Demand Forecasting For Supply Processes in Consideration of Pricing and Market Information
Demand Forecasting For Supply Processes in Consideration of Pricing and Market Information
Demand Forecasting For Supply Processes in Consideration of Pricing and Market Information
We develop a dynamic model that can be used to evaluate supply chain process
improvements, e. g. different forecast methods. In particular we use for evaluation
a bullwhip effect measure, the service level (fill rate) and the average on hold
inventory. We define and apply a robustness criterion to enable the comparison of
different process alternatives, i. e. the range of observation periods above a certain
service level. This criterion can help managers to reduce risks and furthermore
variability by applying robust process improvements. Furthermore we are able
to demonstrate with our research results that the bullwhip effect is an important
but not the only performance measure that should be used to evaluate process
improvements.
Introduction
This paper deals with the evaluation of supply chain process improvements under special con-
sideration of demand forecasting. Forecasting is an important lever of performance manage-
ment of supply chain processes. To fulfill the customer requirements on delivery performance
demand forecasts are necessary, e. g. to acquire resources (capital, equipment, labor). In this
context an important question arises. How should the performance of different demand forecast
alternatives be evaluated under consideration of the supply chain process. Lee et al. (2004)
present the bullwhip effect (i. e. the amplification of demand variability upstream the supply
chain) as a consequence of the use of quantitative forecast methods in multiple echelons in
the supply chain. The main problem is that the forecasts are based on demand information
generated by inventory policies of the downstream stage in the supply chain. Some papers
show how integration of additional demand information improves the accuracy of demand
forecasting (Natter et al., 2007). Price variation is a further cause of the bullwhip effect that
is interesting in the context of our research. Under consideration of an integrated demand and
supply chain management, forecasting and pricing are also related to each other.
1
University of Neuchâtel, Faculty of Economics, Enterprise Institute, Rue A. L. Breguet 1, 2000 Neuchâtel,
Switzerland
2
Cranfield School of Management, Cranfield, Bedford, MK43 0AL, United Kingdom
1
So far, efforts for a joint optimization of the supply (capacity reservation, planned safety
stock and safety time) and demand forecasting have tried to predict customer demand based
on historical demand patterns. These models have neglected additional available information
for demand side forecasting like prices and/or reference prices. Therefore one goal of this
paper is to present a demand forecast and inventory model that can be used to evaluate the
performance improvements based on the integration of this information. The dependencies
between demand forecast quality, and the drivers of inventory holding costs in make to stock
and the drivers of production capacity reservation costs in make to order environments (ideal
utilization for given variability) should be analyzed via this model. Finally, the evaluation of
the effects of forecast errors based on the bullwhip effect and further “necessary” performance
measures will be discussed. In this context a further research question will be answered: are
there any additional performance measures required or do the classical bullwhip effect measures
cover all relevant performance aspects?
In particular, we develop in this work an extended dynamic demand forecast and inventory
model for a two stage supply chain process, assuming purchase decisions are made based on
rational actors. The last partner (retailer) in the supply chain faces consumer demand. We
will illustrate our approach with an empirical reference dataset of an Austrian retail chain
company. We illustrate how the coordinated application of our model can be used to evaluate
performance of the supply process.
In Section we review relevant literature in the field of supply chain management with
focus on the bullwhip phenomenon and demand management with emphasis on extended price
information. Section describes our estimation and evaluation models while in Section the
results with respect to the reference dataset are shown. Section concludes the paper and gives
an outline of further research activities.
Literature Review
We will give a short overview about the relevant articles for our presented research work in the
context of an integrated supply and demand management. Two major topics are discussed,
bullwhip effect and reference prices.
A pointed definition of the bullwhip effect is provided by de Kok et al. (2005): “The bullwhip
is the metaphor for the phenomenon that variability increases as one moves up a supply chain”.
There are different approaches to identifying the causes of the bullwhip effect. Sterman (1989)
sees wrong decisions made by human decision makers as the major cause of the bullwhip
effect, while Lee et al. (2004) show that this effect occurs even in a supply chain where all
decisions are made in a completely rational way. Notwithstanding these different approaches,
it is obvious that the effect can be reduced by better communication between the elements
making up the chain and through the chance for the manufacturer to see the real (unaltered)
customer demand (Lee et al., 2004, Sterman, 1989). Further research studies compare the
benefits of information sharing with lead time reduction. The results obtained show that in
some supply chain settings the reduction in lead time can have a greater impact on supply
chain performance than information sharing (Cachon and Fisher, 2000).
It is necessary to take the dynamic dependency between bullwhip effect and lead time into
account by evaluating process alternatives. It is obvious that the bullwhip effect has disturbing
performance influence up- and downstream the supply chain process. Chen et al. (2000)
investigates the dependencies between forecasting, lead times and information in a simple
supply chain, i. e. the lead time is an input parameter and not calculated within the model.
