The folly of forecasting:
The effects of sales forecast accuracy and bias on inventory and production
decisions under aggregated and disaggregated forecasting regimes
Alexander Brüggen
Maastricht University
[email protected]
+31 (0)43 3884924
Isabella Grabner
Maastricht University
[email protected]
+31 43 38 84629
Karen Sedatole*
Michigan State University
[email protected]
517-432-2919
December 2014
* Corresponding author. We are grateful to senior managers at the firm that supplied
data for the analysis and for extensive access to the research site and its employees. This
study was conducted with the generous support of the Institute of Management
Accountants (IMA). This study benefited from the helpful comments of Shannon
Anderson, Joe Fisher, Judy Whipple, Sriram Narayanan, the participants of the 2013
GMARS conference in East Lansing, and the participants of the Research Conference
on Management Accounting from the Field: Sponsored by Accounting, Organizations
and Society and Journal of Management Accounting Studies.
Electronic copy available at: http://ssrn.com/abstract=2482321
1. Introduction
Budgetary systems, including both the annual “master budget” and the periodic
subsequent forecasts, play two primary roles within business organizations. First, the master
budget plays a control role by providing cost and sales benchmarks against which managerial
performance can be assessed, and second, budgets and updated forecasts play a planning and
coordination role for organizational activities (Hansen and Van der Stede 2004; Libby and
Lindsay 2010; Horngren et al. 2011; Henttu-Aho and Järvinen 2013). While the determinants
and consequences of budget-based control practices are among the most widely studied topics
in management accounting research (Covaleski et al. 2003), there is little accounting research
examining the planning and coordination role of master budgets, and even less research
examining the role that forecasts play in the coordination of activities through time and across
functional units (Selto and Widener 2001; Hansen et al. 2003; Hansen and Van der Stede
2004).
This void in the literature is surprising given that many scholars and practitioners alike
have noted an important organizational tension; namely, the use of budgetary systems for
control leads to “gaming” by managers which undermines the usefulness of those systems for
planning and coordination (Bittlestone 2000; CIMA 2004; Cassar and Gibson 2008;
Sivabalan et al. 2009). This void in the literature is also counter to evidence that indicates
practitioners (i) regard planning uses of budgetary systems as more important than control
uses (Sivabalan et al. 2009), (ii) argue that forecasts are supplanting the budget as the primary
planning and coordination tool, especially in highly uncertain environments where the master
budget quickly becomes obsolete (Bittlestone 2000; CIMA 2009; Vadasz and Lorain 2010;
Hagel 2014; Sivabalan et al. 2009; Ekholm and Wallin 2011), and (iii) perceive shortcomings
in both budgets and forecasts as planning and coordination tools.
The purpose of this study is to examine the coordinating role of sales forecasts in the
1
Electronic copy available at: http://ssrn.com/abstract=2482321
coordination of demand-facing activities of sales and marketing with supply-facing activities
of operations (e.g., production, logistics, supply chain management), a process referred to as
“sales and operations planning” (S&OP) (Oliva and Watson 2011). More specifically, we
examine the effect of sales forecast accuracy and bias on inventory and production decisions.
Sales forecast accuracy is the unsigned deviation of forecasted sales from actual sales.
Forecast bias is the signed forecast error, with positive bias indicating a forecast greater than
actual sales and negative bias indicating a forecast lower than actual sales. We further
examine how these relations differ between an aggregated and disaggregated sales forecasting
regime.
Our research setting is a large agricultural chemical manufacturing organization. Our
site is an ideal setting to analyze effects of a forecasting system on the S&OP process for at
least three reasons. First, the industry is characterized by high volatility in customer demand
ensuing from market (i.e., competitor), economic (i.e., commodity prices), and natural forces
(e.g., weather patterns). As a result, sales forecasting is particularly challenging and sales
forecast accuracy is often compromised. Further, because of the uncertainty, there is
significant information asymmetry between sales managers and operations managers,
increasing the importance of having an effective forecasting system. Second, the industry has
a complex production process and extensive and complex supply chain system where forecast
accuracy is likely to have a significant impact on production plans.
Third, during the period of study, our research site introduced into the S&OP process a
“contingency system” that allowed sales managers to disaggregate from the overall sales
forecast the portion of demand associated with highly uncertain contingent events. Whereas
in the period prior to the introduction of the S&OP contingency system (from now on referred
to as pre-introduction period), sales managers had to incorporate the contingent sales into a
single forecast, in the post-introduction period sales managers are allowed to disaggregate the
2
Electronic copy available at: http://ssrn.com/abstract=2482321
total forecast into two constituent components: an official sales forecast with lower ex ante
uncertainty (i.e., lower variance in expected demand), and a second contingency component
with markedly higher ex ante uncertainty (i.e., higher variance in expected demand). This
feature of the research setting allows us to examine how sales forecast disaggregation affects
the relations among sales forecast accuracy and bias, inventory decisions, and production
decisions.
Our study proceeded in two stages. In the first stage of the study, we conducted
interviews (19 field interviews at corporate headquarters with managers from various
departments and 11 interviews with different managers at a local production site) to gain an
understanding of the budgeting and forecasting process, including the incentives faced by the
relevant decision-makers and the perception of the S&OP contingency system introduction.
We draw on these interviews in interpreting our findings. In the second stage of our study, we
collected from both the production plant and the corporate headquarters stock-keeping-unit
(SKU) level data on sales forecast accuracy, sales forecast bias, finished goods inventory
levels, planned and actual production, as well as various product attribute variables to be used
as controls.
In stark contrast to the extensive theoretic (i.e., analytic and simulation methods)
operations research showing deleterious effects of forecast inaccuracy on operational
outcomes,1 we find no effects of forecast accuracy or forecast bias on either of inventory level
or production plan changes in the pre-introduction period (i.e., before the introduction of the
contingency system). Field interviews suggest this may be due to low perceived credibility of
sales forecasts from the perspective of production planners. Indeed, pre-introduction data
show that forecast accuracy is relatively low (i.e., 12% and 18% for three- and one-month
forecasts, respectively), and positively biased (forecast larger than actual sales), and that
1
For reviews of this literature, see Aytug et al. (2005) and Mula et al. (2006).
3
cross-sectional variation in positive forecast bias is associated with proxies for sales manager
incentives to inflate inventory as a means of avoiding stockouts.
The S&OP contingency system introduced at our field site provided a formal
mechanism whereby sales managers could log contingencies in future product demand, such
as an outbreak of a late season pest, or a possible rise or fall in commodity prices that could
affect market prices. The system allowed the sales managers to provide important detail
regarding the source and timing of the event, the potential volume and revenue impact, and its
probability of occurrence. Whereas in the pre-introduction period, sales managers had to
make judgment calls about whether and when to include potential events into the sales
forecast, once the contingency system was in place, they had a means of communicating
highly uncertain contingent demand outside of the forecast. From the point of view of the
firm, the objective of the contingency system was to improve the reliability of the sales
forecast, the results of which were expected to be a reduction in finished goods inventories
and improved production plan stability.
Consistent with the firm’s objective, we find that sales forecast accuracy improves in
the post-introduction period. Moreover, the improvement in forecast accuracy is associated
with lower inventory levels, suggesting inventory holding cost savings. However, while
forecast accuracy improves in the post-introduction period, we also find some evidence that
the forecast bias actually increases compared to the pre-introduction period. Thus, while
greater forecast accuracy was associated with lower finished goods inventory, this was
partially offset by higher inventory owing to increased positive sales forecast bias.
Importantly, we find that the introduction of the contingency system restored the associations
between sales forecast accuracy and operational outcomes predicted by operations research.
More specifically, higher accuracy in sales forecasts is associated with fewer and smaller
short-term production plan changes, with inventory mediating this association. Taken
4
together, our findings show that sales forecast disaggregation served not only to improve
operational performance in the form of lower inventory and production plan changes
associated with improved sales forecast accuracy, it also improved the overall S&OP process
by restoring coordinated decision-making. That is, our results provide evidence that
operations managers find post-introduction sales forecasts to be a reliable source of
information and, in contrast to the pre-introduction period, incorporate that information into
their decision-making process.
This study makes several important contributions to both the accounting and
operations literatures. The operations literature is replete with analytic and simulation studies
of the effects of sales forecast error on inventory levels and production instability, most of
which aim to identify previously unrecognized optimization opportunities (Boudreau 2003).
This research, however, provides a dearth of empirical evidence of these associations
(Kerkkänen et al. 2009), resulting in a stream of research that is largely devoid of the
organizational context within which decisions are actually made, including the incentives that
decision-makers face and the quality of information at their disposal. We contribute to this
literature by examining the S&OP process within the organizational context.
We also contribute directly to the accounting literature. The accounting budgeting
literature has dealt at great length with the gaming behavior of opportunistic managers in
budget settings (e.g., Chow et al., 1994; Dunk 1993; Fisher et al., 2002; Shields and Shields
1998). Whereas the accounting literature provides extensive evidence of the tendency of
managers to generate pessimistic budgets (i.e., build budget “slack”), we show that
subsequent sales forecasts are actually optimistically biased. Few studies have examined the
judgment aspects of forecasts in an empirical setting (Fildes et al. 2009). Moreover, we
answer calls for more research providing evidence of the largely undocumented consequences
for the users of (potentially inaccurate and biased) budgeting and forecasting information
5
(Jonsson 1998; Wacker and Lummus 2002; Hall 2010).
Further, our study provides rare and unique insights on the effects of sales forecast
disaggregation that serves to provide transparency regarding sales demand uncertainties,
thereby improving the reliability of managers’ forecasts. We show that both intended and
unintended effects emerge and more importantly, we document the restoration of patterns of
associations that operations research predicts through analytic models and simulations, but
has yet to provide much in the way of empirical evidence.
The remainder of the study is organized as follows. The next section discusses the
theoretical background of our study. In Section 3 we describe our research setting and the
insights from our field interviews. Section 4 thereafter describes our data sources and defines
the variables. The results section, Section 5, presents empirical models and the results of our
empirical analysis, and the final section concludes.
