ON pre.shipper = exact_post.shipper
LEFT JOIN prefix_post ON pre.shipper = prefix_post.shipper

Sources

1. Kirchner. 2026. “Name fragmentation as measurement error.” Replication of the ClickHouse snapshot: naive, locked, prefix, contains, legal-norm, held-fixed strings, false-exit queries, and every SQL statement. /research-name-fragmentation. The article body above does not load that file.
2. Kirchner. 2026. “Did COVID permanently raise supplier turnover among large US importers?” Walmart 2017 and Home Depot 2020 appear there as coverage filters. This paper treats those holes as the measurement-error result. /blog/covid-supply-chain.
3. Kirchner. 2026. “Did large US importers replace Chinese factories, or only Chinese ports?” Locked consignee lists and the Walmart 2016–17 name hole as a reason not to take origin-share means at face value. /blog/china-plus-one.
4. Flaaen, Aaron, Flora Haberkorn, Logan Lewis, Anderson Monken, Justin Pierce, Rosemary Rhodes, and Madeleine Yi. 2023. “Bill of lading data in international trade research.” Review of International Economics. What public BOLs can and cannot do relative to Census LFTTD, including the absence of a census firm identifier.
5. S&P Global Market Intelligence. Panjiva / Capital IQ company identifiers on US bills of lading. Proprietary entity resolution sold to universities through WRDS and direct license. This paper does not use or evaluate that matcher.
6. U.S. Customs and Border Protection. Automated Manifest System / ocean bill of lading public filings (19 U.S.C. § 1431; 19 CFR § 103.31).