020-0131

AN EXPERIMENTAL INVESTIGATION OF THE EFFECTS OF SUPPLY UNCERTAINTY ON SUPPLY CHAIN PERFORMANCE

Ancarani A.1, Di Mauro C.2, D’Urso D.3

1Dipartimento di Ingegneria Civile e Ambientale, Università di Catania – - +390957382715

2D.A.P.P.S.I., Università di Catania – - +3909570305216

3Dipartimento di Ingegneria Industriale e Meccanica, Università di Catania - - +390957382712

POMS22 nd Annual Conference

Reno, Nevada, U.S.A.

April 29 to May 2, 2011

Abstract: We present a series of controlled human experiments investigating the impact of supply uncertainty on buyers’ inventory management. The experiments focus on one specific source of SC risk, namely stochastic lead times, within the framework of the “beer game”.

Specifically, we study the impact of stochastic lead times on supply chain performance and on the formation of the bullwhip effect. Next, the impact of players’SCM skills and game experience is investigated.Two between subject treatments were run, a standard beer game (SBG) and a beer game with stochastic lead times (SLT). A total of 104 MBA students and 24 purchasing managers participated. Results show that SLT gives rise to higher costs and a more marked bullwhip effect than the SBG. Botheffectsdecrease with experience in the game. However, they do not disappear, even when real managers are involved in the game.

Keywords: Experiments, Beer Game, SCM

  1. Introduction

Today’s supply chains are more complex than they used to be. There are various reasons for supply chain complexity, such as higher levels of R&D and manufacturing outsourcing, supplier–supplier relationships in supplier networks, increased dependence on supplier capabilities, new technologies, regulatory requirements, shorter product life-cycles due to rapidly changing customer preferences, and international market and production expansion (Wagner and Neshat, 2010). Further, as firms try to reduce costs through the rationalization and reduction of the supply base, the aim to secure an interrupt flow of materials has become more difficult to achieve (Harland et al., 2003). Finally, global trends towards leaner supply chains have taken place at the cost of increased vulnerability due to fewer buffers. Since lean supply chains reduce the margin for human errors, it is more likely that disruptions or supply chain instability result from those errors (Tang, 2006)).

As a consequence, the instability of supply chains has increased (Geary et al., 2006). The Bullwhip Effect (BWE) is a paradigmatic representation of SC instability. The BWE is generally triggered by demand uncertainty (Forrester, 1958), and it entails that as external demand passes upstream through the SC from the downstream to the upper stream levels of the chain, the variance of orders is amplified. This behaviour can imply substantial costs in terms of stock-out costs and in inventory and obsolescence costs.

The BWE is documented empirically in several industries (Blanchard, 1983; Lee et al. 1997, 2000 among many others) and has been studied in a large number of experiments, either numerical or carried out with human subjects. Controlled human experiments have gained importance as a methodology for the study of the BWE since Sterman’s (1989) finding that the BWE is a problem arising mainly as a consequence of human decision making. In particular, Sterman posited that due to the lags in acquiring materials and reacting to changes in demand, individuals involved in SC management amplify unanticipated changes in demand and have a biased perception of the flows in transit through the SC pipeline. The volatility thus generated is further amplified by feedback effects across the network.The experiments conducted with human subjects have further demonstrated that the BWE is especially strong when the SC is non-integrated (Croson and Donohue, 2002), and have pinpointed the advantages of introducing information sharing (Croson and Donohue, 2006) and coordination across the various layers of the SC (Wu and Katok, 2006).

Most of the extant literature on the BWE has focused on the impact of uncertainty in demand, while ignoring the impact of other sources of uncertainty (process side, supply-side, control-side). Among the supply-side sources of uncertainty, lead times variabilityis one of the most relevant, as the assumption of deterministic lead-time is especially restrictive since lead times are not deterministic in many supply chains. Moreover,some authors have posited that the BWE reduction is best enabled via implementation of the principles of smooth material flow and the reduction of actual or perceived shortage risk (Geary et al., 2002; 2006). Only a small number of numerical simulations have investigated the effects of lead time uncertainty on the formation, extent and consequences of the BWE (Chatfield et al., 2004; Truong et al., 2008), showing that stochastic lead times do contribute to worsen BWE.