2. Theoretical background and research question
Sales forecasts and sources of forecast error
The S&OP process involves the translation of information regarding expected demand,
current work-in-process and finished goods inventories, and labor and material availability
into viable production plans that provide for needed product on a timely basis (Bowersox et
al. 2012). The sales demand forecasting system is at the heart of this coordination effort, and
the measure of forecast system effectiveness is how well the forecasts support the
coordination of sales and operations with forecasts that are integrated, accurate, detailed, and
timely (Oliva and Watson 2009; Bowersox et al. 2012). While the annual static budget
typically provides the initial sales forecast, sales managers update forecasts at regular
intervals–typically monthly or quarterly–to reflect the latest information related to competitor
actions, market conditions, customer circumstances, and responses to the organization’s own
promotion activities (Bowersox et al. 2012; Sivabalan et al. 2009).
6
As sales forecasts are based on the prediction of future demand, they are inherently
subject to uncertainty. While statistical algorithms (e.g., exponential smoothing) for
generating sales forecasts are common, they are typically reviewed by sales managers who
adjust the forecast to incorporate other factors expected to affect demand not captured by the
statistical algorithm (Fildes et al. 2009). Thus, sales forecasts are based at least in part on the
judgment of sales managers (Oliva and Watson 2009; Fildes et al. 2009).
Judgment-related forecast error derives from several sources. First, despite their best
efforts, sales managers have limited ability to perform the cognitively demanding process of
incorporating numerous sources of demand information, often with varying levels of
uncertainty (i.e., variance), into a single, comprehensive sales forecast. This limitation leads
to an increase in forecast inaccuracy. We define forecast inaccuracy as the unsigned deviation
of the forecasted sales from the actual sales. Greater forecast accuracy indicates lower
deviation of the sales forecast from actual sales.
Second, sales managers may introduce bias into the sales forecast. This could emanate
from either unintentional cognitive biases or from intentional game-playing by self-interested
sales managers (Kerkkänen et al. 2009). As an example of the former, sales managers may
insufficiently consider current inventory when forecasting demand, resulting in positively
biased forecasts (Bloomfield and Kulp 2013). As an example of the latter, sales managers
may intentionally positively bias forecasts in order to avoid lost sales, which may be
personally costly in the form of bonus reduction. We define forecast bias as a measure of
signed forecast error, with positive bias indicating a forecast greater than actual sales and
negative bias indicating a forecast lower than actual sales.
The effect of sales forecast accuracy and bias on production stability
The sales function and the operations function are inextricably linked via the S&OP
process. Sales managers provide sales forecasts which operations managers match to current
7
inventory levels and production constraints (e.g., machine or labor constraints) in order to
establish an optimal production plan (Bowersox et al. 2012). The objective is to “bring
production through in the required quantity, of the required quality, at the required time, and
at the most reasonable cost” (Younger 1930, p. iii, as quoted in Aytug et al. 2005).
A well-documented finding in operations research is that a decrease in sales forecast
accuracy results in an increase in production instability (e.g., Blackburn et al. 1987; Aytug et
al. 2005; Kadipasaoglu et al. 1995; Bai et al. 2002; Xie et al. 2003; Jeunet 2006; Kerkkänen et
al. 2009; Pujawan and Smart 2012).2 Production instability is the “intensity of revisions or
changes to the production schedule over time” (Pujawan and Smart 2012). Production
instability results in significant costs associated with inventory handling, inventory
obsolescence, labor overtime, increased materials costs, increased freight costs, increased
record-keeping costs, quality failure costs, and lost sales (Lee and Adam 1986; Inman and
Gonsalvez 1997; Wacker and Lummus 2002; Kerkkänen et al. 2009; Pujawan and Smart
2012). Consistent with prior operations research, we expect sales forecast accuracy to be
negatively associated with production instability.
Prior research suggests that forecast bias may also have deleterious effects on
production stability. As described above, while forecast bias can be intentional and reflect the
incentives of the forecaster, it can also be the result of cognitive limitations and be
unintentional. Indeed, the well-documented inter-firm “bullwhip effect” which ultimately
results in production instability for the manufacturer, has been attributed, in part, to positive
bias emanating from the failure of planners all along the supply chain to fully incorporate the
existing supply line (i.e., orders en route or in process) into judgments of future product needs
(Sterman 1989; Croson et al. 2006). Bloomfield and Kulp (2013) similarly provide
experimental evidence that managers, even within a single firm, insufficiently consider
2
This is sometimes referred to as “production schedule nervousness” (c.f., Kadipasaoglu et al. 1995; Pujawan
and Smart 2012).
8
current inventory when forecasting demand, resulting in positively biased forecasts. To the
extent that positive forecast bias results from unintentional cognitive limitations, the bias will
reverse in future periods with the net effect of inducing production instability. That is,
forecast bias can lead to an intra-firm manifestation of the inter-firm bullwhip effect (c.f.,
Burgess et al. 2006; Bloomfield and Kulp 2013). On the other hand, to the extent that sales
forecast bias is intentional it will be more persistent and we would not necessarily expect to
observe an association between forecast bias and production instability. It is thus an empirical
question whether sales forecast bias will result in increased production instability in any given
setting.
The mediating role of inventory
An additional stream of operations literature examines mechanisms to reduce the
negative effects of sales forecast inaccuracy and bias on production stability (e.g., schedule
freezes and lot-for-lot ordering, Blackburn et al. 1987). Of particular importance to this study
is the stream of research in the operations literature documenting the beneficial role of
inventory “buffer stocks” to mitigate the effect of sales forecast inaccuracy on production
instability (e.g., Sridharan and LaForge 1989, 1990; Bai et al. 2002; Enns, S. T. 2002). In
determining the optimal level of inventory, the costs of production instability must be
balanced against increased holding costs of the buffer stocks (Blackburn et al. 1987;
Sridharan and LaForge 1990), a cost for which operations managers are typically held
accountable. This research suggests that inventory levels will mediate the association between
sales forecast accuracy and production instability; that is, operations managers will utilize
increased inventory levels to buffer (i.e., mitigate) the negative effect of forecast inaccuracy
on production stability. We also likewise expect the association between sales forecast bias to
be mediated by inventory levels. To the extent sales forecast bias is unintentional on the part
of the sales managers and results in production instability as the over- or under-forecasts are
9
corrected, operations managers will similarly employ inventory as a buffering mechanism.3
Based on the above discussion of the operations S&OP literature, we formally propose the
associations depicted in Figure 1.
INSERT FIGURE 1
The above-described S&OP research is extensive and dates back to the 1950s (Aytug
et al. 2005). However, the preponderance of this research uses analytical or simulation
methods (Pujawan and Smart 2012; Kerkkänen et al. 2009) and focuses on identifying
previously unrecognized optimization opportunities through inventory management and
production strategies (Boudreau 2003). As such, the problem as historically studied in the
operations literature is viewed in a very mechanistic way, with little role for the individual
cognitive limitations or biases, or the incentive-induced self-interested behavior of the sales
and operations managers. In reality, inventory levels and production schedules are the result
of the decisions and judgments of (cognitively limited and self-interested) operations
managers. In particular, operations managers choose whether and how much to rely on the
sales forecast judgments of (cognitively limited and self-interested) sales managers. It is thus
an empirical question as to whether the fundamental relations between sales forecast error
(accuracy and bias), inventory levels, and production stability established (and depicted in
Figure 1) in the highly mechanistic operations literature will be observed in a field setting.
Thus, while the proposed relations have strong theoretical underpinnings in the operations
literature, we view our provision of empirical evidence of these relations to be an important
contribution.
Having considered the relations among sales forecast accuracy and bias, inventory
levels, and production instability that constitute the S&OP process (Figure 1), in the following
3
Of course inventories will also be positive associated with sales forecast bias that is intentional on the part of
ths sales managers if this intentional bias is at least partly unanticipated by the operations managers. Indeed, that
would be the very reason for the intentional bias.
10
section we consider the role that sales forecast disaggregation will play in these relations.
The role of forecast disaggregation
As described above, sales forecasts are typically formed with statistical algorithms that
are then adjusted based on the judgments of sales managers. Sales managers incorporate into
their judgments a vast amount of information expected to affect customer demand. Some of
this information provides for an ex ante expected demand that is known with relative certainty
(i.e., low variance around the expected value). An example would be an expected demand
increase associated with a planned sales promotion. Other information is highly uncertain
(i.e., high variance around the expected value) and often contingent on factors outside the
control of the firm. As an example, an agricultural chemical firm such as the one we study can
have dramatic demand shifts arising from weather patterns. At issues is how the aggregation
of these two types of demand information into a single sales forecast affects the reliability of
that forecast.
Assume two sources of demand information, each represented as a normally
distributed random variable: X ~ (E[X], Var(X)) and Y ~ (E[Y], Var(Y)), where Var(X) <
Var(Y) and COV(X, Y) = 0. Let Z be the aggregated demand distribution, that is, Z = X + Y.
The aggregated sales forecast, Fa = E[Z] = E[X] + E[Y], is the expected value of the normally
distributed total demand distribution, Z. That is, Z ~ N(Fa, Var(Z)), where:
Var(Z) = Var(X+Y) = Var(X) + Var(Y) + 2Cov(X,Y).
(1)
Given the assumption that the two sources of demand information, X and Y, are not
negatively correlated, the variance of the aggregated demand, Z, is greater than the variance
of the more certain demand information, X. Thus, when sales managers are forced to
aggregate all demand information into a single forecast, there is greater uncertainty around the
11
forecast.4 Operations managers will correctly assess the aggregated forecast as less reliable.
One solution to this problem is to disaggregate the sales forecast into its two
constituent components: the disaggregated official sales forecast, Fd = E[X], which is the
expected value of the demand information of lower uncertainty (X), and a contingency, E[Y],
which is the expected value of the demand information with markedly higher uncertainty (Y).
We contend that the contingency system introduced at our research site (described in more
detail below) provides for a forecast disaggregation of the form described above.
We predict that that sales forecast disaggregation of this type will have three effects on
the S&OP process. First, we predict that sales forecast disaggregation will have a positive
effect on sales forecast accuracy. That is, the lower ex ante uncertainty of the disaggregated
sales forecast, Fd, as compared to the aggregated forecast, Fa, will result in higher ex post sales
forecast accuracy (i.e., sales forecast closer in absolute value to actual sales).
