Tracking Real Time Layoffs with SEC Filings: A Preliminary Investigation

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Tracking Real Time Layoffs with SEC Filings: A Preliminary Investigation Leland D. Crane; Emily Green; Molly Harnish; Will McClennan; Paul E. Soto; Betsy Vrankovich; Jacob Williams 2024-020 Please cite this paper as: Crane, Leland D., Emily Green, Molly Harnish, Will McClennan, Paul E. Soto, Betsy Vrankovich, and Jacob Williams (2024). “Tracking Real Time Layoffs with SEC Filings: A Preliminary Investigation,” Finance and Economics Discussion Series 2024-020. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2024.020. 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.

Tracking Real Time Layoffs with SEC Filings: A Preliminary Investigation* LelandD.Crane EmilyGreen MollyHarnish WillMcClennan PaulE.Soto BetsyVrankovich JacobWilliams† April11,2024 Abstract We explore a new source of data on layoffs: timely 8-K filings with the Securities and andExchangeCommission.Wedevelopmeasuresofboththenumberofreportedlayoff eventsandthenumberofaffectedworkers. Theseseriesarehighlycorrelatedwiththe businesscycleandotherlayoffindicators.Linkingfirm-levelreportedlayoffeventswith WARNnoticessuggeststhat8-KfilingsaresometimesavailablebeforeWARNnotices, and preliminary regression results suggest our layoff series are useful for forecasting. We also document the industry composition of the data and specific areas where the industrysharesdiverge. *WethankTomazCajner,RyanDecker,ChrisKurz,RobertKurtzman,AndersonMonken,andNitishSinha forvaluablefeedback. Opinionsexpressedhereinarethoseoftheauthorsaloneanddonotnecessarilyreflect theviewsoftheFederalReserveSystemortheBoardofGovernors. †AllauthorsareattheFederalReserveBoardofGovernors

1 Introduction Layoffs are among the most closely-watched macroeconomic indicators. Unemployment insurance (UI) claims and the Job Openings and Labor Turnover Survey (JOLTS) layoffs have long been considered early warnings for downturns. More recently, economists have beguntrackingtheWorkerAdjustmentandRetrainingNotification(WARN)systemwhich requires the public announcement of certain layoffs months in advance (Krolikowski and Lunsford,2024). Inthispaperweexploreanalternativetimelylayoffindicatorbasedon8-K filings. Publicly-tradedcompaniesfile8-KswiththeSecuritiesandExchangeCommission(SEC) whentheyneedtodisclosecertaineventstothepublic,andlayoffplansareoftenrecorded in 8-Ks. Some layoffs are recorded under a specific 8-K item number (discussed below) while others are recorded under more general items. The latter require natural language processing to identify; we experiment with sentence embeddings from a workhorse langauge model (BERT) as well as prompting a generative large language model (Llama 2) to identify layoffs. We also explore estimating the quantity of workers laid off by parsing the magnitudesdiscussedinthefilingsandlinkingtoemploymentdatafromCompustat. Theresultingseriesarehighlycorrelatedwiththebusinesscycleandotherlayoffindicators. Thereareonlytworecessionsinthesampleperiod,butthe8-Kseriesclearlycaptures the increase in layoffs for both. Interestingly, the 8-K layoff series have been elevated in 2023;thispatternisnotpresentinotherindicators. Weleaveittofutureworktoexplorethis patterninmoredepth. Wealsopresentpreliminaryevidencethatthe8-Kseriesareusefulforforecastingimportantquantitiessuchastheunemploymentrateandinitialunemploymentinsuranceclaims. These(insample)resultsarerobusttotheinclusionofmanycontrolsincludingWARNnoticedata. The sectoral composition of 8-K announced layoffs is also of interest. We compare the sector shares of 8-K announced layoffs with the composition of publicly-traded firms and 1

theuniverseof(U.S.)firms. Wefindthat—onafirmcountbasis—publicly-tradedfirmsare highlyconcentratedinthemanufacturingsectorascomparedtotheuniverseofpublicand private firms. 8-K layoff events are even more highly concentrated in manufacturing than thedistributionofpublicly-tradedfirmswouldimply. Asignificantpartofthisdivergence isduetothepharmaceuticalindustry: ourevidencesuggeststhatpharmaceuticalfirmsare disproportionatelylikelytobepublicfirms. Publicly-tradedpharmaceuticalfirmsalsotend tobeverysmallandthussubjecttoidiosyncraticshocksthatwouldgeneratelayoffs. Finally, we capitalize on the firm-level nature of the data to link 8-K reported layoffs to the publicly-available WARN Notices, which also record layoffs. We find that—on the linkedsample—neitherdatasourceclearlydominatesintermsoftimeliness. Inmanycases both data sources record the layoff event in the same week, but in a substantial fraction of casesthe8-KsappeartorecordthelayoffeventfourormoreweeksinadvanceoftheWARN notices. WARN notices also arrive weeks earlier than 8-Ks in many cases. We also examine whether the layoffs in announced in the (whole, unlinked) 8-K data would be subject to WARN notice requirements. Layoffs only need to be announced in a WARN notice if theymeetcertainthresholdsbasedonplantsizeandthelayoffsize. Ourevidencesuggests thatmostofthelayoffsinthe8-KdatawouldnotmeettheWARNreportingrequirements, thoughtherearemanycaveatstothiscomparison. Thepaperproceedsasfollows: Section2describestherawdataandtheconstructionof the series. Section 3 presents the series and comparisons to other labor market indicators. Section4explorestheindustrycompositionofthedata,Section5highlightspost-pandemic layoffdynamics,andSection6comparesthedatatoWARNnoticesanddevelopsforecasting results. Section7concludes. 2 Data TheSECrequirespubliccompaniestomakecertainfilingsanddisclosures;forexample,10- K annual reports have been widely studied (Bodnaruk et al. (2015), Sufi (2009), Loughran 2

