Increasing Inventories: The Role of Delivery Times

Layer 1: Overview

This paper documents and explains a previously unreported reversal in U.S. manufacturing inventory trends: after a 25-year secular decline, inventories-to-sales ratios have been rising steadily since 2005. The central claim is that this reversal is driven by the rise of global sourcing, which lengthens and makes more volatile the delivery times of inputs, compelling firms to hold larger buffer stocks. The paper combines new empirical evidence with a calibrated quantitative model to attribute 81% of the post-2005 inventory rise to global sourcing.

The research question is timely: while the efficiency gains from global sourcing are well-documented, the risk implications—particularly for inventory behavior—have received scant attention. The inventory trend reversal itself was previously undocumented. The average U.S. manufacturing firm held 1 month and 4 days of sales as inventories at the lowest point in December 2005; by end of 2019, firms were holding an additional 12 days of sales as inventories. This reversal is present across all NAICS three-digit manufacturing industries (except Paper Manufacturing), all inventory types (finished goods, materials/supplies, work-in-process), public firm data from Compustat, and in the manufacturing sectors of Australia, Canada, Japan, and South Korea. Within inventory types, intermediate-input inventories show the steepest decline and rise, directly implicating input sourcing decisions.

Contemporaneously, the share of foreign inputs in U.S. manufacturing production rose from 13.3% in 1997 to 16.5% in 2018, with approximately 3 percentage points of that increase attributable to inputs from China. The distance traveled by imports rose at an average annual rate of 6% from 1995 to 2018 across the U.S. and four peer countries. Since roughly 80% of Chinese imports arrive via ocean and take approximately 25–35 days in transit (with around 30% of shipments arriving more than one day late), the shift toward Chinese inputs materially increases both the mean and the variance of delivery times. Cross-industry regressions confirm the link: a 10% increase in foreign inputs is associated with a 7% rise in intermediate-input inventories (controlling for industry value added).

To quantify the causal role of delivery times, Carreras-Valle builds a dynamic partial-equilibrium model of final-good firms that source both domestic and foreign inputs, stock inventories, and face iid firm-specific demand shocks. The key methodological innovation is a tractable formulation of stochastic delivery times: a random fraction λ of the ordered inputs arrives within the period and can be used for production, while the remainder arrives in the following period. This setup nests as special cases the fixed one-period lag used in prior literature, while permitting calibration to observed lead-time distributions and enabling comparative statics across the full distribution of delivery times. The model features CES aggregation of domestic and foreign inputs (elasticity σ = 0.8, from Boehm, Flaaen, and Pandalai-Nayar 2017), Cobb-Douglas technology with an input share α = 0.63 (from BEA Input-Output Tables), and monopolistic final-good producers.

The model is calibrated to 1992 U.S. manufacturing and then subjected to two observed trends: (i) a technology channel—decreasing mean and variance of domestic delivery times, calibrated to ISM lead-time data (mean 35 days in 1992, declining thereafter); and (ii) a trade channel—a falling relative price of foreign inputs, calibrated to match the 3 percentage point rise in the Chinese input share, implying an approximately 1% average annual decline in the foreign-to-domestic input price ratio. The model generates the full U-shaped inventory trend as an untargeted prediction, accounting for 50% of the 1992–2004 decline (data: −2.3% per year; model: −1.2% per year) and 81% of the 2005–2018 rise (data: +1.2% per year; model: +1.0% per year).

A key structural decomposition reveals that the total inventory rise is driven entirely by foreign inventories (rising at +1.5% per year), which more than offset the continuing decline in domestic inventories (−0.5% per year). Firms require both channels: the technology channel alone produces initial decline but no subsequent rise; the trade channel alone generates a monotone increase that misses the initial decline. Further, the model decomposes inventory incentives into demand risk (the interaction of positive delivery times with demand volatility) and delivery-time risk (the variance of λ). Demand risk accounts for most of the level of inventories; delivery-time risk accounts for the growth in inventories over time—especially important as firms shift toward foreign inputs subject to frequent delays.

The model also characterizes an aggregate efficiency-volatility tradeoff from globalization. Comparing an economy with the 2018 share of foreign inputs (16%) to one fixed at the 1992 share (13%), output rises 13.9% and the price level falls 2.6% in the more globalized economy, but the standard deviation of prices rises 9.7% and the standard deviation of output rises 12.3%. The share of firms experiencing stock-outs rises from 8% to 12%. Even with higher inventories, firms cannot fully insure against the amplified demand risk, so price and output volatility rise. Results are robust to alternative values of the demand elasticity (ε = 1.5, 4), the input substitution elasticity (σ = 0.6, 0.8, 1.5), and storage costs (δ = 5%, 7.5%, 15%).