Second, we predict that the introduction of disaggregated forecasts will lead to a
decrease in finished goods inventory. By disaggregating the sales forecast from the more
uncertain contingent demand, operations managers can more effectively engage in a
production “postponement” strategy. Postponement refers to “delaying activities in the supply
chain until customer orders are received with the intention of customizing products, as
opposed to performing those activities in anticipation of future orders” (van Hoek 2001). The
delay in transforming work in process inventory to SKU-level finished goods inventory
provides for additional time for uncertainty to resolve and has been shown to be effective in
reducing production instability derived from forecast error (LeBlanc 2007; Garcia-Dastugue
and Lambert 2007). Disaggregated sales forecasts facilitate a postponement strategy by
4
Moreover, prior research shows that when forced to combine relatively certain information with uncertain
information into a single forecast, forecasters fail to fully incorporate the uncertain information (i.e., sales
forecast ≠ E[Z]) (Fildes et al. 2009). At the extreme, a potentially demand-inducing (or reducing) event that is
highly uncertain may be ignored altogether. The result is an aggregated forecast that has less information content
and will again be viewed by operations managers as less reliable.
12
allowing operations managers to delay production of the highly uncertain contingent demand,
while accelerating production of the more certain and reliable portion of the official sales
forecast.
Third, we predict that the introduction of disaggregated forecasts will lead to greater
reliance on the sales forecast by the operations managers in making inventory and production
decisions.5 Increased reliance by operations managers on the disaggregated sales forecast will
manifest in stronger relations among sales forecast accuracy, sales forecast bias, inventory,
and production instability (depicted in Figure 1). Stated formally, we make the following
predictions regarding the pre- versus post-introduction period:
As compared to the pre-introduction period with aggregated sales forecasts, in the
post-introduction period when sales forecasts are disaggregated:
H1: Sales forecast accuracy will be higher.
H2: Finished goods inventory will be lower.
H3: The relations among sales forecast accuracy, sales forecast bias, inventory, and
production instability (depicted in Figure 1) will be stronger.
Note that, because we have no ex ante expectations regarding how the introduction of
sales forecast disaggregation will affect sales managers’ incentives to bias forecasts or their
unintentional tendency toward biased forecasts, we make no prediction regarding the effect of
sales forecast aggregation on the level of sales forecast bias.
3. Research Setting and Field Evidence of the S&OP Process
3.1 Research Setting
To address our research questions empirically we conduct a field study that combines
qualitative and quantitative information into our analyses. In particular, we use field evidence
and archival data from AgroCo, a leader in the agricultural chemical industry, an industry
characterized by high product and production process complexity and uncertain annual
5
Indeed, prior research in accounting finds investors deem disaggregated earnings to be more reliable (Hirst et
al. 2007; Elliott et al. 2011).
13
demand and product life cycles (uncertain time to market/ complex patent structures).
Demand uncertainty is endemic to this industry because demand for agricultural products –
namely, herbicides, fungicides, and insecticides – varies based on weather and grower
plantings driven by – among other things - commodity prices. Moreover, the nature of the
production processes (i.e., the management of complex “active ingredients” with lengthy
formulation lead times) makes production and supply chain planning particularly challenging.
During the period of study, our research site introduced a forecast system innovation, the socalled “S&OP contingency system,” that disaggregated the sales forecast based on the level of
ex ante uncertainty which, in turn, significantly altered the information flow between sales
and operations managers. Thus, the agricultural chemical industry in which AgroCo operates
and the specific firm with its recent system innovation is ideal for examining the interplay
between sales forecasts, forecast revisions and production planning, and the consequences of
an innovation in the planning process.
AgroCo is a globally active firm and has major operations within the US. Our project
focuses on one US-based production facility, which we refer to as AgroPlant, that is the sole
supplier of the products produced on site to serve the global market, allowing us to directly
link global sales forecasts to local production.
3.2 Interview protocol
To obtain a detailed understanding of the impact of sales forecast accuracy on
operational performance as well as the role of the new contingency system, we conducted
field interviews at AgroCo with managers from various corporate departments, such as
Marketing, Forecasting and Supply Chain Planning, and Logistics. We conducted our field
interviews in two series. In early 2011, we interviewed 19 managers at AgroCo who were
familiar with the sales forecasting process, or had an active role in it, in order to get an overall
understanding of the forecasting process and of the drivers of forecast accuracy. From these
14
discussions we developed the interview protocol for the production plant, where we sought to
investigate the perspective of the production side, and analyze the role of forecast accuracy in
production scheduling. In late 2011, we conducted interviews with 11 managers from all
relevant functional areas at AgroPlant, i.e., plant management, production engineering, plant
finance, R&D and quality, asset planning, and vendor scheduling. The objective of the second
interview series was to identify the impact of sales forecast accuracy on the complete
operation cycle of the plant (procurement – production – inventory).
Although we guided our interviews by a standardized set of questions, the interviews
were primarily kept open-ended to allow the interviewees to provide their perspectives on the
determinants and consequences of sales forecast accuracy. The research team was alert not to
ask any leading questions, using the interview protocol as guidance. The interviews were
recorded and transcribed by a professional transcription firm. The interview transcripts were
analyzed by multiple members of the research team to gain consensus over the most important
insights, and our main conclusions were validated with follow-up interviews at the firm. The
archival data were collected either on site following the respective interviews, or provided
digitally at a later point in time. Following the interviews, we held phone meetings if
clarifications about the data were necessary. In the following section we describe our main
insights from our interviews regarding the determinants and consequences of sales forecast
accuracy.
3.3 Field evidence on the forecasting process
Our detailed field interviews revealed a sales budgeting and production planning
process at AgroCo that is very complex and involves many different parties. The interviewees
expressed opinions that there are considerable biases in both the annual budget and the
monthly forecasts driven largely by the different incentives that the parties face. The S&OP
system at AgroCo results in a production plan and a sales budget, both established at different
15
points in time:

Production plans for the next budget year are initially developed by the production
units in July based on older long-term forecasts.

Annual sales budgets for each family of product (i.e., all products from a single
active ingredient), without regard to specific product type or packaging, are
proposed by each business unit representing a geographic marketing region with
input from sales managers and finance managers in October.
Although the production units established an initial production plan and budget in the
prior July (based on older, long-term sales forecasts), the sales budget is often not officially
approved until January of the budget year once finance has reviewed prior year-end sales
results. Although claimed to be a “top-down budget” from Headquarters, the described
bottom-up budget is largely approved as the binding annual budget. As a Business Unit
Marketing Head states:
“The top-down budget from –Headquarters- is 95% or 90%  a number I provided
them to begin with. … There may be a few overviews, but it’s largely our number to
begin with.”
The annual sales budget established in October is binding on the sales managers,
serving as the primary basis for the sales force incentive plan. More specifically, sales and
marketing (i.e., district managers and salespersons) incentive bonuses are based on the
achievement of the business unit sales budget (50%), the corporate sales budget (25%) and a
subjective performance assessment of a qualitative nature. Although forecast accuracy is a
designated input into the subjective assessment, interviewees indicated that its importance to
the overall bonus is low. Thus, taken as a whole, the incentive plan creates implicit incentives
for the commercial side of the business to initially underestimate what can be sold, potentially
resulting in a downward-biased annual sales budget. One of the managers in business
16
planning and forecasting remarked:
“…the sales folks are responsible for helping create [the sales budget] and of course
if their target’s lower, their bonus is higher. So you get some sandbagging just from
the get-go on that.”
This view is further supported by a global supply chain manager:
“You’ve got a commercial organization that says, ‘We’re going to develop a forecast
and set budget for the year. If we exceed that budget from a performance award
standpoint for us. All right. Then so with that kind of metric in place, and that’s what
your reward is going to be based on, what’s the driver for you to provide the most
accurate forecast? Or is it better for you to make sure you provide a forecast or set a
budget that you know you can actually make or exceed?’”
Once the sales budget is approved, ownership is transferred back to the business units
(BUs). Usually, the first revisions to the budget start to take place in January (and not between
October and December). Such forecast revisions take place on a monthly basis following a
standardized protocol.
The demand volatility causing the forecast revisions is claimed to be driven by market
(i.e., competitor), economic (i.e., commodity prices), and natural forces (e.g., weather
patterns). Although these factors can drive forecasts to be revised in either direction, the
interviews suggest that the BUs tend to over-forecast rather than under-forecast throughout
the year, mainly to avoid lost sales and be able to react to supply failures of competitors. As a
marketing manager remarks:
“…given the seasonal nature of our business, we find that if we under forecast, the
likelihood of us having supply gets to be difficult. And by that I’m saying if I were to
raise my forecast substantially 30 days before I needed product, there’s a pretty good
chance I won’t receive all of what I had for demand.”
17
The interviews also revealed that AgroCo benefits from sales taken from (smaller)
competitors that are not able to react to demand swings caused by weather conditions or
revised grower planting choices. This implies that in order to be ready to react to these
contingencies, forecasts are more likely to be positively biased.
3.4 Field evidence on the impact of the system innovation
The interviews at our field site provided us with rich information on the contingency
system innovation that was intended to increase the transparency of sales contingencies; that
is, non-trivial environmental or market (i.e., customer or competitor) demand events for
which sales managers were unable (or unwilling) to commit to the formal forecast but for
which there was a nonzero probability of occurring. The system allows the sales managers to
add a so-called contingency “event” in a separate system outside of the formal forecast but
visible to production planners. In essence, the system allows sales managers to disaggregate
from the overall sales forecast the portion of demand associated with highly uncertain
contingent events, resulting in an official sales forecast that reflects relatively more certain
demand. Each event entry into the system communicates an expected demand quantity effect
(either up or down), the reason for the potential demand effect, and the probability of
occurrence. One of the brand managers explained:
“We generate the events and work with our supply chain planners to make sure that
everybody is aware of what the event is and when it should be triggered. And
generally they’re market driven. Could be a weather event. It could be an acreage
event. It could be a pest event. In the case of soybeans, soybean aphid is a pest that
comes periodically. But when it comes, it’s a huge upside and we never know if it’s
going to come until it’s there. … So we use an event structure to allow us to flag
those”
Once an event is assessed at 60% or higher probability of occurring, the event is added
18
to the official forecast. Whereas this new system allows for better communication of demand
contingencies, it also introduces additional opportunities for gaming on the part of the sales
managers. Indeed, positive events are much more likely to be “triggered” (i.e., entered into the
forecast) than are negative events, reflecting an asymmetrical preference on the part of sales
managers to avoid stockouts more than to avoid excessive production, especially for higher
margin products. As stated by one of the finance managers:
“Well, considering the lost sales opportunity for products with a low margin, yeah you
don’t want a bunch of that sitting around. But some of these seed care products have
[a very high] margin. You’re crazy to miss a sales opportunity because you didn’t
want to have a little bit of carrying cost for that inventory. We’d be better off carrying
a little bit of inventory. We are scrambling right now because we have this huge
opportunity in cotton, and that’s a huge gross margin product, but we didn’t forecast
for it because this is truly an event that was unforeseen.”