and McDonald (2011)). Our focus is on form 8-K, or the “current reports.” These forms are filed irregularly, whenever a reportable event occurs. This usually means within four days of a decision or action, though SEC guidance states that firms may wait to disclose layoffsuntiltheemployeesareinformedabouttheplan(SEC,2023). TheSECspecifiesthat different kinds of events be filed under different numbered item codes, which we use to narrowdownthesetofrelevantfilings. Our analysis uses the item codes and the text of the 8-K filings. We begin with the universeofSECfilings,asavailableonEDGAR.1 Fromeachdocumentweextracttheform type, company identifier (CIK), SIC industry code, company name, filing date, period of report,andthetextofthefiling. Wefurthercleanthetextbyextractingindividualitems. To do this we modify and extend the EDGAR Crawler code of Loukas et al. (2021) (designed forreading10-Ks)toparsetheitemsin8-Ks. Themetadataandtheitemizedtextfromeach filingarethenloadedintoadatabase. Layoffsareoftenreportedunderitem2.05,“CostsAssociatedwithExitorDisposalActivities.” This item—introduced in 2004—requires firms to report the disposal of any longlivedassetsortheterminationofemployeesiftheseactionsleadtothefirmincurring“materialcharges”(SEC,2004).2 Thefilingmustincludetheplannedcourseofaction,estimated costs for various types of expenses related to the action, and an estimate of the resulting charge that will lead to future cash expenditures. Item 2.05 can be used to report many types of disposal costs—such as fees for terminating sales or marketing contracts—but in practiceitappearsthatseverancepaymentsarenearlyalwaysamongthereportedcharges.3 As a result filing an 8-K with an item 2.05 is a very good indicator that layoffs will be occurring at the firm, and we include all items 2.05 as layoff events. Item 2.05 is only present inasmallnumberoffilings: wefindthatin20220.38percentof8-Ksincludedanitem2.05 1WedonotactuallyuseEDGAR,insteadweusethedailyfeedofallfilingsthroughtheSEC’svendorAttain. ThefilingsareinXBRL,anformofXML. 2TheitemrequiresreportingterminationsandseverancecostsasdescribedinFASBASCparagraph420-10- 25-4. 3Seethe2006-01-09filingfromToys“R”Usforanexamplewithseveraltypesofcharges. 3

(a total of 258 filings.) Perhaps because of the relatively small number of filings and its focus on labor item 2.05 has received relatively little attention in the finance and accounting literatures,thoughJungetal.(2016)andLaurion(2020)areexceptions. 2.1 Layoffsreportedunderotheritems Itturnsoutthatlayoffsaresometimesreportedunder8-Kitemsotherthan2.05. Inparticular,wefindlayoffannouncementsunderitem7.01anditem8.01. Item7.01(“RegulationFD Disclosure”) relates to Regulation Fair Disclosure, the SEC rule requiring that firms giving material nonpublic information to certain market participants or shareholders also make a publicdisclosureoftheinformation. Item8.01(“OtherEvents”)canbeusedtosatisfyRegulationFD,orreportotherevents“theregistrantdeemsofimportancetosecurityholders” (SEC, 2004). Items 7.01 and 8.01 are much more common than item 2.05: in 2022 they were present in roughly 24 percent and 22 percent of 8-Ks, respectively. However, unlike item 2.05,items7.01and8.01areusedtoreportmanynon-layoffeventssolayoffsareonlyavery smallfractionoftheitems. Filteringoutthenon-layoff-relatedreportsisanempiricalchallenge. Ourinitialreading of the documents suggested that firms frequently use the same terms when announcing layoffs: theword“layoff”oftenappears,asdoes“reducing”/“reduce”/“reduced”followed by“workforce”. Startingwiththesephrases—andtreatingeachsentenceofeachfilingasa separateobservation—wetakeaniterativeapproach: 1. Wecheckthattheexistingstockoflayoff-relatedphrasesdoesnotgeneratemanyfalse positives: false positives occur when sentences match one of the patterns but are not infactaboutlayoffs. 2. We flag all sentences matching the stock of phrases as true layoffs and fit a flexible modeltopredictwhetheroranotasentence isaboutlayoffsasafunctionofthesentencetext. Themodelreturnsaprobabilitythatagivensentenceisaboutlayoffs. 4

3. we examine the set of sentences which are very likely about layoffs but not already flaggedassuch. Thisexaminationsuggestsnewphrasesthatcanmarklayoffs. Atthis pointtheprocessrepeatsandwereturntostep1withanexpandedstockofphrases. TheflexiblemodelwhichpredictslayoffsisbasedonBERTsentenceembeddings(Devlin et al., 2019). The BERT model converts a sentence to a 768-dimensional vector, meant to capture the semantic meaning of the sentence. We feed the vector into a flexible classifier (XGBoost,ChenandGuestrin(2016)). XGBoostisextremelyfastandstable,makingitwellsuitedtoouriterativeapproach.4 Amongthephrasesweidentifyare“reductioninforce”,“workforcereduction”,“eliminatestaffpositions”,and“reduceheadcount”. Anadvantageofthismethodologyisthat— while the iterative ranking process is not particularly transparent—the final classification ruleisjustalistofpatternsthatsentenceseithermatchordon’tmatch. Alltold,thelayoffs fromitems7.01and8.01accountforabout16percentofthetotal,withitem2.05comprising therest. Toavoidanydoublecountingwetreatallfilingsbyafirm(asdefinedbytheSEC’s CIKidentifier)onagivendayasasingleevent.5 2.2 FurtheranalysiswithLLMs Oneconcernisthattheprocedureoutlinedabovewillonlyidentifysentenceswithparticulargrammaticalpatternsaslayoff-related. Sentencesthatuseterminologyorphrasingvery different from the initial patterns we identify may never be ranked as likely layoffs. To address this issue we experimented with using a large language model to identify layoffs missed by the methodology. In particular, we took a subset of sentences that were scored as fairly likely layoffs (but not classified as layoffs) and prompted the Llama 2 model of 4Addingtothechallengesofidentifyinglayoffs,manyfilingsincludeboilerplatedisclaimersaboutforwardlooking statements. These disclaimers often refer to hypothetical future layoffs that could materially affect thevalueofthecompany. Weflagexamplesofthesestatementsasaseparateclasswhenfittingtheclassifier, anticipatingthatitwouldbehardforanaiveclassifiertodistinguishtheseremarksfromstatementsthatare actual,plannedlayoffs. 5Thereareonly8caseswheremultipleitemsareflaggedonasingleday,theseareallcaseswherearelevant item7.01andarelevantitem8.01appeartogether. 5

