Do High- and Low-Inventory Turnover Retailers Respond...
Transcript of Do High- and Low-Inventory Turnover Retailers Respond...
0
Do High- and Low-Inventory Turnover Retailers Respond Differently to Demand Shocks?
August 2014
Abstract This paper examines the differences in the behaviors of high and low inventory turnover retailers in responding to demand shocks. We identify quantity- and price-responsiveness as two mediating mechanisms that distinguish how high- and low-inventory-turnover retailers (HIT and LIT retailers, respectively) can manage demand shocks. Using quarterly firm-level data of 183 U.S. retailers between 1985 and 2012, we find that HIT retailers are able to respond quickly by changing their purchase quantities in response to demand shocks, while LIT retailers primarily rely on price changes to manage demand shocks. We demonstrate the responsiveness of HIT retailers by showing that they can postpone ordering to a later time period compared to LIT retailers, who react to older demand signals. In addition, we examine the differential implications of these mechanisms on the financial performance of HIT and LIT retailers. On average, HIT and LIT retailers appear to be adept at using quantity- and price- responsiveness to avoid excesses and shortages during demand shocks. However, the negative financial impact of excesses and shortages, when they occur, can be eight times more severe for LIT retailers compared to HIT retailers. Keywords: Inventory Turnover, Price Responses, Quantity Response, Demand Shock, Firm Performance, U.S. Retail Industry, Econometric Analyses
1
1. Introduction
The virtues of high inventory turnover have been expounded for decades. Yet, considerable heterogeneity in
inventory turns across retailers can be observed even in narrowly defined segments. For example, the 80th and
20th percentile of inventory turns in the apparel segment (SIC: 56) in 2012 are 5.84 and 2.68. In other words,
retailers in the 20th percentile require 120% more inventory to generate the same sales as retailers in the 80th
percentile. Recent research in operations management has examined the differences in performance across
high and low inventory turnovers. Researchers have benchmarked inventory turnover (Gaur et al. 2005),
reconciled inventory variation in practice with analytical inventory theory (Rumyantsev and Netessine 2007a,
Bray and Mendelson 2012, Jain et al. 2012, Rajagopalan 2012), and also correlated inventory productivity with
financial performance (e.g., Chen et al. 2007, Rumyantsev and Netessine 2007b, Larson et al. 2011). However,
little research has been done to understand the behavioral differences between high and low inventory
turnover retailers that contribute to the observed performance differences. For example, it is unclear how
these retailers manage demand-side risk and whether there are differences in the way they do so that could
contribute to the observed performance differences in these two groups of retailers.
Consider the reactions of the two groups of retailers in the apparel segment (discussed above) during the
recent economic downturn. Prior to the most recent economic slowdown the Consumer Confidence Index
(CCI) had risen steadily from 2003 and peaked at 111.9 points in July 2007. In the next quarter, the CCI
declined abruptly by 15%. Both groups of retailers experienced similar decline in sales of about 6.3% during
that quarter. We expect both retailers to adjust to this downturn by changing their orders. Consistent with our
expectation, we find that the ratio of purchases to cost-of-goods sold (COGS) for retailers above the 80th
percentile of inventory turns reduced by 30% in that quarter. However, for those retailers with turns in the
bottom 20th percentile, this ratio increased by 30% in that quarter before declining by 44% in the next
quarter. One possible explanation is that the purchases of low inventory turnover retailers responded to the
demand shock with a delay. In contrast, the gross margin of these low inventory turnover retailers declined by
nearly twice the amount compared to their high inventory turnover counterparts suggesting that low
inventory turnover retailers may have reduced prices more aggressively than high inventory turnover retailers.
Together, this example suggests that the group of retailers with higher inventory turns were able to change
their orders quickly (quantity response) to manage demand shocks while low inventory turnover retailers
changed prices (price response) to manage demand shocks.
Inventory theory offers an explanation for this observed difference in behavior. Specifically, the
difference in responses to demand shocks between high inventory turnover (HIT) and low inventory
turnover (LIT) retailers are consistent with the joint inventory-pricing literature. This literature considers how
firms can change their order quantity (quantity response) and/or pricing (price response) to manage demand
uncertainty. Smaller ordering costs (Chen and Simchi-Levi 2004) and shorter lead time (Bernstein et al. 2014)
2
are generally associated with smaller, more frequent orders and less price changes compared to higher
ordering costs and longer lead time, respectively. Since firms with lower set-up costs and shorter lead times
will have higher inventory turnover, ceteris paribus, the HIT retailers may have been able to change their
orders in a timely fashion in response to demand changes. Thus, we expect HIT retailers to rely less on price
responses compared to LIT retailers. Admittedly, lead time and set-up costs are not reported by firms or
observed in their public data. However, the implications of our arguments are testable. Thus, we address the
following questions in this paper: (i) do HIT retailers use more quantity response and less price response
(than LIT retailers) to manage demand shocks? (ii) what are the resulting implications of the differences in
this behavior on the profitability of the two types of retailers?
Using 9,028 quarterly observations from 183 U.S. public retailers from the Compustat database for the
period 1985-2012 and consumer confidence index (CCI) from the Conference Board, and dividing retailers in
each SIC segment into time-invariant high- and low-inventory-turnover groups, we find that:
(i) The inventory purchases of LIT retailers are affected by less recent demand shocks than those of HIT retailers. The former respond to demand shocks occurring two, three and four quarters ago, whereas the latter respond to demand shocks in the current quarter. Thus, HIT retailers appear to have more responsive supply chains than LIT retailers.
(ii) Consistent with less responsive supply chains for LIT retailers, the order variability of LIT retailers is significantly more sensitive to demand shocks compared to that of HIT retailers.
(iii) LIT retailers change their gross margins by significantly larger amounts compared to HIT retailers in response to demand shocks. So, LIT retailers appear to use price response, as opposed to quantity response, to manage demand shocks.
(iv) The quantity response strategy of HIT retailers enables them to avoid excesses and shortages of inventory (as measured by abnormal inventory growth) but LIT retailers incur such abnormal inventory growth in subsequent quarters after a demand shock.
(v) Finally, abnormal inventory growth, when it occurs, is much more detrimental to the financial performance of LIT retailers than HIT retailers. We find that a one-standard-deviation increase (from the mean) in abnormal inventory growth (ABIG) leads to a 0.64% decline in return on assets (ROA) for HIT retailers, and a 1.42% decline in ROA for LIT retailers. A one-standard-deviation decrease (from the mean) in ABIG leads to a 0.42% decline in ROA for HIT retailers, and a much larger 4.29% decline in ROA for LIT retailers. Thus, LIT retailers face more severe penalty when there are excesses and shortages of inventory.
We obtain these results using two different methodologies and two different measures of demand shocks.
In the first methodology, we utilized a vector auto regression (VAR) model to determine the impact of
demand shocks on demand, purchases, gross margin, ABIG, and ROA for the following reasons. Purchases
for a retailer in a quarter are measured using the accounting identity: purchases = ending inventory + cost of
goods sold – beginning inventory. The VAR methodology is advantageous as it handles the simultaneity
among different variables and permits the examination of contemporaneous as well as delayed impacts of
demand shocks on the variables using an impulse response function (IRF) analysis. The VAR models are
extensively used in Economics, Finance, Marketing, and Operations Management for modeling multivariate
3
time series models. We use macroeconomic shocks as proxy for firm-level demand shock in this
methodology. In the second methodology, we use individual reduced form regressions to examine the impact
of demand shocks. In this methodology, we measure independent firm-level demand shocks based on
Martingale Model of Forecast Evolution (MMFE). Both methodologies and measures of demand shocks
support our theory.
This evidence is timely and relevant because demand uncertainty has been increasing in recent years due
to lengthening supply chains, global recession, and macroeconomic events. So, while retail managers are
under competitive pressure to lower the physical costs in their supply chain by sourcing from different parts
of the world, they also worry about the mediation costs due to mismatches in supply and demand arising
from uncertainty (Cave, 2014). The challenge to examining this trade-off between lower physical costs in the
supply chain and increased mediation costs is that the mediation costs are less visible and harder to quantify.
Our paper quantifies mediation costs that arise from less responsive supply chains and shows that the ABIG
is larger for such supply chains and hurts their profitability by much more compared to responsive supply
chains.
Our paper contributes to the academic literature in the following ways. First, this paper contributes to the
recent empirical research that has observed that high inventory turnover retailers have better financial
performance (e.g., Alan et al. (2014), Chen et al. 2007, Rumyantsev and Netessine 2007b) by demonstrating
one reason for the observed difference. This paper shows that high inventory turnover retailers pursue an
order postponement strategy that enables them to manage demand-side risk better than low inventory
turnover retailers who rely on changing prices to manage this risk.
In addition, this paper offers a new explanation for the presence of the ‘earns versus turns’ trade-off in
retailing. The empirical evidence of this inverse relationship between inventory turns and gross margin has
been well documented in prior research (Gaur et al. 2005). The reasons for the presence of this trade-off are
usually argued based on the occurrence of competition, inherent differences in product characteristics across
firms, and the newsvendor logic. According to the first explanation, firms with low inventory turns and low
gross margin will exit the market, those with high inventory turns and high gross margin will erode their
advantage over time, and the average firm in the marketplace will manifest a negative correlation between
inventory turns and gross margin. According to the newsvendor logic, retailers with higher margin have
greater incentives to carry more inventory and will therefore have lower inventory turns. These arguments,
however, assume that other aspects of retail inventory management, such as set-up costs and lead time, do
not vary across firms. Yet, we find contrasting examples of firms that source domestically, e.g., American
Apparel, and those that source from faraway foreign locations, e.g., Gap, for the same market. Our paper
offers a new explanation for this trade-off that is based on substitution of capabilities to respond to demand
shocks with ordering or pricing changes motivated by the joint pricing and inventory management literature.
4
LIT retailers who are unable to react to demand shocks due to their less responsive supply chains will need to
change prices; sustaining greater price volatility would require higher margins. Thus, LIT retailers will have
higher margins compared to HIT retailers, leading to this trade-off. Future research may examine what drives
retailers to adopt an earns versus a turns strategy in the first place.
Third, this paper also contributes to empirical research on inventory turnover performance by showing
differences in the impact of abnormal inventory on financial performance across HIT and LIT retailers. Both
the accounting and operations management literatures have highlighted the importance of the ABIG metric
as a predictor of earnings per share, ROA, and stock prices (Abarbanell and Bushee 1997, Thomas and Zhang
2000; Rumyantsev and Netessine 2007b, Kesavan and Mani 2013). While Rumyantsev and Netessine (2007b)
and Kesavan and Mani (2013) document a non-linear relationship between ABIG and profitability, we show
the nature of this relationship to be moderated by the type of retailer: namely, HIT or LIT. So, while prior
literature has shown the financial performance of all retailers to be impacted by excesses and shortages in
inventory, our paper shows that this impact is much larger for LIT retailers.
2. Hypotheses
In this section, we argue for the difference in behaviors between HIT and LIT retailers during demand
shocks. An important challenge in theorizing about the differences in the behaviors of HIT and LIT retailers
is that the primitives that drive the ability of a retailer to use quantity response or price response to demand
shocks are unobservable. For example, inventory theory suggests that factors such as replenishment lead
time, length of review cycle, and fixed ordering costs play important roles in determining when an order
would be placed. However, using publicly available data it is not possible to determine how HIT and LIT
retailers vary along these different dimensions. To overcome this gap between theory and the observable data,
we assume that HIT retailers have a shorter lead time, a shorter review cycle, and a smaller fixed ordering cost
compared to LIT retailers and generate our hypotheses about their behaviors implied by those underlying
primitives. It is not necessary for HIT retailers to be superior to LIT retailers in each of these dimensions for
our hypotheses to be supported. So, a support for our hypotheses implies that HIT retailers on an average
have a more responsive supply chain due to shorter lead time and/or shorter review cycle and/or smaller
fixed ordering costs compared to LIT retailers. Other factors such as gross margin, sales growth, and firm
size may differ between HIT and LIT retailers as well. Thus, we use control variables from the previous
literature to test the predictions of normative theory as closely as possible.
We use examples of inventory models with nonstationary demand to motivate the hypotheses on
quantity response, and joint inventory-price optimization models for the hypotheses on price response. We
measure quantity response through change in purchases, and price response through changes in gross margin.
Hypotheses 1-3 below pertain to quantity response, Hypothesis 4 to price response, and Hypothesis 5 to the
financial impact of demand shocks on HIT and LIT retailers
5
2.1. Quantity Response
The inventory purchase quantity in a periodic review model is determined by forecasting demand for the
sum of the lead time L and the review cycle R. Therefore, a historical demand shock will affect purchases
through the forecast of future demand over L+R. Intuitively, the longer the lead time or the length of the
review cycle, the larger will be the effect of a demand shock on the forecast of demand, and thereby on the
purchase quantities. Similarly, the higher the fixed ordering cost, the less frequent the orders, which then has
the same effect as the length of the review cycle. For the rest of our hypotheses’ development, we focus on
the effect of lead time. The effect of length of review cycle is similar. We formalize this reasoning below. H1
and H2 differentiate between HIT and LIT retailers with respect to purchases, while H3 formulates the post
facto implication for their inventory levels.
H1: [Order Postponement] The purchases of HIT retailers respond to relatively recent demand shocks, while purchases of LIT retailers respond to older shocks.
H2: [Order Variability] The purchases of LIT retailers change by larger amounts upon the occurrence of demand shocks than do those of HIT retailers.
H3: LIT retailers have more excesses and shortages in inventory due to demand shocks than do HIT retailers.
