Spread Too Thin: The Impact of Lean Inventories

Board of Governors of the Federal Reserve System International Finance Discussion Papers ISSN 1073-2500 (Print) ISSN 2767-4509 (Online) Number 1342 April 2022 Spread Too Thin: The Impact of Lean Inventories Julio L. Ortiz Please cite this paper as: Ortiz, Julio L. (2022). “Spread Too Thin: The Impact of Lean Inventories,” International Finance Discussion Papers 1342. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/IFDP.2022.1342. NOTE: International Finance Discussion Papers (IFDPs) 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 International Finance Discussion Papers Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers. Recent IFDPs are available on the Web at www.federalreserve.gov/pubs/ifdp/. This paper can be downloaded without charge from the Social Science Research Network electronic library at www.ssrn.com.

Spread Too Thin: The Impact of Lean Inventories Julio L. Ortiz ˚ Federal Reserve Board April 2022 Abstract Widespreadadoptionofjust-in-time(JIT)productionhasreducedinventoryholdings. This paper finds that JIT creates a trade-off between firm profitability and vulnerabilitytolargeshocks. Empirically,JITadoptersexperiencehighersalesandless volatilitywhilealsoexhibitingheightenedcyclicalityandsensitivitytonaturaldisasters. Iexplainthesefactsinastructurallyestimatedgeneralequilibriummodelwhere firmscanadoptJIT.Relativetoano-JITeconomy,theestimatedmodelimpliesa1.3% increaseinfirmvalue. Atthesametime, anunanticipatedshockresultsinaroughly 15%deeperoutputcontraction. Thisoccursbecausefirms“stockout”orhoardmaterials. Keywords: Inventoryinvestment. Firmdynamics. Just-in-timeproduction. JELCodes: D25,E22,G30 ˚IwouldliketothankStephenTerry,AdamGuren,andPascualRestrepofortheirmanyinsights,suggestions, and encouragement which greatly shaped this paper. I would also like to thank discussants George Alessandria, Ryan Charhour, and Jun Nie, as well as Pablo Cuba-Borda, Andrea DeMichelis, Sebastian Graves, Matteo Iacoviello, Nils Lehr, Hyunseung Oh, Frank Warnock, and the participants at many seminars and conferences. Lastly, I am grateful to William Wempe and Xiaodan Gao for sharing their data on JITadoption. Theviewsexpressedinthispaperaresolelytheresponsibilityoftheauthorandshouldnotbe interpretedasreflectingtheviewsoftheBoardofGovernorsoftheFederalReserveSystemorofanyother personassociatedwiththeFederalReserveSystem. Email: julio.l.ortiz@frb.gov.

1 Introduction IntheUnitedStates,70%ofmanufacturersusejust-in-time(JIT)production,aleaninventorymanagement philosophy that aims to minimize the time between orders.1 JIT grew in popularity beginning in the early 1980s as firms adopted technologies and practices that allowed them to cut costs associated with managing large material purchases and storing idle stocks. Instead these firms committed to placing smaller and more frequent orders from their suppliers.2 Consequently, lean inventory management is believed to have contributed to the 35% reduction in the aggregate inventory-to-sales ratio between 1980 and 2018.3 Moreover, many commentators and academics have pointed to leaner inventory management practices as one of the reasons for the decline in volatilityofseveralmacroeconomicaggregatesthattookplacebeginninginthe1980s(McConnell andPerez-Quiros,2000;BlanchardandSimon,2001;Kahnetal.,2002). Thispaperoffersanewperspectiveontheroleofleaninventoriesindrivingaggregatefluctuations,findingthatitcancreatemacrofragilityinthefaceofunexpectedshockssuchasCOVID-19. I documentevidenceofatrade-offfromanoveldatasetofJITfirmsandquantitativelyassesstherole that lean production plays at the micro and macro levels in a structurally estimated heterogeneous firmsmodel. I begin by developing an indicator of the adoption of JIT for approximately 200 publicly listed manufacturing firms. Using narrative records from SEC filings and historical archives, I construct an adoption dummy that measures the year when the firm adopted JIT. I then link the measure of JIT to firm-level balance sheet data and document stylized facts relating to JIT prouducers. First, I show that JIT adoption is associated with a 13% decrease in inventory-to-sales ratios and a 9% increase in sales. In addition, JIT firms experience a 7% decline in employment and sales growth volatility. These empirical results, though not causal, are consistent with positive selection into adoptionwhichsubsequentlyyieldsfirm-levelefficiencygainsasinmymodel. 1In2015,theCompensationDataManufacturing&DistributionSurveyfoundthat71%ofsurveyedfirmsemploy leanmanufacturing. Similarly,in2007,theIndustryWeek/MPICensusofManufacturersfoundthat70%ofrespondents hadimplementedleanmanufacturing. 2Ohno(1988)providesadetailedhistoryofJITwhichfirststartedwithToyota’sKanbansystem. 3U.S. Bureau of Economic Analysis, Ratios of nonfarm inventories to final sales of domestic business [A812RC2Q027SBEA], retrieved from FRED, Federal Reserve Bank of St. Louis; https://fred.stlouisfed.org/series/A812RC2Q027SBEA. 1

I then exploit variation external to the firm and document that JIT adopters are exposed to the business cycle and other unexpected aggregate events. At the firm level, sales growth among JIT firms comoves more closely with GDP growth than their non-JIT counterparts. JIT firms are estimatedtobebetween25-30%morecyclicalthannon-JITfirms. Inaddition,JITadoptersexperience a 3% sharper drop in sales when faced with unexpected weather disasters. My analysis points to heightened sensitivity among JIT firms upon the realization of external shocks, indicating that an economycomposedofmoreJITproducersislessresilienttosuchdisturbances. Inlightoftheseempiricalfacts,Ibuildandstructurallyestimateadynamicgeneralequilibrium modelofJITproduction. Themodelfeaturesadistributionoffirmsthatdifferintheiridiosyncratic productivity,inventoryholdings,andinventorymanagementstrategy. Materialsareneededforproductionandcanbeacquiredsubjecttoastochasticfixedordercost. JITfirmsdrawordercostsfrom a distribution that is first order stochastically dominated by that of non-JIT firms. Implementing JITrequiresincurringaninitialadoptioncostandasmallercontinuationcostthereafter. Inagiven period, firms must choose their JIT adoption status, whether to order materials, and how much to produce. I numerically solve and structurally estimate the model via the simulated method of moments (SMM)basedondatafrom1980through2018. Theestimatedmodelsuccessfullyfitsthetargeted moments and is also able to produce non-targeted regression coefficients consistent with those estimatedinthedata. RelativetoacounterfactualeconomywithnoJIT,theestimatedmodelyieldsa welfare gain of 1.4% in consumption equivalent terms.4 In addition, the estimated model delivers a 1.3% increase in measured TFP in the steady state. Intuitively, JIT adoption leads to a reduction in fixed order costs which enables adopters to better align material input usage with realized productivity. As a result, measured aggregate productivity rises as firms smooth out their inventory cycles,yieldingareductioninfirm-levelvolatility,consistentwiththemicrodata. Whereas individual adopters benefit from JIT in normal times, the existence of leaner firms renders the economy more vulnerable to unexpected shocks. I consider an unanticipated supply disruption calibrated to match the drop in real US output during the onset of the COVID-19 pan- 4Thiswelfarefigureiscomparablethoughslightlylowerthanmeasuresofgainsfromtrade(CostinotandRodriguez- Clare,2015). 2

demic. Inresponsetosuchashock,theJITeconomyexperiences: (1)amoregradualdepletionof inventories,(2)ahighershareoffirmsthatfacestockouts,and(3)anincreaseintheshareoffirms that do not adopt JIT. Since JIT firms carry fewer inventories, an unexpected spike in the price of orders makes them more likely to fully exhaust their existing stocks. At the same time, as order costs rise, inventories are suddenly more highly valued, with an increase in the shadow value of inventories within the firm. As a result, producers that do not fully stock out cut back on material inputuseinanefforttodrawinventoriesdownmoreslowly. Theutilizationoffewermaterialinputs in production in the JIT economy due to stockouts and hoarding leads to a sharper drop in output relativetothecounterfactualmodel. My empirical and theoretical analysis quantifies a novel trade-off between the higher profits afforded by JIT production and the higher volatility caused by a leaner inventory system. Firms maybenefitinnormaltimesfrompursuingaleaninventorystrategy,eventhoughanunanticipated adverseshockmaytriggeradeepercrisisinaneconomypopulatedbyalargeshareofJITproducers. Inventory investment has long been of interest to economists as a potential source of macroeconomic volatility.5 Seminal contributions developed production smoothing models (Ramey and Vine,2004;Eichenbaum,1984),stockoutavoidancemodels(Kahn,1987),and(S,s)models(Scarf, 1960; Caplin, 1985) of inventory investment. Khan and Thomas (2007) elegantly models inventoriesinageneralequilibriumenvironmentwithheterogeneousfirmsandbusinesscycleshocks. The authorsfindthatinventoriesplaylittletonoroleinamplifyingordampeningbusinesscycles.6 My modelissimilarwithanendogenousJITadoptiondecisionthatdependsonthefirm’sproductivity, and a focus on large unanticipated shocks. A trade-off emerges in my model because firms do not internalizetheprospectofthelargeshockintheirprivatedecisions.7 In addition, this paper speaks to the inventory management literature. Kinney and Wempe (2002) finds that JIT adopters outperform non-adopters, primarily through profit margins. Nakamuraetal.(1998)aswellasRoumiantsevandNetessine(2008)findsimilarevidence. Inacontributiontothecorporatefinanceliterature,Gao(2018)examinestheroleofJITproductionincorporate 5SeeforinstanceAhmedetal.(2004),McConnellandPerez-Quiros(2000),McCarthyandZakrajsek(2007),Irvine andSchuh(2005),andMcMahonandWanengkirtyo(2015). 6Iacovielloetal.(2011)comestoasimilarconclusionalbeitthroughadifferentmodel. Ontheotherhand, Wen (2011)buildsastockoutavoidancemodelandfindsthatinventoriesarestabilizing. 7Thisresultholdsevenwhenallowingforpartialanticipationoftheshockortheintroductionofstockoutcosts. 3

cashhoarding. Moreover,inthetradecontext,Alessandriaetal.(2010)studyaneconomyfeaturing inventorymanagementproblemsrelatingtodeliverylags. Mypaperprovidesabridgebetweenthe managementliteratureonleaninventoriesandtherichliteratureoninventoriesinmacroeconomics byhighlightinghowJITproductionmattersforaggregateoutcomes. Furthermore, this paper relates to the literature on supply chain disruptions. On the empirical front,IadoptastrategysimilartoBarrotandSauvagnat(2016)todeterminewhetherJITproducers aredisproportionatelyexposedtounexpectedweatherdisasters. Otherempiricalworkhasassessed how shocks propagate through a network of firms. For instance, Carvalho et al. (2021) does this in the context of the 2011 Japanese earthquake. Similarly, Cachon et al. (2007) assesses empirical evidence of the bullwhip effect along the supply chain. From a theoretical perspective, my paper relates to models of heterogeneous firms, sunk costs, and supply chains. Meier (2020) models supplychaindisruptionsinthecontextoftimetobuild. Moreover,ImodeltheJITadoptiondecision in a manner similar to Alessandria and Choi (2007) who model path dependent export decisions. Mypaperexplicitlylinkssupplychaindisruptionstoanimportantsourceofinvestmentatthemacro level,inventoryaccumulation. The rest of the paper is organized as follows. Section 2 documents evidence that is consistent with the stabilizing effects of JIT at the firm level along with the exposure to unexpected shocks thatitengendersatthemacrolevel. Sections3and4developthegeneralequilibriummodeloflean production. IestimatethemodelinSection5. Section6quantifiestheaforementionedmicro-macro trade-offassociatedwithJIT,andSection7concludes. 2 Empirical Patterns Among JIT Firms IfirstdocumentempiricalevidenceindicatingthatJITadoptersaremoreefficientandyetaremore exposed to external shocks. I use this as motivating evidence for the model outlined in Section 3. This analysis will also provide moments and external validation to the model once I structurally estimateit. I gather firm-level information by making use of Compustat Fundamentals Annual data for firms from 1980-2018. I merge these data with information on county-level weather events from 4