Touvron et al. (2023) to classify the sentences as layoffs-related or not.6 Llama 2 is a model similarinspirittoChatGPTexceptthatitcanberunlocallyandissmallerinscale. Seethe Appendixforsomeadditionaldetails. This procedure identifies several hundred additional layoffs beyond the roughly 7,000 previously identified. Interestingly, Llama 2 does appear capable of identifying layoffs in sentences thatuse a wide varietyof linguistic devices. However, manualinspection shows a modest but non-negligible fraction of false positives. Given the complexities of handling these cases we do not include the Llama 2-identified layoffs in our series, though we see this as a promising avenue for future work. In any case, the experiment suggests that the iterativerankingproceduredoescapturethelion’sshareoflayoffs,evenifthereisroomfor improvement. 3 Results Figure1showsthemonthlytimeseriesoflayoffsasmeasuredby8-Kfilings. Thesolidblack lineincludesallitems2.05,andtheitems7.01and8.01whichmatchourstringcomparisons. Whileitem2.05accountsformostoftherelevantfilings,theotheritemsconstituteabout16 percentofthetotal. Thetimeseriesfollowsthebusinesscycle,withboththeGreatRecession and the Covid shock clearly visible. The series also show a pronounced increase in early 2023, discussed more below. Interestingly, the other items (red dashed line) spike sharply duringCovid. Overall, while the 8-K series clearly show an increase during Covid the increase is not as large as other layoff indicators. Figure 2 compares the 8-K layoff announcements with initial UI claims and WARN notices. All series are indexed to average 100 in 2006 to make comparisons easier. In addition, we seasonally adjust the 8-K series using X-13 (there is a small degree of seasonality in the raw data, with more layoff announcements coming in 6We use the 7B chat-tuned version from https://huggingface.co/meta-llama/Llama-2-7b-chat-hf, run locally. 6

June and October than other months.) The first panel shows clearly that UI claims and WARN notices rose far more during Covid than 8-K layoff announcements. The SEC did allow for some leniency in filing early in the pandemic but a more important factor might be the unique nature of the layoffs. The pandemic saw a large wave of temporary layoffs, which likely include little or no severance payments. These kinds of layoffs may not be reportedasoftentotheSEC,reducingthecountsof8-Ks. Thesecondpanelshowsthesame series, but suppresses the 2020 data for WARN and UI claims so the dynamics outside of the pandemic are easier to see. All three series move in lockstep at the onset of the Great Recession. Towards the end of the Great Recession, WARN and 8-Ks move down sharply, while initial UI claims take many years to fall. The dynamics around the Great Recession suggest that 8-K filings may have been a timely indicator of both the business cycle peak andthetrough. While counts of layoff events can be informative, it is also useful to track the number of workers affected. This is not trivial in the 8-K data as firms describe the magnitude of layoffs in different ways, though when there are specific numbers it is generally either a count of workers, or a percentage of the firm’s workforce. The procedure searches for a quantity appearing in proximity to a quantifier (“reduc-”, “decreas-”, “lower-”, “eliminat- ”, or “terminat-”) and a word related to layoffs (“layoff”, “headcount”, or “workforce”). For example, if the 8-K mentioned either “We decreased our headcount by 10” or “The firm experienced a workforce reduction of 5%”, our algorithm would return 10 and 5%, respectively. If a percentage is found, we multiply the percentage by the lagged value of total employee counts, obtained from Compustat. If a cardinal amount is mentioned, we use this value forthe numberof layoffs. If multiplequantities are captured within the8-K, we sum the values to obtain the number of layoffs at the firm-filing level. We seasonally adjustthisseriesaswell. Figure3showsthethreemonthtrailingmovingaverageof8-Kworkerscounts(redline), as well as the other relevant series (which are not averaged.) Again, all are normalized by 7

their2006averagestofacilitatecomparisons. Weuseathreemonthmovingaveragebecause theworkercountcanbeextremelyvolatilemonthtomonth. Whilethethreemonthaverage still displays significant noise, the business cycles are also readily apparent. Interestingly, the 8-K worker count rose by far more in the Great Recession than any of the other series. During Covid it also increased by a factor comparable to initial UI claims, which (as discussed above) were inflated by temporary layoffs. While the 8-K worker count series may benoisy,italsoappearstoreactstronglytothebusinesscycle. 4 Industry Composition In this section we examine how the layoff announcements are distributed across sectors and specific industries. SEC filings include the SIC industry code of the filing firm. As other agencies use NAICS codes instead, we convert the SIC codes to NAICS to facilitate comparisontootherdata. Wherethereisnotaone-to-onecorrespondenceweassignequal weight to each successor NAICS code. Our main comparisons are to firm count shares fromtheCensus’sStatisticsofU.S.Businesses(SUSB)data,establishmentcountsharesfrom the Business Dynamics Statistics (BDS), and firm count shares from Compustat. SUSB and BDScoveressentiallytheuniverseofemployersintheU.S.(bothprivateandpublic),while CompustatonlycoversfirmsthatarepubliclytradedandthusregulatedbytheSEC.These comparisonscomplementtheexistingliteratureondifferencesbetweenthepubliclytraded universeandcomprehensiveadministrativedata.7 Figure4showsthesharesoffirms,establishments,andlayoffeventsacrossNAICSsectors in 2015. We focus on 2015 because this is the most recent year for which SUSB has statisticsbasedonfirmindustry,ratherthanestablishmentindustry,makingitmostcomparabletothefirm-levelindustrycodesinCompustatandtheSECdata. Several things are worth noting. First, the SUSB firm shares and the BDS establishmentshares(darkmagentaandpurple,respectively)arebroadlysimilar. Thissuggeststhat 7ForrecentworkseeTito(2019),DeckerandWilliams(2023),andFlynnandGhent(2023). 8