In light of the discovery of the importance of biases in human decision making for SC performance, it is even more surprising that negligible attention has been paid to the effects of supply uncertainty on the BWE and SC costs. Human experiments on the BWE have shown that human behavior deviates from a model that accounts only for rational interactions and the dynamics of the system, and thus there are grounds for positing that natural aversion to uncertainty may bias the performance of the SC under uncertain lead times. Further, since people learn “to live with risk”, a human experiment can also provide insight into the path that individual behavior follows as they experience a highly variable environment.

To the best of our knowledge a test of the impact of uncertainty in lead time using human subjects is still lacking. In this paper, we explore the behavior of members of a SC in the face of supply uncertainty, that we make operational through uncertainty in lead times, and contrast the performance of a SC with stochastic lead times with that of a SC with deterministic lead times. We carry out the study within the framework of the beer distribution game, a simulated serial supply chain with four echelons (retailer, wholesaler, distributor, and factory).

The research questions we investigate are the following:

  1. What is the impact of stochastic lead times in supply chains?
  2. To what extent the different performance (if any) of the supply chain under known vs. stochastic lead times is caused by inexperience of those who manage the supply chain?
  3. How does out-of-task experience affect performance? Both classroom and managerial experience can be expected to influence judgment and information handling. Task experience gained in the classroom exposes participants to the broad principles underlying inventory control and SC management. Managerial experience provides intensive exposure to practical problems, so it is likely that students and managers behave differently (Bolton et al., 2008).

The paper is organized as follows: Section 2 reviews the relevant literature that underpins the hypotheses tested through the experiment, Section 3 discusses the experimental design, while Section 4 presents the results. Section 5 concludes the paper summarizing the main findings and highlighting implications for future research and SC management.

  1. Factors investigated and hypotheses tested

2.1The bullwhip effect with stochastic lead times

The BWE has been widely studied in the context of the so-called “beer game”. In the classic beer distribution game (Forrester, 1961) the supply chainnormally consists of four echelons (retailer-wholesaler-distributor-factory). Inventory is managed according to the periodic review inventory model (order-up-to). During the game each i-participant, i [1,...,4], places orders, Oi(t), to the immediate upstream supplier and fills downstream customer’s orders, Di(t). At each level, when a buyer places an order a delay of one week (LTI) occurs before this latter is known to the upstream supplier, Di(t) = Oi-1(t-1); a two weeks lead time (LTD) is requested to ship orders to the downstream echelon and the same happens to the factory when beer is brewed (LTP). At each level, goodsreceived at time t, Ri(t), correspond to the ones which were shipped by the upstream supplier two weeks before, Si+1(t-2). During the game each player must respect an inventory balance: Ii(t) = Ii(t-1) + Ri(t) – Si(t), where Ii(t) is the on hand quantity; customer orders are filled if Ii(t) ≥Di(t)otherwise Si(t)Di(t) and backorders occur, Bi(t) = Di(t) –Ii(t-1) + Ri(t).

If external demand is variable, larger oscillations of orders occur as one moves upstream the SC, giving rise to the BWE. Most of the extant literature on the BWE has focused on the standard case described above, i.e. variable customer demand and constant lead times, thus ignoring the impact of other sources of uncertainty on the BWE (process side, supply-side, control-side).