And in a competitive setting when a competitor cannot produce on time:
“So we have the opportunity to fill that gap. And it’s a huge profit margin and we are
making our supply chain go through hoops. But it’s potentially a $XX million upside
or $XX million upside. … We may hold that market share next year, up to 90% of it.
So we’d be fools not to try to get it.”
These interviews thus show that the system can help to react quicker to demand
volatility, but that this system may also come with some unintended effects as it provides
another mechanism in addition to the formal forecast that may exhibit intentional bias.
In the following section, we conduct empirical tests of the determinants of sales
forecast accuracy and bias, and we examine the subsequent effects of accuracy and bias on
operational decisions, i.e., production plans and inventory holdings. We then examine the
effect of the S&OP contingency system introduction on the behavior of the managers
19
involved in the S&OP process.
4. Data Sources and Variable Descriptions
We use monthly data on product level for the period of 2006-2010 to examine the
association between sales forecast accuracy and production changes before and after the
S&OP contingency system introduction, i.e., the forecast disaggregation. Our sample
comprises 2,300 monthly observations of 80 unique products. In order to use a consistent
sample across all analyses, we restrict the sample to those observations for which we have
data on all variables, leaving us with 1,298 product-month observations of 67 unique
products. Data were collected from two major sources. We obtained sales (related) data from
the corporate sales department at AgroCo, including sales forecast accuracy and bias, sales
price, and supply chain complexity. Production data were locally obtained at AgroPlant.
These data include detailed information on production decisions (e.g., quantities, sequencing,
production changes). We limit our analysis to the products of two specific production lines
within AgroPlant for which data about the production process had been thoroughly and
consistently documented over our sample period. Below we define the variables used in our
analysis and the sources for data collection.
4.1 The forecasting system innovation
The S&OP contingency system was introduced in the end of 2007 and
comprehensively implemented in January 2008. We create an indicator variable for the preand post-introduction period. We label POST as 1 in the years 2008-2010 after the
introduction, and 0 otherwise.
4.2 Sales data
The sales data were collected from AgroCo’s global forecast reporting tool, which
produces monthly sales forecast accuracy information on product level as the major KPI.
Sales forecast accuracy (ACCURACY) is the converse of forecast error and constrained to be
20
between 0 and 1. Forecast error is defined as the deviation of the actual sales (AS) from the
forecasted sales (FS), either 1 month or 3 months out, i.e., Error = |AS-FS|/AS. Therefore,
Accuracy = 1 - Error. The constraint to values between 0 and 1 implies that if, AS = FS 
Accuracy = 1; and if Error ≥ 1  Accuracy = 0.
Sales forecast bias (BIAS) reveals patterns of over- and under-forecasting and is
measured as Bias = -1 × (100 – (FS / AS×100)). Positive values are a sign of overforecasting and negative values indicate under-forecasting. We also compute an indicator
variable, POSITIVE, set equal to one for observations indicating a positive bias, and zero
otherwise.
4.3 Production data
Production changes are computed by subtracting Planned Production (PP) in units
from Actual Production (AP), i.e., PROD_CHANGE = AP - PP. To control for scale effects
and determine the effect of percentage deviations of actual production from forecasted
production, we compute ln PROD_CHANGE as the absolute logarithmic deviation of the AP
from PP, i.e., ln PROD_CHANGE = |(log AP + 1) – (log FP + 1)|.6
Forward coverage (FORWARD) is the number of months’ sales (based on actual
subsequent sales) in finished goods (i.e., SKU-level) inventory at the end of the month prior
to the month of production.
4.4 Control variables
We control for variables that are likely to capture sales manager incentives to bias
forecasts. QUANTITY is the budgeted sales quantity in physical units, expressed in liters.
SALES VALUE is the sales price per unit of measurement (e.g., liters, gallons) multiplied by
the individual product pack size (e.g., 100 liters, 2.5 gallons). Gross margin percentage
(GMPCT) is calculated as the ratio of gross margin per unit (i.e., sales price minus standard
6
Given the seasonality of the business and the resulting non-trivial number of observations with zero forecasted
and/or actual production, we add one prior to computing the logarithms.
21
costs) and sales price per unit. These three variables, QUANTITY, SALES VALUE, and
GMPCT, reflect differences in product volume and profitability that may provide incentives to
positively bias forecasts to ensure adequate inventory. In addition, we control for month fixed
effects to allow for varying incentives through the budget cycle.
We also control for a number of additional variables that may affect forecast accuracy
and/or bias. First, we control for forecasting attributes of the product “active ingredient” (AI).
Each product consists of one key component, called the lead AI, that multiple products share.
AIs differ on a number of dimensions that make their derivative products more complex from
a supply chain planning perspective, including the scarcity of the AI and the number of
products derived from the AI. We control for four dimensions of AI complexity as measured
by the firm on a scale of 0-10. AI_LIFECYCLE captures the lifecycle phase for a given AI’s
product portfolio. AIs with new products in their portfolio yield the highest scores on this
variable; AIs with mature products yield low scores. Because forecasting is more difficult
during the product introduction phase, we expect forecast accuracy to be lower for higher
values of AI_LIFECYCLE. Two complexity dimensions that capture competition for AI
resources may give rise to incentives to bias sales forecasts in order to ensure adequate
inventory. AI_PRODUCTS captures the number of products that share a certain AI.
AI_ALLOCATION indicates the degree to which the AI is a scarce resource that needs to be
allocated to the products sharing the AI (high scores), or freely available on the market (low
values). Higher values of AI_LIFECYCLE, AI_PRODUCTS, and AI_ALLOCATION indicate
greater complexity. Lastly, AI_FCASTIBILITY, is the firm’s overall assessment of the ease of
forecasting AI demand, with higher values indicating lower complexity and greater ease of
forecasting.
We further control for whether or not the product is exported with EXPORT, an
indicator variable that equals 1 if the product is sold outside the US, and 0 otherwise. During
22
our period of study, contingency events were entered for only a subset of products. We thus
create an indicator variable, CONTINGENT, that equals 1 for products for which at least one
contingency event had been entered during our period of study, and 0 otherwise. For models
in which the one-month dependent variables of ACCURACY and POSITIVE are modeled, we
also control for the corresponding three-month variable. In all models we control for year
fixed effects. See Appendix for variable definitions.
5. Empirical Results
5.1 Descriptive Statistics
Table 1 provides descriptive statistics for the variables used in the empirical analysis
(Panel A), as well as for the pre- and post-introduction period separately (Panel B). On
average, the 3-month forecast accuracy (ACCURACY 3m) is 14%; 1-month forecast accuracy
(ACCURACY 1m) is 23%. The difference in forecast accuracy is statistically significant (t =
11.33, p < 0.01), consistent with forecast accuracy decreasing with the forecast horizon.
Further, 1-month forecast accuracy is significantly higher after the introduction of the S&OP
contingency system, providing univariate evidence in support of H1. The increase from 18%
to 25% is statistically significant (Panel B, p < 0.01). The same holds for 3-month forecast
accuracy, which increases from 12% to 15% (p < 0.01).
For the entire sample on average, 69% (53%) of 3-month (1-month) forecasts are
positively biased (Panel A). Interestingly, forecast bias also increases after the introduction of
the S&OP contingency system. While the number of positively biased 3-month forecasts
(POSITIVE 3m) increases from 63% to 72% (p < 0.01), 1-month bias (POSITIVE 1m)
increases from 49% to 54% (p < 0.1) (Panel B).
Actual production is, on average, 1,008,000 units lower than the 1-month production
plan. One month production plan changes (PROD_CHANGE 1m) range from actual
production that is 542 million units higher than the plan, to actual production that is 290
23
million units lower than the plan. This large range reflects production runs that are shifted
from one month to another. While actual production is, on average, lower than the 1-month
production plan in the period before the introduction, the opposite holds for the postintroduction period.
INSERT TABLE 1
The main objective of our study is to examine the effects of forecast disaggregation on
forecast accuracy and bias, and the effects on the relations between forecast accuracy and bias
and inventory and production decisions. We first examine whether accuracy (and bias)
increase after the introduction of the S&OP contingency system (H1). Second, we examine
the impact of forecast disaggregation on the level of finished goods inventory (H2). Finally,
we document the impact of sales forecast variability on production plan changes and
inventory before and after the forecast disaggregation to analyze whether the strength of these
relations is affected by the introduction of the system (H3).
5.2 The effect of sales forecast disaggregation on sales forecast accuracy and bias (H1)
To test our first hypothesis stating that forecast accuracy increases in the postintroduction period, we examine whether forecast accuracy (and bias) is systematically
different in the pre- and post-introduction period. In particular, we examine the determinants
of sales forecast accuracy and bias with the following empirical models:
[DV] = 0 + 1 POST + <controls>
(2)
Where [DV] is either 3-month or 1-month ACCURACY or 3-month or 1-month
POSITIVE. We estimate equation (2) using OLS for the ACCURACY dependent variable and
using a logistic regression for the POSITIVE dependent variable. For both models we cluster
standard errors by product.