the establishment distribution across industries is a reasonable proxy for the firm distribution. Second,somesectorsareoverrepresentedinthepublicly-tradeduniverse. Information (NAICS 51), finance and insurance (NAICS 52), and especially manufacturing (NAICS 31- 33) all have higher firm shares in Compustat as compared to SUSB. Third, layoff events from SEC filings are even more concentrated in manufacturing than the Compustat firm counts. Notethatsimilaranalysesforotheryears(notshown)suggestthesebroadpatterns aregenerallystableovertime. The manufacturing sector shares in Figure 4 are particularly striking. Figure 5 focuses onthe3-digitsubsectorswithinmanufacturing,butstillplotstheirshareoftotalbusinesses orlayoffs(i.e. nottheirshareofmanufacturingbusinesses/layoffs). Nearlyeverysubsector isoverrepresentedinthepubliclytradeddata,butchemicals(NAICS325)andcomputers& electronics (NAICS 334) stick out as being particularly overrepresented in Compustat and the SEC filings. Note that while layoffs are very high in NAICS 334 in 2015, this pattern is notconsistentacrossyears. Interestingly,withinthechemicalssubsector,itispharmaceuticalmanufacturing(NAICS3254)thataccountsformuchofthegapbetweentheSUSB/BDS and Compustat/8-Ks. To put it in perspective, according to SUSB about 4% of firms were inmanufacturingin2015,andabout0.03%wereinpharmaceuticalmanufacturing. According to Compustat, about 37% of publicly-traded firms were in manufacturing, and about 11% of publicly-traded firms were in pharmaceutical manufacturing. The latter amounts to overrepresentation by a factor of more than 300. It seems that some factors specific to the pharmaceutical industry—perhaps the regulation of the products or the need to fund research—leadtomorefirmsinthatindustrygoingpublic. Wefocusonrepresentativenessintermsoffirmcounts, since(unweighted)firmcounts arethebestpointofcomparisonforlayoffevents. Itisworthnotingthatmostofthepatterns describedabovealsoholdforemploymentsharestoo,thoughtheoverrepresentationissues aresomewhatmoremoderate. Partofthereasonisthatwhilepharmaceuticalfirmsappear to go public disproportionately, public pharmaceutical firms are also very small: median 9

employment is only about 50 workers. Thus they tend to contribute less to employmentweightedstatistics, thoughtheirsmallsizemayalsomakethemmorevulnerabletolayoffinducing shocks. The large number of very small firms manufacturing may help explain whythesectorhassomanylayoffsreportedin8-Ks. Reweightingthelayoffseriesisleftfor futurework,thoughbelowwetakeastepinthatdirectionbydroppingsometheindustries thatmaybethemostidiosyncratic. 5 Post-Pandemic Dynamics Returning to Figure 1, the 8-K derived layoffs series has a notable peak in early 2023, a period when technology sector layoffs and recession fears were making headlines. The 8- K layoffs series also remained elevated throughout 2023, relative to its pre-pandemic level (Figure 3 shows similar patterns for the count of 8-K laid off workers.) This is interesting because other layoffs series did not change much over this period. Table 1 shows the percentage change in various series relative to their 2019 average. The first column covers 2022Q4 through 2023Q1—the period when the 8-K series spiked. The second row covers theremainderof2023. 8-Klayoffeventsrose146percentin2022Q4-2023Q1relativetotheir 2019 level, and remained 82 percent above their 2019 level for the rest of 2023. In contrast, initial UI claims, JOLTS-reported layoffs, and WARN notices (discussed more below) were allessentiallyflat. Wecaninvestigatethispatternattheindustrylevel,usingtheSICindustrycodesthatare partoftheforms(hereweuseSICcodesinsteadofNAICScodesbecausewearenotlinking to other data sources, and we avoid the complications of bridging codes.) It turns out that muchofthe2023risein8-Klayoffscanbeaccountedforbytwoindustrygroups: computer programming & data processing (SIC 737) and drugs (SIC 283). The former covers many high-tech companies, and the latter corresponds to the pharmaceutical industry discussed earlier. Theseindustrygroupsgenerallyaccountforalargenumberoflayoffs,butalsosaw particularlylargeincreases(inpercentageterms)inlate2022andearly2023. Takentogether 10