A few papers in the last ten years have addressed the problem of the BWE and of SC performance relaxing the assumption of deterministic lead times. Chatfield et al. (2004) use simulation to investigate the effects of stochastic lead times in a k-node SC. In particular, through a factorial design, the effect of various levels of lead time variability is crossed with that of four levels of information quality and updating rules, and absence/presence of information sharing. The authors show that as the variance of lead time increases, in the case of a normally distributed customer demand, BWE worsens. This result is confirmed by Truong et al. (2008) assuming either an AR(1) or an ARMA(1,1) model for customer demand.

Chatfield et al. (2004) further show that the amplification of order variances is higher when historical information on lead times variances is used to update inventory parameters than when this information is not used (due to misperception of the variability or indifference towards uncertainty). This finding seems to be consistent with that of Chen et al. (2000) whoverified that the BWE would not exist if there were no forecast-based orders attempting to capture the latest demand information.

Kim et al. (2006) present a model with stochastic lead time in which the case of information sharing (customer demand is common knowledge for all echelons of the chain) is contrasted with that of no sharing. Results show that the variance of orders increases nearly linearly in echelon stage with information sharing, and exponentially without information sharing. One managerial implication of this result is that the sharing of information on customer demand by all echelons is an effective way to reduce BWE also under conditions of supply uncertainty. However, information sharing per se does not eliminate BWE, which remains higher than under condition of deterministic lead time.

Heydari et al. (2009) try to isolate the impact of lead time uncertainty from that of demand uncertainty by simulating a four-stage SC in which customer demand is constant. Results show that the uncertainty in lead time increases the varianceof orders at each echelon but does not worsen BWE. Further, order variance is positively correlated to the variance of inventory levels and the amounts of stock-out. However, results of this study are not directly comparable with those already discussed, since Heydari et al. (2009) assume that if the supplier holds insufficient inventory to satisfy an order, the unavailable quantity will be lost. Similarly, all delayed orders will be lost. Further, the uppermost level of the chain does not face stochastic lead time.

Chaharsooghi and Heydari (2010) study the impact of two classes of policies (implemented either through supplier selection or investment strategies) meant to reduce supply risk under stochastic lead time: policies aimed at reducing the mean lead time and policies addressing the reduction of the variance. The former make the customer get the product quickly, the latter make the lead time more predictable. Results of simulations carried out by the authors on a four-stage SC suggest that reduction in lead time variance reduces stock-outsize, whereas mean lead time reduction reduces BWE, with the effect of the former on SC performance much greater than the effect of the latter. The simulations are run under the same assumptions (in particular constant demand) of Heydari et al. (2009).

The above discussion suggests the following hypotheses:

Hypothesis 1: in the presence of both customer demand uncertainty and stochastic lead time, the BWE will be higher than in the case when only demand uncertainty is considered and the lead time is deterministic.

2.2The effect of experience and learning-by-doing in BWE experiments

Several experimental studies have focused on the impact of the amount of information about the SC and of information sharing on SC performance and the bullwhip effect using the beer game. For instance, Croson and Donohue (2003) study the impact of Point of Sale (PoS) data sharing and find that information sharing mitigates the bullwhip effect by reducing the oscillations of orders of upstream members. Similar results are obtained by Steckel et al. (2004). Gupta et al. (2002) study the beer game under three different demand scenarios (a step-up function, a S-shaped function, and a S-shaped plus error function) and show that PoS information results in improved performance only in the simplest scenario (the step-up demand). Machuca and Barajas (2004) find that significant savings can accrue from the implementation of an electronic data interchange (EDI) across the entire supply chain. Cantor and Macdonald (2009) show that information availability about the supply line (local vs. system-wide) interacts with the problem-solving approach (abstract/flexible rules versus concrete/fixed rules) in determining the performance of the supply chain.

A different strand of research has focused on the learning process of the actors involved in the game. Wu and Katok (2006) study the effect of learning-by-doing on SC performance and find that experience of the game significantly reduces the bullwhip effect only when it is “system wide”, i.e. it concerns the whole structure of the game, whereas ambiguous effects are reported when role specific experience is given to participants in the experiment. The implication is that training (i.e. repetition of the task) may improve individuals’ knowledge and understanding of the system. However, it does not improve SC performance unless supply chain partners are allowed to communicate and coordinate through knowledge sharing.