The intuition behind the above hypotheses can be understood through a consideration of the classic
“Beer Game” simulation of supply chain management (Sterman 1989). Lee et al. (1997) describe forecast
updating and order batching as two causes of the bullwhip effect simulated in the game. If we expect HIT
retailers to have a shorter lead time and smaller order batches than LIT retailers, then HIT retailers would be
less susceptible to the bullwhip effect. They would see faster response to demand shocks (H1), smaller
variations in purchases (H2), and less excess and shortage of inventory (H3). These hypotheses also follow
the predictions of the literature on the value of postponement, wherein demand uncertainty is associated with
costs of excess and shortage of inventory, and lead time reduction is used as a mechanism for performance
improvement.
We derive the three hypotheses on quantity responses using the models by Lee et al. (2000) for AR(1)
demand, by Graves (1999) for ARIMA(0,1,1) demand. Next, we use each of these models sequentially to
argue these hypotheses. First, consider the model of Lee et al. (2000).1 Demand Dt follows an AR(1) process,
t1tt DdD
where (0,1) and t is a sequence of i.i.d. normally distributed random variables with mean zero and
constant variance. Let the replenishment lead time be r. From Lee et al. (2000), the order quantity Yt after
observing the demand in period t is:
1 Raghunathan (2001) obtains different conclusions from Lee et al. (2000) with respect to the value of information
sharing. Our analysis is limited to deriving the retailer’s order quantity, which is similar for both Raghunathan (2001) and
Lee et al. (2000).
6
1t
1r
t
2r
t D1
)1(D
1
1Y
This order quantity is delivered in period t+r. Representing the delivered quantity as purchases, we obtain:
t
2r
1t
2r
2r
1t
1r
t
2r
rt1
)1(Dd
1
1D
1
)1(D
1
1Purchases
This expression shows that the observed impact of a demand shock t at time t depends on two factors. First,
the magnitude of the impact depends on the lead time. The longer the lead time, the larger is the coefficient
of t, and the larger the impact on purchases. Therefore, LIT retailers, which have longer lead times, will have
larger changes in their purchases for a given demand shock than HIT retailers. Second, the impact of the
shock will be realized on purchases only after the lead time. Thus, purchases of HIT retailers will respond to
more recent demand shocks than will those of LIT retailers. In other words, HIT retailers can postpone their
orders to later periods compared to LIT retailers. Thus, these two effects motivate H1 and H2.
The same inference can be drawn from the ARIMA(0,1,1) demand model of Graves (1999). In this
model, the retailer’s demand during period t, Dt, is represented as:
11 dD
t1t1tt )1(DD for t=2,3,4…
Here, t is a series of i.i.d. shocks with mean zero and variance 2, and is the moving average coefficient.
This coefficient may be interpreted in the following way: t is the shift in demand during future periods due
to shock t. Further, when is equal to zero, demand is expected to follow a stationary i.i.d. process, with
mean d. As increases, the demand during any time period is expected to depend more and more on recent
demand realizations. Again let the replenishment lead time from the manufacturer to the retailer be one
period. In this setup, Graves (1999) proposes a base-stock policy that yields the order quantity during period t
as: Yt = Dt + t. The first component of the ordering quantity represents the amount of replenishment
required to make up for the demand, and the second component adjusts the base-stock level for the change
in forecast over the lead time. Graves (1999) notes that this policy is not optimal, but a reasonable extension
of the base-stock policy to the case of nonstationary demand. This gives us:
1tt1tttrt )1()r1(DrDPurchases
Thus, we find that the effect of the demand shock t in time period t increases with lead time. This implies
that magnitude of changes in purchases of LIT retailers will be greater than HIT retailers yielding H1.
Moreover, HIT retailers will respond to more recent demand shocks than LIT retailers leading to H2.
H3 follows in each of the above cases because retailers with a longer lead time will face more excess or
shortage of inventory before they can recover from a demand shock. Thus, we would expect LIT retailers to
have more excess and shortage in inventory than do HIT retailers following demand shocks.
7
2.2. Price Response
We next argue for a difference in the price responses of HIT and LIT retailers. As alternatives to
adjusting their purchases, retailers can react to demand shocks by changing prices, which would then affect
their gross margins. The theory on joint inventory-price optimization models predicts that retailers with
higher fixed ordering costs who cannot adjust their purchases quickly will rely on price adjustments more
than will retailers with smaller fixed ordering costs. Therefore, we propose the following hypothesis,
comparing LIT and HIT retailers:
H4: The gross margin of LIT retailers is more responsive to demand shocks than is that of HIT retailers.
Chen and Simchi-Levi (2004) present an infinite-horizon, single-period, periodic-review inventory model
with fixed ordering cost, in which a retailer optimizes both price and order quantities to maximize the
expected discounted profit. The authors show the optimality of a stationary (s,S,p) policy in which order
quantities are determined based on the classical (s,S) policy, and the price in each period is determined based
on the inventory position at the start of that period. Applying this model, if LIT retailers have higher fixed
ordering costs than HIT retailers, then they would have a higher order up to level S. Correspondingly, they
would place orders infrequently, greater heterogeneity in inventory position at the start of each period, and
thus, their prices would vary more than HIT retailers’ price from one period to the next. This yields H4.
Chen et al. (2006) consider the impact of setup cost on joint inventory and pricing optimization in
periodic review systems with lost sales, and find that the profit impact of dynamically changing prices
increases with setup cost. Aguirregabiria (1999) uses data from supermarkets to show that fixed ordering
costs are associated with greater price changes.
These papers support our arguments for differences in ordering costs across HIT and LIT retailers
contributing to differences in changes in gross margin across these retailers. We note that the existing
literature has considered fixed ordering cost, but not lead time, in multiperiod joint price-inventory
optimization. The theory on joint inventory-price optimization typically ignores lead time for reasons of
mathematical tractability (Bernstein et al. 2014). So, Bernstein et al. (2014) use a heuristic to show that shorter
lead time is associated with a more stable pricing policy. In summary, the theoretical literature has shown that
longer lead time or higher ordering costs are associated with more price changes. Therefore, we expect LIT
firms to have greater changes in gross margin compared to HIT firms.
2.3. Financial Performance
In H3, we argued that LIT retailers will have more excesses and shortages compared to HIT retailers.
The current literature has shown that abnormal growth in inventory is detrimental to financial performance
8
of retailers. Rumyantsev and Netessine (2007b) use ROA and Kesavan and Mani (2013) use earnings per
share as measures of financial performance to provide evidence of non-linearity in the relationship between
ABIG and profitability. Neither paper examines whether this impact of ABIG on financial performance
varies across retailers.
Excesses and shortages impact financial performance in different ways. Excess inventory increases direct
costs for a retailer due to the capital tied to inventory and physical costs of holding inventory. Excesses could
also force retailers to undertake steep discounting or clearance sales that will result in a decline in gross
margin. In extreme cases, excess inventory can lead to write-offs. Larson et al. (2011) find that retailers with
inventory write-down experienced an average decline in ROA of -15.4%. Shortages, on the other hand, affect
financial performance primarily due to lost sales.
We argue that the impacts of excesses and shortages on financial performance can be mitigated if a
retailer is able to quickly change its purchases. Retailers that have excess inventory in a period can reduce their
replenishments for the next period, thereby limiting the impact of excess inventory to only one time period.
Similarly, retailers who suffer a shortage of inventory can reduce its impact on financial performance by
replenishing their stores during the next period. If HIT retailers have shorter lead times and smaller ordering
costs compared to LIT retailers, then we argue that HIT retailers will be able to mitigate the impacts of
excesses and shortages on financial performance better than LIT retailers.
H5: The financial performance of LIT retailers is more negatively impacted by inventory excesses and shortages than is the financial performance of HIT retailers.
3. Data and Methodology
We obtain data from two public sources: COMPUSTAT and the Conference Board. Quarterly data for
retailer-level variables were obtained from COMPUSTAT for 28 years, 1985-2012. These data correspond to
all retailers required to file financial statements with the SEC. These retailers belong to one of the eight two-
digit SIC codes numbered 52 to 59, which correspond to the retail sector. These SIC codes cover the
following segments within retail: construction and home improvement (SIC = 52), department store (SIC =
53), groceries and produce (SIC = 54), automobile dealers (SIC = 55), clothing and accessories (SIC = 56),
furniture and white goods (SIC = 57), restaurants and eating outlets (SIC = 58), and others (SIC = 59: drug
stores, direct retailers, bookstores, stationery, florists, optical stores, news vendors, etc.). In keeping with
Kesavan et al. (2010), we decided not to use retailers from SIC 58 (restaurants and eating outlets) and 55
(automobile dealers) in our analyses because retailers in these industries have a significant service component;
consequently, inventory management is only partly related to their performance. Additionally, we dropped
from our analyses retailers categorized as SIC 54 (grocery stores) because we use macroeconomic shocks as
9
proxy for unobservable demand shocks and the grocery segment is acyclical. In other words, the correlation
between macroeconomic shocks and demand shocks are likely to be low in this industry2
Past research has measured macroeconomic shocks through several different approaches, each being
appropriate for a given empirical setting: Lamey et al. (2007) use Gross Domestic Product (GDP); Kesavan
and Kushwaha (2014) use GDP, Personal Consumption Expenditure (PCE), and Customer Confidence
Index (CCI); and Doms and Morin (2004) and Lemon and Portniaguina (2006) use University of Michigan’s
Index for Consumer Sentiment (ICS) as well as CCI. Past research (see Bram and Ludvigson 1999; Ludvigson
2004) has shown that CCI is better at predicting economic growth and expenditure behavior of consumers
for most product categories vis-à-vis ICS and managers use it in their planning cycles (Carroll et al. 1994,
Ludvigson 2004 and Souleles 2004). So we choose CCI series to measure macroeconomic demand shocks3.
In order to perform the empirical analysis, we only use retailers with fiscal year end of December and
January so that they are aligned in the release of macroeconomic information. Retailers with fiscal year end in
December and January constitute about 68% of the overall population. We adjust all of the firm-level
variables with consumer price index in order to control for inflation.
Table 1 summarizes data sources and measurements of different variables. We generated the following
independent and dependent variables for our analyses. Inventory Turns (ITit) is measured as the ratio of cost-
of-goods-sold to average inventory. Purchases (PURCHit) in a given quarter are obtained using the following
accounting identity: purchases = ending inventory + cost of goods sold – beginning inventory. While H1 and
H2 deal with purchase quantity, we are limited by data availability to measure purchases in dollar amounts.
This follows standard approach in the literature (Cachon et al. 2007, Larson et al. 2011).
--Insert Table 1 about here--
We use ABIG as a proxy for excesses and shortages, where ABIGit is defined as inventory growth (IGit)
minus sales growth (SGit) with respect to the same quarter in the previous year. This measure has been used
in the accounting literature (Lev and Thiagarajan 1993; Abarbanell and Bushee 1997), operations management
literature (Rumyantsev and Netessine 2007b; Kesavan and Mani 2013), and has been found to be used in
practice (Raman et al. 2005). We treat ABIG>0 as a proxy for excess inventory, and ABIG<0 as a proxy for
shortages.
2 This can be seen as follows. We extract cyclical component of PCE and industry sales by passing the logged value of these series through a Hodrick-Prescott filter. Co-movement elasticity is the relationship between the cyclical components of two series. Co-movement elasticity measures percentage change in cumulative industry sales for each 1% unexpected change in PCE. The co-movement elasticity for SIC 54 in our data set is .0108 (p>.10). Thus, the demand shocks encountered by retailers in SIC 54 are uncorrelated with those observed more generally, throughout the larger economy. 3 The CCI series is reported monthly but financial measures are only available at quarterly interval. We use mean CCI across three months of a quarter to arrive at quarterly CCI measure. Subsequently, we test the sensitivity of our results to firm specific demand shock generated using MMFE model.
10
Finally, we use Return on Assets (ROAit) as the measure of financial performance and Gross Margin
(GMit) as proxy for price. Though gross margin can also change as a result of change in input costs, we
assume that retailers’ demand shocks do not have a differential impact on the input costs of HIT retailers
compared to LIT retailers. So, we expect the differences in gross margins across HIT and LIT retailers to be a
proxy for the relative changes in prices.
We trim the top 1% and bottom 1% of observations based on Purchases, ABIG, and gross margin. This
approach ensures that our analyses are not unduly influenced by extreme outliers.
3.1. Classification of HIT and LIT retailers
The VAR methodology that we employ to test our hypotheses requires us to group retailers into two
time-invariant groups (HIT vs. LIT) so we can perform a split-sample analysis. We perform the grouping in
the following way. First we classified retailers in every quarter into two groups based on whether their rank on
inventory turns within their SIC segment falls in the top 75th percentile or bottom 25th percentile. Then we
classified those retailers that appeared in the top 75th percentile for more than 75% of the quarters they were
presented as HIT retailers and those retailers that appeared in the bottom 25th percentile for more than 75%
of the quarters as LIT retailers. The rest of the retailers were not used for the main analysis but they were
included in robustness checks where we reclassify HIT and LIT retailers based on whether their inventory
turn was above or below the median for the segment in that quarter. We begin with 13,227 quarterly
observations from 357 retailers and used our sorting procedure to obtain 9,028 quarterly observations across
183 retailers. We summarize key variables in our data by each industry in Table 2 and by their inventory turn
classification in Table 3. As seen in Table 3, the average quarterly inventory turn for high sample (HIT) is
1.65, and is significantly greater than that for low sample (LIT), i.e. .63.
--Insert Tables 2 and 3 about here--
The two main concerns with using time-invariant classification of retailers into HIT and LIT groups are
that if retailers frequently change their positions between these groups and, more importantly, if this
switching is driven by demand shocks then this method of classification is neither appropriate nor exogenous
to the dependent variables being examined. We use three tests to rule out such confounding effects. First, we
examine if past demand shocks explain whether a retailer is classified as HIT or LIT. We use random effects
panel data Probit model to estimate group membership of the key dependent variable: probability of a retailer
belonging to HIT classification, as a function of past financial performance, size, and demand shocks. The
results are reported in Table 4. The results suggest that while past performance and assets are significant
determinants of retailer’s relative inventory turn position in their industry, past demand shocks play no role.