Table 1: JIT Adoption and Firm Profitability (1) (2) Inventory-to-sales Sales Adopter -0.128*** 0.090*** (0.044) (0.027) Fixedeffects Firm,IndustryˆYear Firm,IndustryˆYear Firms 5,017 5,017 Observations 45,768 45,768 Note: ThetablereportspanelregressionresultsfromCompustatAnnualFundamentalsbasedonregression(1). The regressorofinterestisthefirm-yearspecificadoptionindicator. Firmageinthesampleisspecifiedasacontrolvariable. Four-digitSICcodesarespecifiedintheindustry-by-yearfixedeffects. Standarderrorsareclusteredatthefirmlevel. Thestandarddeviationsofthedependentvariablesare0.82and2.21, respectively. ***denotes1%significance, ** denotes5%significance,and*denotes10%significance. the National Oceanic and Atmospheric Administration (NOAA) with specific links from Barrot andSauvagnat(2016). Inaddition,IdevelopanewmeasureofJITadoptionamongpubliclytraded manufacturersbyextendingpreviousworkintheliterature(KinneyandWempe,2002;Gao,2018). This is done through an exhaustive analysis of news reports and SEC filings. Following the literature, I search these documents for key words such as “JIT,” “just-in-time,” “lean manufacturing,” “pullsystem,”and“zeroinventory.” Ithenanalyzeeachofthesedocumentstoconfirmtheyearof adoptionandtoensurethatthefirminquestionimplementsJITratherthananysupplierspotentially mentionedintheannouncements. Inall,mydatasetidentifiestheyearsinwhichapproximately200 Compustat manufacturers adopted JIT.8 More than half of observed JIT producers adopt prior to 1990,andnearlyalloftheadoptersinmysampleadoptJITbefore2000. Myfinalsampleconsists ofanunbalancedpanelofaboutfivethousanduniquemanufacturingfirmsspanningtheaforementionedtimeperiod. AppendixAprovidessummarystatisticsofthedata,anditalsocorroboratesthe empiricalresultsusinganalternatemeasureofJITbasedonstructuralbreaksininventoryholdings. Using these data, I document four facts about JIT adopters. First, JIT adoption is associated 8The data on JIT adoption could be subject to measurement error. First, there are potentially false negatives in thecross-section(i.e. JITfirmsthatarenotpickedupinthetextsearchandwhicharesubsequentlyassignedasnon- JITfirms). IaccountforthispossibilitywhenmodelingJITbyincorporatingaparameterthatgovernstheobserved frequencyofadoption. Section5discussesthisinfurtherdetail. Second,therecouldpotentiallybemeasurementerror inthereportedyearsofJITadoption. AppendixAprovidesvalidatingevidenceofmyJITmeasurebydemonstrating thatinventoryholdingsdeclinepreciselyintherecordedyearofadoption. 5

withbothlowerinventoryholdingsandhighersaleswithinfirms.9 Iestimate: y “ γadopter `X1 β `δ `δ `ν , (1) ijt ijt ijt jt i ijt where y is an outcome variable for firm i belonging to 4-digit SIC manufacturing industry j in ijt year t. I specify the outcomes to be log inventory-to-sales ratio and log sales. The regressor of interest,adopter ,isatime-varyingindicatorforwhetherfirmiisaJITadopterinagivenyear. ijt Table 1 reports the regression results.10 Adopters experience a 13% decrease in inventory-tosalesratiosanda9%increaseinsales. Theresultsimplyachangeof-16%and4%ofonestandard deviation in the outcomes, respectively. The regression results allude to the benefits of JIT in the model. Facing lower fixed order costs, adopters hold fewer inventories in favor of placing smaller more frequent orders. Upon shrinking their inventory stocks, adopters also incur fewer carrying costs. Thesecostreductionsleadadopterstoallocatemoreresourcestoproduction. Second,JITadoptersexperiencelessmicrovolatility. Iestimatethefollowingregression: y “ γadopter `βy `δ `η , (2) ijt ijt ijt´1 jt ijt wherey nowdenotesarolling5-yearstandarddeviationofsalesgrowthandemploymentgrowth ijt for firm i in industry j in year t. Table 2 reports the results. Adopters see a roughly 7% decline in sales and employment growth volatility. This is consistent with the stabilizing role that JIT plays in the model. Due to the lower fixed order costs, firms smooth out their inventory cycles which moderatesthevariabilityofotheroutcomesaswell. I next document facts relating to firm-level exposure brought on by JIT, exploiting aggregate variationandexaminingsensitivitytoasetofspecificeventssuchasmacrofluctuationsandweather disasters. The regression results accord with the model in that adopters are less insured against unanticipated disruptions, and an economy with more JIT firms is more exposed to unexpected 9FigureA1plotstotalinventoryholdingsbytypebasedonmysample. Aggregateandindustry-levelinventory-tosalesdatasimilarlyshowthatinputinventorieshavedeclinedsincethe1980s. Withthatsaid,inventoryholdingshave recentlyrisen,particularlyfollowingthelasttworecessions. Iviewthisasconsistentwiththenotionthatfirmsreassess risksassociatedwithcarryingfewerinventoriesfollowinglargeshocks. 10AppendixAprovidesadditionalresultsrelatingJITtohighersalesperworkerandmorepreciseforecastsdevised bymanagersabouttheirownfirms’earnings. 6

Table 2: JIT Adoption and Firm Volatility (1) (2) Std. salesgrowth Std. employmentgrowth Adopter -0.065*** -0.068*** (0.009) (0.019) Fixedeffects IndustryˆYear IndustryˆYear Observations 10,710 10,710 Note: ThetablereportspanelregressionresultsfromCompustatAnnualFundamentalsbasedonregression(2). The regressorofinterestisthefirm-yearadoptionindicator. Alagofthedependentvariableisspecifiedasacontrol. FourdigitSICcodesarespecifiedintheindustry-by-yearfixedeffects. Standarderrorsareclusteredatthefirmlevel. *** denotes1%significance,**denotes5%significance,and*denotes10%significance. aggregateshocks. Third,JITadopterstendtobemorecyclical. Iquantifythisviaregressionsthatinteractadoption withGDPgrowth: “ ‰ y “ γ adopter `γ GDPgrowth `γ adopter ˆGDPgrowth `X1 β `δ `ε , ijt 1 ijt 2 t 3 ijt t ijt j ijt (3) whereXdenotesasetofcontrols. Thecoefficientγ measurestheextenttowhichJITfirmsexhibit 3 morecyclicality. Table3reportstheregressionresults. Basedoncolumn(1),a1%increaseinGDP growth is associated with a roughly 1.6% increase in sales growth among non-adopters. Adopters experienceanadditionalsalesgrowthincreaseof0.47%abovethisbaseline. Turningtocolumn(2), a1%increaseinGDPgrowthisassociatedwitha1.6%increaseinemploymentgrowthamongnonadopters, with a further 0.39% increase in employment growth among adopters. Taken together, adoptersarearound25-30%morecyclicalthannon-adopters. Fourth, JIT adopters are more sensitive to local weather events. I examine this by estimating thefollowingregression: “ ‰ y “ ψ adopter `ψ disaster `ψ adopter ˆdisaster `X1 β `δ `δ `ω . (4) ijt 1 ijt 2 ijt 3 ijt ijt ijt i t ijt The“disaster”regressorisanindicatorforasevereweathereventoccurringinagivenyear. Icollect 7

Table 3: JIT Adoption and Cyclicality Salesgrowth Employmentgrowth GDPgrowth 1.625*** 1.561*** (0.287) (0.256) AdopterˆGDPgrowth 0.467** 0.393** (0.199) (0.188) Controls Yes Yes FixedEffect Industry Industry Observations 34,502 34,502 Note: ThetablereportsregressionresultsfromCompustatAnnualFundamentalsbasedonregression(3). TheindependentvariableofinterestistheinteractionbetweentheadopterindicatorandGDPgrowth. Controlvariablesinclude firmageinthesample,cash-to-assets,sales-per-worker,aswellastheadoptionindicator. Four-digitSICfixedeffects arespecified. Standarderrorsareclusteredatthefirmlevel. ***denotes1%significance,**denotes5%significance, and*denotes10%signficance. information on county-level weather disasters from NOAA and link these disasters to public firm headquarter zip codes via the aforementioned Barrot and Sauvagnat (2016) links. Table 4 reports the estimation results. On average, a given weather event in my sample predicts an additional 3% declineinJITfirmsalesandemploymentrelativetonon-JITfirms.11 Takentogether,thedatasuggestthatJITadoptersbenefitfromhigherprofitsandsmootheroutcomes. Atthesametime,adoptionisassociatedwithheightenedexposuretoaggregatefluctuations and unanticipated shocks as proxied by local weather disasters. My model of heterogeneous firms withanendogenousJITadoptiondecisioncanexplainthesepatterns. Themodelalsoallowsmeto quantitativelyassesstheimpactofJITamidanunanticipatedmacrodisaster,somethingthatcannot becapturedbyfirmlevelregressions. 3 A Model of Just-in-Time Production Having illustrated the essence of the trade-off in the data, I next build the full general equilibrium modelwhichwillprovidequantitativestatementsabouttheimplicationsofJIT.Themodelissimilar in spirit to Khan and Thomas (2007) and Alessandria and Choi (2007), embedded with JIT and 11SimilarconclusionsaredrawnwhenlinkingfirmstotheirprimarysuppliersusingtheCompustatSegmentfiles, andestimatingthedifferentialeffectthataweathereventoriginatingupstreamhasonthedownstreamfirmbasedonits JITstatus. AppendixAreportstheseresults. 8

Table 4: JIT Adoption and Sensitivity to Local Disasters Sales Employment Disaster -0.012** -0.012** (0.005) (0.005) AdopterˆDisaster -0.029* -0.030* (0.017) (0.017) Controls Yes Yes FixedEffects Firm,Year Firm,Year Observations 43,123 43,123 Note: The table reports weather event regressions from a sample of Compustat firms based on regression (4). The independent variable of interest is the interaction between the adoption indicator and the disaster indicator. Control variables include capital investment rate, sales per worker, ratio of cost of goods to sales, finished goods inventory holdings,adopterindicator,andthedisasterindicator. Standarderrorsareclusteredatthefirmlevel. ***denotes1% significance,**denotes5%significance,and*denotes10%significance. ultimatelyincorporatinglargeunanticipateddisastersratherthantraditionalbusinesscycleshocks. Arepresentativehouseholdhaspreferencesoverconsumptionandleisure. Thehouseholdsupplies its labor frictionlessly to the two sectors of the economy: the intermediate goods sector and thefinalgoodssector. Arepresentativeintermediategoodsfirmproducesmaterialsbyusinglabor and capital. In addition, a continuum of heterogeneous final goods firms make use of labor and materialstoproduceusingadecreasingreturnstoscaletechnology. Finalgoodsproducersareheterogeneousinidiosyncraticproductivity,inventorystocks,andJITadoptionstatus. Allmarketsare perfectlycompetitive. The representative household is endowed with one unit of time in each period and values consumptionandleisureaccordingtothefollowingpreferences:12 UpC ,H q “ logpC q`φp1´H q, t t t t where φ ą 0 denotes the household’s labor disutility. Total hours worked is denoted by H and t labor is paid wage, w . In addition to wage income, the household earns a dividend each period t fromownershipoffirms,D ,andchoosessavingsonaoneperiodrisklessbond,B ,giveninterest t t`1 12Rogerson(1988)microfoundsthesepreferencesinamodelofindivisiblelaborandlotteries. 9

rateR . Therepresentativehousehold,facingnoaggregateuncertainty,maximizesitsutility: t`1 ÿ8 max βtUpC ,H q, t t Ct,Ht,Bt`1t“0 subjecttoitsbudgetconstraintwhichholdsforallt, C `B ď R B `w H `D . t t`1 t t t t t Theparameterβ P p0,1qisthehousehold’ssubjectivediscountfactor. The representative intermediate goods firm produces materials using capital K and labor L t t accordingto: FpK ,L q “ KαL1´α. t t t t Capitalevolvesaccordingtoinvestmentwithatime-to-buildconstraint: K “ p1´δqK `I , t`1 t t where δ P p0,1q is the depreciation rate of capital. Taking prices as given, the problem of the intermediategoodsfirmis: max q FpK ,L q´w L ´K `p1´δqK t t t t t t`1 t Kt`1,Lt whereq denotesthepriceoftheintermediategood. t Finally,acontinuumoffinalgoodsfirmsproduceusingmaterials,m ,andlabor,n ,according t t adecreasingreturnstoscaletechnology: y “ z mθmnθn, θ `θ ă 1, t t t t n m whereidiosyncraticproductivityevolvesasanAR(1)inlogs: logpz q “ ρ logpz q`σ ε , ε „ Np0,1q. t`1 z t z t t 10

Figure 1: Decisions of Final Goods Firms Note: Thefiguresummarizestheorderofthedecisionsmadebyfinalgoodsfirmswithinaperiod. Materials are drawn from the firm’s existing inventory stock, s , to use in production. Final goods t firms procure new materials from the intermediate goods firm subject to a stochastic fixed order costdrawnfromaknowndistribution. Figure 1 details the final goods producers’ decision-making timeline. Each period is broken into three stages. A producer enters the period with realized productivity, z , inventory stock, s , t t and adoption status, a . In the first stage, the producers decide whether or not to adopt JIT. If a t producerdoesnotentertheperiodasacontinuingadopter,itmustpayc inordertoinitiallyadopt. s Alternatively,iftheproducerenterstheperiodasanadopter,itmustpayasmallercontinuationcost 0 ă c ă c inordertomaintainitsstatusasaJITproducer. f s Intuitively,adoptingJITrequiresthataplantrepurposeitsshopfloor,enterintolong-termcontractswithsupplierstofulfillordersinatimelyfashion,andpossiblyevenpurchasenewtechnologies to share information with suppliers. The sunk setup cost encompasses all of these one-time costs. The continuation cost embodies smaller costs for suppliers to participate in timely delivery, costs of training labor on JIT best practices, and greater attention or communication required to shareinformationwithsuppliers. In the next stage, producers learn their order costs, ξ „ Fpξq, and decide whether or not to place an order, o . JIT producers face a more favorable order cost distribution, Epξ q ď Epξ q. t A NA Lastly, following the adoption and the order decisions, final goods producers decide how much to produce. I characterize the final goods firms’ problem in terms of inventory stocks rather than specific 11

order or material input choices. In particular, if a firm enters the period with inventory stock s , t its target inventory stock is denoted by s˚. This means that any orders, if placed, are defined as t o “ s˚ ´ s . Following the order decision, suppose that inventory stock sr is carried into the t t t t productionstage. Materialsusedinproductionarethendefinedasm “ sr ´s wheres refers t t t`1 t`1 to the inventory stock carried forward into the next period. In what follows, I suppress the time subscriptandinsteaddenotenextperiodvariableswithaprime. Stage1: AdoptionDecision A final goods producer begins the period with pz,s,aq, faces labor-denominated adoption costs tc ,c u,andendogenousprices,p,q,andw. ThefirmfirstdecideswhethertoadoptJIT.Notethat s f theadoptionstatusisabinaryoutcome. Thevalueofadoptingis: " ż ż * VApz,s,aq “ max ´pwcpaq` VOpz,s,1,ξqdFpξ q, VOpz,s,0,ξqdFpξ q , (5) A NA where $ ’ ’ & c ifnoJIT(a “ 0) s cpaq “ ’ ’ % c ifJIT(a “ 1), f and VOpz,s,a,ξq refers to the firm’s value in the second stage. Order costs are assumed to be distributeduniformly,Fpξq “ Upξ,ξq.13 Thefirm’soptimaladoptionpolicy,a1pz,s,aq,solves(5) Stage2: OrderDecision Given the firm’s order cost draw, ξ, also denominated in units of labor, it then decides whether to place an order, o. If the firm is an adopter, its order cost distribution is first order stochastically dominatedbythoseofnon-adopters. Thevalueinthesecondstageis " * VOpz,s,a,ξq “ max ´pwξ `pqs`V˚pz,s,a,ξq,VPpz,s,aq , (6) 13InAppendixD,Iconsideranalternateordercostdistribution(KhanandThomas,2016)whichdeliversquantitativelysimilarresults. 12