these industry groups accounted for more than 65 percent of the increase in 8-K layoffs. However,droppingtheseindustriesandredoingthecalculations(row2ofTable1),wefind thatthepurgedseriesstillrisesconsiderablyinlate2022and2023,apatternatoddswiththe other indicators. We leave this for future investigation, but note that public companies are different from the universe of firms along many dimensions, including industry (as noted above),sizeandage. 6 Comparison with WARN Notices The Worker Adjustment and Retraining Notification (WARN) Act requires that firms with 100 or more workers notify their employees—and the public—in advance of certain layoff events. Notification must be given if (1) a plant is closing that would displace 50 or more workers, (2) 500 or more workers are to be laid off from an establishment, or (3) 50-499 workers are to be laid off and the total is at least 33 percent of the location’s workforce, for firms with more than 100 workers. The law requires that notification be given 60 days beforethelayoffsoccur, thoughsomestateshaveadditionalnotificationrequirements. While WARN notices cover many layoffs, some significant layoffs may not trigger a notice. For example,Okta’sFebruary2023layoffof300workersonlyamountedtofivepercentoftheir workforce.8 Asaresult, WARNnoticesmaybebetteratcapturingplantclosurethanother layoffs.9 There are clear advantages and disadvantages to both WARN and the SEC filings. Importantly,WARNcoversbothprivateandpubliccompanies,andisonlytriggeredbylayoffs occurring the U.S. In contrast, layoffs only show up in the SEC filings if the firm is public andifthelayoffsmeetthebarformaterialdisclosure. Theyalsoincludestaffinothercountries. However, SEC filings can cover layoffs that fall below the employment threshold for WARN. 8Seehere. 9SeethediscussioninKrolikowskiandLunsford(2024)andGAO(2019). 11

Wecancomparethesizeoflayoffsannouncedin8-KstotheWARNreportingthresholds togetabettersenseofthelikelyoverlapinfilings. Firmsoftenreportthesizeofthelayoffin the8-K,eitherasapercentofemploymentorasthenumberofworkersaffected. The8-Kcan be linked to information on firm employment in Compustat (which is ultimately derived from other SEC filings.) With the firm’s employment level and the layoff as a percent of employmentinhand,wecanjudgeiftheeventsatisfiestheWARNact’sthresholds. There are many caveats to this exercise. In addition to the coverage issues mentioned aboveCompustatemploymentfiguresareonlyavailableannuallyandarethoughttohave some degree of measurement error. Also, WARN notices relate to employment and layoffs at a single establishment whereas SEC filings are for the firm as a whole. Finally the extractionoflayoffquantitiesfromthe8-KsreliesonNLPtechniquesthatcanintroduceerror. Nonetheless,webelieveitisinformativetoexaminetheresults. Figure6showsthedensityoflayoffsannouncedin8-Ks,ingrayscale. Thedistributionis centeredaroundfirmswith1,000-2,000workers,withlayoffsoflessthan10percent. Thered regionwithdashedbordersshowstheareawhereWARNnoticeswouldberequired.10 Itis evidentthatmost8-KannouncedlayoffswouldnotbepickedupbytheWARNsystem. In thissampleonlyabout23percentofthelayoffsmeettheWARNfilingthreshold(assuming thatnoplantclosingswereinvolved.) ThisisconsistentwiththebeliefthatWARNisbetter atcapturingplantclosingsthanothermasslayoffevents. Thoughnodiscontinuityisvisible ontheWARNfilingthreshold,itisanopenquestionwhetherfirmsstrategicallychoosethe leveloflayoffstoavoidthemandateddelaytheWARNactrequires. 6.1 Firm-levellinking WARNnoticesareusefulinpartbecausetheycangivesignalwellinadvanceoftheactual layoffs: atleast60days. Thetimingof8-Kannouncementsrelativetothedateoftheactual 10TherearefourdistinctregionsintermsofthebindingconditionsforWARNnotices.Betweenzeroand100 workers,nonoticesarerequired(exceptforplantclosings).Between100and150workers,alllayoffsofatleast 50workersmustbereported.Between150and1,500workers,alllayoffsabove33percentofemploymentmust bereported.Finally,above1,500workersalllayoffsabove500workersmustbereported. 12

layoffs is less clear. The SEC’s rules generally require that an 8-K be filed within four business days of the triggering event. Further, when 8-Ks are filed to satisfy Regulation FD the rules are even stricter, requiring simultaneous disclosure or at most a lag of one business day.11 While the firm’s commitment to a restructuring would generally be considered the triggeringeventforafiling,thereisanexception,inthatfirmsmaywaittofileanitem2.05 untiltheaffectedemployeesarenotified.12 Thismeansthatfirmsmightbeabletowaittofile the 8-K until they tell the employees, which would presumably be at the same time as any WARNnoticeisfiled. Inpractice,itisunclearhowmanyfirmswaitthislong. Anecdotally, some8-Ksmakeitclearthatthelaidoffemployeesarebeingnotifiedsimultaneously,while manyothersaremorevagueaboutthetiming. Othersexplicitlystatethatthespecificindividuals affected have not yet been determined.13 This suggests that layoffs are sometimes announcedin8-KsbeforecorrespondingWARNnoticesarefiled,oratleastsimultaneously. We test this hypothesis by linking WARN notices to 8-K layoff announcements. Both datasetsrecordthecompany’sname,whichweusetobuildthelink. Thematchingisfuzzy, butourgoalisonlytogetasubsetoffirmswhereweareconfidentofthematch. Wenormalize the names (standardizing case and removing common suffixes like “Inc.”), and require thatthefirstseveralcharactersmatchexactly. Inadditionwerequirethatthefirmonlyhave asinglelayoffeventreportedinthecalendaryeartoavoidlinkingdistinctlayoffs. The resulting dataset has 285 linked layoffs that appear in both the WARN notices and the8-Kfilings. Figure7showsahistogramofthedateofthe8-Kfiling,relativetothedateof theWARNnotice,inweeks. Themassaroundzeroshowsthatthelayoffisoftenannounced in both datasets in the same week: roughly 22 percent of the linked events are reported in thesameweekinbothdatasets. Interestingly,thereissignificantvarianceaswell. About25 percent of linked layoffs appear in the 8-Ks four weeks or more in advance of the WARN notification. On the other end of the spectrum, about 18 percent of the WARN notices are 11Seehttps://www.ecfr.gov/current/title-17/chapter-II/part-243 12SeeQuestion109.01inSEC(2023). 13See,forexample,the2023-01-03filingfromPegasystemsInc 13