Second, we permit HIT and LIT classification to vary over time and generate a transition matrix (see
Table 5) to see how often such transition is observed in the data. The results suggest that on 17% of instances
retailers’ transition between groups. However, most of these transitions are observed between adjacent
11
groups (i.e. Low to Medium and vice-a-versa, Medium to High and vice-a-versa). Less than 1% of transitions
are observed between two extreme groups (i.e. LIT and HIT). Thus, our analysis is not unduly affected by
retailers who may belong to both groups at different points in time. We perform further tests to rule out that
these transitions are driven by demand shocks.
We model the probability of a retailer transiting between two different groups (i.e. between diagonals and
off diagonals in Table 5). Again, we use random effects panel data Probit model to estimate group
membership of the key dependent variable: probability of retailer transitioning between groups, as a function
of past financial performance, size, and demand shocks. The results of this analysis are reported in Table 6.
We find that the impact of demand shocks on probability of a retailer transiting between groups is not
statistically significant. These tests establish that although demand shocks may impact inventory turns of the
retailers, their relative position amongst their peers remains unaffected. Therefore, our HIT and LIT
classification is exogenous to the demand shocks.
--Insert Table 4, 5, & 6 about here--
3.2. Model Specification to test Hypotheses H1-H4
We use a vector auto regression (VAR) model to determine the impact of demand shocks on demand,
purchases, gross margin, abnormal inventory growth, and ROA as they allow us to account for simultaneity
among them and their lagged effects. For example, increase in purchases may lead to reduction in gross
margins, which in turn, may lead to higher demand. In addition, several variables are influenced by their own
past realizations as well as past realizations of other variables. For example, purchase decisions should take
into account past purchases, demand, and gross margin. The VAR models are similar in spirit to simultaneous
equation models but provide the following additional benefits compared to a simultaneous equation system:
(a) the VAR models allow for dynamic interactions between the variables thereby improving model fit, (b) the
VAR models permit policy simulations through an impulse response function (IRF) analysis, and (c) the VAR
models enable us to identify dynamic impacts such as delayed response and persistence. Stock and Watson
(2001) provide additional details on VAR models.
The VAR models are extensively used in Economics, Finance, and Marketing for modeling multivariate
time series. For example: in Economics, the methodology was promoted by Nobel laureate Christopher Sims
in his seminal paper (1980) and has since been applied to study relationships between macroeconomic
variables such as interest rates, unemployment, inflation, exchange rates, and economic growth (see Stock and
Watson 2001 for a review). In Finance, the methodology has been used for understanding relationships
between stock prices, dividends, earning (Campbell and Shiller 1988); international transmission of stock
market movements (Eun and Shim 1989); information content of stock trades (Hasbrouck 1991); and
monetary policy and stock returns (Thorbecke 1997). In Marketing, the methodology has been adopted to
examine the relationships between consumer demand, advertising spending, promotional expenditure, and
12
customer profitability (see Dekimpe and Hanssens 1999); product quality and profitability (Jacobson and
Aaker 1987); new product introduction, sales promotion, and firm value (Pauwels et al. 2004); and word of
mouth and social networking (Trusov et al. 2009). More recently, the technique has found utility in
Operations Management literature for studying relationships between inventory investment and other firm
decisions (see Wu and Chen 2010; Kesavan and Kushwaha 2014)
The first step in model specification is to determine whether VAR models need to be specified in levels or
changes. Consistent with Levin et al. (2002), we performed the panel unit root tests to identify if variables are
stationary or evolving. Before performing the test, we use the logarithmic transformation for purchases
(LPURCH), gross margin (LGM), and demand (LCOGS) to account for scale differences (Gaur et al. 2005).
The unit root tests suggest that all the variables except customer confidence index and ABIG are not
stationary (see Web Appendix A for detailed results). ABIG is already fourth differenced by definition.
Therefore, we specify all variables, except CCI and ABIG, as the fourth differenced specification.
Furthermore, the tests suggest that when differenced these variables are stationary. The choice of using
fourth differencing rather than conventionally used first differencing is driven by seasonality considerations.
The quarterly public filing data is not seasonally adjusted and retail industry exhibits strong seasonality.
The second step, in model specification requires two other choices: identifying the endogenous variables
and specifying the number of lags for those endogenous variables (order of the model). We use the Granger
causality test to determine choice of endogenous variables (Enders 1995), see Web Appendix B for details. Our
results indicate that sales, purchases, gross margin, abnormal inventory growth, return on assets and demand
shocks are Granger-caused by other variables in the system and are therefore endogenous.4
The third step is determining optimal order of lags of the variables to be included in the model. These lags
act as instruments for identifying the system of equations specified above. In order to determine the optimal
lag length, we used the Schwarz Bayesian Information Criterion (SBIC) that consistently estimates the lag
structure that minimizes the sum of squared errors by taking into account model complexity (Schwarz 1978),
see Web Appendix C for details. The suggested order of the VAR model is four.
In matrix formulation the VAR model for each retailer i from industry segment j can be written as:
4 Granger Causality test is a test of predictive causality. It tests whether including a regressor in the system improves
forecasting accuracy of time series significantly. Thus a variable X Granger-causes Y if including X (and its lags)
improves the forecasting accuracy of Y over and above what is provided by the lagged values of Y.
13
ROA
it
ABIG
it
GML
it
URCHPL
it
OGSCL
it
CCI
it
6j
5j
4j
3j
2j
1j
lit
lit
lit
lit
lit
lt
4
1l
l
66
l
65
l
64
l
63
l
62
l
61
l
56
l
55
l
54
l
53
l
52
l
51
l
46
l
45
l
44
l
43
l
42
l
41
l
36
l
35
l
34
l
33
l
32
l
31
l
26
l
25
l
24
l
23
l
22
l
21
l
16
l
15
l
14
l
13
l
12
l
11
60
50
40
30
20
10
it
it
it
it
it
t
ROA
ABIG
MGL
URCHPL
OGSCL
CCI
ROA
ABIG
MGL
URCHPL
OGSCL
CCI
(1)
where, l stands for number of lags of each endogenous variable to be included. We are primarily interested in
examining the impact of demand shock (CCI
it ) on the rest of the variables. The second and third rows
explain changes in demand (measured as change in log cost of goods sold – ΔLCOGSit) and purchases
(measured as change in log purchases – ΔLPURCHit), respectively. The fourth, fifth, and sixth rows explain
changes in gross margin (ΔLGMit), abnormal inventory growth, and change in return on assets (ΔROAit),
respectively. We control for industry segment specific fixed effects () in our system of equations. The ’s are
white noise residuals which are distributed MVN (0,Σ). The lag terms act as instruments to identify the
system. In the above system, 11, λ22, λ33, λ44, λ55, and λ66 are the carryover effects (lag terms) for the
endogenous variables.
The last step in the VAR methodology is the use of impulse response functions (IRFs) to examine the
impact of demand shock (CCI
it ) on the rest of the variables. An impulse response analysis is frequently
undertaken after model estimation since interpreting the coefficients of a VAR model is often problematic
due to multicollinearity amongst lags of variables (Sims 1980). This is consistent with previous work in
operations management where, Wu and Chen (2010; p. 1375) recommend that “Because of the complicated
dynamics in VAR, impulse responses are more informative than the estimated VAR coefficients or R2 statistics, which typically
go unreported.” An impulse response is the forecasted response of a system of variables to a unit (or one
standard deviation) exogenous shock in another variable. The procedure in using IRF analysis for VAR
models is as follows. We first estimate the system of equations as specified in the VAR model. Next, we
predict the change in value of other endogenous variables over the next 10 quarters due to a one standard
deviation shock to CCI in the current quarter. The statistical significance of the impulse response weights are
assessed by examining the t-statistics associated with the forecasted values of the dependent variable (Sims
1980). We use Cholesky’s degrees of freedom adjusted decomposition to generate relevant IRFs. We use
analytical standard errors to test the significance of policy simulations. We use a conservative +/- two
standard deviation band for evaluating statistical significance of impulse responses (Sims and Zha 1999).
We exemplify the calculations of forecasted responses for system of equations in Web Appendix D. The
Granger Causality test results reported in Web Appendix B suggest that: (a) demand shocks Granger-causes all
firm specific variables; (b) change in demand Granger-causes change in purchases but not the other way
around; (c) change in purchases Granger-cause change in gross margin and vice-a-versa; (d) change in
14
demand, purchases, and gross margin Granger-cause ABIG; and (e) ABIG Granger-causes ROA and not
vice-a-versa. Consistent, with these results we specify following causal ordering for generating IRFs:
Demand Shock Demand Purchases Gross Margin Abnormal Inventory Growth Return on Assets.
From the IRFs we calculate the immediate and total impact of demand shock on other variables in the
system. The immediate impact is operationalized as the impulse response weight in the concurrent time period.
The total impact is operationalized as the sum of effects of impulse response weights until equilibrium (i.e.,
mean reversion or new trend) is reached. We exemplify these impacts using Figure 1. In interest of parsimony
we only highlight the impact of demand shocks on purchases. The solid lines represent the immediate impact
of demand shock in time period ‘t’ on purchases in time period ‘t’. This impact includes the direct impact as
well as the one that permeates through change in demand. The dashed lines represent the delayed impact of
demand shock in time period ‘t’ on purchases in time periods ‘t+1’ and ‘t+2’. Again, this effect includes the
direct impact as well as the one that permeates through demand as well as recursive relationships between
demand shocks, demand, and purchases. The sum total of immediate and delayed impact is the total impact.
--Insert Figure 1 about here--
In the VAR methodology, the moderating impact of a variable is examined by using a split-sample
approach (see Kesavan and Kushwaha 2014 for a similar approach). We perform VAR analysis for HIT and
LIT sub-samples separately and compare the results.
3.3. Model Specification to test Hypothesis H5
In H3 we argued that demand shocks would create different amounts of excesses and shortages for HIT
and LIT retailers. H5 predicts that the impact of such ABIG on profitability would be different for HIT and
LIT retailers. Because such an impact is likely to be non-linear (Rumyantsev and Netessine 2007b), we cannot
use the VAR analysis. Instead, consistent with past research (Kesavan and Mani 2013), we account for the
nonlinearity in the relationship between ABIG and ROA using linear and quadratic terms for ABIG in the
following model specification:
ittiit
2
it2it10it ZABIGABIGROA (2)
Here, α are response parameters, Zit is the vector of control variables including linear and quadratic terms of
gross margin, CCI, and lagged ROA. To account for retailer and quarter-industry specific unobserved
heterogeneity in the above equation, we include retailer dummies (i) and quarter-industry (t) dummies.
Since we hypothesize a nonlinear impact of ABIG on ROA, and ABIG takes both positive and negative
values, comparing statistical significance of coefficients α1 and α2 across the HIT and LIT subsamples may
not reveal the range of ABIG values in which ROA differs across the two types of retailers. Bollen and Stein
(1990) find that in large sample the bootstrap distribution of an estimator is close to that assumed with
classical methods. They also suggest that such bootstrap distribution of an estimator is appropriate for
nonlinear and indirect effects. We formally test the difference in the impact of ABIG on ROA between HIT
15
and LIT retailers by performing 1,000 bootstraps of linear and quadratic coefficients from their two standard
deviation asymptotic intervals. We compare the mean of fitted ROA values across the 1,000 bootstraps
between the two subsamples in the observed ABIG range, and test for the statistical significance of the
difference between the means to evaluate the range of ABIG values in which HIT and LIT retailers are
different from each other. This constitutes the test of H5.
4. Results
4.1. Results: Hypotheses 1-4
As discussed previously, we generate IRFs and calculate immediate and total effects of demand shocks for
hypotheses testing. The IRF and their associated effects for our key variables of interest are reported in
Figures 2a-2j.For completeness, we report the coefficients from estimation of two VAR models in Web
Appendix E. These coefficients along with variance-covariance matrix are used for generating IRFs discussed
below. As mentioned earlier, the standard errors are not meaningful due to the high level of multicolinearity
amongst the lagged variables.
--Insert Figure 2 about here--
Demand Shock Demand. We find that a one standard deviation (1) increase in demand shock leads to
0.0049 (p<0.05) and 0.0050 (p<0.05) increase in realized demand (i.e. cost-of-goods-sold) in the HIT and LIT
sample respectively. These coefficients imply that demand increases by 0.49% and 0.50% for HIT and LIT
retailers compared to their demand four quarters back, respectively.5 These effects are immediate with no
significant persistence. After the quarter in which the demand shock occurs, the demand change quickly
reverts to zero. Additionally, the impact of demand shocks on two samples is uniform (i.e. 0.0049≈0.0050).
This suggests that difference in purchase behavior across two types of retailers is not attributable to different
demand shifts.
Demand Shock Purchases. We find that for a 1 increase in demand shock, an HIT retailer increases its
purchase by 0.0080 (p<0.05). This effect is immediate with no significant persistence. On the contrary, for a
1 increase in demand shock, the immediate increase in purchases is not statistically significant for the LIT
sample (0.0097, p>0.10). However, the impact of demand shock on purchases for the LIT sample is felt
beyond the first quarter and is persistent for up to four quarters. As seen in the IRF in Figure 2d the impact
of this demand shock is statistically significant in second through fourth quarter. This supports H1 that the
impact of demand shocks on purchases will be delayed for LIT retailers as the orders from LIT retailers are
delivered (as purchases) with a longer lag due to slower responsiveness of its supply chain compared to HIT
retailers.