wherethevalueofplacinganorderis14 „  V˚pz,s,a,ξq “ max ´pqs˚ `VPpz,s˚,aq , (7) s˚ěs andVPpz,s,aqisdefinedbelow. Thefirm’sorderproblemdeliversathresholdrule. Inparticular, a firm places an order if and only if the order cost draw is lower than a threshold order cost: ξ ă ξ˚pz,s,aqwhere pqs`V˚pz,s,aq´VPpz,s,aq ξ˚pz,s,aq “ . (8) φ Stage3: ProductionDecision UponchoosingitsJITstatus,decidingwhethertoplaceanorder,andpotentiallyselectinganorder size, the firm then makes a production decision. Suppose that a firm enters the production stage withinventorystocksrsuchthat: $ ’ ` ˘ ’ & s˚ z,s,a1pz,s,aq iforderplaced sr“ ’ ’ % s ifnoorderplaced. In the production stage, the firm selects labor, npz,sr,s1,aq, and materials, psr´ s1q, to maximize profits. Itsvaluefunctionintheproductionstageis: “ ‰ VPpz,sr,aq “ max πpz,sr,s1,aq`βE VApz1,s1,a1q (9) s1Pr0,srs where „  c πpz,sr,s1,aq “ p znpz,sr,s1,aqθnpsr´s1qθm ´wnpz,sr,s1,aq´ m s12 (10) 2 are period profits. The end of period inventory stock is denoted by s1, and c is a convex carrying m costofstoringunusedinventory.15 A final goods producer is said to stock out if it enters the period with no inventories, s “ 0, 14Theconstraintontheorderdecisionallowsforonlypositiveorders. Inparticular,themodelabstractsawayfrom inventoryliquidation. 15ThequadraticcarryingcostassumedissimilartoLuoetal.(2021) 13

andchoosestonotplaceanorder. Withoutanyinventories,thefirmhasnomaterialinputstodraw from when making its production decision. As a result, the firm forgoes production in that period. The producer can flexibly restart production in the future conditional on a favorable productivity realizationandordercostdraw.16 4 Analyzing the Model The endogenous adoption decision allows the model to replicate important features of the data, namely,higherprofitabilityandreducedmicrovolatilityamongJITfirms. SinceimplementingJIT comesatarelativelylargesunkcost,notallfirmsoptimallychoosetoadoptJIT.Figure2plotsthe adoption frontiers for JIT and non-JIT producers. The shaded area in the lower right corner represents the region of the state space in which non-JIT firms choose to adopt JIT. This illustrates the positive selection into adoption implied by the model. Moreover, the scope for initiating adoption is decreasing in inventory stocks as the value of adopting is higher among firms that are closer to stockingout. At the same time, a producer is likely to remain an adopter conditional on already being one. This is because the continuation cost of retaining JIT is smaller than the initial sunk cost. Hence, theendogenousadoptiondecisionexhibitspersistence. ThelargerstripedareainFigure2confirms thisintuition. OnlytheleastproductiveadopterswillopttoabandonJIT.Furthermore,thescopefor exitingadoptionisincreasingininventoryholdings. Theselectiondetailedherecouldcontributeto thepatternsamongJITfirmsdocumentedinthedata. Inparticular,thedecisiontoadoptJITreflects afavorableproductivityrealizationwhich,whencoupledwithloweraverageordercosts,leadsfirms toreduceinventorystocksandincurfewercarryingcoststherebygeneratingmoreoutput. Figure3showstheprobabilityofplacinganorderasafunctionofproductivity. Consistentwith thedecisiontoselectintoadoption,orderprobabilitiesareincreasinginproductivityanddecreasing ininventoryholdings. Moreover,thebenefitsofJITadoptioncanbeunderstoodbycomparingthe two panels. Across both inventory levels, the probability of placing an order is higher for adopters 16InAppendixD,Iconsideraneconomywithstockoutcostswhichdissuadeproducersfromallowings “ 0,with littleimpactontheheadlineresultsdiscussedinSection6. 14

Figure 2: Adoption Frontiers Note: ThefigureplotstheadoptionfrontieramongJITandnon-JITfirms. Thesolidshadedareaplotstheregionof thestatespaceinwhichnon-JITfirmsselectintoadoption. Thestripedareaalongwiththeshadedareajointlydenote theregionofthestatespaceinwhichexistingJITfirmschoosetoremainadopters. sincetheyfaceloweraverageordercosts. Asaresult,adoptersinthemodelplacesmallerandmore frequentorders. Thisisconsistentwiththereductionininventoryholdingsamongadopters. Figure 4 plots material usage as a function of productivity. Material inputs are increasing in productivity and inventory holdings. Firms with very low inventory stocks will tend to exhaust their remaining inventories regardless of their level of productivity. As a result, the flat lines in these policies reflect endogenous decisions to fully utilize existing inventory stocks in production. Furthermore, adopters make greater use of materials when producing thereby raising output. Becauseadopterscanrestockmoreflexibly,duetothelowerordercosts,theyexhausttheirinventory stocks more often. As a result, production among JIT firms tends to be uninterrupted despite their lowerinventoryholdings. Boththeorderthresholdandthematerialinputpolicyreflectatreatment effectthatallowsfirmstoproduceatlowercostswhichinturnraisesfirmsalesfollowingadoption. A comparison of outcomes between economies that differ only in the option to adopt JIT confirmsthemodel-impliedbenefitstoleanproduction: highersalesandlessvolatility. Figure5visu- 15

Figure 3: Order Probabilities Note: The figure plots the probability of placing an order in the order stage as a function of productivity. Panel (a) plotstheprobabilitiesamongnon-adoptersandpanel(b)plotstheprobabilitiesforadopters. Thesolidredlinereflects ahighinventoryestablishmentinthemodelwhilethedashedbluelinereflectsalowinventoryestablishment. alizes the results from such an exercise. The figure plots a plant’s simulated path in both models. Theplantineacheconomyfacesthesameproductivityrealizations. Upon adopting JIT, the establishment retains its status as an adopter through the rest of the simulated path despite lower productivity realizations in the latter periods. This enables the establishment to undertake production despite holding fewer inventories. The cost savings associated withJITallowthefirmtoredirectitsresourcestoproductionratherthanorderplacingorinventory storage. Asaresult,salesarehigheramongJITfirms. Furthermore,uponadoptingJIT,theplant’ssimulatedpathforordersissmoothedconsiderably relative to the economy without adoption. This illustrates the insight that JIT mutes the inventory cycle. Because adopters face lower fixed order costs, their target inventory stocks are lower in the JIT model and the frequency of placing an order increases. The smoother path for orders also smooths firm sales which can explain the lower variance of outcomes among adopters in the data. Lastly,thefourthpanelofthefigureconfirmsthatJITproducersenjoy,onaverage,highersales. 16

Figure 4: Material Usage Note: The figure plots material usage policy functions in the production stage as a function of productivity. Panel (a)plotsthepolicyamongnon-adoptersandpanel(b)plotsthepolicyforadopters. Thesolidredlinereflectsahigh inventoryestablishmentinthemodelwhilethedashedbluelinereflectsalowinventoryproducer. 5 Structural Estimation I structurally estimate the model using the micro data analyzed in Section 2. The estimated model captures important features of the firm-level data including the adoption frequency, levels of and covariances between inventories and sales, and spikes in inventory holdings. Importantly, the estimated model allows me to quantify benefits to JIT in normal times as well as the vulnerabilities thatitengenderstounanticipatedmacroshocks. ThecomprehensivesearchoffirmfinancialsandpublicstatementsensuresthatthedataonJIT adoptiondonotincludefalsepositives. However,informationonJITimplementationisconstrained to what is reported in these records. To allow for the possibility that JIT is more widespread than theempiricalfrequencyofadoptioninmysample,Iusethestructureofthemodelinordertoinfer patternsofadoption. Idosobydefiningaparameter,τ P p0,1q,thatgovernstheshareofobserved non-adoptersfromasimulatedpaneloffirms.17 17Asinmysample,afirminthemodelissaidtobeanadopterifatleastoneofitsestablishmentsadoptsJIT.Upon simulatingapaneloffirms,ashareτ,aredesignatednon-adoptersirrespectiveoftheirtrueadoptionstatus. 17

Figure 5: Adoption Mutes the Order Cycle Note: The figure plots the path of a selected establishment in the unconditional simulation. The top panel plots the (shared)pathofidiosyncraticproductivityacrossbothmodels. Thesecondpanelplotstheplant’sJITadoptionstatus, thethirdpanelplotsorders,thefourthplotssales. There are 16 parameters in the model. I first externally fix seven parameters to match standard targets in the literature. Table 5 details the annual calibration. The discount factor, β is set to be consistent with a real rate of 4%. Capital depreciation is set to match the average investment rate intheNBER-CESmanufacturingdatafrom1980-2018(approximately6.4%). Thematerialshare, θ ,issettomatchthematerialshareintheNBER-CESdatabase,andthecapitalshare,α,isfixed m tomatchthecapital-outputratio. Theparameterθ issettomatchaneconomy-widelaborshareof n 0.65. Theleisurepreferenceiscalibratedsothatthehouseholdworksone-thirdofthetime. Finally, InormalizethelowerboundoftheordercostdistributionforJITproducerstozero,consistentwith thecalibratedlowerboundinKhanandThomas(2007).18 18Rather than fix the lower support of both order cost distributions to zero, I instead include ξ in the SMM NA procedure. Since non-JIT firms are expected to face higher average order costs, this approach allows me to more flexibly capture a higher first moment in the non-adopters order cost distribution without necessarily requiring the variancetobehigheraswell. 18

Table 5: External Parameterization Description Parameter Value Notes DiscountFactor β 0.962 Realrateequalto4% Capitaldepreciation δ 0.064 NBER-CES(1980-2018) Materialshare θ 0.520 NBER-CES(1980-2018) m Capitalshare α 0.420 NBER-CES(1980-2018) Laborshare θ 0.190 Laborshareequalto0.65 n Labordisutility φ 2.500 Onethirdoftotalhoursworked Ordercostlowerbound(adopters) ξ 0.000 LowerboundinKhanandThomas(2007) A Note: Thetablereportsthesevencalibratedmodelparameters. 5.1 Simulated Method of Moments ` ˘ 1 The parameter vector to be estimated is θ “ ρ σ ξ ξ ξ c c c τ . These paramz z NA NA A s f m eters residing in θ govern the exogenous productivity process, the order and adoption costs, the carryingcost,andtheshareofobservednon-JITfirms. Themodelhasnoclosedformsolution,so I solve it using standard numerical dynamic programming techniques detailed in Appendix B. To parameterizethemodel,IemploySMM(DuffieandSingleton,1993;Bazdreschetal.,2018). This isdonebycomputingasetoftargetedmomentsinthemodelandminimizingtheweighteddistance betweentheempiricalmomentsandtheirmodel-basedanalogs. Specifically, I target 11 moments to estimate the nine parameters. My estimator is therefore an overidentified SMM estimator. The first targeted moment is the empirical frequency of adoption. Of the remaining ten moments, five are specific to JIT firms and five to non-JIT firms. These five moments,whicharethesameacrossbothtypesoffirms,are: themeaninventory-to-salesratio,the covariancematrixofinventory-to-salesratiosandlogsales(whichdeliverthreemoments),andthe frequency of positive inventory-to-sales ratio spikes, defined as instances in which the inventoryto-sales ratio exceeds 0.20.19 I specify the asymptotically efficient choice of the weighting matrix whichistheinverseofthecovariancematrixofthemoments. 19TheempiricalmomentsarelistedinthethirdcolumnofTable7. 19

5.2 Informativeness of Moments While the targeted moments jointly determine the parameters to be estimated, there are nonetheless moments that are especially informative in pinning down certain parameters. I discuss their informativenessinturn. Idiosyncraticproductivitypersistencemostlyinformsthecovariancebetweeninventory-to-sales andlogsales. Forinstance,anincreaseρ willsmoothoutfirmsalesandinventoryholdings. Since z sales and inventory-to-sales covary negatively, an increase in productivity persistence delivers a more negative covariance between inventory-to-sales and log sales. Moreover, idiosyncratic productivity dispersion mostly affects variances, as an increase in σ results in more dispersed outz comesamongproducers. Theordercostsarestronglyrelatedtothefirstandsecondmomentsofinventory-to-salesratios. An increase in the lower bound of the non-adopter order cost distribution leads to a lower level of inventoryholdingsamongnon-JITproducers. Asthelowerboundoftheordercostsincreases,nonadopters face an ever higher average order cost. Intuitively, an increase in expected non-JIT order costs raises the returns to adoption. Due to positive selection into adoption, the remaining pool of non-adoptersislessproductive,meaningthattheirtargetinventorystocksarerelativelylower. An increase in the upper support of the order cost distribution for non-adopters raises both the first and second moment of order costs. As a result, an increase in the upper bound will raise the variance of inventory-to-sales ratios for non-adopters. On the other hand, an increase in the upper support of the order cost distribution for adopters leads to higher inventory-to-sales ratios among adopters. While some less productive producers switch out of JIT, the remaining firms raise their target inventory stocks in an effort to lengthen the time between orders. As a result, inventory-tosalesamongexistingadoptersrises. Theadoptioncostsarelargelyinformedbythecovariancebetweeninventoryholdingsandsales andthevarianceofsales. Anincreaseinthesunkcostofadoptionweakensthecovariancebetween inventory-to-salesandlogsalesamongadopters. Becauseahighersunkcostreducesthearearepresentingtheadoptionfrontierfornon-JITproducersinFigure2,onlythemostproductiveproducers willselectintoadoption. Thesehighlyproductiveproducers,whenfacedwithloweraverageorder 20

costs, substantially reduce their target inventory-to-sales ratios. Furthermore, the variance of inventory holdings also decline leading to a looser covariance between inventory-to-sales and sales. On the other hand, the continuation cost of adoption affects the variance of sales. In particular, a highercontinuationcostofadoptionreduceslikelihoodofremaininganadopterconditionalonalreadybeingone. Themarginalproducer,whichislessproductiveandmorebloated,willtherefore switch out of adoption. As a result, the pool of non-adopters faces a wider range of endogenous outcomes since there is now a larger set idiosyncratic productivity realizations that are consistent withbeinganon-JITproducer. Hence,non-JITfirmsseeariseinthevarianceoflogsales. The convex storage cost affects inventory holdings and spike rates as expected. In particular, highercarryingcostsleadfirmstoleanoutacrosstheeconomysothatinventory-to-salesandspike rates fall among adopters and non-adopters alike. This also implies that the variance of inventory holdingsfallsacrossallfirmsamidariseincarryingcosts. Atthesametime,theoverallvarianceof logsalesrisesassomefirmscanflexiblyoperateandgeneratesalesintheleanerenvironmentwhile other firms cannot. Finally, a rise in the the share of observed non-adopters reduces the frequency ofadoption,asexpected. FigureC1inAppendixCoutlinesthesekeymonotonicrelationshipsbetweenthemomentsand the parameters. In addition, Figure C2 helps assess the sources of identification by reporting the sensitivity of each of the nine parameters to changes in a given moment, based on Andrews et al. (2017). Thesefiguresconfirmtheintuitionlaidoutabove. 5.3 Estimation Results Table 6 reports the estimated parameters, all of which are precisely estimated. The technology parameters, ρ and σ , are consistent with parameterizations in the literature (Khan and Thomas, z z 2008;HennessyandWhited,2007;Meier,2020),collectivelyrangingfrom0.68-0.89and0.02-0.12 respectively. My estimates imply a more persistent and less dispersed idiosyncratic productivity processthanthatestimatedinClementietal.(2015)whichisattributabletothefactthatmysample consistsofpublicmanufacturerswhoarelargerandolderthantheuniverseofmanufacturers. Thelowerboundoftheordercostdistributionamongnon-JITproducersis0.008whiletheupper 21