morethanfourweeksaheadofthe8-Kfilings. Theseresultssuggestthatneitherdatasource clearlydominatesintermsoftimeliness. It is understandable that 8-Ks would sometimes be filed well in advance of WARN notices,forexample,afirmmaydeterminethatlayoffsaregoingtohappenwithinsixmonths, butmaynothavefinalizedtheexactnumberandlocationofthelayoffs. Inthiscasethefirm would file an 8-K immediately, but can wait to file a WARN notice, so long as they file 60 days before the layoffs. Consistent with this story, manually inspecting the 8-Ks filed well before the linked WARN notices shows that firms often expect the process to take several fiscalquarters,morethanenoughtimetowaitseveralweeksandthenfileWARNnotices. ItislessclearwhatishappeningwhentheWARNnoticespredatethe8-Kssignificantly. After all, if the decision to do layoffs was a material event it should have been reported promptly. Examining the 8-Ks manually shows that a mix of factors are at play. Some of these cases are from the early pandemic, when the SEC allowed additional time to make filings. Inothercases,thefirmisinvolvedinacomplicatedrestructuring,soitappearsthat separatelayoffs/plantclosuresarebeingreportedineachdatasource. 6.2 Regressions Inthissectionweprovidepreliminaryquantitativeevidenceonthepredictivepowerofthe 8-K data. Our exercise follows the VAR specification of Krolikowski and Lunsford (2024) (which in turn is based on Barnichon and Nekarda (2012)). Specifically, we regress various indicatorsonthe8-Kseriesalongwithcontrols. Thedependentvariablesofinterestarethe change in the log of the unemployment rate ( ∆ lnur ), the monthly log level of initial UI t claims (lnuic ), and the log job separation rate lns , which is calculated as in Krolikowski t t and Lunsford (2024). The controls include two lags each of these variables, plus various combinations of the (change in the) 8-K series and the (change in the) WARN notice series. Weruntheregressionsfortheperiod2006-2019. The first four columns of Table 2 have the log change in unemployment as the depen- 14

dent variable, ∆ lnur t . Both the 8-K layoff event counts ( ∆ 8Kevents t−1 , ∆ 8Kevents t−2 ) and ∆ ∆ the laid off work counts ( 8Kworkers t−1 , 8Kworkers t−2 ) have statistically significant relationships with the dependent variable. These relationships appear weakened when we include the WARN series as well (columns three and four), as may be expected. However, ∆ the second lag of 8Kevents remains highly statistically significant event when the WARN data are included. Broadly similar patterns are apparent for the other dependent variable, though the association with 8-K seems weakest for the log separation rate lns . Both for t ∆ lnur andlnuic thesecondlagof ∆ 8Keventsseemstohavethestatisticallystrongestand t t mostrobustrelationship. We take these regressions as suggestive evidence that the 8-K data has forecasting signal for important labor market quantities, complementing the existing stock of indicators. Important future tasks include testing out-of-sample performance and performance over differentsampleperiods. 7 Conclusion SEC filings contain a wealth of information, much of it reported in a timely fashion. We have focused on a particular dimension, the announcement of layoffs in 8-K filings. These announcements track UI claims and WARN notices tightly during the Great Recession, though their behavior during and after the pandemic has deviated from other indicators. While neither WARN notices nor 8-Ks dominate with respect to timeliness, we find suggestive evidence that 8-Ks record a significant number of layoff events that would not be pickedupviaWARNnotices. Wealsofindsuggestiveevidencesthethe8-Kseriesareusefulforforecasting, evenwhenWARNdataareincluded. Topicsforfutureresearchinclude better understanding the reason 8-K layoff events did not rise more during the pandemic, exploring why the series has been elevated in 2023, and evaluating the utility of series for forecasting/nowcastingrecessionsasinBergeetal.(2016). Inongoingresearchweexplore theroleofseverancepaymentsinexplainingwhetherornotlayoffsarereportedin8-Ks. 15

The richness of the data allows for interesting comparisons as well as other topics for future research. Linking to WARN notices lets us better understand what kinds of layoffs arelikelycapturedbyeachdatasources,andtherelativetimingoftheannouncements. The industry patterns suggest some unusual dynamics in the pharmaceutical industry, which warrant further investigation. Finally, large language models have proven useful even in ourlimitedexperimentation,andmaybeusefulforadditionalanalysis. 16

References Barnichon,RegisandChristopherNekarda,“TheInsandOutsofForecastingUnemployment: Using Labor Force Flows to Forecast the Labor Market,” Brookings Papers on EconomicActivity,092012,2012,83–131. Berge,TravisJ.,NitishR.Sinha,andMichaelSmolyansky,“WhichMarketIndicatorsBest ForecastRecessions?,”FEDSNotes2016-08-02,BoardofGovernorsoftheFederalReserve System(U.S.)August2016. Bodnaruk, Andriy, Tim Loughran, andBillMcDonald, “Using10-K TexttoGaugeFinancialConstraints,”TheJournalofFinancialandQuantitativeAnalysis,2015,50(4),623–646. Chen, Tianqi and Carlos Guestrin, “XGBoost: A Scalable Tree Boosting System,” in “Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining” KDD ’16 Association for Computing Machinery New York, NY, USA 2016,pp.785–794. CompustatNorthAmerica.S&PGlobalMarketIntelligence,CapitalIQPlatform,WhartonResearchDataServices,http://wrds.wharton.upenn.edu/. Decker, Ryan A. and Jacob Williams, “A Note on Industry Concentration Measurement,” FEDSNotes2023-02-03,BoardofGovernorsoftheFederalReserveSystem(U.S.)February 2023. Devlin,Jacob,Ming-WeiChang,KentonLee,andKristinaToutanova,“BERT:Pre-training ofDeepBidirectionalTransformersforLanguageUnderstanding,”2019. Flynn, Sean J. and Andra Ghent, “Does Main Street Benefit from What Benefits Wall Street?,”JournalofFinancialandQuantitativeAnalysis,2023,pp.1–37. 17