5 LCOGSit = LCOGSit - LCOGSit-4 = Log(COGSit/COGSit-4). For HIT retailers the impact of 1 demand shock on demand is 0.0049 which translates to e(0.0049) for ratio of year-over-year change in demand. Thus for HIT retailers the ratio of year-over-year change in demand is 1.0049, i.e. the demand increases 0.49% over same quarter last year.
16
Additionally, the persistence of effect suggests that the response time of LIT retailers can be potentially
as long as four quarters. The total effect of 1 increase in demand shock on purchases for LIT sample
(0.0529, p<0.01) is greater (Diff = 0.0449, p<0.01) than that for the HIT sample. Thus in response to demand
shocks the LIT retailers increase their purchases by 4.6% more than their HIT counterparts. These two
results support H2 which had argued that order variance will be larger for LIT retailers compared to HIT
retailers.
Demand Shock Gross Margin. The impact of 1 increase in demand shock on gross margin is positive but
not significant for HIT retailers (see Figure 2e). However, LIT retailers increase their gross margin in
response to demand shocks (0.0043, p<0.01). This impact is only immediate (i.e. concurrent quarter) with no
persistence. This is along the expected direction as changing gross margin does not require advanced planning
unlike placing orders where lead time considerations are important. These findings support H3.
Demand Shock ABIG. Figure 2g and 2h suggest that both HIT and LIT retailers do not face immediate
increase in ABIG in response to demand shocks. This is inconsistent with H4. It suggests that even though
LIT retailers are able to respond slowly in changing their purchases, they are able to change gross margin
quickly to avoid ABIG in the current quarter. However, as shown above LIT retailers’ orders get delivered as
purchases after one quarter. So, consistent with this earlier finding, we observe abnormal inventory growth in
subsequent quarters as shown in Figure 2h. Thus, the impact of demand shocks on ABIG of LIT retailers
occurs after the current quarter but persists up to the fourth quarter.
The total impact of the demand shocks on ABIG of LIT retailers is significant (0.0304, p<0.01) and
greater than that for HIT retailers (Diff=0.0326, p<0.01). Thus quantity responsiveness mechanism adopted
by HIT retailers for mitigating impact of demand shocks appears to be more effective in preventing ABIG
compared to the price responsiveness mechanism adopted by the LIT retailers.
Demand Shock ROA. Figure 2i and 2j suggest that both HIT and LIT retailers face immediate changes
in firm performance, measured as return on asset, in response to demand shocks. The immediate (i.e.
HIT=0.0047≈LIT=0.0053) and total (i.e. HIT=0.0141≈LIT=0.0119) effect of demand shocks on ROA for
two samples is uniform. Thus the baseline impact of demand shocks on ROA is consistent across both HIT
and LIT retailers, after controlling for ABIG.
4.2. Results: Hypothesis 5
Results of the impacts of ABIG on ROA are reported in Table 7. In both sub-samples, we find support
for an inverted-U relationship between ABIG and ROA. This is consistent with the nonlinear relationships
demonstrated in Rumyantsev and Netessine (2007b) and Kesavan and Mani (2013). For the LIT sample, the
coefficients of the linear (0.1071, p<0.01) and quadratic terms (–0.8712, p<0.01) for ABIG are significant. For
the HIT sample, only the coefficient for the quadratic term is significant (–0.0693, p<0.01).
17
--Insert Table 7 about here--
To quantify the impact of ABIG on ROA of HIT and LIT retailers, we examine the effect of one
standard deviation change in ABIG from its mean values on the ROA of both types of retailers. The mean
() and standard deviation () of ABIG for HIT retailers in our sample are -0.023 and 0.349, respectively. For
HIT retailers, a 1 increase in ABIG at the mean leads to a decrease in ROA by 0.63%. Similarly, for these
retailers, a 1 decline in ABIG at the mean leads to a decrease in ROA by 1.08%. The mean and standard
deviation of ABIG for LIT retailers in our sample are –0.021 and 0.261, respectively. For LIT retailers, a 1
increase in ABIG leads to a 2.44% decrease in ROA and a 1 decline in ABIG leads to a 9.95% decrease in
ROA. Thus for a 1 increase (decrease) in ABIG the ROA of LIT retailers is 1.81% (8.87%) less than their
HIT counterparts.
We formally test these differences in the impact of ABIG on HIT and LIT retailers by performing 1,000
bootstraps of linear and quadratic coefficients from their two standard deviation asymptotic interval. The
mean values and associated standard errors for these bootstraps are reported in Table 8. We plot the mean of
these bootstrapped ROA values against ABIG for HIT and LIT retailers in Figure 3. The figure suggests that
the impact of ABIG on ROA is larger for LIT retailers compared to HIT retailers. We find that departure
from mean ABIG in either direction is associated with lower ROA. Importantly, we find that the implications
of these inventory excesses and shortages in terms of declining ROA are much more severe for LIT retailers
than for HIT retailers. Specifically, for the 2 increase in ABIG values, the decrease in ROA of LIT retailers
is about five times more severe than that for HIT retailers. The magnitude of this difference is much larger
for a 2 decline in ABIG values, where we observe the impact on LIT retailers to be close to eight times
more than that on HIT retailers. Thus, we find support for H5.
--Insert Table 8 and Figure 3 about here--
Overall, when synthesized together the results suggest that: (a) HIT and LIT retailers, on an average, face
similar demand expansion (or contraction) in uncertain economic times. (b) Shorter response times permit
HIT retailers to make smaller changes in purchases to mitigate the uncertainty of demand shocks. On the
contrary, LIT retailers have to make larger changes to their purchases because of their longer response time.
(c) While HIT retailers use quantity responsiveness to mitigate the impact of demand shocks, LIT retailers
use price responsiveness to do the same. (d) LIT retailers experience higher ABIG when faced with demand
shocks. This impact is sticky and can persist for as long as four quarters. (e) While, the impact of demand
shocks on firm performance is consistent across HIT and LIT retailers, the negative impact of ABIG on firm
performance is more severe and long lasting for LIT retailers vis-à-vis their HIT counterparts.
18
4.3. Robustness Checks
Alternate Methodology. We replace the VAR methodology used to test Hypotheses 1-4 with an alternate
methodology to validate the robustness of our results. Further, we use macroeconomic shocks as a proxy for
demand shocks in the VAR methodology. In the alternate methodology, we generate firm-level demand
shocks directly based on the Martingale Model of Forecast Evolution (MMFE) (Hausman 1969, Heath and
Jackson 1994), in which the difference in successive forecasts for a time period is used as a measure of
demand shock. We generate demand signals of different quarterly lead times using these shocks and examine
the signal(s) to which HIT and LIT retailers react. In recent research, Bray and Mendelson (2012) use a
similar MMFE methodology to decompose the bullwhip into a series of bullwhips based on information lead
time of the demand signals.
Let , ,ri qD denote the forecast made in quarter q-r for quantity demanded in quarter q for retailer i.
Following the MMFE convention, Di,q,0, represents the actual demand. Since the difference in forecasts from
two successive periods capture the new information that arrived during this time period (Heath and Jackson
1994), we define the demand shock faced by retailer i in quarter q-r as the difference1r,q,ir,q,i DD . This shock
represents the demand signal with lead time r.
Next we explain the model for forecasting quarterly demand for retailers. Since demand is typically not
observed at the firm level, it is conventional to use cost-of-goods-sold (COGS) as a proxy for the size of
demand at cost (example, Bray and Mendelson 2012). We estimate the following autoregressive demand
forecast model of order four at five discrete lagged time periods with r ∈ (1, 5) for each retailer i:
r,q,ir,q5i
3
0t
trq,i1t,i0iq,i FCOGSCOGS
where Fq,r represents the macroeconomic forecast for personal consumption expenditure (PCE) for quarter q
released in quarter q-r.
The fitted values from above regressions are used to obtain Di,q,r. Since we make comparisons across
retailers, we compute the lead-r demand signals after normalizing i.e.,
1r,q,i
1r,q,ir,q,i
r,q,iD
DDDS
. We measure
five demand shocks corresponding to r = 0...4, which are encountered by retailer i in quarters q, q–1, q–2, q–
3, and q-4, respectively.
We use these demand shocks to examine Hypotheses H1-H4. The detailed methodology and the results
are provided in Web Appendix F. We find support for all Hypotheses H1-H4. One advantage of this
methodology is that we can interpret the magnitude of the impact of demand shocks on different variables,
which was harder in the case of VAR models. For example, we find that the impact of 1% demand shock at
the mean values on change in purchases amounts to increases of $1.68m and $6.61m in YOY purchases for
19
HIT and LIT retailers respectively. Thus, LIT retailers increase their purchases four times as much as their
HIT counterparts. We also find that LIT retailers change their gross margin 2.5 times more than HIT
retailers.
In summary, we find that even with an alternate methodology that employs a different model
specification and different measures of demand shocks, we obtain consistent results that support our
contention that HIT and LIT retailers differ in their responses to demand shocks.
Alternate Classification of Retailers in High and Low Group. Recall, we classify retailers in high, medium, and
low inventory turn group, though we use only the high and low group for hypotheses testing. To ensure a
conservative test of our hypotheses we utilize the entire sample by classifying retailers into high (low)
inventory turn group if they are above (below) the median inventory turns in their industry for at least 3/4th
of the quarters that they are present in the sample. This permits us to utilize data from more than 300 retailers
(vis-à-vis 183 before) to test our hypotheses. The immediate and total impact from the IRFs from this sample
are reported in Web Appendix G. The results are directionally consistent with, albeit slightly weaker than, those
reported in Figure 2. Thus our results are not sensitive to alternate classification schemes.
Face Validity of Findings. Since we find that when faced with demand shocks, LIT retailers face higher
ABIG which subsequently has stronger detrimental impact on financial performance, it is likely that their (a)
long term financial performance and (b) survival rate should be inferior to those of their HIT counterparts.
To test the face validity of our findings we conduct two tests. First, we generate monthly portfolio returns of
HIT and LIT retailers for ten year window (2003-2012). We permit portfolio rebalancing at the end of each
month. We plot the value of HIT and LIT portfolios in Web Appendix H for $100 invested in each of them on
01/01/2003. We find that over the 10 year window the HIT portfolio outperforms the LIT portfolio by a
significant margin. By 12/31/2012, the value of HIT and LIT portfolios is $205 and $132 respectively. This
result is consistent with Alan et al. (2014).
Second, we examine rate of bankruptcies for two different types of retailers using the UCLA-LoPucki
Bankruptcy Database.6 From the database we extract bankruptcies filed by publicly traded US retailers
between 1985 and 2011. We find 36 instances of bankruptcies of which 18 are from LIT retailers and seven
are from HIT retailers. The remaining 11 are filed by medium inventory turn retailers. Thus half of all
bankruptcies are filed by LIT retailers which constitute 29% of our sample. On the other hand HIT retailers
which constitute 23% of the sample form only 19% of the bankruptcies filed. The fact that LIT retailers have
inferior stock market returns and lower survival rate provides additional evidence in support of our findings.
6 http://lopucki.law.ucla.edu/
20
5. Conclusions, Limitations, and Future Work
We use a vector auto regression (VAR) methodology to discern differences in ordering and pricing
behavior across HIT and LIT retailers. We find that HIT retailers are able to quickly respond to demand
shocks by changing their purchase quantity. LIT retailers appear to depend primarily on price changes to
manage demand shocks. Consistent with slower response time in their supply chains, LIT retailers need to
change their order quantities by more compared to HIT retailers. This results in relatively more ABIG for
LIT retailers. Finally, we find that such ABIG has a much larger detrimental effect on the financial
performance of LIT retailers compared to HIT retailers.
Our empirical analysis suggests areas for future work in theoretical and empirical research. First, lead time
may be an important predictor in the choice of price and quantity response to demand shocks. Most
theoretical papers in joint price-inventory optimization assume lead time to be zero. Incorporating lead time
in these models may help to reflect reality and help bridge the gap between theoretical research and practical
application. Finally, while our paper does not distinguish between ordering cost, lead time and review cycle,
future research may develop a methodology that permits estimation of the differential impacts of each of
these factors on the ability to respond to demand shocks.
The managerial implication of our paper is in quantifying the value of supply chain responsiveness in the
retail setting. While it is known that market mediation costs due to excesses and shortages would increase in
the presence of demand volatility for firms with less responsive supply chains, limited empirical evidence
exists about these costs at the firm-level (Randall and Ulrich 2001 is a notable exception). Recent moves by
U.S. manufacturers to source products from Mexico instead of China have been motivated by concerns about
responsiveness (Cave 2014). While price responsiveness might appear like a substitute to quantity
responsiveness to manage demand shocks, our paper suggests that price responsiveness as a strategy to
manage demand shocks may be inferior to being able to have quantity responsiveness. Thus, retailers with
less responsive supply chains cannot merely depend upon price changes through markups and markdowns to
manage demand shocks. Our paper uses data from about 300 retailers over a 25-year period to show that the
financial performance of HIT retailers are better than LIT retailers, in part due to the ability of HIT retailers
to postpone orders.
Our examination of HIT and LIT retailers also sheds new light on trade-offs faced by these managers.
Prior empirical research used observable factors such as gross margin, size, and growth to explain differences
in inventory turns across retailers. Our research, however, shows that HIT and LIT retailers may also differ
on unobservable factors such as ordering cost, lead time, and review cycle. These appear to be critical in
predicting how retailers react to demand shocks. Both ordering cost and lead time are drivers of the
21
“bullwhip effect,” the amplification of order variability as one moves upstream in a supply chain. Our finding
that orders from LIT retailers are more variable than those from HIT retailers suggests that supply chains of
LIT retailers could have large bullwhip effects. Future research may examine if there is a difference in the
magnitude of bullwhips generated by LIT retailers and HIT retailers.