Table 6: Estimated Parameters Description Parameter Estimate Idiosyncraticproductivitypersistence ρ 0.851 z (0.002) Idiosyncraticproductivitydispersion σ 0.022 z (0.001) Ordercostlowerbound(non-adopters) ξ 0.008 NA (0.001) Ordercostupperbound(non-adopters) ξ 0.451 NA (0.006) Ordercostupperbound(adopters) ξ 0.060 A (0.006) Sunkcostofadoption c 0.201 s (0.002) Continuationcostofadoption c 0.073 f (0.003) Carryingcost c 1.037 m (0.009) Observedshareofnon-adopters τ 0.952 (0.001) Note: Thetablereportstheestimatedparameterswithstandarderrorsinparentheses. supportoftheordercostdistributionamongnon-adoptersis0.45. Thisupperboundisestimatedto be an order of magnitude larger than that of adopters, implying that JIT firms place orders that are about45%smallerthanthoseofnon-JITfirms,indicatingasizablereturntoadoptionforthosewho can initiate it. Furthermore, the adoption cost estimates suggest a substantial amount of hysteresis in the adoption decision. In particular, firms pay a continuation cost that is slightly more than one third of the original sunk cost. Conditional on being an adopter, the probability of remaining an adopter is 94%. This estimate is similar to estimates of the sunk cost of exporting, which place theprobabilityofremaininganexporterconditionalonalreadybeingoneat87%(Alessandriaand Choi,2007). Theestimatedcarryingcostisabout20%ofthevalueofinventories,anon-negligible amountthatpreventsfirmsfromstoringtoomanyinventoriesacrossperiods. Lastly,theestimated share of observed non-adopters implies that the mass of JIT establishments in the model’s steady stateisabout0.40. Given that I target 11 moments to estimate the nine parameters, the model is overidentified andwillnotexactlymatchtheempiricalmoments. Withthatsaid,theestimatedmodelfitsthedata 22

Table 7: Model vs. Empirical Moments Moment Model Data Mean(inventory-salesratio|adopter) 0.101 0.094 (0.005) Mean(inventory-salesratio|non-adopter) 0.122 0.146 (0.002) Std(inventory-salesratio|adopter) 0.059 0.054 (0.001) Corr(inventory-salesratio,logsales|adopter) -0.132 -0.098 (0.001) Std(logsales|adopter) 0.219 0.206 (0.005) Std(inventory-salesratio|non-adopter) 0.074 0.161 (0.001) Corr(inventory-salesratio,logsales|non-adopter) -0.307 -0.282 (0.001) Std(logsales|non-adopter) 0.267 0.296 (0.002) Spike(inventory-salesratio|adopter) 0.059 0.045 (0.012) Spike(inventory-salesratio|non-adopter) 0.156 0.188 (0.005) Frequencyofadoption 0.047 0.042 (0.004) Note: Thetablereportsmodel-basedandempiricalmomentswithstandarderrorsinparentheses. well. Table7comparesthe11targetedmomentsgeneratedbythemodelwiththeirempiricalvalues. Importantly, the model replicates important features between adopters and non-adopters. Relative to non-JIT firms, adopters hold fewer inventories as a share of their sales. In addition, adopters are broadly characterized by less variable outcomes and a looser association between inventoryto-sales ratios and log sales. Lastly, adopters exhibit fewer spikes in inventory holdings relative to theirsales. 5.4 Non-targeted Moments To further assess the estimated model’s ability to match the patterns present in the data, I run empiricalregressionsbasedonapanelofsimulatedfirmsfromboththeestimatedandcounterfactual models. The results are reported in Table 8. The regressions in Panel A are identical to those in 23

Table 8: Model-Based Regressions PanelA:Levels Inventory-to-sales Sales Data -0.128(0.044) 0.090(0.027) Model -0.180(0.009) 0.060(0.002) PanelB:Volatility Salesgrowth Employmentgrowth Data -0.065(0.009) -0.068(0.019) Model -0.046(0.003) -0.041(0.003) Note: Thetablereportsempiricalandmodel-basedpanelregressionsatthefirmlevelfromtheestimatedandcounterfactualmodelswithstandarderrorsinparentheses. PanelAreportsregressionresultsasinTable1. PanelBreports regressionresultsasinTable2. Table1whiletheregressionsinPanelBareidenticaltothoseinTable2. Followingadoption,theestimatedmodelisabletosuccessfullyreproducereductionsininventoryto-sales ratios. The OLS coefficient from the estimated model resides within the 95% confidence interval of the empirical point estimate. In addition, the estimated model predicts a quantitatively similar increase in sales among adopters. Moreover, the estimated model predicts reductions in firmvolatilityof4-5%amongadopters,closetothe6-7%estimateddeclinesinthedata. With precisely estimated parameters delivering a broadly successful fit to the data, I can now exploitthisstructureasalaboratoryforquantitativeexperiments. 6 Quantifying the Trade-off Iproceedtoquantifythetrade-offbetweenthelong-rungainstoJITandthevulnerabilitytounanticipateddisastersthatJITexposes. Ifirstexaminethemodel’ssteadystatetocharacterizethebenefits of lean production. I then analyze the dynamics of the estimated economy following a COVID-19 disaster. A natural benchmark against which to compare the estimated model is a world in which JIT adoptionisnotpossible. Idefinesuchacounterfactualbysolvingaversionoftheestimationmodel with adoption cost parameters c and c fixed to be prohibitively large such that no adoption takes s f 24

Table 9: Long-Run Aggregates Across Models Output Order frequency Order size Price of orders 9.64 48.45 -19.25 4.76 Inventory stock Firm value Measured TFP Welfare -35.80 1.30 1.31 1.43 Note: The table reports steady state values of the estimated model relative to the counterfactual model, in percent deviations. place. InAppendixC,IconductasubsampleanalysisinwhichIseparatelyestimatethemodelfor theyears1980-1989and1990-2018,definingtheformerperiodastherelevantcounterfactual. The resultsfromthisexercisearequalitativelysimilartotheanalysisinthissection. 6.1 Steady State AcomparisonbetweenthetwomodelspointstosizablegainsassociatedwithJITadoption. Table9 reports the steady state in the estimated model relative to the counterfactual economy in percent deviations. The prevalence of JIT in the estimated model delivers a 9-10% increase in output and impliesthatsmallerandmorefrequentordersplacedsuchthatorderdemandrises. As expected, inventory holdings fall in the estimated model. The reduction in inventories is due to a decrease in target inventory stocks across all producers.20 Relative to the counterfactual, the estimated model delivers a 40% decline in the aggregate inventory-to-sales ratio, close to the observed35%declineintheratioofnonfarminventoriestofinalsalesfrom1980-2018. Inaddition, firm value rises by about 1.3% in the estimated model. For reference, the literature measures firm valuelossesof2%duetobiasesinmanagerialbeliefs(Barrero,2020)and3%duetoCEOturnover frictions(Taylor,2010). Welfareintheestimatedmodelis1.43%higherinconsumptionequivalent terms,amagnitudewhichresidesbetweenthecostsofmanagerialshort-termism(Terry,2017)and staticgainstotrade(CostinotandRodriguez-Clare,2015). Ordercostsareasourceofdispersioninthemodel. Ideallyfirmswouldliketoholdnomaterial inventories, instead placing orders and fully utilizing them when producing every period. In an efforttominimizethenumberoftimesthefixedordercostsareincurred,producersholdnon-zero 20Non-JITproducersalsoreducetheirinventorytargetsduetotheriseinthepriceoforders. 25

inventories. For this reason, the estimated JIT adoption model implies an increase in measured TFP. With more adoption, a greater number of producers operate subject to lower order costs. At the aggregate level, this implies that resources are reallocated to high marginal product producers. In essence, firms place more frequent orders and therefore have the flexibility to better align their material usage with their realized micro productivity realizations. The estimated model implies thatJITadoptionraisesmeasuredTFPbyapproximately1.3%. 6.2 Effects of an Unanticipated Disaster I next show that despite enjoying higher profits and smoother firm-level outcomes, an economy populated by lean producers is more vulnerable to an unexpected disaster. To do this, I introduce aggregateproductivityintotheproductionfunctionforintermediategoods. O “ AKαL1´α Whereas in the steady state A “ 1, in a disaster episode A unexpectedly falls below one.21 I shock A so as to match the 3.4% drop in real GDP between 2019 and 2020. After this onetime unforeseen shock, I trace the endogenous outcomes in the JIT economy. I then repeat the exercise for the counterfactual economy, keeping the shock to A the same across both economies. Figure 6 displays the output response to this unexpected disaster. In addition, Figure 7 reports the keydifferencesinendogenousresponsesbetweenthetwomodelsamidthedisaster. Overall, the JIT economy sees a roughly 0.40 percentage point excess output contraction amid thedisaster,amountingtoaround13%morethantheoutputlostinthecounterfactualmodel. During anunexpecteddisaster,theshadowvalueofinventoriesrisesleadingtoaspikeinstockoutsandan overall drop in the likelihood of placing an order. Though both economies experience a decline in inventoryholdings,theJITeconomyexperiencesarelativeincreaseininventoriessincetheleaner firms draw their stocks down more slowly. Due to this hoarding-like behavior, firms in the JIT 21ConsistentwiththeburgeoningliteraturestudyingCOVID-19,Imodelthedisasterasanunanticipatedevent(Arellanoetal.,2020;Espinoetal.,2020). InAppendixD,Ishowthatmyquantitativeresultsarerobusttoallowingfor someanticipation. AppendixDalsoprovidesrobustnesschecksincludingdifferentparameterizations,alternatedisaster severities,andtheinclusionofstockoutcostswhichserveasamotiveforfimstoraisetheirinventorytargets. 26

Figure 6: Deeper Crisis with More Adoption Note: Thefigureplotstheoutputresponsetoaproductivityshockthatmatchesthe3.40%annualdeclineinrealGDP in2020. modelmakeuseoffewermaterialinputsinproduction,causingsalestocontractmoresharply. Aseeminglyminordifferenceininventorymanagementstrategiesacrossthetwomodelsdeliversasubstantialdifferenceintheextenttowhichtheeconomyfallsintocrisisamidadisaster. The excess output loss amounts to approximately $100 billion, a figure comparable to the funds allocated to state and local governments following the passage of the CARES Act.22 Lean inventory management therefore plays a meaningful role in determining the vulnerability of the economy to unanticipated shocks. During widespread unanticipated supply disruptions, inventories can serve asastabilizingforce. 6.3 The JIT Trade-off Havingexaminedtheeffectsofleaninventorymanagementontheeconomyinnormaltimesaswell as amid a COVID-19-magnitude disaster, I next trace out frontiers that illustrate the micro-macro 22CoronavirusAid,Relief,andEconomicSecurityAct,H.R.748,116thCongress(2020). 27

Figure 7: More Stockouts and Inventory Hoarding Note: The figure plots the responses of key endogenous variables over the course of the simulated disaster in the estimatedeconomyrelativetothecounterfactualeconomy(inpercentagepoints). trade-off associated with JIT for a range of counterfactual economies. These frontiers point to an economically important trade-off and imply that inventory management is an important source of aggregatefluctuationsamidlargeunexpectedshocks. Panel (a) of Figure 8 plots the trade-off between firm value and the magnitude of the GDP contraction on impact for several counterfactual economies, each differing in steady state mass of JIT firms. The points on the curve each refer to a specific parameterized economy, traced out by varyingtheadoptioncosts,c andc . Theredcircledenotestheestimatedeconomyandtheorigin s f denotestheno-JITeconomy. ThepanelshowsthataveragefirmvalueriseswithmoreJITadoption, at the risk of elevated vulnerability to a shock. A 1.3% increase in firm value comes at the cost of an13%sharperGDPcontraction. Panel (b) of Figure 8 plots a similar trade-off, this time comparing steady state welfare gains with the magnitude of the GDP contraction. The curve again slopes upward, as welfare gains are increasing in adoption while the extent to which the economy is vulnerable to an unanticipated shockalsorises. A1.4%increaseinwelfarecomesatthecostofan13%sharperGDPcontraction. Forreference,thesameincreaseinwelfarewouldariseinamodelwithnoJITanda10%reduction 28

Figure 8: Steady State Gains vs. Macro Vulnerability (a)FirmValue (b)Welfare Note: Panels (a) and (b) plot the magnitude of GDP contraction amid a disaster on the horizontal axis. Panel (a) plotsthefirmvaluegainsintheJITeconomy’ssteadystaterelativetotheno-JITcounterfactual. Panel(b)plotsthe consumption-equivalentwelfaregains. Eachpointrepresentsadifferentcounterfactualeconomy, withtheestimated economydenotedbytheredcircleandtheno-JITeconomycoincidingwiththeorigin. Thesunkcostparameterspc ,c q s f arevariedinordertogeneratethesetofcounterfactualeconomies,andthecurvesarepolynomialinterpolationsthese counterfactuals. in economy-wide order costs. The ranges of this frontier imply an economically large trade-off betweenlong-rungainstoJITandmacrovulnerability. 6.4 Welfare Implications of JIT The exercise in the previous section underscores the vulnerabilities associated with JIT amid the realization of unexpected aggregate shocks. I next examine the welfare implications of JIT. In the calibrated exercise above, JIT remains welfare-improving. This implies that a social planner would not want to reduce the prevalence of JIT in spite of the added volatility brought on amid an unanticipatedshock. While alternative shock severities are capable of reversing this result, I find that such disasters must be far more severe than the simulated COVID-19 shock in the previous exercise. Figure 9 plots the welfare gains to JIT across a range of shock sizes. In order for the planner to prefer a no- JITworld,thenegativeproductivityshocktotheordersproducermustbealmost14%,anorderor magnitude higher than the calibrated productivity shock denoted by the red circle. As a result, the 29