2
Boute et al. (2007) show for a two-echelon supply chain how the bullwhip effect could be
dampened under consideration of lead times. Hosoda and Disney (2006) analyze the variance
amplification in a three echelon supply chain model under the assumption of a first-order
autoregressive consumer demand. In this context they were able to obtain analytical results.
Motivated by their research we investigate the influence of demand modelling on supply chain
process performance by comparing different demand models.
By observing actual and past prices customers develop price expectations, or reference prices,
which become the benchmark against which current prices are compared (Popescu and Wu,
2007). Observed prices lower than the customers’ reference price are perceived to be low by the
customer and therefore lead to higher demand compared to a setting where price expectation
and observed price are equal. Observed prices higher than the reference price consequently
result in reduced customer demand. Traditional simplistic demand functions, where demand
D(p) is a decreasing function in price p are extended to demand functions D(p, r) decreasing
in price p and increasing in reference price r.
In the literature there are numerous reference pricing models used for e. g. forecasting of
future demand, describing actual demand patterns or optimization of prices and/or inventory
decisions. A recent review is given e. g. in Mazumdar et al. (2005) and Briesch et al. (1997)
presented empirical results of different reference price models and show that using “past prices
of a brand is the best model of reference price”. Exponential smoothing of reference prices
(eq. 1) is very common in literature (Winer, 1986, Greenleaf, 1995, Kopalle et al., 1996, Fibich
et al., 2003, Popescu and Wu, 2007) and empirically validated (Briesch et al., 1997).
Since reference prices are not directly observable, they have to be calculated or estimated.
It is possible to use
• historical sales and price data to estimate reference prices using time series analysis,
• if historical data are not available or difficult to collect reference prices could be estimated
once by using conjoint analysis (see Jedidi and Zhang, 2002, for a similar model to
estimate reservation prices) and updated using the formation mechanism in Eq. (1) or
While parts of the literature deal with dynamic pricing to optimize current and future profits
(e. g. Greenleaf, 1995, or Kopalle et al., 1996, developed pricing models for retailers and
manufactures), we concentrate in our work on the influence of the price effect on demand
forecasting and assume pricing decisions to be exogenously.
Popescu and Wu (2007) provide analytical insights in the reference effect on demand. They
describe a Gain-Loss Asymmetry which implies that customers react differently on discounts
or surcharges. For a loss averse customer, a surcharge reduces demand more than a discount
of the same magnitude increases demand and vice versa. Empirical evidence for risk averse
pricing behaviour can be found in Ho and Zhang (2004).
3
7 70
sales (right)
price (left)
6
60
5
50
40
30
2
20
1
0 10
2001 2001.5 2002 2002.5 2003 2003.5 2004
4
where (rt − pt )+ = max(rt − pt , 0) is the positive gap between reference price and observed
price and (rt − pt )− = min(rt − pt , 0) is the negative reference price gap. The separation of the
positive and negative gap allows us to model the risk preferences (risk aversion) of customers.
The updating of reference price r follows Eq. (1) but with periodically updated αt :
Note that a similar reference price model can be found e. g. in Natter et al. (2007). An
updating mechanism of reference prices based on Eq. (1) is used.
We compare the performance of the following five estimation models. Two of them include
reference price r and price p while the other two depend on price p only, the simple moving
average model 0 used by Chen et al. (2000) is evaluated as benchmark. Furthermore we
compare the performance of models using auto-regressive terms with non-time-series models.
Parameters β1t are expected to be negative implying decreasing d in p while β2t and β3t are
expected to be positive. The estimation is done using an OLS technique with a finite grid
of α in 0.01 steps evaluated against the minimum sum of squared errors by using the GNU
Regression, Econometric and Time-series library “Gretl” (Cottrell and Lucchetti, 2007). The
R2 measures are similar to other empirical studies of this type (see e. g. Winer 1986).
Model 0 Simple moving average model for n periods used by Chen et al. (2000). Average
goodness of fit: R2 = 0.06, e2 = 13943.
P
Pn
i=1 Dt−i
D̂t = (4)
n
5
Retailer product demand
Supplier orders • base stock
(Manufacturer ) • cycle stock Consumer
• production capacity • safety stock
• production time shipments • Forecast accuracy shipments
• observation periods service level
lead time
supply risk (lead time variability )
mainly dependent on lead time which is influenced by forecast accuracy. The retailer uses an
adapted base stock inventory policy.
Formal description The supplier has to apply a make-to-order production strategy. There-
fore, we use a general distributed queuing model (G/G/1) for the supplier (one product is
produced at one assembly line which is reserved for the retailer) to determine the replenish-
ment lead time L for the retailer (eq. 10) (Hopp and Spearman, 1996). This is an estimation
of the “real” lead time. For the production process only approximate values of the parameters
of lead time can be computed, like the mean lead time which is composed of a variability
component, a utilization component and a capacity component. The queuing model (G/G/1)
seems to be a quite hard approximation of reality. However, an empirical study by Klassen
and Menor (2007) show that it is a good indicator if no other lead time data are available.