The effect of learning-by-doing has also been explored in the context of the newsvendor problem. Bolton and Katok (2008) find that knowledge gained through personal experience leads to a significant improvement in performance. Similarly, Ben-Zion et al. (2008) find that game experience is important in improving profits and leads to stationary orders. Although in the earlier stages of the game orders exhibited a bias towards mean demand, this effect was significantly reduced as the game unfolded. Bostian et al. (2008) test and find evidence in favour of a learning model with recency effects, i.e. in which subjects respond to recent gains and losses, and inertia. This model adds to previous studies since it incorporates demand chasing in a framework where expected payoffs also matter. Gavirneni and Isen (2010) carry out a verbal protocol analysis to explore the reasons underlying recency effects. They observe that participants had difficulty dealing with the abstractness of the task and tried to identify anchors for their decision-making process, the most common of which was average demand. Further, the players focused on the basic information relevant to the decision and ignored some of the advanced information that would have helped them make a better decision.

Finally, performance in ordering decision may be tied to professional experience of the players. Bolton et al. (2008), comparing the ordering behaviour of students at different level of education with that of expert procurement managers, find the anchoring bias toward average demand is observable whatever the level of experience the subjects have. Rather than professional experience, it is the exposure to varying levels of information and task training that has a significant effect on performance.

Thus, on the basis of the above literature, we formulate the following hypotheses:

Hypothesis 2: task experience in presence of demand variability reduces the bullwhip effect both under conditions of variability of lead times and with deterministic lead time.

Hypothesis 3: Out of task experience has a lower impact than task experience on SC performance.

  1. Experimental design
  2. Students’ experiment

In this experiment, two treatments were run. The first treatment (SBG hereafter) reproduced a beer game with four echelons (i = 1,...,4), i.i.d. normally distributed external demand with parameters known to all echelons (µ = 100, σ = 20), known and constant lead times equal to 3 (LTI = 1, LTD = 2). This design differs from Sterman’s experiments, in which the retail demand is completely unknown and non-stationary, and is represented by a simple step-function whereby demand starts at 4 units and jumps to 8 units after the eighth game period. However, subsequent studies by Croson and Donohue (2006) have shown that even with stationary and known demand distributions, the BWE arises. Thus, we expect the BWE to arise also in our setting.

In the second treatment (SLT henceforth) lead times of all suppliers in the chain (including the factory’s brewery) are uniformly distributed in the interval (1, 2,3) periods. Thus, in addition to demand uncertainty, players face also supply uncertainty stemming from stochastic lead times.

In both treatments, an order placed with the supplier can be partially fulfilled with a continuous distribution, depending on the supplier’s inventory availability.The histories of incoming demands, of past shipments and past purchases are available to each player. From this information, the history of lead times can be worked out in SLT. Also, because of stochastic lead times, in SLT order cross-over can occur.

Summing up,the game can be assumed to mimic a non-integrated supply chain in which each buyer has a single supplier; no information sharing about actual demand, inventories, backlogs, and own lead times, is allowed among SC participants; and the retailer is the only echelon of the chain that observesexternal demand.

Behaviour in both treatments was observed for a number of periods (T), from 36 up to 50. Players were not informed of the final period of the gameto avoid end-of-game behaviour that might triggerover- or under-ordering.

Each echelon began with an initial inventory levelIi(t=1) =300, outstanding orders Oi(t= 0, -1) =100 for the previous two periods,and an incoming shipment Si(t=2, 3) =100 in the following twoperiods. All experimentsalso used the same random number seed to generatedemand, i.e., Di(t), t=1, …..T was identical acrossgroups. This allowed us to isolate variations due toordering behaviour from variations due to differentdemand streams.