Figure 9: Disaster Severity and Welfare Note: ThefigureplotsthewelfaregainsintheJITeconomyrelativetothecounterfactualagainstthesizeoftheunexpectedshock. TheredcircledenotesthewelfaregainsforJITunderthecalibratedshockintheprevioussection. estimatedmodelimpliesthatJITremainswelfare-improving,evenamidaCOVID-19-likeshock.23 7 Conclusion At the firm level, it pays to be lean. I provide empirical evidence of the benefits of JIT inventory management among public manufacturers. Upon adopting JIT, firms hold fewer inventories, and observe higher sales and smoother outcomes. JIT firms, however, appear to be more cyclical and susceptibletodisasterepisodes. Inaheterogeneousfirmsmodelinwhichthemostproductivefirms adopt JIT, lean production raises long-run firm value by 1.3% and welfare by 1.4%. At the same time, JIT elevates firm vulnerability due to low inventory buffers. Amid an unexpected supply disruption,outputintheestimatedJITeconomycontractsroughly15%morethanacounterfactual economy with no JIT. Adoption, therefore, gives rise to an important and previously unquantified trade-offwhichimpliesthatinventoriescanmatterforaggregatefluctuations. Economistsinterested 23FutureworkstudyingalternativedriversofJITsuchasinvestorpressureandimperfectcompetition,orformally modelingJITinamoregeneralnetworkstructure,couldreachdifferentwelfareconclusions. Whereastheunderlying driversofJITcanmatterforwelfare,theoutcomeofJIT,whichisleanness,mattersforthetrade-off. 30

in understanding fluctuations within firms, and the responsiveness of the economy to aggregate shocks,shouldplaycloseattentiontobothinventoriesandmanagementpractices. 31

References Ahmed, Shaghil, Andrew Levin, and Beth Anne Wilson (2004), “Recent U.S. Macroeconomic Stability: Good Policies, Good Practices or Good Luck?” Review of Economics and Statistics, 86,824–832. Alessandria, George and Horag Choi (2007), “Do Sunk Costs of Exporting Matter for Net Export Dynamics?” TheQuarterlyJournalofEconomics,122,289–336. Alessandria, George, Joseph Kaboski, and Virgiliu Midrigan (2010), “Inventories, Lumpy Trade, andLargeDevaluations.”AmericanEconomicReview,100,2304–2339. Andrews, Isaiah, Matthew Gentzkow, and Jesse M. Shapiro (2017), “Measuring the Sensitivity of ParameterEstimatestoEstimationMoments.”TheQuarterlyJournalofEconomics,132,1553– 1592. Arellano, Cristina, Yan Bai, and Gabriel Mihalache (2020), “Deadly Debt Crises: COVID-19 in EmergingMarkets.”WorkingPaper. Barrero,JoseMaria(2020),“TheMicroandMacroofManagerialBeliefs.”WorkingPaper. Barrot,Jean-NoëlandJulienSauvagnat(2016),“InputSpecificityandthePropagationofIdiosyncraticShocksinProductionNetworks.”TheQuarterlyJournalofEconomics,131,1543–1592. Bazdresch, Santiago R., JayKahn, and Toni M.Whited (2018), “Estimating andTesting Dynamic CorporateFinanceModels.”ReviewofFinancialStudies,31,322–361. Blanchard,OlivierandJohnSimon(2001),“TheLongandLargeDeclineinU.S.OutputVolatility.” BrookingsPapersonEconomicActivity,32,135–174. Brown, R.L., J. Durbin, and J.M. Evans (1975), “Techniques for Testing the Constancy of RegressionRelationshipsoverTime.”JournaloftheRoyalStatisticalSociety,37,271–285. Cachon, Gerard, Taylor Randall, and Glen Schmidt (2007), “In Search of the Bullwhip Effect.” ManufacturingandServiceOperationsManagement,4,457–479. 32

Caplin, Andrew S. (1985), “The Variability of Aggregate Demand with (S,s) Inventory Policies.” Econometrica,53,1395–1409. Carvalho, Vasco, Makoto Nirei, Yukiko Saito, and Alrieza Tahbaz-Salehi (2021), “Supply Chain Disruptions: EvidencefromtheGreatEastJapanEarthquake.”QuarterlyJournalofEconomics, 136. Clementi,Gianluca,RuiCastro,andYonsooLee(2015),“Cross-SectoralVariationintheVolatility ofPlant-LevelIdiosyncraticShocks.”JournalofIndustrialEconomics,63,1–29. Costinot, Arnad and Andres Rodriguez-Clare (2015), “Chapter 4. Trade Theory with Numbers: QuantifyingtheConsequencesofGlobalization.” Davis, Stephen J., John Haltiwanger, Ron Jarmin, and Javier Miranda (2006), “Volatility and DispersioninBusinessGrowthRates: PubliclyTradedversusPrivatelyHeldFirms.”NBERMacroeconomicsAnnual,21,107–180. Duffie,DarrellandKennethJ.Singleton(1993),“SimulatedMomentsEstimationofMarkovModelsofAssetPrices.”Econometrica,61,929–952. Eichenbaum,MartinS.(1984),“RationalExpectationsandtheSmoothingPropertiesofInventories ofFinishedGoods.”JournalofMonetaryEconomics,14,71–96. Espino,Emilio,JulianKozlowski,FernandoM.Martin,andJuanM.Sanchez(2020),“Seigniorage andSovereignDefault: TheResponseofEmergingMarketstoCOVID-19.”WorkingPaper. Gao, Xiaodan (2018), “Corporate Cash Hoarding: The Role of Just-in-Time Adoption.” ManagementScience,64,4471–4965. Hennessy, Christopher A. and Toni M. Whited (2007), “How Costly is External Financing? EvidencefromaStructuralEstimation.”JournalofFinance,62,1705–1745. Iacoviello, Matteo, Fabio Schiantarelli, and Scott Schuh (2011), “Input and Output Inventories in GeneralEquilibrium.”InternationalEconomicReview,52,1179–1213. 33

Irvine, Owen F. and Scott Schuh (2005), “Inventory Investment and Output Volatility.” InternationalJournalofProductionEconomics,93-94,75–86. Kahn,JamesA.(1987),“InventoriesandtheVolatilityofProduction.”AmericanEconomicReview, 77,667–679. Kahn, James A., Margaret M. MConnell, and Gabriel Perez-Quiros (2002), “On the Causes of the IncreasedStabilityoftheU.S.Economy.”EconomicPolicyReview,8,183–202. Khan, Aubhik and Julia Thomas (2007), “Inventories and the Business Cycle: An Equilibrium Analysisof(S,s)Policies.”AmericanEconomicReview,97,1165–1188. Khan, Aubhik and Julia Thomas (2008), “Idiosyncratic Shocks and the Role of Nonconvexities in PlantandAggregateInvestmentDynamics.”Econometrica,76,395–436. Khan, Aubhik and Julia Thomas (2016), “(S,s) Inights Into The Role of Inventories in Business CyclesandHighFrequencyFluctuations.”WorkingPaper. Kinney, Michael R and William F Wempe (2002), “Evidence on the Extent and Origins of JIT’s ProfitabilityEffects.”TheAccountingReview,77,203–225. Luo,Yulei,JunNie,XiaowenWang,andEricYoung(2021),“ProductionandInventoryDynamics UnderAmbiguityAversion.”FederalReserveBankofKansasCity,ResearchWorkingPaperno. 21-05. MacQueen,JamesB.(1966),“AModifiedDynamicProgrammingMethodforMarkovianDecision Problems.”JournalofMathematicalAnalysisandApplications,14,38–43. McCarthy,JonathanandEgonZakrajsek(2007),“InventoryDynamicsandBusinessCycles: What HasChanged?” JournalofMoney,CreditandBanking,39,591–613. McConnell, Margaret M. and Gabriel Perez-Quiros (2000), “Output Fluctuations in the United States: What Has Changed Since the Early 1980’s?” American Economic Review, 90, 1464– 1476. 34

McMahon, Michael and Boromeus Wanengkirtyo (2015), “Beyond Inventory Management: The BullwhipEffectandtheGreatModeration.”WorkingPaper. Meier,Matthias(2020),“SupplyChainDisruptions,TimetoBuildandtheBusinessCycle.”WorkingPaper. Nakamura,Maso,SadaoSakakibara,andRogerSchroeder(1998),“AdoptionofJustinTimeManufacturingMethodsatU.S.andJapanese-OwnedPlants: SomeEmpiricalEvidence.”IEEETransactionsonEngineeringManagement,45,230–240. Ohno,Taiichi(1988).InToyotaProductionSystem: BeyondLarge-ScaleProduction,Productivity Press. Ploberger, Werner and Walter Kramer (1992), “The CUSUM Test With OLS Residuals.” Econometrica,60,271–285. Porteus, Evan L. (1971), “Some Bounds for Discounted Sequential Decision Processes.” ManagementScience,18,7–11. Ramey,ValerieA.andDanielJ.Vine(2004),“TrackingtheSourceoftheDeclineinGDPVolatility: AnAnalysisoftheAutomobileIndustry.”NBERWorkingPaper10384. Rogerson, Richard (1988), “Indivisible Labor, Lotteries and Equilibrium.” Journal of Monetary Economics,21,3–16. Roumiantsev, Serguuei and Serguei Netessine (2008), “Should Inventory Policy Be Lean or Responsive? EvidenceforUSPublicCompanies.”WorkingPaper. Scarf,HerbertE.(1960),“TheOptimalityof(S,s)PoliciesintheDynamicInventoryProblem.”In MathematicalMethodsintheSocialSciences,196–202,StanfordUniversityPress,Stanford. Tauchen, George (1986), “Finite State Markov-Chain Approximations to Univariate and Vector Autoregressions.”EconomicsLetters,20,177–181. Taylor, Lucian A. (2010), “Why Are CEOs Rarely Fired? Evidence from Structural Estimation.” JournalofFinance,65,2051–2087. 35

Terry,Stephen(2017),“TheMacroImpactofShort-Termism.”WorkingPaper. Wen, Yi (2011), “Input and Output Inventory Dynamics.” American Economic Journal: Macroeconomics,3,181–212. Young,EricR(2010),“SolvingtheIncompleteMarketsModelwithAggregateUncertaintyUsing theKrusell-SmithAlgorithmandNon-StochasticSimulations.”JournalofEconomicDynamics andControl,34,36–41. 36

Appendix A Empirics This section provides summary statistics of the data used in Section 2. The section also includes furtherdetailsontheJITadopterssample,theweatherregressionresults,andanalternativemeasure ofJITamongpublicfirms. A.1 Sample Construction Mydatacomefromthreesources. First,ImakeuseofannualCompustatdatatoobtaininformation onfirm-levelinventoryholdingsaswellassales. Second,IgatherdataonJITadoptionbyreviewing firm financials and financial news. Lastly, I collect county-level weather event data from the NationalOceanicandAtmosphericAdministrationandmapthemtofirmheadquarterzipcodes. CompustatData I make use of Compustat Fundamentals Annual data from 1980-2018. I keep only manufacturers (four-digitSICcodesbetween2000-4000). Inaddition,Idropfirmyearsinwhichacquisitionsexceed5%oftotalassets(toavoidinfluenceoflargemergers). Tomitigateforanymeasurementerror, I keep only those firms with non-missing and positive book value of assets, number of employees, totalinventories,andsales. All variables are winsorized at the top and bottom 0.5% of the empirical distribution. Sales growthandemploymentgrowtharedefinedasinDavisetal.(2006)24: x ´x g “ i,t i,t´1 it 1px `x q 2 i,t i,t´1 Lastly because the focus of the paper is on JIT, an input inventory concept, I define the relevant measure of inventories to be the sum of raw material and works in process (invrm+invwip). As indicatedbyFigureA1inputinventoriesaretheprimarycontributortothedeclineinoverallinventoryholdingsinmysamplesince1980,whichcoincideswiththeprevalenceofJIT. This empirical definition also accords with the structural model developed in the main text in 24Thesegrowthratesarethesameaslogchangesuptoasecond-orderapproximation. 37

Figure A1: Compustat Inventory-to-Sales Ratios 120 100 80 60 40 001=0891 ,xednI Total inventory/sale RM inventory/sale FG inventory/sale WIP inventory/sale 1980 1990 2000 2010 2020 Year Note: Thefigureplotsaggregateinventory-salesratiosfrommysampleofCompustatfirms,byinventory-type. RM= “rawmaterials,”WIP=“worksinprocess,”andFG=“finishedgoods.” whichproducerscarrystocksofinputsacrosstime. Myfinalsampleconsistsof5,017uniquefirms. TableA1reportssummarystatisticsforthevariablesused. AdoptersDataset First, I obtained data from JIT adopters, kindly provided to me by William Wempe (from his joint workwithMichaelKinney),andXiaodanGao. ThesedataincludetheyearsinwhichaCompustat manufacturer was identified to have adopted JIT (via Form 10-K filings, press releases, among other communications. See Kinney and Wempe (2002) and Gao (2018) for further details). After verifyingthesedata,Iconductedaseparatesearchanduncoveredanadditionalsetoffirms(reported in Table A2). After linking these identified firm-years to those in my Compustat dataset, I am left with a total of 185 idenified adopters in the manufacturing sector. Figure A2 plots the empirical CDFoftheadoptersinmysampleovertime. 38