Layer 2: Deep Dive

What is the core identification strategy for the empirical relationship between foreign inputs and inventories, and what are the main threats?

The paper uses panel regressions of log inventories on log imported inputs with industry and year fixed effects, covering NAICS three-digit manufacturing industries from 1997 to 2018. The industry fixed effects absorb time-invariant industry characteristics that correlate with both import intensity and inventory needs; year fixed effects absorb common macro trends. The main threat is omitted variables that are industry-time varying: for instance, a demand boom that simultaneously induces firms to import more and stock more could generate a spurious correlation. The author partially addresses this by controlling for value added, showing the elasticity falls from 0.59 to 0.35 for total inventories (and from 0.72 to 0.42 for input inventories) but remains positive and significant. The author also presents results separately for inputs from China specifically, where a 10% increase in Chinese inputs is associated with 2–5% higher inventories (Table 1, columns 7-8), and replicates results with three independent data sources (WIOD, OECD I-O Tables, U.S. Census end-use classification), all showing consistent findings.

What is the main mechanism by which delivery times raise inventory holdings, and how does it differ from the delivery-time risk mechanism?

The primary mechanism is demand risk exposure: because firms must order inputs before demand is realized, and because a share of the order only arrives in the following period, longer delivery times reduce a firm’s ability to respond to the current period’s demand shock using new orders. Firms therefore hold buffer inventories to bridge the gap. This mechanism operates even when delivery times are positive but deterministic (the dashed line in Figure 15), and it accounts for most of the level of inventories. The secondary mechanism is delivery-time risk: since the fraction λ that arrives is itself stochastic, firms also hold inventories to insure against low-λ realizations (input shortfalls). Delivery-time risk contributes less to the level of inventories but accounts for a disproportionate share of the growth in inventories over time, because growth accelerates as firms shift toward foreign inputs—subject to more frequent ocean-shipping delays—come to dominate the input mix. The model separates the two by running a scenario with deterministic but positive delivery times (demand risk only) against the full stochastic model.

How does the paper model delivery times, and what is novel about this approach relative to the literature?

The paper introduces a tractable stochastic delivery-time specification in which a firm-specific iid fraction λ drawn from an input-specific log-normal distribution G_i(μ_λ, σ_λ) arrives within the period and is available for production, while (1−λ) of the order arrives at the start of the next period and is added to the following period’s inventory. The literature had largely assumed a fixed deterministic one-period lag (all inputs arrive exactly one period later). One exception is Alessandria, Kaboski, and Midrigan (2010b), who model a binary probability-of-arrival (either the entire order arrives now or next period); Carreras-Valle’s formulation allows a stochastic share to arrive, which accommodates heterogeneous delivery time distributions across inputs and enables direct calibration to observed lead-time data from ISM and Freightos. This flexibility permits the paper to match different mean and variance profiles for domestic versus foreign inputs and to study how marginal changes in the delivery-time distribution affect sourcing and inventory choices.

What data sources are used, and how are the key variables constructed?

Inventory and sales data come from the U.S. Census Bureau’s Manufacturers’ Shipments, Inventories, and Orders (M3) survey, matched to NAICS three-digit industries (monthly, 1992–2018; petroleum sector NAICS 324 excluded). Firm-level inventory data are from WRDS Compustat. Imported input shares by country of origin are constructed from U.S. Census Bureau import data (retrieved from Schott 2008), apportioned using BEA Input-Output Tables following the BEA’s own import-matrix methodology: the share of imports from country i used as inputs in industry j is assumed proportional to country i’s share of total U.S. imports in that sector. This is robust to using WIOD and OECD I-O tables. Domestic delivery times are from the ISM Manufacturing PMI, adjusted to remove foreign transit times using Chinese transit data, then smoothed with the Hodrick-Prescott filter. Foreign delivery times are calibrated to Freightos ocean-shipping data for the U.S.–China route (25 days to West Coast, 35 days to East Coast, combined average 30 days plus 35 days domestic transit). Distance of imports uses CEPII Gravity dataset population-weighted distance weighted by dollar value of imports by origin country.

How does the model generate the inventory trend as an untargeted moment, and what does it miss?