22
References
Abarbanell, J., B. Bushee. 1997. Fundamental analysis, future earnings, and stock prices. J. Accounting Research. 35: 1-24.
Aguirregabiria, V. 1999. The dynamics of markups and inventories in retailing firms. The Review of Economic Studies. 66: 275-308.
Alan, Yasin, George P. Gao, and Vishal Gaur. Does Inventory Productivity Predict Future Stock Returns? A Retailing Industry Perspective. Management Science. Forthcoming.
Ang, A., G. Bekaert, M. Wei. 2007. Do macro variables, asset markets, or surveys forecast inflation better? J. Monetary Economics. 54: 1163-1212.
Bernstein, F., Li, Y., and K. Shang. 2014. A simple heuristic for joint inventory and pricing problems with lead time. Fuqua Working Paper.
Bollen, K.A., R. Stine. 1990. Direct and indirect effects: Classical and bootstrap estimates of variability. Sociological Methodology. 20. 115-140.
Bram, J., S.C. Ludvigson. 1998. Does consumer confidence forecast household expenditure? A sentiment index horse race. Economic Policy Review. 4(2): 59-78.
Bray, R., H. Mendelson. 2012. Information transmission and the bullwhip effect. Management Sci. 58 860-875.
Cachon, G. P., T. Randall, G. M. Schmidt. 2007. In search of the bullwhip effect. Manufacturing Service Oper. Management 9(4) 457–479.
Campbell, J.Y., R.J. Shiller. 1988. Stock prices, earnings, and expected dividends. J. Finance. 43(3): 661-676.
Carroll, C.D., J.C. Fuhrer, D.W. Wilcox. 1994. Does consumer sentiment forecast household spending: If so why? American Economic Rev. 84(5): 1397-1408.
Cave, Damien. 2014. As ties with China unravel, U.S. companies head to Mexico. May 31, 2014. Retrieved June 2, 2014 from http://www.nytimes.com/2014/06/01/world/americas/as-ties-with-china-unravel-us-companies-head-to-mexico.html?hp&_r=1
Chen, H., M.Z. Frank., O.Q.Wu. 2007. U.S. retail and wholesale inventory performance from 1981 to 2004. Manufacturing & Service Oper. Management. 9(4): 430-456.
Chen, X, D. Simchi-Levi. 2004. Coordinating inventory control and pricing strategies with random demand and fixed ordering cost: The infinite horizon case. Oper. Research. 52(6) 887-896.
Chen, Y., S. Ray., Y. Song. 2006. Optimal pricing and inventory control policy in periodic-review systems with fixed ordering cost and lost sales. Naval Research Logistics. 53(2) 117-136.
Dekimpe, M.G., D.M. Hanssens. 1999. Sustained spending and persistent response: A new look at long-term marketing profitability. J. Marketing Research. 36(4): 397-412.
Doms, M.E., N.J. Morin. 2004. Consumer sentiment, the economy, and the news media. FRB of San Francisco Working Paper No. 2004-09.
Enders. W. 1995. Applied Econometric Time Series. John Wiley & Sons: New York.
Eun, C.S., S. Shim. 1989. International transmission of stock market movements. J. Financial and Quantitative Analysis. 24(2): 241-256.
Fisher, M.L. 1997. What is the right supply chain for your product? Harvard Business Review. March-April, 105-116.
Gaur, V., M.L. Fisher, A. Raman. 2005. An econometric analysis of inventory turnover performance in retail services. Management Sci. 51(2) 181-194.
23
Granger, C.W.J., P. Newbold. 1974. Spurious regressions in econometrics. J. Econometrics. 2: 111-120.
Graves, S.C. 1999. A single-item inventory model for a nonstationary demand process. Manufacturing Service Oper. Management. 1(1) 50-61.
Hasbrouck, J. 1991. Measuring the information content of stock trades. J. Finance. 46(1): 179-207.
Hausman, WH. 1969. Sequential decision problems: A model to exploit existing forecasters. Management Sci, 16(2): 93-111.
Heath, D.C., P.L. Jackson. 1994. Modeling the evolution of demand forecasts with application to safety stock analysis in production/ distribution systems. IIE Transaction. 26(3) 17-30.
Jacobson R., D.A. Aaker. 1987. The strategic role of product quality. J. Marketing. 51(4): 31-44.
Jain, N., K. Girotra., S. Netessine. 2011. Managing global sourcing: Inventory performance. Working paper, INSEAD.
Kesavan, S., V. Gaur, A. Raman. 2010. Do inventory and gross margin data improve sales forecasts for US public retailers? Management Sci. 56: 1519-1533.
Kesavan, S. T. Kushwaha. 2014. Differences in retail inventory investment behavior during macroeconomic shocks: Role of service level. Production Oper. Management. Forthcoming.
Kesavan, S., V. Mani. 2013. The relationship between abnormal inventory growth and future earnings for US public retailers. Manufacturing Service Oper. Management. 15(1) 6-23.
Lamey, L., B. Deersnyder, M.G. Dekimpe, J.B.E.M. Steenkamp. 2007. How business cycles contribute to private-label success: Evidence from the U.S. and Europe. J. Marketing 71(1): 1-15.
Larson, C.R., D. Turcic, F. Zhang. 2011. Inventory write-downs, sales growth, and ordering policy: An empirical investigation. Working Paper, Washington University.
Lee, H., P. Padmanabhan, S. Whang. 1997. Information distortion in a supply chain: The bullwhip effect. Management Sci. 43: 546-558.
Lee. H.L., K.C. So, C.S. Tang. 2000. The value of information sharing in two-level supply chain. Management Sci. 46(5) 626-643.
Lemon, M., E. Portniaguina. 2006. Consumer confidence and asset prices: Some empirical evidence. Review of Financial Studies. 19(4): 1499-1529.
Lev, B., and S. R. Thiagarajan. "Fundamental information analysis." Journal of Accounting research 31 (1993): 190-190.
Levin, A., C.F. Lin, C.S.J. Chu. 2002. Unit root tests in panel data: Asymptotic and finite-sample properties. J. of Econometrics. 108: 1–24.
Ludvigson, S.C. 2004. Consumer confidence and consumer spending. J. Economic Perspective. 18(2): 29-50.
Pauwels, K., J. Silva-Risso, S. Srinivasan, D.M. Hanssens. 2004. New products, Sales promotion, and firm value: The case of automobile industry. J. Marketing. 68(4): 142-156.
Raghunathan, S. 2001. Information sharing in a supply chain: A note on its value when demand in nonstationary. Management Sci. 46(5) 626-643.
Rajagopalan, S. 2012. Factors driving inventory levels at US retailers. Working paper, University of Southern California.
Raman, A., V. Gaur., S. Kesavan. 2005. David Berman. Harvard Business School Case 605-081.
Randall, T., K. Ulrich. 2001. Product variety, supply chain structure, and firm performance: Analysis of the U.S. bicycle industry. 47(12): 1588-1604.
24
Rumyantsev, S., S. Netessine. 2007a. What can be learned from classical inventory models? A cross-industry exploratory investigation. Manufacturing Service Oper. Management. 9(4): 409-429.
Rumyantsev, S., S. Netessine. 2007b. Should inventory policy be lean or responsive? Working Paper, University of Pennsylvania.
Schwarz, G. Estimating dimensions of a model. The Anals of Statistics. 6(2): 461-464.
Sims, C.A. 1980. Macroeconomics and reality. Econometrica. 48(1): 1-48.
Sims, C.A., T. Zha. 1999. Error bands for impulse responses. Econometrica. 67: 1113:1155.
Souleles. N.C. 2004. Expectations, heterogeneous forecast errors, and consumption: Micro evidence from the Michigan consumer sentiment surveys. J. Money Credit & Banking. 36(1): 39-72.
Sterman, J.D. 1989. Modeling managerial behavior: Misperceptions of feedback in a dynamic decision making experiment. Management Sci. 35 321-339.
Stock, J.H., M.W. Watson. Vector autoregressions. J. Economic Perspectives. 15(4): 101-115.
Thomas, J., and X.J. Zhang. 2000. Identifying unexpected accruals: a comparison of current approaches. J. Accounting and Public Policy. 19: 347–376.
Thorbecke, W. 1997. On stock market returns and monetary policy. J. Finance. 52(2): 635-654.
Trusov, M., R.E. Bucklin, and K. Pauwels. 2009. Effect of word-of-mouth versus traditional marketing: Findings from a internet social networking site. J. Marketing. 73: 90-102.
Wu, O.Q., H. Chen. 2010. Optimal control and equilibrium behavior of production-inventory systems. Management Sci. 56(8): 1362-1379.
25
Table 1: Variables Description
Variables Measurement R
aw V
aria
ble
s
(Dat
a So
urc
e:
Co
mp
ust
at)
CPI Consumer price index as measure of inflation
INVTit CPI adjusted inventory (INVTQ)
COGSit CPI adjusted cost of goods sold (COGSQ)
SALESit CPI adjusted sales (REVQ)
NIit CPI adjusted net income in quarter (NIQ)
ATit CPI adjusted total assets (ATQ)
En
do
gen
ous
Var
iab
les Consumer Confidence Index
Macroeconomic shock
CCIt
CCI
it
Demand LCOGSit = Log(COGSit)
Purchases LPURCHit = Log[INVTit + COGSiq - INVTit-1]
Gross Margin Log [(SALESit - COGSit)/COGSit]
Abnormal Inventory Growth ABIGit = [(Iit – Iit-4) / Iit-4]–[ (COGSit – COGSit-4)/ COGSit-4]
Return on Assets ROAit = NIit/ATit
Med
iati
ng
Var
iab
le Inventory Turns ITit = COGSit/((INVTit + INVTit-1)/2)
HITi or LITi Classification
Retailer is classified as HITi if ITi,t-5 > 75th percentile for at least 3/4th of quarters Retailer is classified as LITi if ITi,t-5 < 25th percentile for at least 3/4th of quarters
26
Table 2: Data Summary: Different Industries
SIC 52 SIC 53 SIC 56 SIC 57 SIC 59 Overall Sample
Description Building Materials &
Hardware General
Merchandise Stores Apparel &
Accessories Stores Home Furniture &
Furnishing Stores Miscellaneous
Retailers
Example of Retailers
Home Depot, Lowe’s, National Lumber &
Supply, Tractor Supply & Co.
Kohl’s, J.C. Penney, Macy’s, Target
Aeropostale, Footlocker, Gap,
Stein Mart
Bombay Co., Linen N Things, Restoration
Hardware, Williams - Sonoma
Build-a-Bear, CVS, Staples,
Toys-r-Us
No. of Retailers 25 77 89 22 144 357
No. Of Observations 1,140 3,071 4,449 861 5,033 13,227
Inventory Turn (ITit) 1.17 .89 1.06 .87 1.30 1.11
(.60) (.39) (.48) (.35) (1.29) (.87)
Purchases (PURCHit in $m)
652.42 876.99 258.84 104.54 271.23 414.78
(1801.11) (1909.34) (513.28) (116.62) (779.16) (1175.81)
Cost of Goods Sold (COGSit in $m)
638.78 865.01 255.59 101.41 265.99 408.20
(1772.99) (1887.28) (511.58) (114.44) (762.83) (1160.12)
Ratio of PURCHit to COGSit
1.02 1.05 1.04 1.05 1.06 1.05
(.20) (.30) (.25) (.21) (.30) (.28)
Gross Margin (GMit) .28 .30 .36 .37 .35 .34
(.12) (.08) (.11) (.10) (.13) (.12)
Abnormal Inventory Growth (ABIGit)
.00 .02 .01 .01 .02 .01
(.28) (.36) (.26) (.15) (.15) (.14)
Return on Assets (ROAiq)
.0051 .0004 .0082 .0046 -.0027 .0021
(.0680) (.0927) (.0661) (.0672) (.0932) (.0840) Note: Mean (Standard deviation)
27
Table 3: Data Summary: HIT versus LIT
HIT Sample LIT Sample
Example of Retailers Amazon, Claries, Gap, J.C. Penney, Mays
Department Store, Petsmart, Office Depot, QVC, Target
Big Lots, Borders, Dicks Sporting Goods, Dillards, Eddie Bauer,
Footlocker, Toys R Us
Number of Retailers 81 102
Number of Observations 4,208 4,820
Frequency Distribution by Industries (52, 53, 56, 57, 59)
6, 21, 23, 4, 27 5, 25, 29, 4, 39
Inventory Turn (ITit) 1.65 .63
(.99) (.24)
Purchases (PURCHit in $m) 512.78 204.26
(1,166.14) (374.31)
Cost of Goods Sold (COGSit in $m) 506.25 200.98
(1,151.81) (385.09)
Ratio of PURCHit to COGSit 1.02 1.07
(.15) (.33)
Gross Margin (GMit) .52 .55
(.15) (.14)
Abnormal Inventory Growth (ABIGit)
.01 .01
(.30) (.24)
Return on Assets (ROAiq) .0075 -.0001
(.0622) (.0760)
Table 4: Probability of Classification in HIT (versus LIT) Group
Dependent Variable: P(HITit) Coeff. S.E.
ROAit-1 .9070 .3848
ASSETSit-1 (in $m) .0512 .0189
CCIt-1 .0011 .0012
Intercept 1.0073 .5426
Quarter Effects None Significant
Industry Segment Effects One Significant
Sample Size N 9,052
No. of Retailers 183 Note: All bold coefficients have p<.05 and italicized coefficients have p<.10.
Table 5: Transition between Different Inventory Turn Classification
Transition Matrix LIT (Low) MIT (Medium) HIT (High)
LIT (Low) 3,819 61 0
MIT (Medim) 780 3,778 790 HIT (High) 70 824 4,432
28
Table 6: Probability of Transition between Different Inventory Turn Classifications
Dependent Variable: P(TRANSITIONit)
Coeff. S.E.