Table A1: Compustat Summary Statistics Mean Median StandardDeviation 25% 75% Employmentgrowth -0.001 0.000 0.244 -0.083 0.095 Inventory-to-sales 0.142 0.103 0.024 0.063 0.167 Capitalinvestmentrate 2.271 1.921 1.910 0.549 3.551 Logsales 4.513 4.483 2.209 2.961 6.038 Salesgrowth 0.058 0.053 0.302 -0.060 0.168 Logcash-to-assets -3.014 -2.815 1.555 -4.015 -1.830 Logsalesperworker 4.913 4.916 0.779 4.409 5.422 Logemployment -0.400 -0.481 1.924 -1.839 0.956 Ageinsample 10.824 9.000 8.527 4.000 15.000 Note: The table reports summary statistics for the relevant variables in estimation in the main text. The sample is constructedfromCompustatFundamentalsAnnualfilesfor1980-2018. Sampleconsistsof5,017uniquefirms. Table A2: Additional JIT Adopters Firm Compustat gvkey FordMotors 4839 GeneralMotors 5073 Dell 14489 Motorola 7585 NCR 7648 SunriseMedical 10185 Tellelabs 10420 VanDornCo 11101 DonnellyCorp 14462 Tuscarora 14578 Selectron 17110 HoneywellInc 5693 ADCTelecommunications 1013 Sunbeam 1278 Boeing 2285 Campbell 2663 CascadeCorporation 2802 Caterpillar 2817 Note: ThetablereportstheadditionalJITadoptersthatwereaddedtotheoriginalsetofadopters. 39

Figure A2: Adopters by Year 1 .8 .6 .4 .2 0 sretpoda fo FDC laciripmE 1980 1990 2000 2010 Year Note: ThefigureplotstheempiricalcumulativedensityfunctionforJITadoptioninthesample. 40

Whereas in the model I account in part for cross-sectional measurement error, measurement errorintheyearinwhichJITisrecordedtohaveoccurred,couldalsobeaconcern. If,forinstance, afirmadoptsJITinagivenyear,butdoesnotannouncethatitisaJITfirmuntilasubsequentyear, then the primary measure of JIT utilized in the main text would be subject to an additional form of measurement error. While such measurement error would imply that my reported estimates are attenuated,FigureA3providesevidencethattherecordedyearsofadoptionareaccurate. Irunthe followingregression: y “ βadopt `δ `δ `ε ijt ijt jt i ijt where the outcomes of interest are the level and first difference in inventory-to-sales ratio, and adopt is an indicator taking on a value of one only in the recorded year of adoption. Industryijt by-year and firm fixed effects are also specified. The figure plots 95% confidence intervals for a three-yearwindowaroundtherecordeddateofadoption,andshowsthatinventoryholdingsdecline intheyearofadoption. 41

Figure A3: Validation of JIT Indicator .1 .05 0 −.05 −.1 −.15 selas−ot−yrotnevni no tceffE −3 −2 −1 0 1 2 3 Years since JIT adoption Note: ThefigureplotstheestimatedeffectofJITadoptiononthelevelofinventory-to-sales. 95%confidencebands aredisplayedalongisdepointestimates. A.2 Additional Results on JIT Adoption and Firm Profitability Beyond the results reported in Table 1, Table A3 reports additional reduced-form estimates of JIT and firm profitability. Column 1 reports the results from a regression of sales per worker, a basic proxy for productivity, on the JIT adoption indicator as well as fixed effects and other controls. Theresultindicatethatfollowingadoption,JITfirmsexperiencea5%increaseinsalesperworker relativetonon-JITfirms. Furthermore, column 2 considers the relationship between JIT adoption and a firm’s forecast accuracy. I first merge IBES Guidance data with my JIT adoption data set. The IBES Guidance database provides information on managers’ forecasts about their own firm outcomes. I focus on one-yearearningsforecasts. AftermergingthesedatawithmyJITadoptiondataset,Iobtainasampleof453uniquemanufacturingfirmsspanningtheyears1995-2018. Thesamplemeanofforecast errorsis0.037withastandarddeviationof1.431. Icomputesquaredforecasterrorsastherelevant measureofforecastaccuracy. Ithenregressthesquarederrorsontheadoptionindicatoraswellas 42

Table A3: Additional Results on JIT Adoption and Firm Profitability (1) (2) Salesperworker Squaredforecasterror Adopter 0.051** -0.515** (0.023) (0.227) Fixedeffects Firm,IndustryˆYear Industry,Year Observations 45,357 2,243 Note: ThetablereportspanelregressionresultsfromCompustatAnnualFundamentalsbasedonregression(1). The regressorofinterestisthefirm-yearspecificadoptionindicator. Controlvariablesinthefirstcolumnincludefirmage insample,firmsize,andsharesoutstanding. Standarderrorsareclusteredatthefirmlevel. Thestandarddeviationsof thedependentvariablesare0.78and2.49,respectively. ***denotes1%significance,**denotes5%significance,and *denotes10%significance. industry and time fixed effects. The results indicate that following adoption, JIT firms observe a -0.515declineintheirsquaredforecasterrors,aroughly20%standarddeviationreduction. Jointly, thesefactssuggestthatJITfirmsaremoreproductiveandarebetterabletopredicttheirprofitability. ThemodelimpliestheJITproducersobserveanincreaseinsalesperworker,andtheimproved accuracylendssupporttomodelingJITthroughareductioninordercostssubjecttoadoptioncosts. A.3 Additional Weather Event Regressions Inthissection,IfurtherexaminethesensitivityofJITfirmstoweatherevents. Ratherthanconsidering weather events that directly hit the firm’s headquarters, I instead focus on upstream weather events. Todoso,IutilizetheCompustatSegmentfilesinordertolinkcustomerandsupplierfirms. TheStatementofFinancialAccountingStandards(SFAS)No. 131requirespublicfirmstodisclose the identity of any customer representing more than 10% of total sales. After linking downstream JITandnon-JITfirmstotheirmajorsuppliersbasedontheinformationdisclosedthroughthisregulation, I proceed to link the upstream suppliers to weather events realized in the zip codes where theyareheadquartered. Theseriesofweatherevent-to-supplier-to-customer-toJITstatuslinksconsiderably reduce the sample size to roughly 200 public manufacturers. Nevertheless, I estimate a negative and significant effect of upstream weather events on downstream firm sales and employ- 43

Table A4: JIT Adoption and Sensitivity to Local Disasters Sales Employment Upstreamdisaster -0.080** -0.070* (0.034) (0.036) Adopterˆupstreamdisaster -0.085* -0.071* (0.047) (0.040) Controls Yes Yes FixedEffects Firm,Zip,Year Firm,Zip,Year Observations 1,139 1,139 Note: The table reports weather event regressions described in the text. The independent variable of interest is the interactionbetweentheadoptionindicatorandtheaveragenumberofupstreamweathereventsexperiencedbyadownstreamfirms’suppliers. Controlvariablesincludecustomerandsupplierage,propertydamageassociatedwithweather events,customerandsupplierfinishedgoodsinventories,salesperworker,averageorderbacklogforsuppliers,adopter indicator, andthedisasterindicator. Standarderrorsareclusteredatthefirmlevel. ***denotes1%significance, ** denotes5%significance,and*denotes10%significance. ment,withgreatersensitivityamongdownstreamJITproducers. Irunthefollowingregression: y “ ψ adopter `ψ disaster `ψ radopter ˆdisaster s`X1 β `δ `δ `ω , ijt 1 ijt 2 ijt 3 ijt ijt ijt i t ijt where, as before, the subscripts refer to firm i belonging to industry j in year t. The regression resultsarereportedinTableA4. Theupstreamdisastervariableistheaveragenumberofdisasters experienced by a downstream firms’ major suppliers in a given fiscal year. A unit increase in the averagenumberofdisastersexperiencedbyafirm’ssupplierspredictsa7-8%declineinfirmssales andemployment,witharoughlysimilarexcessdeclineforJITfirms. 44

A.4 Alternative Measure of JIT Adoption Toexploretherobustnessofmytext-basedmeasureofJITadoption,Idevelopanalternatemeasure ofadoptionamongpubliclytradedfirmsbytakingatimeseriesapproach. Theapproachedisbased on the intuition that JIT adopters commit to leaning out. Since they commit to reducing inventory holdings,itstandstoreasonthatuponadoptingJIT,afirmwouldexperienceastructuralbreakinits meaninventory-to-salesratio. IimplementacumulativesumofresidualstesttodetectJITadoption among the majority of firms in my sample (Ploberger and Kramer, 1992; Brown et al., 1975). To operationalizethestructuralbreaktest,Inarrowmyfocustofirmswithatleastfiveyearsworthof consecutiveobservations. Inthissection,Iverifythatthestructuralbreakapproachdeliverssimilar empirical results to the ones reported in Section 2. An appealing aspect of this approach is that it successfully detects JIT adoption among the firms identified through the text-based approach, as wellasawidervarietyoffirms(i.e. manufacturersandretailers.). CUSUMApproachandValidation I implement a CUSUM test of OLS residuals for each individual firm time series of inventory-tosales.25 In particular, I am interested in detecting structural breaks in the mean level of inventoryto-sales. Let y be a firm’s inventory-to-sales ratio in year t, this approach requires constructing a t p teststatisticbasedontheOLSresiduals,rp “ y ´β , t t 0 ř Tz rp Bpzq “ t?“1 t σp T where z denotes the break period. When |Bpzq| exceeds the relevant critical value, then a break is statisticallydetected.26 In all, this alternate dataset identifies the years in which approximately 560 firms adopted JIT. My final sample consists of an unbalanced panel of 4,005 unique firms spanning the same period as in the main text, and a wider range of industries. Due to the spell length restriction required to implementtheCUSUMtestwithOLSresiduals,thecompositionoffirmsacrossthesampleinthe 25IalsoimplementedastandardCUSUMtestandobtainedqualitativelysimilarresults. 26Iselectasignificancelevelof10%. 45

Table A5: JIT Adoption and Firm Profitability (1) (2) Inventory-to-sales Sales Adopter -0.629*** 0.164*** (0.021) (0.039) Fixedeffects Firm,IndustryˆYear Firm,IndustryˆYear Observations 37,266 37,266 Note: ThetablereportspanelregressionresultsfromCompustatAnnualFundamentalsbasedonregression(1). The regressorofinterestisthefirm-yearspecificadoptionindicator. Firmageinthesampleisspecifiedasacontrolvariable. Four-digitSICcodesarespecifiedintheindustry-by-yearfixedeffects. Standarderrorsareclusteredatthefirmlevel. Thestandarddeviationsofthedependentvariablesare0.84and2.08, respectively. ***denotes1%significance, ** denotes5%significance,and*denotes10%significance. maintextandthesampleusedinthissectiondiffer. However,oftheJITadoptersthatresideinboth datasets,thestructuralbreakapproachpicksuproughly72%ofthe(text-based)JITadopters. EmpiricalFacts Using these data, I revisit the four facts about JIT adopters presented in the main text. First, JIT adoption is associated with both lower inventory holdings and higher sales. Based on Table A5 , adopters experience a much sharper 63% decrease in inventory-to-sales ratios and a 16% increase in sales. The results imply a change of -75% and 8% of one standard deviation in the outcomes, respectively. Themagnitudesinthetablearelargerthanthosereportedinthemaintext,mainlydue to the fact that adopters are defined as firms that experience sufficiently large drops in inventory holdingssoastotriggerarejectionofthenullhypothesis. TableA6onceagainindicatesthatJITadoptersexperiencelessmicrovolatility. Basedonthese estimates,adoptersseeaslightlylower2-3%decreaseinmeasuredvolatility. Third, JIT adopters are more cyclical, as shown in Table A7. The table indicates that a 1% increase in GDP growth is associated with a roughly 0.78% increase in sales growth among nonadopters. Adopters experience an additional sales growth increase of 0.18% above this baseline. Similarly,a1%increaseinGDPgrowthisassociatedwitha0.77%increaseinemploymentgrowth, witha0.16%furtherincreaseforadopters. 46

Table A6: JIT Adoption and Firm Volatility (1) (2) Std. salesgrowth Std. employmentgrowth Adopter -0.032*** -0.022*** (0.008) (0.007) Fixedeffects IndustryˆYear IndustryˆYear Observations 14,647 14,647 Note: ThetablereportspanelregressionresultsfromCompustatAnnualFundamentalsbasedonregression(2). The regressorofinterestisthefirm-yearadoptionindicator. Alagofthedependentvariableisspecifiedasacontrol. FourdigitSICcodesarespecifiedintheindustry-by-yearfixedeffects. Standarderrorsaredoubleclusteredatthefirmand yearlevels. ***denotes1%significance,**denotes5%significance,and*denotes10%significance. Finally, Table A8 shows that JIT adopters are more sensitive to local weather events. While slightly lower, the point estimate in the first column implies that a local weather event predicts abouta3%declineinfirmsales. 47

Table A7: JIT Adoption and Cyclicality Salesgrowth Employmentgrowth GDPgrowth 0.781*** 0.774*** (0.165) (0.124) AdopterˆGDPgrowth 0.176* 0.155* (0.099) (0.091) Controls Yes Yes FixedEffect Industry Industry Observations 30,016 30,016 Note: ThetablereportsregressionresultsfromCompustatAnnualFundamentalsbasedonregression(3). TheindependentvariableofinterestistheinteractionbetweentheadopterindicatorandGDPgrowth. Controlvariablesinclude firmageinthesample,cash-to-assets,sales-per-worker,aswellastheadoptionindicator. Four-digitSICfixedeffects arespecified. Standarderrorsareclusteredatthefirmlevel. ***denotes1%significance,**denotes5%significance, and*denotes10%signficance. Table A8: JIT Adoption and Sensitivity to Local Disasters Sales Employment Disaster 0.009 0.009 (0.006) (0.006) AdopterˆDisaster -0.021* -0.022* (0.012) (0.012) Controls Yes Yes FixedEffects Firm,Year Firm,Year Observations 27,779 27,779 Note: The table reports weather event regressions from a sample of Compustat firms based on regression (4). The independent variable of interest is the interaction between the adoption indicator and the disaster indicator. Control variables include capital investment rate, sales per worker, ratio of cost of goods to sales, finished goods inventory holdings,adopterindicator,andthedisasterindicator. Standarderrorsareclusteredatthefirmlevel. ***denotes1% significance,**denotes5%significance,and*denotes10%significance. 48