GAO,“TheWorkerAdjustmentAndRetrainingNotificationAct: RevisingtheActandEducationalMaterialsCouldClarifyEmployerResponsibilitiesandEmployeeRights,”GAO- 03-10032019. Jung, Boochun, Byungki Kim, Woo-Jong Lee, and Choong-Yuel Yoo, “Corporate Layoffs andAccountingConservatism,”WorkingPaper2016-001,KAISTCollegeofBusinessJanuary2016. Krolikowski, Pawel M. and Kurt G. Lunsford, “Advance Layoff Notices and Aggregate JobLoss,”JournalofAppliedEconometrics,2024,n/a(n/a). Laurion, Henry, “Implications of Non-GAAP Earnings for Real Activities and Accounting Choices,”JournalofAccountingandEconomics,2020,70(1),101333. Loughran,TimandBillMcDonald,“WhenIsaLiabilityNotaLiability? TextualAnalysis, Dictionaries,and10-Ks,”TheJournalofFinance,2011,66(1),35–65. Loukas, Lefteris, Manos Fergadiotis, Ion Androutsopoulos, and Prodromos Malakasiotis, “EDGAR-CORPUS: Billions of Tokens Make The World Go Round,” in Udo Hahn, Veronique Hoste, and Amanda Stent, eds., Proceedings of the Third Workshop on Economics and Natural Language Processing, Association for Computational Linguistics Punta Cana, DominicanRepublicNovember2021,pp.13–18. SEC, “Additional Form 8-K Disclosure Requirements and Acceleration of Filing Date,” 69 FR155932004. , “Exchange Act Form 8-K,” https://www.sec.gov/divisions/corpfin/guidance/8kinterp.htm2023. Sufi,Amir,“BankLinesofCreditinCorporateFinance: AnEmpiricalAnalysis,”TheReview ofFinancialStudies,2009,22(3),1057–1088. 18

Tito, Maria D., “Trade in Goods and Services: Measuring Domestic and Export Flows in Buyer-SupplierData,”FEDSNotes2019-06-26,BoardofGovernorsoftheFederalReserve System(U.S.)June2019. Touvron, Hugo, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami,NamanGoyal,AnthonyHartshorn,SagharHosseini,RuiHou,HakanInan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, YinghaiLu,YuningMao,XavierMartinet,TodorMihaylov,PushkarMishra,IgorMolybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom, “Llama 2: Open FoundationandFine-TunedChatModels,”2023. 19

Percentchangerelativeto2019 2022Q4-2023Q1 2023Q2-2023Q4 8-KLayoffsEvents 145.56 81.72 8-Ks,ex. SICcodes283and737 68.78 41.14 InitialUIClaims −2.21 3.77 JOLTSLayoffs −9.48 −10.97 WARNNotices −10.74 7.45 Note:Allentriesarethepercentdifferencebetweentheaverageforthespecificedperiodandtheaveragefor2019. Source: BLS, Department of Labor, SEC, Krolikowski and Lunsford (2024), authors’ calculations Table1: LayoffIndicators,ChangeRelativeto2019 20

tciunl tsnl trunl∆ )51( )41( )31( )21( )11( )01( )9( )8( )7( )6( )5( )4( )3( )2( )1( 910.0 540.0 380.0 ∗∗231.0 520.0 ∗∗830.0 1−tstneveK8∆ )430.0( )230.0( )360.0( )250.0( )020.0( )710.0( ∗∗∗280.0 ∗∗∗001.0 740.0 ∗180.0 ∗∗∗650.0 ∗∗∗560.0 2−tstneveK8∆ )130.0( )920.0( )250.0( )640.0( )810.0( )510.0( 780.0 ∗∗321.0 340.0− 460.0 430.0 ∗∗∗360.0 1−tsrekrowK8∆ )060.0( )850.0( )380.0( )480.0( )420.0( )020.0( ∗201.0 ∗∗921.0 160.0 ∗041.0 330.0 ∗∗450.0 2−tsrekrowK8∆ )450.0( )350.0( )670.0( )170.0( )320.0( )120.0( 071.0 791.0 ∗∗832.0 ∗∗∗155.0 ∗∗184.0 ∗∗∗735.0 ∗∗∗141.0 ∗∗021.0 ∗∗∗761.0 1−tNR(cid:100)AW∆ )801.0( )121.0( )611.0( )971.0( )581.0( )661.0( )350.0( )250.0( )940.0( 690.0 480.0 061.0 122.0 780.0 722.0 960.0 230.0 ∗190.0 2−tNR(cid:100)AW∆ )490.0( )001.0( )201.0( )371.0( )691.0( )361.0( )250.0( )450.0( )150.0( 789.0 789.0 689.0 689.0 789.0 318.0 618.0 118.0 008.0 708.0 264.0 984.0 654.0 934.0 574.0 2R 589.0 589.0 589.0 589.0 589.0 697.0 697.0 697.0 697.0 697.0 514.0 514.0 514.0 514.0 514.0 2RledomenilesaB nekateraselbairavtnednepedllA.9102-6002:doirepelpmaS.sisehtnerapnisrorredradnatstsuboR.01.0<p*,50.0<p**,10.0<p***:etoN .slortnocsatciunldna,tsnl,trunl∆fosgalowtedulcnisnoitacfiicepsllA.snoitamrofsnartriehtesudna)4202(drofsnuLdnaikswokilorKmorf .slortnocesehtylnognisunoissergerehtfo2Rehtsevig”2RenilesaB“ secitoNNRAWdnaseireSK-8 :snoissergeRgnitsaceroF :2elbaT 21

Count of Layoff Events 150 All layoffs 150 Item 2.05 only Items 7.01 and 8.01 100 100 50 50 0 0 2007 2011 2015 2019 2023 Note:Seriesarenotseasonallyadjusted Source:SEC,KrolikowskiandLunsford(2024),authors’calculations Figure1: 8-KLayoffEvents 22