The model is calibrated to match only two 1992 moments (the level of input inventories over output and the share of foreign inputs in 1992). The time path of inventories from 1992 to 2018 is then entirely untargeted. Given the estimated paths of domestic delivery times (declining from ISM data) and the relative price of foreign inputs (declining at roughly 1% per year to match the observed import share), the model generates a U-shaped inventory trend qualitatively and quantitatively similar to the data. The main shortcoming is timing: the model’s inventory reversal begins around 2003, two years ahead of the 2005 reversal in the data. The author attributes this gap to China’s WTO accession in 2001 feeding into the model’s trade channel immediately, whereas in reality adjustment lags and other factors may have delayed the full inventory response. The model accounts for 50% of the initial decline and 81% of the subsequent rise, leaving room for other forces including changes in demand volatility (e.g., rising trade-policy uncertainty, Amazon’s market penetration), improvements in inventory-storage technology, and the low-interest-rate environment.

What heterogeneity is documented across industries and types of inventories?

The inventory trend is present across all NAICS three-digit manufacturing industries except Paper Manufacturing (NAICS 322, which represents only 3% of total manufacturing inventory). Import-intensive industries show the largest growth in inventories: sorting industries into terciles by average imported-input intensity (1997–2018), the most import-intensive group shows the largest decline and the sharpest subsequent rise in both total and intermediate-input inventories. Among the three inventory types, intermediate-input inventories (materials/supplies + work-in-process) show the steepest decline and steepest rise, consistent with sourcing decisions being the primary driver. Finished-goods inventories also rise but less sharply. The cross-sectional slope between imported-input intensity and inventories is 0.3 for total inventories and 0.9 for intermediate-input inventories. The trend is also present in four other countries’ manufacturing sectors (Australia, Canada, Japan, South Korea), with the distance of imports rising at 6% per year on average across these countries, suggesting the phenomenon is a global consequence of globalization.

What are the robustness checks?

The paper presents extensive robustness. For the empirical inventory trend: the U-shaped pattern holds when including the petroleum and coal sector (NAICS 324), when excluding the transportation sector (NAICS 336), and using the long-horizon NBER-CES Manufacturing Industry Database from 1958 (annual, 6-digit NAICS). The positive relationship between imported inputs and inventories is robust to using WIOD, OECD I-O Tables, and the U.S. Census end-use classification as alternative data sources, and appears consistently in both cross-sectional and time-series regressions. For the quantitative model: the inventory trend is robust to alternative values of the final-good elasticity of substitution (ε = 1.5 and 4), the domestic/foreign input substitution elasticity (σ = 0.6, 0.8, 1.5), and storage costs (δ = 5%, 7.5%, 15%). The qualitative proposition that inventories increase with delivery times is proved formally for the full multi-input model (Appendix C), not just for the simplified one-input version.

How does this paper relate to and differ from the closest prior work on inventories?

The paper builds directly on Khan and Thomas (2007) and Alessandria, Kaboski, and Midrigan (2010a) for the theoretical framework of inventories in general equilibrium. It departs from both by introducing stochastic and heterogeneous delivery times rather than a fixed one-period lag. The earlier literature on the inventory decline (Ohno 1988 just-in-time; Feinberg and Keane 2006; Dalton 2013; Shirley and Winston 2004; Li and Li 2013; Cui and Li 2018) focused exclusively on the downward trend attributed to improvements in transportation and information technology. This paper is the first to document the reversal and to introduce a model that accommodates both the decline and the subsequent rise through opposing forces. The inventory-import nexus has been documented in firm-level data for Chilean firms (Alessandria, Kaboski, and Midrigan 2013) and Indian firms (Khan and Khederlarian 2020), but this paper is the first to show the relationship across U.S. manufacturing industries and to tie it explicitly to China’s WTO accession and the delivery-time channel. It also contributes to the global supply chain risk literature (Baldwin and Freeman 2022) by quantifying inventories as the instrument firms use to absorb that risk.

What does the efficiency-volatility tradeoff finding imply for policy, and what are its scope conditions?

The model’s key aggregate implication is that globalization—access to cheaper foreign inputs—raises output and lowers prices on average, but simultaneously raises macroeconomic volatility because longer delivery times amplify demand shocks. Specifically, moving from the 1992 to the 2018 import share raises average output 13.9% and lowers the average price level 2.6%, but raises price volatility by 9.7% and output volatility by 12.3%. The share of firms in stock-out (constrained) rises from 8% to 12%. This tradeoff is not negated by the endogenous inventory response: firms do hold more inventories with globalization, but optimal inventory holdings leave some demand states unmet because insuring fully against all demand shocks is prohibitively costly. Policy implications are cautionary: reshoring or restricting imports to reduce delivery-time risk would reduce volatility but at the cost of lower average output and higher prices. The scope conditions are important: the model abstracts from labor reallocation, firm entry/exit, foreign-firm productivity dynamics, and consumer welfare under price variability. The calibration is to U.S. manufacturing 1992–2018, and the foreign input price trend is modeled as a single composite (China-focused) reduction, so the quantitative results may not generalize to settings where trade partners differ substantially.