Appendix B Model B.1 Order Threshold for Final Goods Firm The firm’s problem delivers a threshold rule for placing an order. In particular, a firm places an orderifandonlyiftheordercostdrawislowerthanathresholdordercost: ξ ă ξ˚pz,s,aqwhere r pqs`V˚pz,s,aq´VPpz,s,aq ξpz,s,aq “ (11) φ and ` ` ˘ ˘ r ξ˚pz,s,aq “ min max ξ,ξpz,s,aq ,ξ (12) B.2 Intermediate Goods Firm Inrecursiveform,thevalueoftheintermediategoodsfirmis: " „  * WpKq “ max p qKαL1´α `p1´δqK ´K1 ´wL `βWpK1q K1,L DuetotheCobb-Douglasproductiontechnologyassumed,theintermediategoodsfirm’svalue canbeexpressedasalinearfunctionoftheaggregatecapitalstock. Giventhis,onecansolveforq analytically. Inparticular,therelativepriceoftheintermediategoodis: ˆ ˙ ˆ ˙ 1´βp1´δq α w 1´α q “ βα 1´α B.3 Equilibrium Anequilibriumisasetoffunctions, (cid:32) ( VA,VO,V˚,VP,W,s˚,s1,ξ˚,a1,K,L,p,w,q,Γ , µ suchthat: 49

1. Thehousehold’sfirstorderconditionshold: 1 p “ , w “ φC. C 2. Theintermediategoodsfirmfirstorderconditionshold: ˆ ˙ K α p “ βW1pK1q w “ p1´αqq . L 3. VA,VO,V˚,VP solvethefinalgoodsfirm’sproblem. 4. Marketforfinalgoodsclears: ż ż ż ż C “ ypz,s˚,s1,a,ξqdFpξ˚qdµpz,s,aq` ypz,s,s1,a,ξqr1´dFpξ˚qsdµpz,s,aq´I. 5. Marketforordersclears: ż ż O “ rs˚pz,s,a,ξq´ssdFpξ˚qdµpz,s,aq. 6. Marketforlaborclears: ż ż ż ż H “ npz,s˚,s1,ξqdFpξ˚qdµpz,s,aq` npz,s,s1,a,ξqr1´dFpξ˚qsdµpz,s,aq ż „ż  ż „  ξ˚pz,s,aq qp1´αq α 1 ` ξdFpξq dµpz,s,aq` a1pz,s,aqrp1´aqc `ac sdµpz,s,aq` K. s f w 0 7. Theevolutionofthedistributionoffirmsisconsistentwithindividualdecisions: ż ż ż Γ pz,s,aq “ 1 dµpz,s,aqdFpξqdΦpε q µ A z Apz1,s1,a1,ξ,ε ;µq “ tpz,s,aq|s1pz,s,a,ξ;µq “ s1,z1 “ ρ z `σ ε ,a1pz,s,a,ξ;µq “ a1u z z z z Φpxq “ Ppε ď xq, z 50

andthecapitalstockevolvesaccordingto K1 “ p1´δqK `I. B.4 Numerical Solution The model is solved using methods that are standard in the heterogeneous firms literature. The exogenousproductivityprocessisdiscretizedfollowingTauchen(1986)whichallowsmetoexpress the AR(1) process for log firm productivity as a Markov process. I select N “ 11 grid points for z idiosyncraticproductivity. Furthermore,IselectN “ 200gridpointsfortheendogenousinventory s holdings state. Finally, the binary adoption state implies that the discretized model has 4,400 grid points. I solve for the policy functions via value function iteration which is accelerated by the use of the MacQueen-Porteus error bounds (MacQueen, 1966; Porteus, 1971). This acceleration method makesuseofthecontractionmappingtheoremtoobtainboundsforthetrue(infinitehorizon)value function. These bounds are used in order to produce a better update of the value function. The ergodic distribution of firms is obtained via nonstochastic simulation as in Young (2010). This histogram-based method overcomes sampling error issues associated with simulating individual firmsinordertoobtainthestationarycross-sectionaldistribution. Operationally,Isolvethemodelbyinitiatingaguessofthefinalgoodsprice,p . Ithencompute 0 q and w given the guess p . From here, I solve the firm’s problem via value function iteration 0 0 0 and then obtain the ergodic distribution. Using the policies and ergodic distribution, I compute aggregates and the associated market clearing error using the household’s optimality condition. I updatethepricebasedonthiserrorusingbisection. For the unexpected shock exercise, I implement a shooting algorithm. I first set the duration of the disaster to be a predetermined length T, so that the model returns to steady state at T ` 1. Based on this, I solve the final goods firms problem backwards, obtaining a set of time-indexed policy functions. Using these policies, I then push the distribution of final goods firms forward. Withthetime-indexedpoliciesandweightsinhand,Icomputeaggregatesateachpointintimeand 51

iterateonpricesuntiltheordersmarketclears. 52

Appendix C Estimation Inthissection,Idetailtheestimationofthemodelandprovideadditionalresultsrelatingtoidentification. Thefinalsubsectionreportstheheadlineresultsbasedonanalternatecounterfactual. C.1 Simulated Method of Moments ` ˘ 1 The parameter vector to be estimated is θ “ ρ σ ξ ξ ξ c c c τ . Estimating θ z z NA NA A s f m requires making a guess θ , solving my plant-level model, and simulating a panel of firms from 0 which I compute the different moments. I define a firm to be composed of ten plants and simulate a panel of firms roughly eight times the size of the panel in Compustat. A firm is defined to be an adopter if at least one of the ten plants adopt JIT, consistent with the classification of JIT firms in my sample. I discard the first 25 years of simulated data so as to minimize the impact of initial values. IthencollectthetargetedempiricalmomentsinastackedvectormpXqwhichcomesfrom my Compustat sample. I next stack the model-based moments, which depend on θ, in the vector p mpθq. FinallyIsearchtheparameterspacetofindtheθ thatminimizesthefollowingobjective ` ˘ ` ˘ 1 min mpθq´mpXq W mpθq´mpXq θ whereW istheoptimalweightingmatrix,definedtobetheinverseofthecovariancematrixofthe moments. I obtain the covariance matrix via a clustered bootstrap, allowing for correlation within firms. Iestimatetheparametervectorviaparticleswarm,astandardstochasticglobaloptimization solver. p Thelimitingdistributionoftheestimatedparametervectorθ is ? p d Npθ´θq Ñ Np0,Σq where ˆ ˙„ˆ ˙ ˆ ˙ 1 Bmpθq 1 Bmpθq ´1 Σ “ 1` W S Bθ Bθ and S is the ratio of the number of observations in the simulated data to the number of observa- 53

Figure C1: Monotonic Relationships Note: The figure plots the changes in select moments to changes in the parameters, in percentage points relative to momentatestimatedparametervalues. tions in the sample.27 I obtain the standard errors by computing the secant approximation to the partial derivative of the simulated moment vector with respect to the parameter vector. Given the discontinuitiesinducedbythediscretizedstatespace,Iselectastepsizeof1%. C.2 Identification The 11 moments jointly determine the nine parameters that reside in vector θ. To supplement the discussiononmonotonerelationshipsfromthemaintext,FigureC1reportsthemonotonerelationships between selected moments and parameters. Figure C2 reports the sensitivity of each of the nine parameters to changes in each of the moments. These results come from an implementation p ofAndrewsetal.(2017). Inparticular,thesensitivityofθ tompθqis „ˆ ˙ ˆ ˙ ˆ ˙ Bmpθq 1 Bmpθq ´1 Bmpθq 1 Λ “ ´ W W Bθ Bθ Bθ 27S issettobeapproximately8. 54

I then transform this matrix so as that the interpretation of the coefficients is the effect on each parameterofaonestandarddeviationchangeintherespectivemoments. 55

Figure C2: Sensitivity 0.1 0.04 0.05 0.02 0 -0.05 0 -0.1 -0.15 -0.02 M(is M A ( ) isNA) V C (is (i A sA ) ,lsA) V(ls C V A ( ( ) i i s s N N A A , ) lsNA V ) (lsNA) spA spNA F(A) M(is M A ( ) isNA) V C (is (i A sA ) ,lsA) V(ls C V A ( ( ) i i s s N N A A , ) lsNA V ) (lsNA) spA spNA F(A) 0.02 0.06 0.01 0.04 0 0.02 -0.01 0 -0.02 -0.03 -0.02 M(is M A ( ) isNA) V C (is (i A sA ) ,lsA) V(ls C V A ( ( ) i i s s N N A A , ) lsNA V ) (lsNA) spA spNA F(A) M(is M A ( ) isNA) V C (is (i A sA ) ,lsA) V(ls C V A ( ( ) i i s s N N A A , ) lsNA V ) (lsNA) spA spNA F(A) 0.1 0.1 0.05 0 0 -0.1 -0.05 -0.2 M(is M A ( ) isNA) V C (is (i A sA ) ,lsA) V(ls C V A ( ( ) i i s s N N A A , ) lsNA V ) (lsNA) spA spNA F(A) M(is M A ( ) isNA) V C (is (i A sA ) ,lsA) V(ls C V A ( ( ) i i s s N N A A , ) lsNA V ) (lsNA) spA spNA F(A) 10-3 4 0.2 2 0.1 0 -2 0 -4 -0.1 -6 -0.2 M(is M A ( ) isNA) V C (is (i A sA ) ,lsA) V(ls C V A ( ( ) i i s s N N A A , ) lsNA V ) (lsNA) spA spNA F(A) M(is M A ( ) isNA) V C (is (i A sA ) ,lsA) V(ls C V A ( ( ) i i s s N N A A , ) lsNA V ) (lsNA) spA spNA F(A) 0.04 0.02 0 -0.02 -0.04 -0.06 M(is M A ( ) isNA) V C (is (i A sA ) ,lsA) V(ls C V A ( ( ) i i s s N N A A , ) lsNA V ) (lsNA) spA spNA F(A) 56 Note: ThefigureplotssensitivityestimatesasinAndrewsetal.(2017). Theseestimatesdescribethechangesineach oftheeightparameterstoaonestandarddeviationincreaseineachmoment.

C.3 Subsample Estimation Inthissection,Iprovideestimatesofthemodelacrosstwosub-periodswherethebaselineestimates encompasstheyears1990-2018andthecounterfactualisestimatedfromovertheyears1980-1989. This alternate counterfactual offers a different point of comparison, a benchmark world in which JITisnotabsentbutsimplylessprevalent. TableC1reportstheestimatedparametersfromthetwomodels. Inordertohighlightdifference relating directly to the incentives to adopt JIT, I re-estimate only the two adoption cost parameters for the earlier sample. As a result, the counterfactual holds firm level technologies fixed as well as theordercosts,carryingcost,andmeasurementerrorestimates. Table C1: Parameterizations for Subsamples Description Parameter 1980-1989 1990-2018 Idiosyncraticproductivitypersistence ρ 0.867 0.867 z - (0.003) Idiosyncraticproductivitydispersion σ 0.021 0.021 z - (0.001) Ordercostlowerbound(non-adopters) ξ 0.049 0.049 NA - (0.0004) Ordercostupperbound(non-adopters) ξ 0.450 0.450 NA - (0.004) Ordercostupperbound(adopters) ξ 0.095 0.095 A - (0.001) Sunkcostofadoption c 0.284 0.225 s (0.004) (0.001) Continuationcostofadoption c 0.080 0.072 f (0.001) (0.001) Carryingcost c 1.239 1.239 m - (0.006) Observedshareofnon-adopters τ 0.938 0.938 - (0.0001) Note: Thetablereportstheestimatedparametersforthesubsamples(standarderrorsinparentheses). Parameterswere estimated by targeting the same 11 moments. All nine parameters are estimated for the 1990-2018 sample whereas onlyc andc areestimatedforthe1980-1989sample. s f The parameters are precisely estimated. The upfront sunk cost of adoption estimated from the 1980’s sample is higher at around 26% of the sunk cost estimated in the later sub-sample. The relatively lower sunk cost today implies that it has become easier to initiate JIT production. The 57

Table C2: Model vs. Empirical Moments Moment 1980-1989 1990-2018 Model Data Model Data Mean(inventory-salesratio|adopter) 0.112 0.111 0.097 0.092 (0.007) (0.005) Mean(inventory-salesratio|non-adopter) 0.145 0.162 0.116 0.139 (0.003) (0.002) Std(inventory-salesratio|adopter) 0.053 0.055 0.059 0.054 (0.001) (0.001) Corr(inventory-salesratio,logsales|adopter) 0.068 -0.213 -0.151 -0.097 (0.001) (0.001) Std(logsales|adopter) 0.235 0.153 0.225 0.209 (0.004) (0.005) Std(inventory-salesratio|non-adopter) 0.071 0.153 0.073 0.164 (0.002) (0.002) Corr(inventory-salesratio,logsales|non-adopter) -0.212 -0.292 -0.340 -0.278 (0.001) (0.001) Std(logsales|non-adopter) 0.310 0.276 0.277 0.305 (0.003) (0.003) Spike(inventory-sales|adopter) 0.035 0.076 0.051 0.042 (0.028) (0.013) Spike(inventory-sales|non-adopter) 0.246 0.239 0.137 0.167 (0.007) (0.006) Frequencyofadoption 0.051 0.012 0.061 0.054 (0.002) (0.005) Note: The table reports the model-based moments and the empirical moments for the estimated subperiod models. Standarderrorsofmomentsaredisplayedinparentheses. estimatedperperiodcontinuationcostsofremaininganadopterareabout11%higherintheearlier sample. While the probability of remaining an adopter conditional on already being one remains at around 95% across both periods, the steady state mass of adopters inthe earlier sample is about 35%lowerthanthelatersample. Table C2 reports the fit of the models. As expected, the baseline 1990-2018 sample provides a more successful fit than the constrained 1980-1989 counterfactual. Table C3 reports the steady state comparisons across models, similar to Table 9. Unsurprisingly, the difference in long-run aggregatesisattenuatedwhenthecounterfactualfeaturessomeadoption. Forinstance,steadystate outputrisesbyabout3.4%inthe1990-2018periodrelativeto1980-1989. Nonetheless,theresults 58

arequalitativelythesame: output,measuredTFP,andwelfareallriseasmorefirmsadoptJIT. Table C3: Long-Run Aggregates Across Models Output Order frequency Order size Price of orders 3.40 14.67 -7.68 1.78 Inventory stock Firm value Measured TFP Welfare -17.38 1.84 0.91 1.11 Note: The table reports steady state values of the estimated model relative to the counterfactual model, in percent deviations. Finally, Figure C3 reproduces the trade-off exercise in Section 6 of the main text. The figure plots two points. One of these points, denoted by the red ‘+’, illustrates the trade-off discussed in the main text. In particular, when estimating the model with the full 1980-2018 sample and comparing it to a no-JIT counterfactual, the firm value gains are at around 1.3% while the excess output contraction amid a one-year disaster is around 13%. The blue ‘x’ repeats this exercise but for the alternative baseline and alternative counterfactual from the sub-sample analysis. In other words, comparing the steady states of the model estimated from 1990-2018 to the counterfactual encompassing 1980-1989, I find that firm value gains to JIT are slightly higher, at about 1.8%. The slightly higher gains to adoption based on this exercise are attributed to the higher estimated adoption and order costs for non-JIT producers relative to the baseline estimates in the main text. Furthermore,whencalibratinganunanticipateddisastertogeneratea3.40%contractionthe1990- 2018 model, the excess output contraction relative to the contraction in the 1980’s model is about 15%closetothebaselineestimate. 59