INSERT TABLE 2
Results are reported in Table 2. Hypothesis H1 predicts that forecast accuracy
24
increases after the introduction of the contingency system. Hence, we expect a positive and
significant coefficient on POST in Model 1 (3-month accuracy) and Model 2 (for 1-month
accuracy) in Table 2. Consistent with expectations, both 3-month (p<0.10) and 1-month
(p<0.01) forecast accuracy is significantly higher after the forecast disaggregation, providing
support for H1. Interestingly, the likelihood of both 3-month (Model 3) and 1-month (Model
4) forecast to be positively biased also increases with the introduction of the contingency
system, indicating that the new system does not fulfill one of the firm’s stated goals which is
to reduce sales managers’ opportunities and incentives to report upward biased forecasts. In
follow-up interviews, supply chain managers speculated that the increase in 3-month bias
after the system introduction might be a favorable consequence of the S&OP contingency
system. Their reasoning was that the system may force managers to incorporate contingent
events earlier in the forecast than they otherwise would have (i.e., the contingent demand will
show up in the 3-month forecast instead of as a last minute addition to the 1-month forecast).
This however does not explain the increase in 1-month bias that is also observed.
Turning to the control variables, the results for sales forecast bias (Models 3 and 4)
provide evidence suggesting sales managers seek to ensure adequate supply of those products
having greatest impact on their ability to achieve sales targets. Three-month forecasts are
more likely to be positively biased for products with larger budgeted volume. By contrast, 1month forecasts are more likely to be positively biased for higher margin products. Bias,
however, is not significantly more positive for higher value products, on average. Further,
interesting insights are gleaned from the monthly pattern exhibited in the likelihood of a
positive forecast bias as reported in Model 3. As compared to December (the omitted month
variable), 3-month forecasts are more likely to be positively biased in March when the sales
budget is usually finalized. Over the following three months (April-June), the likelihood of
positively biased forecasts increases, likely reflecting efforts by the commercial side of the
25
organization to ensure adequate product supply. On the contrary, in August and September,
when the budgeting process for the following year has already started, sales forecast bias is
significantly lower, indicating that sales managers’ incentives to sandbag their annual budget
mitigate their incentives to positively bias sales forecasts in order to convey a more
conservative picture. These results are reflective of behaviors on the part of the commercial
side of the business aimed at ensuring products are available while, at the same time, seeking
to maintain or otherwise induce lower sales performance targets. A manufacturing business
process manager described this phenomenon as follows:
“The sales and commercial group [are] sometimes reluctant to put everything in a
forecast before the budget is locked down because they’re afraid they’ll get asked to
make even more that they won’t be able to deliver. So there’s that kind of thing that
goes on sometimes at particular times of the year… Conservative bias and then it sort
of marches up through time… But now when you get back into the June/July/August
you’ll start seeing some of the behaviors come back because of the finance cycle
again…They’re protecting the second half, which is the baseline for next year. So
behaviors you’re right is probably more biased.”
5.3 The effect of sales forecast disaggregation on inventory (H2)
H2 predicts that finished goods inventory decreases in the post-introduction period. To
test this hypothesis, we estimate the following interaction model:
FORWARD = 0 + 1 POST + 2 ACCURACY (3m)
+ 3 ACCURACY (3m) X POST + 4 ACCURACY (1m)
+ 5 ACCURACY (1m) X POST +6 POSITIVE (3m)
+ 7 POSITIVE (3m) X POST + 8 POSITIVE (1m)
+ 9 POSITIVE (1m) X POST+ <controls>
(3)
For the ease of interpretation, we mean-center ACCURACY (1m) and ACCURACY (3m).
The results of the estimation of equation (3) are provided in Model 1, Panel A, Table
3. As can be seen, the coefficient on POST in Model 1 is positive and significant, indicating
26
that for average levels of accuracy (i.e., mean-centered ACCURACY = 0) and when 3-month
and 1-month forecast is not positively biased (i.e., POSITIVE = 0), inventory is higher in the
post-introduction period. However, untabulated analysis conducted at the plant level shows
that, although on average finished goods inventory has increased, work in process increased to
a greater extent indicating that there is a shift from finished goods to work-in-process
inventory. More specifically, the ratio of finished goods to work in process inventory
decreases in the post-introduction period as compared to the pre-introduction period. Thus,
while the level of finished goods inventory does not decline in an absolute sense, finished
goods inventory declines relative to work in process inventory, consistent with the
postponement strategy underlying H2. We interpret this as partial support for H2.
5.3 The effect of sales forecast disaggregation on the associations between sales forecast
accuracy, inventory levels, and production plan changes (H3)
H3 examines whether the extent to which production planners incorporate sales
forecasts into production decisions. More specifically, we predict that the relations between
sales forecast accuracy, inventory, and production stability depicted in Figure 1 will be
stronger in the post-introduction period. Accordingly, we estimate the associations between
sales forecast accuracy, inventory levels, and production plan changes before and after the
introduction of the S&OP contingency system. We test this hypothesis with equation (3)
above and equation (4):
Ln PROD_CHANGE = 0 + 1 POST + 2 ACCURACY (3m)
+ 3 ACCURACY (3m) X POST + 4 ACCURACY (1m)
+ 5 ACCURACY (1m) X POST +6 POSITIVE (3m)
+ 7 POSITIVE (3m) X POST + 8 POSITIVE (1m)
+ 9 POSITIVE (1m) X POST + 10 FORWARD
+ 11 FORWARD X POST + <controls>
(4)
Again, for the ease of interpretation, we mean-center ACCURACY (1m) and
ACCURACY (3m). The results of the estimation of equation (4) are provided in Model 2,
Panel A, Table 3. Figure 2 provides an overview of the results of the estimations of equations
27
(3) and (4). The associations predicted for the pre-introduction period are presented in Panel
A of Figure 2; the post-introduction associations are presented in Panel B. For both figures,
solid lines indicate statistically significant associations; dotted lines indicate insignificant
associations.
INSERT TABLE 3
INSERT FIGURE 2
Models 1 and 2 of Table 3, Panel A document that in the pre-introduction period (i.e.,
when POST = 0) coefficients on ACCURACY (3m) and ACCURACY (1m) are not statistically
different from zero, indicating that sales forecast accuracy affects neither inventory levels nor
production changes in the period before the forecast disaggregation. However, as indicated in
Model 2, inventory level (FORWARD) is negatively associated (p <0.01) with 1-month
production plan changes, reflecting the long-recognized “buffering” role that inventory is
intended to play. Overall, in stark contrast to prior operations research, we find little evidence
of generally accepted relations among sales forecast accuracy, inventory levels, and
production stability. We interpret this as evidence that operations managers tend to ignore the
sales forecasts and, instead, base production plans on their own historical data. This is likely
the result of having observed inaccurate and biased sales forecasts over an extended period of
time.
Consistent with our predictions, however, the results for the post-introduction period
show a different picture. As predicted, we find that the increased transparency of contingent
events provided by the S&OP contingency system restored the associations between sales
forecast accuracy and operational outcomes predicted by operations research. In particular, we
find that, ACCURACY (3m) has a direct negative effect on LN_PRODCHANGE (1m) (p <
0.05). That is, higher accuracy in sales forecasts results in fewer and smaller short-horizon
production plan changes. Three-month forecast accuracy (inaccuracy) thus has a favorable
28
(unfavorable) impact on 1-month production plan stability. Moreover, the improvement in
three-month forecast accuracy is associated with lower inventory levels (p < 0.05), which in
turn has a significantly negative effect on LN_PRODCHANGE (1m) (p < 0.05). Thus, in the
post-introduction period following the sales forecast disaggregation, we find empirical
support for the generally accepted relations among sales forecast accuracy, inventory levels,
and production stability in the operations literature depicted in Figure 1.7 Note, however, that
we also find higher (p < 0.05) inventory holdings associated with positive 3-month forecast
bias, POSITIVE (3m), possibly reflecting an unintentional consequence of the contingency
system.8
To corroborate our finding that the commonly held relations from operations research
depicted in Figure 1 present only after the system introduction, we rerun the analysis
separately for the samples before and after the forecast disaggregation. Further, this
subsample analysis also allows us to directly test the proposed mediating role of inventory in
the relationship between sales forecast accuracy and production plan stability. Following
Baron and Kenny (1986), there are three steps necessary to establish a mediation effect,
reflected in the following three equations.
Ln PROD_CHANGE = 0 + 1 ACCURACY (3m) + 2ACCURACY (1m)
+3 POSITIVE (3m) + 4 POSITIVE (1m) + <controls>
(5)
FORWARD = 0 + 1 ACCURACY (3m) + 2ACCURACY (1m) +3 POSITIVE (3m)
+ 4 POSITIVE (1m) + <controls>
(6)
Ln PROD_CHANGE = 0 + 1 ACCURACY (3m) + 2 ACCURACY (1m)
+3 POSITIVE (3m) + 4POSITIVE (1m) + 5 FORWARD + <controls>
(7)
7
For the formal test of this proposed mediation effect we refer to the subsample analysis reported in Panel B or
Table 3.
8
Surprisingly, despite its large size, the interaction coefficient ACCURACY (3m) X POST in the estimation of
LN_PRODCHANGE (1m) is not significant. A closer look at the results reveals that this lack of significance is
not driven by a lack of difference between the subsamples, but the lack of a relationship in the period before the
system introduction, resulting in a large standard error and a very wide confidence interval of the ACCURACY
(3m) coefficient, which is then used as the benchmark for the comparison with the significant coefficient in the
post-system introduction sample.
29
We estimate these models for the periods before and after the forecast disaggregation,
respectively. The results of the estimation of equation (5) - (7) are provided in Panel B of
Table 3. Overall, the results are in line with the interaction models. As expected, the
coefficient on ACCURACY (3m) in the estimation of LN_PRODCHANGE (1m) (Model 5) is
close to zero, with a p-value of close to 1, indicating that sales forecast accuracy does not
affect production changes in the period before the forecast disaggregation. On the contrary,
the corresponding coefficient for the post-introduction period (Model 6) is negative and
significant (p<0.05), providing evidence for the restoration of the association between
forecast accuracy and production stability after the new S&OP contingency system has been
introduced.