ROAit-1 -.4191 .1878
ASSETSit-1 (in $m) -.6670 .6170
CCIt-1 -.0002 .0006
Intercept -.6833 .1505
Quarter Effects Two Significant
Industry Segment Effects One Significant
Sample Size N 13,983
No. of Retailers 357 Note: All bold coefficients have p<.05 and italicized coefficients have p<.10.
Table 7: Results: Retailer Performance Model
Dependent Variable - ROAit
Coefficient HIT Full Model
LIT Full Model
Coeff. S.E. Coeff. S.E.
CCIt α1 -.0001 .0001 .0000 .0002
ABIGiqt α 2 .0032 .0076 .1071 .0184
(ABIGit)2 α 3 -.0693 .0105 -.8712 .0456
ΔGMit α 4 .5305 .0631 .4655 .1209
(ΔGMit)2 α 5 -4.2094 .7338 -1.7303 1.6465
ROAit-1 α 6 -.2149 .0139 -.2858 .0199
Intercept α 0 .0199 .0056 .0102 .0101
QTR*SIC Dummies Yes Yes
Note: All bold coefficients have p<.05 and italicized coefficients have p<.10. HIT and LIT classification based on t-2.
Table 8: Results: Bootstraps for Comparing Non Linear Effects
Note: All bold coefficients have p<.05 and italicized coefficients have p<.10. HIT and LIT classification based on t-1.
ABIG HIT LIT Difference (HIT – LIT)
Mean S.E. Mean S.E. Mean S.E.
-.89 -.0571 .0118 -.7847 .0455 -.7276 .0332
-.68 -.0340 .0079 -.4778 .0283 -.4438 .0208
-.47 -.0169 .0048 -.2463 .0155 -.2294 .0115
-.27 -.0057 .0024 -.0900 .0068 -.0844 .0051
-.06 -.0004 .0005 -.0091 .0012 -.0087 .0009
.15 -.0011 .0013 -.0035 .0034 -.0025 .0025
.36 -.0077 .0034 -.0733 .0100 -.0656 .0075
.57 -.0202 .0061 -.2183 .0203 -.1981 .0150
.77 -.0387 .0095 -.4387 .0348 -.4000 .0255
.98 -.0631 .0139 -.7343 .0538 -.6712 .0393
29
Figure 1: Graphical Representation of Impulse Response Functions from VAR Specification
‘t’ ‘t+1’ ‘t+2’
In interest of parsimony we only highlight the impact of Demand Shocks on Demand and Purchases. Other effects similarly follow.
Immediate impact of demand shock in ‘t’ on purchases in ‘t’ (direct as well as through change in demand)
Delayed impact of demand shock in ‘t’ on purchases in ‘t+1’ and ‘t+2’ (direct as well as through change in demand)
+ Total Impact
Demand
Shock
Demand
Demand
Shock
Demand
Shock
Demand Demand
Purchases Purchases Purchases
30
Figure 2: Impulse Response Function of Demand Shocks on Retailer Outcomes
HIT Retailers LIT Retailers
Dem
an
d S
ho
ck
D
eman
d
(a)
Immediate Impact: .0049 (.0023); p<.05
Total Impact: .0049 (.0023); p<.05
(b)
Immediate Impact: .0050 (.0019); p<.05
Total Impact: .0050 (.0019); p<.05
Dem
an
d S
ho
ck
P
urc
hase
s
(c)
Immediate Impact: .0080 (.0031); p<.05
Total Impact: .0080 (.0031); p<.05
(d)
Immediate Impact: .0097(.0056); p>.05
Total Impact: .0529 (.0103); p<.01
Dem
an
d S
ho
ck
G
ross
Marg
in
(e)
Immediate Impact: .0028 (.0024); p>.10
Total Impact: .0028 (.0024); p>.10
(f)
Immediate Impact: .0043 (.0019) ); p<.05
Total Impact: .0043 (.0019) ); p<.05 Note: X-Axis - Quarters; Y-Axis – Sales (a/b), Purchases (c/d); Gross Margin (e/f); the solid line represents response to a one S.D. increase in demand shocks. The dotted lines represent +/-2.S.D. band.
-0.015
-0.01
-0.005
0
0.005
0.01
0.015
1 2 3 4 5 6 7 8 9 10
Response
P+2S.D.
M-2S.D.
-0.015
-0.01
-0.005
0
0.005
0.01
0.015
1 2 3 4 5 6 7 8 9 10
Response
P+2S.D.
M-2S.D.
-0.02
-0.015
-0.01
-0.005
0
0.005
0.01
0.015
0.02
0.025
1 2 3 4 5 6 7 8 9 10
Response
P+2S.D.
M-2S.D.
-0.02
-0.015
-0.01
-0.005
0
0.005
0.01
0.015
0.02
0.025
1 2 3 4 5 6 7 8 9 10
Response
P+2S.D.
M-2S.D.
-0.01
-0.005
0
0.005
0.01
1 2 3 4 5 6 7 8 9 10
Response
P+2S.D.
M-2S.D.
-0.01
-0.005
0
0.005
0.01
1 2 3 4 5 6 7 8 9 10
Response
P+2S.D.
M-2S.D.
31
HIT Retailers LIT Retailers
Dem
an
d S
ho
ck
A
BIG
(g)
Immediate Impact: -.0022 (.0033); p>.10
Total Impact: -.0022 (.0033); p>.10
(h)
Immediate Impact: .0037 (.0020); p<.10
Total Impact: .0304 (.0052); p<.01
Dem
an
d S
ho
ck
R
OA
(i)
Immediate Impact: .0047 (.0017); p<.01
Total Impact: .0141 (.0033); p<.01
(j)
Immediate Impact: .0053 (.0016); p<.01
Total Impact: .0119 (.0028); p<.01
Note: X-Axis - Quarters; Y-Axis – ABIG (g/h), RoA (i/j); the solid line represents response to a one S.D. increase in demand shocks. The dotted lines represent +/-2.S.D. band.
Figure 3: Abnormal Inventory Growth and Return on Assets
Note: Mid 98% values are used for plotting. Above illustrations are based on bootstrap values from Table 8.
-0.02
-0.015
-0.01
-0.005
0
0.005
0.01
0.015
0.02
1 2 3 4 5 6 7 8 9 10
Response
P+2S.D.
M-2S.D.
-0.02
-0.015
-0.01
-0.005
0
0.005
0.01
0.015
0.02
1 2 3 4 5 6 7 8 9 10
Response
P+2S.D.
M-2S.D.
-0.015
-0.01
-0.005
0
0.005
0.01
0.015
1 2 3 4 5 6 7 8 9 10
Response
P+2S.D.
M-2S.D.
-0.015
-0.01
-0.005
0
0.005
0.01
0.015
1 2 3 4 5 6 7 8 9 10
Response
P+2S.D.
M-2S.D.
-.90
-.80
-.70
-.60
-.50
-.40
-.30
-.20
-.10
.00
.10
-1.00 -.50 .00 .50 1.00
RO
A
ABIG
HITLIT
i
Web Appendix A
Table WAA1: Levin, Lin and Chu Unit Root Test for Stationarity of Series
Dependent Variable in Column is Granger Caused by Variable in Row
Level (p-value)
Fourth Differenced
(p-value)
CCI -8.84
Log Demand (LCOGS) -1.76 -4.36 Log Purchases (LPURCH) -1.55 -9.29 Log Gross Margin (LGM) -1.26 -89.16 Abnormal Inventory Growth (ABIG) -9.49 Return on Assets (ROA) -1.67 -53.49
Note: p<.01 (Bold and Italicized); p<.05 (Italicized)
ii
Web Appendix B
Table WAB1: Results of Granger Causality Test
Variable in Column is Granger Caused by Variable in Row
CCI
Log Demand
(LCOGS)
Log Purchases
(LPURCH)
Log Gross Margin
(LGM)
Abnormal Inventory Growth (ABIG)
Return on Assets
(ROA)
CCI 5.79 6.48 6.52 8.82 6.65
Log Demand
(LCOGS) 1.56 318.56 20.93 54.65 .53
Log Purchases
(LPURCH) 1.51 3.42 98.19 122.54 1.67
Log Gross Margin
(LGM) 2.72 49.11 2.33 11.40 2.81
Abnormal Inventory Growth
(ABIG) .95 299.41 131.70 19.68 24.09
Return on Assets
(ROA) 9.14 1.29 11.81 1.62 1.76
Note: Chi Square; p<.01 (Bold and Italicized); p<.05 (Italicized)
iii
Web Appendix C
Table WAC1: Results of Lag Order Selection Criteria
Lag Length HIT Sample LIT Sample
0 4.1533 3.9819 1 -1.2841 -1.9390 2 -1.2723 -1.9125 3 -1.3287 -1.9344 4 -1.3597* -1.9465* 5 -1.2894 -1.8617 6 -1.2240 -1.7651 7 -1.1882 -1.8141
Note: Schwarz Information Criterion
iv
Web Appendix D
Derivation of Impulse Response Function and their Standard Error
Let us consider a simple VAR system with two variables A and S, both being endogenous to each other.
t
t
S
A
1t
1t
t
t
S
A
S
A
(D1)
To derive IRF let us set 00tt SA before t (D2)
10 at t
00 after t
At t the impact is
1
0
1
0
0
0
S
A
t
t
(D3)
At t+ 1 the impact is
0
0
1
0
S
A
1t
1t (D4)
At t+2 the impact is
0
0
S
A
2t
2t (D5)
If SA (S Granger Causes A) the structural equations in D1 can be specified as:
tA tA1t1tt SAS (D6a)
tS tS1t1t SA (D6b)
Where, captures the contemporaneous impact of S on A which recovered though the residual covariance matrix as follows. Expanding the above system leads to the following reduced form specification:
tAtA3t32t21t13t32t21t1t ....SSS......AAAS (D7a)
tStS3t32t21t13t12t21t1 ....SSS....AAA (D7b)
The above set of recursive equations can be specified in matrix form as:
t
t
S
A
3t
3t
33
33
2t
2t
22
22
1t
1t
11
11
t
t....
S
A
S
A
S
A
S
AT
(D8)
Where,
10
1T
10
1T 1
Rearranging the above leads to
v
t
t
S
A1
3t
3t
33
331
2t
2t
22
221
1t
1t
11
111
t
tT....
S
AT
S
AT
S
AT
S
A
(D9)
Where,
ii
iiii
ii
ii
11
111
10
1T
and
t
t
t
tt
t
t
S
A
S
SA
S
A1T
The variance covariance matrix of the impulses is given by:
)(E)(E
)(E)(E2
SSA
SA
2
A
ttt
ttt
(D10)
2
2
222
222
S
A
SS
SSA
0
0
10
1
1
SA
1 T),(VARCOVARTtt
vi
Web Appendix E
Table WAE1: Coefficients from HIT Sample
CCI Coef. (S.E.)
LCOGSQ_4FD Coef. (S.E.)
LPURCHQ_4FD Coef. (S.E.)
LGM_4FD Coef. (S.E.)
ABIG Coef. (S.E.)
ROA_4FD Coef. (S.E.)