What alternative explanations for the inventory rise does the paper consider or rule out?

The paper acknowledges three alternative forces that could contribute to the post-2005 inventory rise but are not modeled: (1) increasing demand volatility (e.g., from Amazon’s market penetration or rising trade-policy uncertainty), which would raise the value of inventories through the demand-risk channel; (2) improvements in inventory storage technology, which lower the cost of holding inventories; and (3) the low-interest-rate environment post-2008, which reduces the opportunity cost of holding inventories. The paper argues these are not the focus and that the delivery-time channel alone can explain 81% of the rise, leaving a residual 19% for which these other factors could account. The demand variance is held constant in the benchmark, so any time-varying demand risk that coincided with the post-2005 period is absorbed into the unexplained residual. The model is described as flexible enough to accommodate and quantify these forces if desired.

What is the model’s treatment of the technology channel and how is it calibrated?

Technology improvements are modeled as a steady reduction in the mean and variance of the domestic delivery-time distribution, proxying for advances in transportation infrastructure (high-speed rail, road investment, air freight) and information technology (just-in-time management, ERP systems). The mean of domestic delivery times is calibrated to ISM monthly data on average commitment lead times for production materials and maintenance/operation supplies, adjusted for the growing foreign input share (subtracting the fraction of ISM-reported lead times attributable to Chinese ocean transit), then smoothed with an HP filter. The mean starts at 35 days in 1992 and declines thereafter, with a mild uptick after 2003–2004. The variance is treated as a fixed proportion of the mean, so it co-moves with the mean. In the model this declining domestic delivery time reduces the value of holding domestic inventories, generating the observed decline in the domestic component of the inventory ratio (−0.5% per year over the full period). When only this channel is simulated (holding foreign input shares fixed at 1992 levels), inventories initially decline but then stagnate or rise only slightly—the trade channel is required to produce the full post-2005 acceleration.

Key Concepts

Delivery time (λ): In the model, the fraction of an input order that arrives within the current period and is available for production, where 1−λ arrives at the start of the following period. Calibrated as λ = max(0, 1 − delivery_days/T) where T = 90 days per quarter. A lower λ means longer delivery times and greater exposure to demand shocks.

Global sourcing: The practice of firms sourcing production inputs from distant foreign locations to exploit cost advantages, specifically the substitution of domestic inputs for cheaper inputs from countries such as China. In this paper it is the primary driver of rising inventories after 2005, because foreign inputs carry longer and more volatile delivery times than domestic alternatives.

Delivery-time risk: The volatility component of the delivery-time shock: because λ is drawn from a distribution with positive variance, firms face uncertainty about what fraction of an order will arrive in the current period. Distinct from demand risk (uncertainty about the quantity demanded). Delivery-time risk accounts primarily for the growth of inventories over time as reliance on volatile-delivery foreign inputs increases, rather than for the level of inventories.

Inventory-intensive inputs: Inputs—primarily foreign inputs in this paper’s framework—that by virtue of their long and/or volatile delivery times require firms to hold a disproportionately large stock of inventories per unit of input used. Foreign inputs from China are inventory-intensive because ocean transit averages 25–35 days and is subject to frequent delays.

Stock-out: An event in which a firm’s available input inventory (on-hand stock plus the fraction of the current order that arrives in time) is insufficient to satisfy its realized demand. When a stock-out occurs, the firm raises its price until the consumer is willing to demand only what the firm can supply. Longer delivery times increase stock-out frequency: the share of constrained firms rises from 8% to 12% as the economy moves from 1992 to 2018 import shares.

Efficiency-volatility tradeoff: The aggregate implication of globalization in the model: a higher share of cheaper foreign inputs lowers average prices and raises average output (the efficiency gain), but simultaneously raises the volatility of prices and output because longer delivery times amplify demand shocks and increase stock-out frequency. Inventories partially but incompletely offset this volatility increase.

Technology channel vs. trade channel: Two opposing forces shaping the delivery-time distribution over 1992–2018. The technology channel (improvements in transportation and information technology) reduces the mean and variance of domestic delivery times, lowering inventory incentives. The trade channel (China’s WTO accession and rising productivity driving down foreign input prices) shifts the input mix toward foreign inputs with longer and more volatile delivery times, raising inventory incentives. Both channels are necessary to reproduce the observed U-shaped inventory trend.