Also, the suggested buffering (mediating) role of inventory is only present in the postintroduction period. As reported in Models 2, 4 and 6 in Panel B of Table 3, all three
conditions for a (partial) mediation are fulfilled in the post-introduction period. In the first
step of our test of mediation (Model 2), we establish a significant relation between the
dependent variable, LN_PRODCHANGE (1m), and the independent variable being mediated,
ACCURACY (3m). In the second step (Model (4)), we document an association between the
independent variable being mediated, ACCURACY (3m), and the mediator, FORWARD. In
the third step (Model 6), we show that the mediator, FORWARD, is associated with the
outcome variable of interest, LN_PRODCHANGE (1m), and that the introduction of the
mediator variable alters the direct association of interest (i.e., that between ACCURACY (3m)
and LN_PRODCHANGE (1m) in Model (2)).
Finally, since the subsamples differ significantly in size, we conduct a further
robustness test to rule out that the difference in significance between the subsamples is driven
by the difference in statistical power. To do so, we limit the post-introduction sample to (1)
only those products that were also produced in the pre-system introduction period, and (2)
30
eliminate the year of system launch, leaving us with a sample of 389 observations as
compared to 370 observations in the period before the system introduction. The untabulated
results confirm that our analysis is robust to this procedure.
Taken together, our findings show that the forecast disaggregation served not only to
improve operational performance in the form of lower inventory and production plan changes
associated with improved sales forecast accuracy, it also improved the overall S&OP process
by restoring coordinated decision-making. That is, our results provide evidence that
operations managers find post-introduction sales forecasts to be credible source of
information and, in contrast to the pre-introduction period, incorporate that information into
their decision-making process. In particular, disaggregation of sales forecasts facilitates a
postponement strategy, indicated by a decrease in finished goods inventory relative to work in
process inventory. We also find, however, an unexpected increase in sales forecast bias
following the S&OP contingency system introduction. In the post-introduction period, 72%
(54%) of three-month (one-month) sales forecasts are positively biased, a significant increase
over pre-introduction levels. We further document that increased three-month sales forecast
bias leads to higher inventory levels, offsetting any reduction in inventory levels attained as a
result of higher sales forecast accuracy.
6. Conclusion
In this study, we use field and archival methods to examine the sales and operations
planning (S&OP) system within a large manufacturing organization. While the operations
literature is replete with analytic and simulation studies of S&OP systems, this research fails
to account for the role that incentives, biases, and organizational dynamics play in the S&OP
process. Based on 30 field interviews with managers from a large agro-chemical firm, we
augment our empirical analysis with an understanding of the process of sales forecasting and
production planning at our research site.
31
Our study adds to the accounting literature and combines insights from the accounting
and the operations research literature. While the operations research literature typically
focuses on statistical demand models as a means of generating accurate forecasts, the
accounting literature has dealt at great length with budgeting and the gaming behavior of
opportunistically acting managers using their informational advantages to serve their own
benefits. There is to our knowledge no study that examines the effect of gaming behavior by
sales managers making sales forecasts, and how potentially biased budget and forecast
information is incorporated into real decisions made by the users of this information, i.e., the
production managers. Whereas the operations research literature tends to ignore the
possibility of gaming behavior in forecasting and planning, the accounting literature
emphasizes its universal existence, but has so far not explored its consequences on operational
decisions. Our study thus adds to the literature by quantifying the impact of sales forecast
accuracy on inventory and production decisions.
Results of our study also have managerial implications. Our study shows that
production managers anticipate forecast error and possibly even forecast bias. Over time, they
may begin to ignore forecasts from sales managers and, instead, rely on their own historical
information. This could potentially lead to significant inefficiencies if sales forecasts do,
indeed, have imbedded in them useful information regarding future customer demand. Our
documentation of the change in behavior following the S&OP contingency system
introduction suggests that increased transparency and a reduction in forecast uncertainty can
restore the credibility of forecasts, thereby improving the organization’s ability to meet
demand in volatile markets. Further, disaggregation of sales forecasts facilitates a
postponement strategy, indicated by a decrease in finished goods inventory relative to work in
process inventory, which can be beneficial for a firm acting in a volatile market.
Our study is subject to several limitations. First, we do not have detailed information
32
on the incentive system and compensation formulas for individual sales managers and are
unable to include direct measures of incentives in our archival analysis of sales forecast
accuracy and bias. Nevertheless, we derive insights on the incentive system based on our field
interviews and use these insights to facilitate our understanding and interpretation of our
findings.
Further, our study is based on one firm which limits the generalizability of the specific
findings related to effect magnitudes. However, the field insights related to a common form of
incentivizing sales managers (i.e., sales bonus paid for beating budget targets) and the nature
of the response of production managers likely generalize to other settings, in particular those
with volatile demand and uncertainty and information asymmetry between sales and
production managers. Demand uncertainty and volatility is common to many industries
(Fisher et al., 1994), including high-tech semiconductor and consumer electronics industries
and the pharmaceutical industry (Wu et al., 2005). In such settings, information asymmetry
between sales departments and production departments are relatively high as sales managers
are closer to the market and production departments cannot rely on historic data in their
planning. Our findings thus generalize to many settings and provide insight into how the
dynamics between sales and production managers manifest in organizational decisionmaking.
33
References
APICS/Cap Gemini Ernst & Young/Oracle Oracle. 2003. The adaptive supply chain:
postponement for profitability. Accessed on 8/15/2014.
http://www.oracle.com/us/products/applications/018543.pdf
Aytug, Haldun, X. Zhao, J. Xie, and J. Leung J. 2005. Executing production schedules in the
face of uncertainties: A review and some future directions. European Journal of
Operational Research 161: 86-110.
Bai, X, J. S. Davis, J. J. Kanet, S. Cantrell, and J. W. Patterson. 2002. Schedule instability,
service level and cost in a material requirements planning system. International Journal of
Production Research 40 (7): 1725-1758.
Barratt, M. and R. Barratt. 2011. Exploring internal and external supply chain linkages:
Evidence from the field. Journal of Operations Management 29 (5): 514.
Bittlestone, R. 2000. Tackling the Numbers Game. Accountancy 125: 46.
Blackburn, J. D., D. H. Kropp, and R. A. Millen. 1987. Alternative approaches to schedule
instability: A comparative analysis. International Journal of Production Research. 25
(12): 1739-1750.
Bloomfield, R. J., and S. L. Kulp. 2013. Durability, transit lags, and optimality of inventory
management decisions. Production and Operations Management 22(4): 826.
Bonner, S. E. 2008. Judgment and decision making in accounting. Upper Saddle River, NJ:
Pearson Education, Inc.
Boudreau, J., W. Hopp, J. O. McClain, and J. L. Thomas. 2003. On the interface between
operations and human resources management. Manufacturing & Service Operations
Management 5 (3): 179.
Bowersox, D., D. Closs, and M. B. Cooper. 2012. Supply Chain Logistics Management. New
Yort: McGraw-Hill/Irwin.
Burgess, K., P. J. Singh, R. Koroglu. 2006. Supply chain management: Astructured literature
review and implications for future research. International Journal of Operations &
Production Management 26 (7): 703-729.
Cachon, G. P. and M. Fisher. 2000. Supply chain inventory management and the value of
shared information. Management Science 46 (8): 1032-1048.
Cassar, G and B. Gibson. 2008. Budgets, internal reports, and manager forecast
accuracy. Contemporary Accounting Research 25 (3): 707-37.
Chen, F. Z. Drezner, J. K. Ryan, and D. Simchi-Levi. 2000. Quantifying the bullwhip effect in
a simple supply chain: The impact of forecasting, lead times, and information.
Management Science 46 (3): 436–443.
Chow C. W., J. C. Cooper, and W. S. Waller. 1988. Participative budgeting: Effects of a
truth-inducing pay scheme and information asymmetry on slack and performance. The
Accounting Review 63 (1): 111-122.
Chow, C. W., M. K. Hirst, and M. D. Shields. 1994. Motivating truthful subordinate
reporting: An experimental investigation in a two-subordinate context. Contemporary
Accounting Research 10: 699-720.
CIMA. 2004. Better Budgeting: A report on the Better Budgeting forum from CIMA and
ICAEW. London: Chartered Institute Of Management Accountants (CIMA).
Covaleski, M. A., J. H. Evans III, J. L. Luft, and M. D. Shields. 2003. Budgeting research:
Three theoretical perspectives and criteria for selective integration. Journal of
Management Accounting Research 15: 3-49.
Croson, R. and K. Donohue. 2002. Experimental economics and supply-chain
management. Interfaces 32 (5): 74-82.
34
Croson, R. and K. Donohue. 2006. Behavioral causes of the bullwhip effect and the observed
value of inventory information. Management Science 52 (3): 323-36.
Dunk A. S. 1989. Budget emphasis, budgetary participation and managerial performance: a
note. Accounting, Organizations and Society 14 (4): 321-324.
Dunk A. S. 1993. The effect of budget emphasis and information asymmetry on the relation
between budgetary participation and slack. The Accounting Review 68: 400-410.
Ekholm, B. G. and J. Wallin. 2011. The impact of uncertainty and strategy on the perceived
usefulness of fixed and flexible budgets. Journal of Business Finance & Accounting 38
(1): 145-64.
Enns, S. T. 2002. MRP performance effects due to forecast bias and demand uncertainty.
European Journal of Operational Research 138 (1): 87-102.
Fildes, R., P. Goodwin, M. Lawrence, and K Nikolopoulos. 2009. Effective forecasting and
judgmental adjustments: an empirical evaluation and strategies for improvement in
supply-chain planning. International Journal of Forecasting 25: 3-23.
Fisher M.L., J. H. Hammond, W. R. Obermeyer, and A. Raman. 1994. Making supply meet
demand in an uncertain world. Harvard Business Review 72 (3): 83.
Fisher, J. G., L. A. Maines, S. A. Peffer, and G. B. Sprinkle. 2002). Using budgets for
performance evaluation: Effects of resource allocation and horizontal information
asymmetry on budget proposals, budget slack, and performance. The Accounting Review
77(4): 847-865.
Fisher, J. G., W. T. Mitchell, S. A. Peffer, and R. A. Webb. 2013. Detecting misreporting in
subordinates’ budgets. Working paper.
Forrester, J. 1958. Industrial dynamics: a major breakthrough for decision makers. Harvard
Business Review 36 (4): 37–66.
Hagel, J. 2014. How to better connect planning, forecasting, and budgeting. Journal of
Accountancy 217 (4): 20-21.