CCI(-1) 1.1703 0.0008 0.0010 -0.0000 -0.0008 0.0009
(0.0176) (0.0003) (0.0004) (0.0003) (0.0008) (0.0003)
CCI(-2) -0.3693 0.0007 0.0006 -0.0001 -0.0010 0.0005
(0.0268) (0.0005) (0.0006) (0.0005) (0.0009) (0.0002)
CCI(-3) 0.3751 0.0007 0.0003 -0.0003 0.0018 0.0001
(0.0273) (0.0004) (0.0006) (0.0005) (0.0010) (0.0003)
CCI(-4) -0.2532 0.0004 0.0007 0.0001 -0.0009 -0.0004
(0.0185) (0.0003) (0.0004) (0.0003) (0.0006) (0.0002)
LCOGSQ_4FD(-1) -1.5413 0.6686 0.7223 -0.0423 0.2735 -0.0255
(1.8459) (0.0301) (0.0410) (0.0309) (0.0594) (0.0225)
LCOGSQ_4FD(-2) -0.3637 0.0528 -0.0611 -0.0160 -0.1692 0.0381
(1.9774) (0.0322) (0.0440) (0.0331) (0.0636) (0.0241)
LCOGSQ_4FD(-3) -0.0313 0.1000 0.2964 -0.0679 -0.3015 0.1148
(1.9635) (0.0320) (0.0437) (0.0328) (0.0632) (0.0240)
LCOGSQ_4FD(-4) -0.4202 -0.1985 0.1543 0.0159 0.4851 -0.0377
(1.4765) (0.0241) (0.0328) (0.0247) (0.0475) (0.0180)
LPURCHQ_4FD(-1) 2.0947 0.1371 0.0456 0.0244 -0.1138 0.0011
(1.2304) (0.0200) (0.0274) (0.0206) (0.0396) (0.0150)
LPURCHQ_4FD(-2) -0.5397 -0.0002 0.1066 -0.0194 0.1565 0.0230
(1.2300) (0.0200) (0.0273) (0.0206) (0.0396) (0.0150)
LPURCHQ_4FD(-3) -0.4034 -0.0272 -0.0628 0.0761 0.1544 -0.0664
(1.2364) (0.0201) (0.0275) (0.0207) (0.0398) (0.0151)
LPURCHQ_4FD(-4) 0.8888 0.0234 -0.4595 0.0202 -0.3367 -0.0427
(1.1533) (0.0188) (0.0256) (0.0193) (0.0371) (0.0141)
LGM_4FD(-1) -0.0159 0.1157 0.1022 0.5186 0.0853 0.0063
(1.3518) (0.0220) (0.0301) (0.0226) (0.0435) (0.0165)
LGM_4FD(-2) -0.6047 0.1049 0.1321 -0.1217 -0.0893 0.0282
(1.4766) (0.0241) (0.0328) (0.0247) (0.0475) (0.0180)
LGM_4FD(-3) -0.1960 -0.0030 0.0657 0.1820 -0.0089 0.0645
(1.5346) (0.0250) (0.0341) (0.0257) (0.0494) (0.0187)
LGM_4FD(-4) -1.5411 0.0484 0.0276 -0.3327 0.0595 -0.1532
(1.2480) (0.0203) (0.0277) (0.0209) (0.0402) (0.0152)
ABIG(-1) 0.7037 0.0499 0.0232 -0.0150 0.6857 -0.0144
(0.6049) (0.0099) (0.0135) (0.0101) (0.0195) (0.0064)
ABIG(-2) -0.1449 -0.0202 -0.0598 0.0047 -0.0384 -0.0081
(0.6617) (0.0108) (0.0147) (0.0111) (0.0213) (0.0085)
ABIG(-3) 0.5391 0.0085 0.0353 -0.0203 -0.0582 -0.0085
(0.6537) (0.0107) (0.0145) (0.0109) (0.0210) (0.0080)
ABIG(-4) -0.5012 0.0064 -0.0051 0.0111 -0.0162 -0.0023
(0.4781) (0.0078) (0.0106) (0.0080) (0.0154) (0.0058)
ROA_4FD(-1) -0.9762 0.0236 0.1514 -0.0184 0.1662 -0.6507
(1.5264) (0.0249) (0.0339) (0.0255) (0.0491) (0.0186)
ROA_4FD(-2) -2.5154 0.0244 0.1746 0.0184 -0.0618 -0.4716
(1.8167) (0.0296) (0.0404) (0.0304) (0.0585) (0.0222)
ROA_4FD(-3) -4.4261 0.0596 0.2176 0.0320 0.0046 -0.3445
(1.8361) (0.0299) (0.0408) (0.0307) (0.0591) (0.0224)
ROA_4FD(-4) -3.4479 0.0670 0.1258 -0.0079 -0.0646 -0.0550
(1.6135) (0.0263) (0.0359) (0.0270) (0.0519) (0.0197)
C 7.6727 0.0089 0.0110 0.0153 -0.0291 0.0041
(0.6889) (0.0112) (0.0153) (0.0115) (0.0222) (0.0084)
Adj. R-squared 0.8926 0.6772 0.5474 0.2732 0.4409 0.3344
vii
Table WAE2: Coefficients from LIT Sample
CCI Coef. (S.E.)
LCOGSQ_4FD Coef. (S.E.)
LPURCHQ_4FD Coef. (S.E.)
LGM_4FD Coef. (S.E.)
ABIG Coef. (S.E.)
ROA_4FD Coef. (S.E.)
CCI(-1) 1.1871 0.0008 0.0004 0.0003 0.0004 0.0008
(0.0180) (0.0003) (0.0009) (0.0001) (0.0003) (0.0003)
CCI(-2) -0.4227 0.0006 0.0013 0.0001 0.0004 0.0004
(0.0278) (0.0004) (0.0006) (0.0002) (0.0005) (0.0002)
CCI(-3) 0.4442 0.0002 0.0011 -0.0001 0.0010 -0.0003
(0.0286) (0.0004) (0.0005) (0.0004) (0.0005) (0.0003)
CCI(-4) -0.2833 0.0002 0.0009 -0.0002 0.0012 -0.0001
(0.0192) (0.0003) (0.0004) (0.0003) (0.0004) (0.0002)
LCOGSQ_4FD(-1) -1.5181 0.7633 0.6982 0.0713 -0.0615 -0.0201
(2.6280) (0.0351) (0.0842) (0.0349) (0.0473) (0.0289)
LCOGSQ_4FD(-2) -1.8899 0.0434 -0.1569 -0.0247 0.0048 -0.0537
(2.9794) (0.0398) (0.0954) (0.0396) (0.0536) (0.0328)
LCOGSQ_4FD(-3) 0.4545 0.0235 0.4258 -0.0200 0.1569 0.0439
(2.9486) (0.0394) (0.0944) (0.0392) (0.0531) (0.0324)
LCOGSQ_4FD(-4) 1.6619 -0.0669 -0.0108 -0.0306 0.0373 -0.0065
(1.9728) (0.0264) (0.0632) (0.0262) (0.0355) (0.0217)
LPURCHQ_4FD(-1) 0.5003 0.0716 0.0975 -0.0188 -0.0112 0.0196
(0.9321) (0.0125) (0.0299) (0.0124) (0.0168) (0.0102)
LPURCHQ_4FD(-2) -0.3071 0.0125 0.0411 -0.0051 0.0086 0.0076
(0.9365) (0.0125) (0.0300) (0.0124) (0.0169) (0.0103)
LPURCHQ_4FD(-3) 0.6963 0.0216 -0.0365 -0.0147 -0.0158 -0.0027
(0.9117) (0.0122) (0.0292) (0.0121) (0.0164) (0.0100)
LPURCHQ_4FD(-4) 0.0496 -0.0688 -0.4509 0.0160 -0.0825 0.0112
(0.6818) (0.0091) (0.0218) (0.0091) (0.0123) (0.0075)
LGM_4FD(-1) -0.9463 0.1566 0.1589 0.5630 -0.0639 0.0137
(1.5292) (0.0205) (0.0490) (0.0203) (0.0275) (0.0168)
LGM_4FD(-2) -0.5027 -0.0378 -0.0536 0.0453 0.0341 -0.0096
(1.8665) (0.0250) (0.0598) (0.0248) (0.0336) (0.0205)
LGM_4FD(-3) 0.4801 -0.0119 -0.0181 -0.0800 0.0225 0.0151
(1.9800) (0.0265) (0.0634) (0.0263) (0.0356) (0.0218)
LGM_4FD(-4) -1.0640 0.1363 0.1354 -0.4030 0.0303 0.0132
(1.7500) (0.0234) (0.0561) (0.0233) (0.0315) (0.0192)
ABIG(-1) -1.2614 0.2814 0.1032 -0.0060 0.3389 -0.0354
(1.5684) (0.0210) (0.0502) (0.0208) (0.0282) (0.0172)
ABIG(-2) 0.3094 -0.0454 -0.2263 -0.0365 0.0925 -0.0305
(1.9082) (0.0255) (0.0611) (0.0254) (0.0343) (0.0160)
ABIG(-3) 1.5017 -0.0122 0.1370 0.0252 0.0396 -0.0321
(1.8734) (0.0251) (0.0600) (0.0249) (0.0337) (0.0166)
ABIG(-4) 0.2182 0.1113 -0.1395 -0.0309 -0.3009 -0.0320
(1.5306) (0.0205) (0.0490) (0.0203) (0.0275) (0.0168)
ROA_4FD(-1) 1.6902 0.1272 0.3869 0.0143 0.0064 -0.7036
(1.9346) (0.0259) (0.0620) (0.0257) (0.0348) (0.0213)
ROA_4FD(-2) 3.4037 0.1749 0.4537 0.0229 -0.0221 -0.5263
(2.2743) (0.0304) (0.0728) (0.0302) (0.0409) (0.0250)
ROA_4FD(-3) 3.8101 0.1425 0.4603 0.0333 0.0160 -0.3576
(2.2101) (0.0296) (0.0708) (0.0294) (0.0398) (0.0243)
ROA_4FD(-4) -1.5529 0.1188 0.3346 -0.0143 0.0747 -0.0325
(1.8733) (0.0251) (0.0600) (0.0249) (0.0337) (0.0206)
C 7.3916 0.0146 0.0125 0.0260 -0.0209 -0.0175
(0.7071) (0.0095) (0.0227) (0.0094) (0.0127) (0.0078)
Adj. R-squared 0.8975 0.6707 0.3080 0.4768 0.2860 0.3188
viii
Web Appendix F
Alternate Methodology
In the main paper we used a VAR methodology to test Hypotheses H1-H4. In this Appendix, we use a
reduced form methodology to test Hypotheses H1-H4. H1 and H2 are based on the impact of demand
shocks on purchases, while H3 and H4 are based on the impact of demand shocks on abnormal inventory
growth and gross margin, respectively.
Calculation of demand shocks
Our methodology to measure demand shocks is based on the Martingale Model of Forecast Evolution
(MMFE) (Hausman 1969, Heath and Jackson 1994), in which the difference in successive forecasts for a time
period is used as a measure of demand shock. We generate demand signals of different quarterly lead times
using these shocks and examine the signal(s) to which HIT and LIT retailers react. In other words we use the
information lead time to gauge the physical lead time of high and low inventory turn retailers. In recent
research, Bray and Mendelson (2012) use a similar MMFE methodology to decompose the bullwhip into a
series of bullwhips based on information lead time of the demand signals.
Let , ,i q rD denote the forecast made in quarter q-l for quantity demanded in quarter q for retailer i.
Following the MMFE convention, , ,0i qD, represents the actual demand. Since the difference in forecasts
from two successive periods capture the new information that arrived during this time period (Heath and
Jackson 1994), we define the demand shock faced by retailer i in quarter q-r as the difference , , , , 1i q r i q rD D .
This shock represents the demand signal with lead time r.
Next we explain the model for forecasting quarterly demand for retailers. Since demand is typically not
observed at the firm level, it is conventional to use cost-of-goods-sold (COGS) as a proxy for the size of
demand at cost (example, Bray and Mendelson 2012). We estimate the following autoregressive demand
forecast model of order four at five discrete lagged time periods with l ∈(1, 5) for each retailer i:
(Eq F1)
3
, 0 , 1 , 5 , , ,
0
i q i i t i q r t i q r i q r
t
COGS COGS F
where Fq,r represents the macroeconomic forecast for personal consumption expenditure (PCE) for quarter q
released in quarter q-r. We use a 4th-order autoregressive model because of considerations of seasonality in
retail demand and model fit criteria.
ix
The fitted values from above regressions are used to obtain , ,i q rD Since we make comparisons across
retailers, we compute the lead-r demand signals after normalizing i.e., , , , , 1
, ,
, , 1
i q r i q r
i q r
i q r
D DDS
D
. We measure
five demand shocks corresponding to l = 0...4, which are encountered by retailer i in quarters q, q–1, q–2, q–3,
and q-4, respectively. Figure WAF1 illustrates the temporal sequencing of demand shocks and orders placed.
--Insert Figure WAF1 about here--
Model Specification to test H1-2
The model specification to estimate the impact of demand shocks on purchases is as follows:
(Eq F2)
4
1 ,0 1 , , 5 , 4 ( ),
1
iq o q r i q r iq i q j i q iq
r
LP AS DS LGM W
where Δ represents year-over-year (YOY) change (i.e. ΔLPq= LPq-LPq-4), L indicates that the respective
variables have been logged, DSi,q,r is demand shock in quarter q-r in the forecast for quarter q, GM denotes
gross margin, W are a lagged set of control variables, x are the response coefficients, are the coefficients of
control variables, j(i),q are industry-quarter dummies for the industry j(i) that firm i is classified in, and ξ is a
normally distributed random error. We avoid using to capture the contemporaneous demand shock
as it would cause the coefficients’ estimates to be biased because both the demand shock and the dependent
variable are functions of cost-of-goods-sold. Instead, we include the actual macroeconomic shock in that
quarter, (measured as the change in personal consumption expenditure), as it would be correlated with
the demand shock experienced by retailers.
Since purchases in a given quarter are the result of orders placed earlier, we use actual macroeconomic
shock from quarter q and demand shocks from quarters q-1 to q-4 to examine the lead times of orders for
retailers. A significant coefficient for , ,i q rDS would indicate that retailers placed orders in quarter q-r for
quarter q. We expect purchases of long lead time retailers to be correlated with older shocks compared to
those of short lead time retailers. This is the basis of our identification strategy.
The control variables include: ΔLCOGSq-4 as a proxy for the YOY change in demand, ΔLIq-4 is the YOY
change in ending inventory to control for replenishment effects, ΔLATq-4 represents YOY changes in assets,
to control for such impacts on inventory purchases as new store openings and existing store closings, and
ΔICOSTq-4 represents YOY change in inventory carrying cost, to control for the cost of financing inventories
(Rumyantsev and Netessine 2007a).
, ,0i qDS
,0qAS
x
Logarithms of variables are used in order to account for differences in scale across firms in the data set.
YOY change in variables is used to account for seasonality and eliminate firm-wise fixed effects that would be
present in a levels model. To account for industry- and time-specific unobserved heterogeneity, we include
industry-quarter (j(i),q) dummies in the above equation. To account for state dependence, we include a lagged
change in purchases (ΔLPi,q-4). Since quantity and pricing decisions are made jointly we include
contemporaneous change in gross margin (ΔLGMq) as an explanatory variable in the above model. We
resolve the endogeneity of contemporaneous change in gross margin using lagged value (ΔLGMq-4) as
instrument.
In H1 we proposed that HIT retailers have a shorter lead time than LIT retailers. The statistical
significance of different shock coefficients across the two samples would indicate the relative importance of
lead time. For example, if 1, the coefficient for the latest shock, is significant in the HIT sample but not in
the LIT sample, that would suggest HIT retailers have a shorter lead time compared to LIT retailers. In H2,
we proposed that the ordering behavior of HIT retailers is significantly less impacted by demand shocks
compared to that of LIT retailers. The difference between the sum of coefficients 1 through 5 for the HIT
and LIT subsamples provides a test of H2.