Figure C3: Trade-off from Sub-Sample Analysis Note: Thefigureplotstwopointsofthetrade-offdescribedinthemaintext. Thered‘+’illustratesthetrade-offbetween the1980-2018modelandano-JITcounterfactual. Theblue‘x’plotsthetrade-offarisingfromthesub-sampleanalysis. 60

Appendix D Robustness In this section I provide different robustness checks relating to the JIT trade-off presented in the main text. I begin by examining the sensitivity of the trade-off to different parameter values. I then consider a more severe unanticipated shock. Following this exercise, I analyze the role that anticipation plays in the headline results. Next, I study a version of the model that incorporates stockoutcostswhichserveasamotiveforfirmstocarrymoreinventoriesinnormaltimes. Lastly, Ientertainanalternateordercostdistribution. D.1 Alternate Parameterizations Table D1: Robustness Parameterizations Description Parameter Value Value Idiosyncraticproductivitypersistence ρ 0.950 0.550 z Idiosyncraticproductivityvolatility σ 0.100 0.010 z Ordercostlowerbound(non-adopters) ξ 0.000 0.050 NA Ordercostupperbound(non-adopters) ξ 0.600 0.300 NA Ordercostupperbound(adopters) ξ 0.100 0.050 A Carryingcost c 1.300 0.800 m Note: Thetablereportsthealternateparameterizationschosentocomputethefirmvalue-aggregatecontractiontradeoffassociatedwithJIT. Table D1 reports a number of different parameter specifications. I vary all parameters in different directions with the exception of the adoption costs which trace out the frontier displayed in Figure 8. FigureD1plotsthegapfirmvaluegainsagainstthesizeoftheexcesscontractionamida disasterbetweentheJITandno-JITeconomies. Thered‘+’denotestheheadlinefigureinthemain textwhiletheblue‘x’reportstheresultsfromthealternateparameterizations. Acrossallspecifications,thefirmvaluegainsarerobustlycoupledwithexcessoutputcontractions,demonstratingthat themicro-macrotrade-offconsistentlyremainsthroughouttherangeofcounterfactualexercises. 61

Figure D1: Trade-off Across Alternate Parameterizations Note: ThefigureplotsthefirmvaluegainsagainstthesizeoftheexceessGDPcontractionintheJITmodelrelative totheno-JITcounterfactual. Thered‘+’denotestheheadlinefindinginthemaintextwhiletheblue‘x’reportsthe trade-offsarisingfromthedifferentparameterizationsdetailedinTableD1. D.2 Disaster Size Figure D2 reports the excess output contraction in the estimated economy relative to the counterfactual for a larger disaster, mimicking the contractions observed in the UK and France in 2020. Similar to the baseline findings, the estimated JIT economy contracts more sharply than the counterfactual with no-JIT adoption amid an unexpected drop in A. Over the course of the disaster, the estimated JIT economy contracts by roughly 10% while the counterfactual contracts by 8.7%, implyingthattheJITeconomyexperiencesa15%largercontraction. 62

Figure D2: Larger Supply Disruption Note: Thefigureplotstheevolutionofoutputacrossbothmodelsforadisasterthatdeliversa10%contractioninthe estimatedmodelonimpact. D.3 Incorporating Partial Anticipation In this subsection, I allow for there to be uncertainty as to whether the disaster occurs in period t. This uncertainty is fully resolved following t regardless of whether or not the disaster comes to pass. LetλdenotetheprobabilitythatthelargeaggregateshocktoAisrealizedattimet. Recallthat finalgoodsfirmsfacethefollowingproblemintheproductionstage: “ ‰ VPpz,sr,aq “ max πpz,sr,s1,aq`βE VApz1,s1,a1q s1Pr0,srs Inperiodt´1,however,theexpectationisnotonlytakenacrossidiosyncraticproductivityrealizationsbutacrosstherealizationofthedisasteraswell: VApz1,s1,a1q “ λVA pz1,s1,a1q`p1´λqVApz1,s1,a1q Disaster SS Theintermediategoodsproducersimilarlyfacesuncertaintyoverwhetherthedisasterwillcome 63

topass. Asaresult,thefirstorderconditiongoverningtheoptimalinvestmentchoicebecomes: „  BWpA1,K1q BWp1,K q SS p “ β λ `p1´λq BK1 BK SS Ievaluatethedynamicsamidthedisastershockbynumericallyimplementinganalgorithmsimilar to the unanticipated case. I begin with an initial guess for prices and work backwards to obtain a sequence of time-indexed value and policy functions. With these in hand, I proceed to a forward step in which I push the distribution of firms forward across time utilizing the optimal policies from the backward step. From here, I compute aggregates, check for market clearing, and update thepricesuntilconvergencetoaspecifiedtolerance. Figure D3 plots three relevant quantities. On the right axis, I plot the changes in inventory and capital stocks in the period leading up to the shock. As the likelihood that the widespread supplydisruptionwillcometopassrises,weobserveanincreaseineconomy-wideinventorystocks accumulated by firms in anticipation of the shock. Intuitively, with the prospect of a widespread disaster on the horizon, firms will optimally hold added precautionary stocks of inventories. At the same time, we observe a rise in capital investment which leads to a higher K1 in period t´1. Theincreaseincapitalinvestmentservestopartiallybluntthespikeinthepriceofordersamidthe disaster. On the left axis, I plot the excess output contraction experienced in the estimated economy relative to the counterfactual (in percent). Importantly, despite the added precautionary inventory holdings among firms, and the increase in capital investment, there is still a sizable excess drop in output across the two economies, indicating that the JIT trade-off documented in the main text is robust to the anticipation modeled here. Intuitively, this is due to the fact that the distribution of individualfirmoutcomesistruncatedontheleft. Theworstcasescenarioforfirmsinthemodelis stocking out and earning zero profits. As a result, even with partial anticipation, firms do not fully appreciatethelargenegativeshockthatmightcometopass. 64

Figure D3: JIT Trade-off Robust to Anticipation Note: Thedots,whichcorrespondtotheleftaxis,displaytheexcessoutputcontraction(relativetotheno-JITcounterfactual)fordifferentdisasterprobabilties. Ontherightaxis,thebarplotreportsthepercentincreaseininventoryand capitalstockspriortotheshock,duetoanticipation. 65

D.4 Incorporating Stockout Costs The model in the main text assumes that firms have the option to “stock out” and simply forgo productioninagivenperiodiftheydonotreceivefavorablez andξ realizations. Inthiscaseprofits arezero. Inreality,however,stockoutscanbecostlyforfirms,particularlyifonetakestheviewthat firms accumulate customer capital or goodwill. In the event of a stockout, firms might risk losing their customer base, or otherwise developing a reputation for poor management. In this section, I exploretherolethatcostlystockoutsplayinquantifyingtheJITtrade-off. Intuitively, when it is costly for firms to stockout, they will carry more inventories with them anddrawexistingstocksdownmoreslowly. Imodelthestockoutcostinthefirm’sproductionstage decision: “ ‰ VPpz,sr,aq “ max πpz,sr,s1,aq`βE VApz1,s1,a1q (13) s1Pr0,srs where „  c πpz,sr,s1,aq “ p znpz,sr,s1,aqθnpsr´s1qθm ´wnpz,sr,s1,aq´ m ps1q2 (14) 2 are period profits. A “stock out” occurs when sr“ 0, so material input usage is zero and the firm produces no output. Rather than earning profits equal to zero amid a stockout, πpz,0,0,aq “ 0, I assumethatfirmsmustpayastockoutcost,c ą 0.28 so Ire-estimatethemodelwiththisadditionalparameterforatotaloftenparametersestimatedby targeting the same 11 moments. Table D2 reports the estimation results and Table D3 reports the model fit. Compared to the estimated model in the main text, economy-wide inventory stocks in thestockoutcostmodelareroughly20%larger. Figure D4 produces a figure similar to Figure C3, where I compare the headline findings in the main text with those implied by a model featuring stock out costs. Since stockout costs raise a firm’s motive to carry more inventories, the gains to JIT are slightly less pronounced. With stockout costs, the firm value gains to JIT in normal times amount to roughly 1.2%. At the same time, the unanticipated shock now implies an excess output contraction of about 10% in the JIT economy relative to the counterfactual. All things considered, the stockout cost model delivers a 28Thisisessentiallyageneralizationofthebaselinemodelinwhichc “0. so 66

quantitativelysimilartrade-offtotheonedocumentedinthemaintext. Table D2: Stockout Cost Model Estimates Description Parameter Estimate Idiosyncraticproductivitypersistence ρ 0.813 z (0.001) Idiosyncraticproductivitydispersion σ 0.029 z (0.002) Ordercostlowerbound(non-adopters) ξ 0.006 NA (0.001) Ordercostupperbound(non-adopters) ξ 0.307 NA (0.006) Ordercostupperbound(adopters) ξ 0.047 A (0.003) Sunkcostofadoption c 0.132 s (0.005) Continuationcostofadoption c 0.070 f (0.001) Carryingcost c 1.092 m (0.018) Observedshareofnon-adopters τ 0.960 (0.0001) Stockoutcost c 0.621 so (0.008) Note: Thetablereportstheparameterizationusedtodefinethecounterfactualmodel. 67

Table D3: Stockout Cost Model Fit Moment Model Data Mean(inventory-salesratio|adopter) 0.106 0.094 (0.005) Mean(inventory-salesratio|non-adopter) 0.140 0.146 (0.002) Std(inventory-salesratio|adopter) 0.053 0.054 (0.010) Corr(inventory-salesratio,logsales|adopter) -0.080 -0.098 (0.001) Std(logsales|adopter) 0.213 0.206 (0.005) Std(inventory-salesratio|non-adopter) 0.141 0.161 (0.001) Corr(inventory-salesratio,logsales|non-adopter) -0.301 -0.282 (0.001) Std(logsales|non-adopter) 0.282 0.296 (0.002) Spike(inventory-salesratio|adopter) 0.032 0.045 (0.012) Spike(inventory-salesratio|non-adopter) 0.189 0.188 (0.005) Frequencyofadoption 0.039 0.042 (0.004) Note: Thetablereportsmodel-basedandempiricalmomentswithstandarderrorsinparentheses. 68

Figure D4: Trade-off with Stockout Costs Note: ThefigureplotstheJITtrade-offacrossthebaselinemodel(red‘+’)andthestockoutcostmodel(blue‘x’). 69

D.5 Alternative Order Cost Distribution In this section I examine the robustness of the headline findings to the order cost distribution assumed. Ratherthanassumingthatordercostsareuniformlydistributed,hereIassumethattheyare rightskewedasinKhanandThomas(2016). Inparticular,letξ „ Bpa,bqwithsupportrξ,ξs. The probabilitydensityfunctionofthisfour-parameterbetadistributionis px´ξqa´1pb´xqb´1 fpx;a,b,ξ,ξq “ , pξ ´ξqa`b´1Bpa,bq where Bpa,bq is the beta function. I set a “ 5 and b “ 2. I then define the location parameters as before. Inparticular,thelowerboundofordercostsforadoptersiszero,ξ “ 0,andtheremaining A threetξ ,ξ ,ξ uareestimatedfromthedata. NA NA A TheparameterestimatesarereportedinTableD4andthemodelfitisreportedinTableD5. Table D4: Beta Cost Distribution Model Estimates Description Parameter Estimate Idiosyncraticproductivitypersistence ρ 0.810 z (0.002) Idiosyncraticproductivitydispersion σ 0.023 z (0.0001) Ordercostlowerbound(non-adopters) ξ 0.029 NA (0.001) Ordercostupperbound(non-adopters) ξ 0.285 NA (0.001) Ordercostupperbound(adopters) ξ 0.012 A (0.002) Sunkcostofadoption c 0.213 s (0.003) Continuationcostofadoption c 0.112 f (0.002) Carryingcost c 1.308 m (0.005) Observedshareofnon-adopters τ 0.953 (0.0001) Note: Thetablereportsparameterestimateswithstandarderrorsinparentheses. TableD6reportsthelongrunaggregatesintheJITeconomyrelativetothecounterfactual. The 70

Table D5: Beta Cost Distribution Model Fit Moment Model Data Mean(inventory-salesratio|adopter) 0.100 0.094 (0.005) Mean(inventory-salesratio|non-adopter) 0.123 0.146 (0.002) Std(inventory-salesratio|adopter) 0.059 0.054 (0.001) Corr(inventory-salesratio,logsales|adopter) -0.124 -0.098 (0.001) Std(logsales|adopter) 0.210 0.206 (0.005) Std(inventory-salesratio|non-adopter) 0.074 0.161 (0.002) Corr(inventory-salesratio,logsales|non-adopter) -0.319 -0.282 (0.001) Std(logsales|non-adopter) 0.266 0.296 (0.002) Spike(inventory-salesratio|adopter) 0.054 0.045 (0.012) Spike(inventory-salesratio|non-adopter) 0.158 0.188 (0.005) Frequencyofadoption 0.047 0.042 (0.004) Note: Thetablereportsmodel-basedandempiricalmomentswithstandarderrorsinparentheses. Table D6: Long-Run Aggregates Across Models Output Order frequency Order size Price of orders 9.30 51.75 -13.01 3.37 Inventory stock Firm value Measured TFP Welfare -24.37 1.80 0.53 1.46 Note: The table reports steady state values of the estimated model relative to the counterfactual model, in percent deviations. 71

steady state results with beta distributed order costs are similar to those with uniform order costs. FigureD5plotsthetrade-offassociatedwiththismodelrelativetotheonereportedinthemaintext. Basedontheestimatedordercostparameters,thegainstoJITarelarge,leadingtoa1.8%increase infirmvalue. Figure D5: Trade-off with Beta Order Cost Distribution 72