Hall, M. 2010. Accounting information and managerial work. Accounting, Organizations and
Society 35 (3): 301-315.
Hannan, R. R., F. Rankin, and K. Towry. 2010. Flattening the organization: The effect of
organizational reporting structure on budgeting effectiveness. Review of Accounting
Studies 15 (3): 503-536.
Hansen, S. C., D. T. Otley, and W. Van der Stede. 2003. Practice developments in budgeting:
an overview and research perspective. Journal of Management Accounting Research 15:
95-116.
Hansen, S. C., and W. A. Van der Stede. 2004. Multiple facets of budgeting: an exploratory
analysis. Management Accounting Research 15: 415–439.
Hansen, S. C. 2011. A theoretical analysis of the impact of adopting rolling budgets, activitybased budgeting, and beyond budgeting. European Accounting Review 20 (2): 289–319.
Hartmann, F. G. H. 2000. The appropriateness of RAPM: Toward the further development of
theory. Accounting, Organizations & Society 25 (4/5): 451-482.
Henttu-Aho, T. And J. Järvinen. 2013. A field study of the emerging practice of beyond
budgeting in industrial companies: An institutional perspective. European Accounting
Review. 22 (4): 765-785.
Horngren, C. T., S. M. Datar, and M. V. Rajan. 2011. Cost Accounting: A Managerial
Emphasis, 14th Edition. New York: Prentice Hall
Inman, R. R., and D. J. A. Gonsalvez. 1997. The causes of schedule instability in an
automotive supply chain. Production and Inventory Management Journal 38 (2): 26-31.
Jeunet, J.. 2006. Demand forecast accuracy and performance of inventory policies under
multi-level rolling schedule environments. International Journal of Production
Economics 103 (1): 386.
35
Jonsson, S. 1998. Relate management accounting research to managerial work! Accounting,
Organizations and Society 23 (4): 411 (4)
Kadipasaoglu, S. N. and V. Sridharan. 1995. Alternative approaches for reducing schedule
instability in multistage manufacturing under demand uncertainty. Journal of Operations
Management 13 (3): 193.
Kerkkanen A., J. Korpela, and J. Huiskonen. 2009. Demand forecasting errors in industrial
context: Measurement and impacts. International Journal of Production Economics 118:
43-48.
Lee, T. S., and E, E. Adam, Jr. 1986. Forecasting error evaluation in material requirements
planning (MRP) production inventory systems. Management Science 32 (9): 1186.
Lee, H. L., V. Padmanabhan, and S. Whang. 1997. The bullwhip effect in supply
chains. Sloan Management Review 38 (3): 93.
Lee, H. L., V. Padmanabhan, and S. Whang. 2004. Information distortion in a supply chain:
The bullwhip effect."Management Science 50 (12): 1875-86.
Libby, T. and R. M. Lindsay. 2010. Beyond budgeting or budgeting reconsidered? A survey
of North-American budgeting practice. Management Accounting Research 21: 56-75.
Merchant, K. .1985. Budgeting and the propensity to create budgetary slack. Accounting,
Organizations and Society 10 (2): 201-210.
Mula J., R. Poler, J. P. Garcia-Sabater, and F. C. Lario. 2006. Models for production planning
under uncertainty: A review. International Journal of Production Economics 103 (1):
271-285.
Oliva. R., and N. H. Watson. 2009. Managing functional biases in organizational forecasts: A
case study of consensus forecasting in supply chain planning. Production and Operations
Management 18 (2):138-151.
Oliva R., and N. H. Watson. 2011. Cross functional alignment in supply chain planning: A
case study of Sales & Operations Planning. Journal of Operations Management 29 (5):
434-448.
Pujawan, I. N. and A. U. Smart. 2012. Factors affecting schedule instability in manufacturing
companies. International Journal of Production Research 50 (8): 2252.
Schweitzer, M. E. and G. P. Cachon. 2000. Decision bias in the newsvendor problem with
known demand distribution: Experimental evidence. Management Science 46 (3): 404–
420.
Selto, F. H., and S. K. Widener. 2004. New directions in management accounting research:
Insights from practice, in Advances in Management Accounting, Volume 12. Emerald
Group Publishing Limited.
Shields, J. F., and M. D. Shields. 1998. Antecedents of participative budgeting. Accounting,
Organizations and Society 23: 49-76.
Sridharan, V., and R. L. LaForge. 1989. The impact of safety stock on schedule instability,
cost and service. Journal of Operations Management 8 (4): 327.
Sridharan, V., and R. L. LaForge. 1990. On using buffer stock to combat schedule
instability. International Journal of Operations & Production Management 10 (7): 37.
Sivabalan, P., P. Booth, T. Malmi, and D. A. Brown. 2009. An exploratory study of
operational reasons to budget. Accounting and Finance 49 (4): 849.
Sterman, J. 1989a. Modeling managerial behavior: Misperceptions of feedback in a dynamic
decision making experiment. Management Science 35 (3): 321–339.
Sterman, J. 1989b. Misperceptions of feedback in dynamic decision making. Organizational
Behavior and Human Decision Processes 43 (3): 419–434.
Vadasz, Tony. 2005. Analysis: Budgeting and forecasting. Accountancy 135 (1341): 59.
Van Hoek, R. I. 2001. The rediscovery of postponement: A literature review and directions
for research. Journal of Operations Management 19: 161-184.
36
Wacker, J. G., and R. R. Lummus. 2002. Sales forecasting for strategic resource planning.
International Journal of Operations & Production Management 22 ( 9): 1014 – 1031.
Waller, W. 1988. Slack in participative budgeting: The joint effect of a truth-inducing pay
scheme and risk preferences. Accounting, Organizations and Society 13 (1): 87-98.
Wu, D. Y. and E. Katok. 2006. Learning, communication, and the bullwhip effect. Journal of
Operations Management 24 (6): 839.
Wu, S. D., M. Erkoc, and S. Karabuk. 2005. Managing capacity in the high-tech industry: A
review of literature. The Engineering Economist 50: 125-158.
Young, S. M. 1985. Participative budgeting: The effects of risk aversion and asymmetric
information on budgetary slack. Journal of Accounting Research 23: 829-842.
Xie, J., X. Zhao, and T. S. Lee. 2003. Freezing the master production schedule under single
resource constraint and demand uncertainty. International Journal of Production
Economics 83 (1): 65-84.
Younger, J. 1930. Work routing in production including scheduling and dispatching. New
York: Ronald Press Company.
37
Figure 1
Theoretical Model
38
Figure 2
Summary of Results
Panel A: Pre-introduction Period
Panel B: Post-introduction Period
39
Table 1
Descriptive Statistics
Panel A: Full Sample (N = 1,298)
Variable
POST
ACCURACY (3m)
ACCURACY (1m)
BIAS (3m)
BIAS (1m)
POSITIVE (3m)
POSITIVE (1m)
PROD_CHANGE (1m)
Ln PROD_CHANGE (1m)
FORWARD
QUANTITY (1,000s)
SALES VALUE
GMPCT
AI_LIFECYCLE
AI_PRODUCTS
AI_ALLOCATION
AI_FCASTIBILITY
EXPORT
M3
M4
M5
M6
M7
M8
M9
M10
M11
Y07
Y08
Y09
Mean
0.715
0.139
0.232
6.008
2.656
0.692
0.529
-1,008.260
1.003
8.090
7.827
6.594
7.180
4.650
0.069
0 .111
0.106
0.095
0.087
0.076
0.084
0.087
0.082
0.084
0.177
0.173
0.307
Std. Dev.