Model Specification to test H3
In H3 we proposed that LIT retailers are more susceptible to excesses and shortages of inventory vis-à-
vis HIT retailers when faced with unexpected demand shocks. We use the following model specification to
estimate the impact of actual demand shock on ABIG after controlling for all other variables that may drive
ABIG:
(Eq F3)
As explained earlier, we use actual macroeconomic shock as a proxy for actual demand shock in this
regression. The control variables in Yi,q-1 include ABIGi,q-1, ΔLATi,q-1, and ΔICOSTi,q-1. To account for
simultaneity between abnormal inventory growth and pricing decisions, we include contemporaneous change
in gross margin (ΔLGMi,q) in the model. We resolve this endogeneity using lagged value (ΔLGMiq-4). To
account for industry- and time-specific unobserved heterogeneity, we include quarter-industry dummies (δj(i),q)
in the above equation. The difference between the estimates of 1 for the HIT and LIT subsamples is a test
of H3.
Model Specification to test H4
In H4, we proposed that LIT retailers are likely to be more price-responsive than HIT retailers when
faced with demand shocks. The model specification to study the impact of actual demand shocks on gross
margin, after controlling for all factors that might predict change in gross margin, is as follows:
, 0 1 ,0 2 , , 1 ( ), ,i q q i q i q j i q i qABIG AS LGM Y
xi
(Eq F4)
Here, γ are the response parameters, Xi,q-1 are the control variables, are the coefficients of control variables,
ηj(i),q are the industry-quarter dummies, and i,q is the normally distributed random-error term. The control
variables include lagged change in (log) assets (ΔLATi,q-1), (log) gross margin (ΔLGMi,q-1), inventory holding
cost (ΔICOSTi,q-1), and contemporaneous change in purchases (ΔLPi,q). We include contemporaneous change
in purchases to account for simultaneity in price and quantity decisions. We resolve the endogeneity of ΔLPi,q
using lagged value (ΔLPi,q-2) as instrument. We do not use demand shocks from prior quarters because the
lead time associated with changing prices is usually negligible. The direct impact of macroeconomic demand
shocks ASq,0 on gross margin is captured by coefficient γ1. The difference between the estimates of
coefficient γ1 for HIT and LIT subsamples is the test of H4.
Results
Estimation of demand shocks of HIT and LIT retailers: The summary statistics for demand shocks are provides
in Tables AH1 and AH2. For the 100 quarters in our data window (1985–2009), the mean values of the
forecasted demand shocks 𝐷𝑆𝑖,𝑞,1, 𝐷𝑆𝑖,𝑞,2, 𝐷𝑆𝑖,𝑞,3, and 𝐷𝑆𝑖,𝑞,4 for the entire sample are -0.59%, -0.29%, -
0.17%, and -0.38%, respectively. Further the demand shocks across HIT and LIT retailers are comparable in
magnitude, as shown in Table WAF1. Finally, we find that the correlation across forecasted demand shocks is
low, as shown in Table WAF2, indicating that the shocks are independent of each other, i.e., they contain
new information.
--Insert Table WAF1 & WAF2 about here--
Next we present the results of purchase, ABIG, and gross margin models in Table AH4-6, respectively.
Results: Quantity Response
From Table WAF3, we observe that HIT retailers react to only the current macroeconomic shock
(1=.0023, p<.10) and to the last forecast update (2=.0037, p<.05). LIT retailers do not react to these two
shocks (1=.0008, p>.10, and 2=.0029, p>.10), showing that they have longer lead times. Instead, they react
to older shocks that occur two quarters (3=.0185, p<.01), three quarters (4=.0110, p<.01), and four quarters
(5=.0002, p<.01) ago. Since LIT retailers do not respond to demand shocks in the current and the preceding
quarter, our estimates imply that their lead time is longer than three months. This evidence supports H1 and
suggests that LIT retailers have longer lead times compared to HIT retailers.
--Insert Table WAF3 about here--
We find that the impact of shocks on purchases for LIT retailers is .0334 (p<.01), and for HIT retailers is
.0088 (p<.05). The difference in magnitude of these coefficients across the two subsamples suggests that LIT
retailers react more strongly to demand shocks compared to their HIT counterparts (diff = .0246, p<.01). The
, 0 1 ,0 2 , , 1 ( ), , .i q q i q i q j i q i qLGM AS LP X
xii
impact of 1% demand shock at the mean values on change in purchases amounts to increases of $1.68m and
$6.61m in YOY purchases for HIT and LIT retailers respectively. Thus LIT retailers increase their purchases
four times as much as their HIT counterparts. This result is interesting as it presents evidence that order
variability of LIT retailers is significantly more sensitive to demand shocks than that of HIT retailers. Thus,
we find support for H2.
We note that the coefficients’ estimates of the control variables are in the expected direction: the greater
the closing inventory, the lower is the inventory investment made in the current quarter; and the greater the
change in sales and assets, the greater the inventory investment in the current quarter. Thus, retailers with a
higher growth rate and greater increase in assets are likely to make more purchases.
Abnormal Inventory Growth Model. Next consider the impact of actual demand shocks on abnormal
inventory growth for HIT and LIT retailers. The results of nested ABIG models are reported in Table
WAF4. The coefficient 1 measures the impact of actual macroeconomic shock encountered in quarter q on
ABIG during quarter q. The impact of actual demand shock on ABIG for LIT retailers is (.0003, p>.10),
while that for HIT retailers is (.0008, p>.10). The difference between the sum of coefficients of the HIT and
LIT subsamples (diff = .0005, p>.10) is not significant. Thus we do not find support for H3. This result is
consistent with our VAR analysis where we found that demand shocks do not create excesses and shortages
in the contemporaneous quarter. However, the VAR analysis showed that LIT retailers have abnormal
inventory growth in the subsequent quarters as their orders get delivered with a delay.
--Insert Table WAF4 about here--
Results: Price Response
In H4, we argued that LIT retailers are likely to change their prices more than HIT retailers to manage
demand shocks. Table WAF5 shows that the coefficient, 1, for LIT retailers is .0026 (p<.01), indicating that
demand shocks are associated with significant price changes for LIT retailers. We find that HIT retailers also
change their gross margin when faced with demand shocks (.0011, p<.05). A comparison of the magnitudes
of coefficients across HIT and LIT retailers (diff = .0015, p<.01) emphasizes the significantly large impact of
demand shock on the gross margin of LIT retailers compared to HIT retailers. Thus we find support for H4.
--Insert Table WAF5 about here—
In summary, our results are consistent with those reported in the VAR analysis in the main section of the
paper. We find that with an alternate methodology that employs a different model specification and different
measures of demand shocks, we obtain consistent results that support our contention that HIT and LIT
retailers differ in their responses to demand shocks.
xiii
Table WAF1: Summary Statistics: Forecasted Demand Shock
Overall HIT Sample LIT Sample
Mean(%) S.D. Mean(%) S.D. Mean(%) S.D.
𝐷𝑆𝑖,𝑞,1 -..587 7.501 -.426 7.370 .-.751 7.629
𝐷𝑆𝑖,𝑞,2 -.292 6.424 -.342 6.373 -.242 6.474
𝐷𝑆𝑖,𝑞,3 -.167 6.236 -.333 6.137 -.194 6.335
𝐷𝑆𝑖,𝑞,4 -.384 9.646 -.367 9.435 -.536 9.856
Note: All bold coefficients have p<.05 and italicized coefficients have p<.10. We remove 5% of outliers on both ends.
Table WAF2: Correlation Between Forecasted Shocks
𝐷𝑆𝑖,𝑞,1 𝐷𝑆𝑖,𝑞,2 𝐷𝑆𝑖,𝑞,3 𝐷𝑆𝑖,𝑞,4
𝐷𝑆𝑖,𝑞,1 1.00
𝐷𝑆𝑖,𝑞,2 .03 1.00
𝐷𝑆𝑖,𝑞,3 -.05 .01 1.00
𝐷𝑆𝑖,𝑞,4 -.01 -.03 -.02 1.00
Table WAF3: Results: Purchases Model
Dependent Variable - ΔLPiq
Coefficient HIT Full Model
LIT Full Model
Coeff. S.E. Coeff. S.E.
𝐴𝑆𝑞,0 1 .0023 .0014 .0008 .0017
𝐷𝑆𝑖,𝑞,1 2 .0037 .0016 .0029 .0027
𝐷𝑆𝑖,𝑞,2 3 .0006 .0014 .0185 .0016
𝐷𝑆𝑖,𝑞,3 4 .0021 .0016 .0110 .0013
𝐷𝑆𝑖,𝑞,4 5 .0001 .0012 .0002 .0000
ΔLGMiqPRED 6 -.6291 .0813 -1.2704 .1852
ΔLCOGSiq-4 7 .0207 .0307 .2042 .0318
ΔLIiq-4 8 -.0355 .0205 -.0677 .0389
ΔLATq-4 9 .1439 .0224 .1843 .0375
ΔICOSTiq-4 10 .0460 .1526 -.0788 .1820
ΔLPiq-4 11 .0407 .0263 -.3314 .0204
Intercept 0 -.0236 .0929 -.0186 .1281
QTR*SIC Dummies Yes Yes
Chi Square (d.f.) 246.01 (29) 326.85 (29)
.0088 .0032 .0334 .0038
Note: All bold coefficients have p<.05 and italicized coefficients have p<.10. HIT and LIT classification based on q-5.
xiv
Table WAF4: Results: Abnormal Inventory Growth Model
Dependent Variable - ABIGiq
Coefficient HIT Full Model
LIT Full Model
Coeff. S.E. Coeff. S.E.
𝐴𝑆𝑞,0 1 .0008 .0011 .0003 .0009
ΔLGMiqPRED
2 -.5305 .0710 -.6244 .1089
ΔLATiq-1 3 -.0348 .0139 -.0553 .0135
ΔICOSTiq-1 4 .1579 .1265 .2099 .1145
ABIGiq-1 5 .4927 .0132 .5245 .0116
Intercept 0 -.0161 .0691 -.0468 .0804
QTR*SIC Dummies Yes Yes
Chi Square (d.f.) 1256.63 (24) 1876.01 (24)
Note: All bold coefficients have p<.05 and italicized coefficients have p<.10. HIT and LIT classification based on q-2.
Table WAF5: Results: Gross Margin Model
Dependent Variable - ΔLGMiq
Coefficient HIT Full Model
LIT Full Model
Coeff. S.E. Coeff. S.E.
𝐴𝑆𝑞,0 γ1 .0011 .0005 .0026 .0003
ΔLPiq PRED
γ 2 -.0048 .0237 -.0272 .0103
ΔLATiq-1 γ3 .0023 .0057 -.0055 .0035
ΔICOSTiq-1 γ 4 .0786 .0540 .0367 .0317
ΔLGMiq-1 γ 5 .1244 .0075 .1377 .0076
Intercept γ 0 .0109 .0326 -.0167 .0223
QTR*SIC Dummies Yes Yes
Chi Square (d.f.) 296.15 (24) 345.93 (24)
Note: All bold coefficients have p<.05 and italicized coefficients have p<.10. HIT and LIT classification based on q-2.
xv
Figure WAF1: Temporal Sequencing of Events and Modeling Choices
xvi
Web Appendix G
Table WAG1: Alternate Classification of Retailers in High and Low Group
Effect HIT_ALT Sample LIT_ALT Sample Diff
Demand Shock
Demand
Immediate:.0059 (.0020); p<.01
Total: :.0059 (.0020); p<.01
Immediate: .0066 (.0020), p<.01
Total: .0066 (.0020), p<.01
Immediate: n.s.
Total: n.s.
Demand Shock
Purchases
Immediate: .0122 (.0029); p<.01
Total:.0212 (.0043); p<.01
Duration: 1Q to 2Q
Immediate: .0104 (.0063); n.s.
Total:.0422 (.0099); p<.01
Duration: 2Q to 4Q
Immediate: n.s.
Total:.0210 (.0107); p<.05
Demand Shock
Gross Margin
Immediate: .0016 (.0035); n.s.
Total: .0016 (.0035); n.s.
Immediate: .0055 (.0024); p<.05
Total: .0055 (.0024); p<.05
Immediate: n.s.
Total: n.s.
Demand Shock
ABIG
Immediate: -.0002 (.0050); n.s.
Total: -.0002 (.0050); n.s.
Immediate: .0009 (.0035); n.s.
Total:.0164 (.0084); p<.05
Duration: 2Q to 4Q
Immediate: n.s.
Total: .0166 (.0098); p<.10
Demand Shock
ROA
Immediate: .0038 (.0017); p<.05
Total: .0150 (.0034); p<.01
Immediate:.0037 (.0012); p<.01
Total: .0096 (.0024); p<.01
Immediate: n.s.
Total: n.s.
ABIG ROA Immediate: -.0080 (.0012); p<.01
Total: -.0080 (.0012); p<.01
Immediate: -.0115 (.0017); p<.01
Total: -.0179 (.0028); p<.01
Immediate: -.0035 (.0021); p<.10
Total: .0099 (.0030); p<.01
xvii
Web Appendix H
Figure WAH1: Portfolio Returns: HIT versus LIT Sample
$-
$50
$100
$150
$200
$250 D
ec-0
2
Mar
-03
Jun
-03
Sep
-03
Jan
-04
Apr-
04
Jul-
04
Oct
-04
Feb
-05
May
-05
Au
g-0
5
Dec
-05
Mar
-06
Jun
-06
Oct
-06
Jan
-07
Apr-
07
Jul-
07
Nov
-07
Feb
-08
May
-08
Au
g-0
8
Dec
-08
Mar
-09
Jun
-09
Oct
-09
Jan
-10
Apr-
10
Au
g-1
0
Nov
-10
Feb
-11
May
-11
Sep
-11
Dec
-11
Mar
-12
Jul-
12
Oct
-12
Port
foli
o V
alu
e
Month-Year
HIT Portfolio
LIT Portfolio