Indexed Layoffs Indexed Layoffs, Suppressing 2020 for UI Claims and WARN Index, 2006=100 Index, 2006=100 1500 1500 300 300 8−K Layoff Events Initial UI Claims WARN Notices 1000 1000 200 200 500 500 100 100 0 2007 2011 2015 2019 20230 0 2007 2011 2015 2019 20230 Note:Allseriesareseasonallyadjusted. Allseriesaredividedthroughbytheir2006averageandmultipliedby 100.“WARNNotices”isthemeasureoflayoffsimpliedbythe“WARNfactor”fromKrolikowskiandLunsford (2024). Source:DepartmentofLabor,SEC,KrolikowskiandLunsford(2024),authors’calculations Figure2: Comparisonof8-KLayoffs,UIClaims,andWARNNotices Indexed Layoffs Index, 2006=100 1500 1500 8−K Layoff Events Initial UI claims WARN Notices 8−K Workers Laid Off 1000 1000 500 500 0 0 2007 2011 2015 2019 2023 Note:Allseriesareseasonallyadjusted. Allseriesaredividedthroughbytheir2006averageandmultipliedby 100.8-KWorkersLaidOffisatrailingthreemonthmovingaverage. Source:DepartmentofLabor,S&P,SEC,KrolikowskiandLunsford(2024),authors’calculations Figure3: 8-KWorkersLaidOff 23

Ag., Forestry, Fishing (11) Mining (21) Utilities (22) Construction (23) Manufacturing (31−33) Wholesale Trade (42) Retail Trade (44−45) Transportation and Warehousing (48−49) Information (51) Finance and Insurance (52) Real Estate and Rental and Leasing (53) Prof. and Tech. Services (54) Management (55) Admin., Support and Waste Services (56) Educational Services (61) Health Care and Social Assistance (62) SUSB Firm Share Arts, Entertainment, and Recreation (71) BDS Estab share Accommodation and Food Services (72) Compustat Firm Share Other Services (ex. Pub. Admin.) (81) Form 8−K Layoffs Share 0 20 40 60 Note: Barsshowthepercentofthetotalrelevantquantity(businesscountorlayoffeventcount)inthatNAICS sector.SUSBfirm-leveldataisclassifiedbytheNAICSofthefirmasawhole. Source:CensusBureau,S&P,SEC,authors’calculations. Figure4: SharesofBusinessesorLayoffEventsAcrossSectors,2015 24

Food (311) SUSB Firm Share Beverage and Tobacco Product (312) BDS Estab share Textile Mills (313) Compustat Firm Share Textile Product Mills (314) Form 8−K Layoffs Share Apparel (315) Leather and Allied Product (316) Wood Product (321) Paper (322) Printing and Related Support Activities (323) Petroleum and Coal Products (324) Chemical (325) Plastics and Rubber Products (326) Nonmetallic Mineral Product (327) Primary Metal (331) Fabricated Metal Product (332) Machinery (333) Computer and Electronic Product (334) Electrical Equipment (335) Transportation Equipment (336) Furniture and Related Product (337) Miscellaneous (339) 0 5 10 15 20 Note:Barsshowthepercentofthetotaleconomy-wide(notmanufacturing-specific)relevantquantity(business countorlayoffeventcount)inthatthree-digitindustry. TheSUSBfirm-leveldataisclassifiedbytheNAICSof thefirmasawhole. Source:CensusBureau,S&P,SEC,authors’calculations. Figure5: ManufacturingIndustrySharesofTotalBusinessesorLayoffs,2015 25

.8 .6 .4 .2 0 tnemyolpme fo noitcarf ,ezis ffoyaL Density of layoffs WARN Notice Required 10 50 250 1,250 6,250 31,250 Firm employment, log scale Note:Xaxisisthecountofworkersatthefirmonalogscale.Yaxisisthemagnitudeofthelayoffasafactionof baseemployment(i.e.,thevalueonthexaxis). Source:SEC,S&P,authors’calculations. Figure6: Densityof8-KReportedLayoffs 26

.2 .15 .1 .05 0 snoitavresbo fo noitcarF −32 −16 −8 −4 0 4 8 16 32 8−K date relative to WARN notice date, weeks Note:Histogramplotsthedistributionofthe8-KfilingdateminustheWARNnoticefilingdate.Eachbarisone week. Source:SEC,https://layoffdata.com/,authors’calculations. Figure7: Histogramof8-KDateRelativetoWARNDate 27

A Llama 2 LLM Details Weusingthe7billionparameterversionofLlama2,specificallythemodelmeta-llama/Llama-2-7b-chat-hf fromtheHuggingfaceHub. Weusethefollowingprompt: <s>[INST] <<SYS>> You are a world-class PhD economist, knowledgable on business topics. <</SYS>> We are interested whether a firm is announcing layoffs. Layoffs include any firings or involuntary elimination of existing employee positions. We do not consider temporary furloughs, hiring freezes, or the departure of executives to be layoffs, unless they are accompanied by additional firings. Consider the following excerpt a firm’s SEC filings: Text: "{x}" Based on the text, is the firm definitely announcing layoffs or otherwise firing existing employees, according to the definition above? Briefly explain your reasoning step by step. Towards the end of your answer include either the exact phrase "Therefore, yes, the firm is laying people off." or "Therefore, no, the passage does not confirm the firm is laying people off." [/INST] where x is a sentence drawn from an 8-K filing. Then we parse the responses for the desired phrases. The hope is that by asking the LLM to reason first and deliver the final answerlastitwillbenefitfromchainofthoughtreasoning. Weleaveitforfutureworktousemoresophisticatedmethods,suchasassessingthetokenprobabilitiesdirectlyinsteadofrelyingontheLLMtoproduceexactstringsformatching. 28