Min
0.452
0
0.268
0
0.318
0
30.897
-65.094
14.654
-21.999
0.462
0
0.499
0
24,332.740 -542,000.000
2.736
0
8.032
0
[omitted for confidentiality]
[omitted for confidentiality]
[omitted for confidentiality]
2.885
4
1.664
2
3.749
0
3.934
1
0.254
0
0.3148
0
0.308
0
0.293
0
0.282
0
0.266
0
0.277
0
0.282
0
0.274
0
0.277
0
0.382
0
0.379
0
0.461
0
40
Max
1
1
1
675.158
310.142
1
1
29,0000.000
13.203
35.000
10
10
10
10
1
1
1
1
1
1
1
1
1
1
1
1
1
Panel B: Pre- versus Post-introduction Period
Variable
ACCURACY (3m)
ACCURACY (1m)
BIAS (3m)
BIAS (1m)
POSITIVE (3m)
POSITIVE (1m)
PROD_CHANGE
Ln PROD_CHANGE
FORWARD
QUANTITY (1,000s)
SALES VALUE
GMPCT
AI_LIFECYCLE
AI_PRODUCTS
AI_ALLOCATION
AI_FCASTIBILITY
EXPORT
Pre-introduction
Period
(POST = 0)
N = 370
Mean
SD
0.122
0.180
5.071
2.753
0.630
0.492
-4,591.892
0.888
7.836
89.175
0.010
0.796
7.827
6.546
7.811
4.351
0.016
0.244
0.281
38.770
18.412
0.484
0.501
37,988.710
2.690
7.562
154.259
0.032
0.174
2.888
1.271
3.283
3.901
0.126
*** p<0.01, ** p<0.05, * p<0.10, two-tailed
41
Post-introduction
Period
(POST = 1)
N = 928
Mean
SD
0.146
0.252
6.382
2.617
0.717
0.543
420.559
1.049
8.191
183.119
0.036
0.899
7.828
6.613
6.929
4.769
0.091
0.277
0.329
27.145
12.865
0.451
0.498
15,709.230
2.754
8.214
377.575
0.132
1.041
2.885
1.798
3.893
3.943
0.287
Difference
0.024
0.072
1.311
-0.136
0.087
0.051
5,012.451
0.161
0.354
93.943
0.026
0.103
0.001
0.067
-0.882
0.418
0.074
***
***
***
*
**
***
***
***
***
***
Table 2
Determinants of Forecast Accuracy and Positive Bias
[DV] = 0 + 1 POST + <controls>
VARIABLES
POST
Controls
QUANTITY
SALES VALUE
GMPCT
AI_LIFECYCLE
AI_PRODUCTS
AI_ALLOCATION
AI_FCASTIBILITY
M3
M4
M5
M6
M7
M8
M9
M10
M11
EXPORT
CONTINGENT
ACCURACY (3m)
POSITIVE (3m)
Y07
Y08
Y09
(1)
(2)
(3)
ACCURACY ACCURACY POSITIVE
(3m)
(1m)
(3m)
0.219*
0.256***
1.111***
(2)
3.037
(4)
POSITIVE
(1m)
0.480*
1.616
Odds
Ratio
Odds
Ratio
0.000***
0.242
0.024
0.010
0.029
-0.001
-0.023**
0.198**
0.143*
-0.182**
-0.168*
-0.192*
0.062
0.193*
-0.155
-0.195*
0.105
-0.007
0.000
-0.159*
0.000
-0.002
0.001
0.004
-0.000
0.056
0.106*
0.047
-0.060
-0.111
-0.026
-0.001
0.000
-0.014
0.103*
0.034
0.950***
0.002***
-0.346
0.013
0.080***
-0.126**
0.042*
-0.031
0.614***
0.817***
0.911***
0.659**
0.191
-0.567**
-0.443*
-0.007
0.030
0.086
-0.038
1.002
0.707
1.014
1.083
0.882
1.043
0.969
1.848
2.264
2.486
1.933
1.211
0.568
0.642
0.993
1.031
1.089
0.962
-0.000
-0.381
0.161**
0.161***
0.075
-0.060**
-0.075**
0.256
0.046
0.113
0.170
0.570*
0.216
0.065
0.134
0.129
-0.011
0.129
1.000
0.683
1.174
1.175
1.078
0.942
0.927
1.292
1.047
1.120
1.185
1.768
1.241
1.067
1.144
1.138
0.909
1.138
0.208*
0.093
-0.171**
0.072
-0.135***
-0.148***
0.922***
-0.314
-0.143
2.515
0.730
0.867
2.121***
0.723**
-0.244
0.279
8.339
2.061
0.784
1.322
Constant
-0.719***
-0.338***
-0.343
0.709
-3.174***
0.042
Observations
Pseudo R2
N
N-Left-censored
N-Clusters
1,298
4.97%
1,298
921
67
1,298
17.00%
1,298
709
67
1,298
8.17%
1,298
67
*** p<0.01, ** p<0.05, * p<0.10, two-tailed
42
1,298
18.00%
1,298
67
Table 3
Determinants of Inventory Levels and Production Plan Changes
Panel A: Pooled Analysis
FORWARD = 0 + 1 POST + 2 ACCURACY (3m) + 3 ACCURACY (3m) X POST
+ 4 ACCURACY (1m) + 5 ACCURACY (1m) X POST +6 POSITIVE (3m)
+ 7 POSITIVE (3m) X POST + 8 POSITIVE (1m) + 9 POSITIVE (1m) X POST + <controls>
Ln PROD_CHANGE = 0 + 1 POST + 2 ACCURACY (3m) + 3 ACCURACY (3m) X POST
+ 4 ACCURACY (1m) + 5 ACCURACY (1m) X POST +6 POSITIVE (3m)
+ 7 POSITIVE (3m) X POST + 8 POSITIVE (1m) + 9 POSITIVE (1m) X POST
+ 10 FORWARD + 11 FORWARD X POST + <controls>
(1)
VARIABLES
POST
Coefficient
sum
(pre + post)
FORWARD
3.416***
ACCURACY (3m)
ACCURACY (3m) X POST
ACCURACY (1m)
ACCURACY (1m) X POST
2.330
-4.358*
-2.046
0.160
POSITIVE (3m)
POSITIVE (3m) X POST
POSITIVE (1m)
POSITIVE (1m) X POST
0.661
1.284
2.442*
-1.316
-2.028**
-1.886*
1.945***
1.126
FORWARD
FORWARD X POST
(2)
Ln
PROD_CHANGE
2.923
1.076
-4.120
-2.057
4.113
0.473
-0.053
2.091
-3.267**
-0.475***
0.196
Controls
QUANTITY
SALES VALUE
GMPCT
AI_LIFECYCLE
AI_PRODUCTS
AI_ALLOCATION
AI_FCASTIBILITY
EXPORT
CONTINGENT
M3
M4
M5
M6
M7
M8
M9
M10
M11
Y07
Y08
Y09
-0.003***
-3.027
-0.323
-0.329**
0.967*
0.163
0.255*
-6.478***
0.637
-1.860**
-0.628
-0.312
-2.111***
-1.419
-0.059
0.836
0.788
-0.191
1.987
-4.890***
-2.434**
0.004***
4.729*
-1.134*
-0.076
-0.013
0.076
0.080
-0.569
1.952*
-1.689
-4.192***
-1.445
-0.794
-0.726
1.140
-1.073
-2.707**
-2.291*
2.901**
0.880
0.195
Constant
-0.384
-7.049**
Observations
Pseudo R2
N
N-Left-censored
N-Clusters
1,298
3.05%
1,298
74
67
1,298
5.19%
1,298
1017
67
*** p<0.01, ** p<0.05, * p<0.10, two-tailed
43
Coefficient
sum
(pre + post)
-3.044**
2.056
0.42
-1.176
-0.279***
(3)
(4)
Panel B: Subsample Analysis
Ln PROD_CHANGE = 0 + 1 ACCURACY (3m) + 2ACCURACY (1m) +3 POSITIVE (3m) + 4 POSITIVE (1m) + <controls>
FORWARD = 0 + 1 ACCURACY (3m) + 2ACCURACY (1m) +3 POSITIVE (3m)+ 4 POSITIVE (1m) + <controls>
Ln PROD_CHANGE = 0 + 1 ACCURACY (3m) + 2 ACCURACY (1m) +3 POSITIVE (3m) + 4POSITIVE (1m) + 5 FORWARD + <controls>
(1)
(2)
(3)
Step 1
VARIABLES
ACCURACY (3m)
ACCURACY (1m)
POSITIVE (3m)
POSITIVE (1m)
FORWARD
Controls
QUANTITY
SALES VALUE
GMPCT
AI_LIFECYCLE
AI_PRODUCTS
AI_ALLOCATION
AI_FCASTIBILITY
EXPORT
CONTINGENT
Y09
Ln
PROD_CHANGE
PRE
-1.024
-0.389
-0.281
1.987
0.012***
-0.141
-7.112***
0.229
-0.287
-0.150
-0.086
-47.682
5.776*
(4)
(5)
Step 2
Ln
PROD_CHANGE
POST
-2.456*
2.328*
-0.189
-1.399
0.004***
4.816***
-1.096**
0.053
-0.170
0.057
0.017
0.898
0.994
0.701
Constant
-3.573
-5.646**
Observations
Pseudo R2
N
N-Left-censored
N-Clusters
370
6.42%
370
305
28
928
3.80%
928
712
66
FORWARD
PRE
2.067
-1.836
-0.343
3.165**
-0.000
-27.692*
2.310
-0.601***
1.097
-0.046
0.667
-48.485
-1.185
1.004
*** p<0.01, ** p<0.05, * p<0.10, two-tailed
44
(6)
Step 3
FORWARD
POST
-2.091**
-1.897*
1.841***
1.152
-0.003***
-2.814
-0.299
-0.239
0.829
0.288**
0.152
-5.879***
0.975
-2.431**
4.551
370
4.56%
370
11
28
(5)
(6)
(7)
928
3.31%
928
63
66
Ln
PROD_CHANGE
PRE
0.062
-1.449
-0.239
3.336**
-0.565***
0.012***
-12.707
-6.230**
-0.073
-0.024
-0.062
0.170
-49.378
5.662*
-2.744
370
8.98%
370
305
28
Ln
PROD_CHANGE
POST
-2.986**
†
2.009
0.309
-1.078
-0.265***
0.004***
4.360*
-0.899**
-0.005
-0.032
0.132
0.031
-0.156
1.274
0.233
-4.642*
928
5.04%
928
712
66
Appendix: Variable Definitions
Variable
Definition
POST
0: pre-introduction period 2006-2007
1: post-introduction period 2008-2010
ACCURACY
Sales forecast accuracy (ACCURACY) is the converse of forecast error and
constraint to be between 0 and 1. Forecast error is defined as the deviation
of the Actual Sales (AS) from the Forecasted Sales (FS), either 1 month or 3
months out, i.e., Error = |AS – FS|/AS. Therefore, Accuracy = 1 – Error.
The constraint to values between 0 and 1 implies:
AS = FS  Accuracy = 1;
Error ≥ 1  Accuracy = 0.
POSITIVE
0: no error (fcst=actual sales) or under forecast (fcst<actual sales);
1: over forecast (fcst>actual sales)
PROD_CHANGE
the deviation of Actual Production (AP) from Planned Production (PP), i.e.,
PROD_CHANGE = AP - PP
Ln PROD_CHANGE
the absolute deviation of logged Actual Production (AP) from logged
Planned Production (PP), i.e., Ln PROD_CHANGE = |Ln (AP + 1) –
Ln(PP+1)|
FORWARD
number of months’ sales it takes to sell (based on actual subsequent sales)
in inventory at the end of the month prior to the month of production.
SALES_VALUE
sales price per unit of measurement (e.g., liters, gallons) multiplied by the
pack size (e.g., 100 liters, 2.5 gallons)
QUANTITY
budgeted sales quantity in physical units
GMPCT
ratio of gross margin per unit (i.e., sales price minus standard costs) and
sales price per unit
AI_PRODUCTS
captures the number of products that share a certain AI. Higher values
indicates greater complexity.
AI_ALLOCATION
indicates the degree to which the AI is freely available on the market or a
scarce resource that needs to allocated to the products sharing the AI.
Higher values indicates greater complexity.
AI_LIFECYCLE
represents the complexity of the lifecycle management, with new (mature)
products being most (least) complex and yielding the highest (lowest) score.
Higher values indicates greater complexity.
AI_FCASTIBILITY
represents the difficulty of forecasting AI demand, with higher values
indicating lower complexity. Lower values indicates greater complexity.
EXPORT
equals 1 if the product is sold outside the United States, and 0 otherwise
CONTINGENT
equals 1 for all products for which at least one event was entered in the
post-introduction period
45