Therefore, the relevant input parameters are about
• the arrival and service process, e. g., arrival rate Ra and service time Tp ,
• forecast accuracy (standard deviation of the forecast error / mean forecast error, ca) and
The most relevant output parameter is L = waiting time Ta + service time Tp but queue length
Ia and utilization ρ (eq. 9) are still relevant, e. g. to determine the right utilization to fulfill
the customer requirements.
ρ = Ra Tp (9)
ρ2 ca2 + cp2
Ia = · (10)
1−ρ 2
Ia
L= + Tp = Ta + Tp (11)
Ra
The retailer uses an adapted base stock policy as presented in the following equations. Fur-
thermore, we explain some restrictions of the operations which have to be taken into account.
The theoretical safety stock st (eq. 18) is calculated like the classical base stock policy (Neale
et al., 2003) with SF denoting the safety factor. In general, for each planning period an order
is placed to boost the inventory position (inventory currently on-hand plus on-order) up to the
base stock level.
6
The base stock yt (eq. 12) determination takes the demand forecast, D̂t , for period t into
account. This forecast is based on the different forecast alternatives described in Section .
A further requirement is to calculate the average replenishment lead time for the retailer L̄t
(eq. 13) at time t, the replenishment lead time standard deviation for the retailer σLt (eq. 14),
the average demand D̄t (eq. 15) per time period t and the average demand standard deviation
σDt (eq. 16) based on the observations from the previous n periods.
yt = D̂t R + L̄t + st (12)
Pt
i=t−n+1 Li−1
for t > n,
L̄t = n P (13)
t
(n−t)L0 + i=1 Li−1
for 1 ≤ t ≤ n
n
r P
t 2
i=t−n+1 ( i
L −L̄i )
for t > n,
n−1
σLt = r Pt 2
(14)
2+
i=1 ( i
(n−t)a0 L −L̄i )
for 1 ≤ t ≤ n
n−1
Pt
i=t−n+1 Di−1
for t > n,
D̄t = n P (15)
t
(n−t)D̄0 + i=1 Di−1
for 1 ≤ t ≤ n
n
r P
t
(e
i=t−n+1 i−1
)2
for t > n,
n−1
σDt = r Pt with et = Dt − D̂t (16)
(n−t)e2+
0 (e
i=1 i−1
)2
for 1 ≤ t ≤ n
n−1
Additionally to the planning parameters, we also present the calculation of the observed
values for the base stock yto (eq. 17) at time t. In addition to the safety stock st (eq. 18), the
observed inventory position yto represents the connection to the supplier because deliveries of
the supplier xt (eq. 19) increase the yto and the fulfillment of customer demand Dt decreases
the yto . Waiting replenishment orders, W ROt , which have not been arrived at the inventory yet
have to be also taken into account (eq. 20). Deliveries of the supplier in period t are identical
to the retailer orders xt−1−Lt (eq. 19). The service level of the retailer (fill rate) SLt (eq. 21)
is calculated under consideration of the observed safety stock sot (eq. 23) and demand Dt .
We calculate the average on-hand inventory OHIt (eq. 22) based on the observed safety
stock sot .
(
o
yt−1 + xt−Lt − Dt−1 for t > 0
yto = (17)
y0 − W RO0 for t = 0
r
st = SF 2 R + L̄ + σ 2 D̄ 2
σD (18)
t Lt t
7
Pt
i=t−Lt xi−1
for t > 1 + L0
t−1
(L0 − t + 1) D̄0 + for 1 ≤ t ≤ 1 + L0
P
W ROt = xi (20)
i=0
L0 · D̄0 for t = 0
Pt o
i=1 max (Dt − st , 0)
SLt = 1 − Pt for t ≥ 1 (21)
i=1 Dt
Pt o
i=1 max (si , 0)
OHIt = for t ≥ 1 (22)
t
cout
W = with
cin
σ (Dout (t, t + T )) σ (Din (t, t + T )) (24)
cout = , cin =
µ (Dout (t, t + T )) µ (Din (t, t + T ))
Motivated by Boute et al. (2007) we extended our model by introducing a bullwhip damp-
ening parameter ζ. This adoption causes the following modification of the retail order xt for
any t > 0 (ζ = 1 for t = 0):
Model parameters
Input parameters are based on discussions with managers of different suppliers and retailers.
The initial values are based on a carefully conducted sensitivity analysis. The verification as
well as validation results are used for calibrating of the systems dynamics model, e. g. lead time
initial value L0 , forecast error initial value e0 or average demand initial value D0 (see table 1).
The time interval ∆t and the duration of the simulation T are motivated by the empirical data,
by the forecast process and the requirements to show dynamic effects over multiple periods.
8
Table 1: Standard input parameters of the analyzed supply chain process
D̄0 average demand initial value 39 consumer orders
a20 squared lead time volatility initial value 10
ζ bullwhip dampening parameter 0.5
ca demand volatility σDt /D̄t
cp production time volatility 0
∆t time interval 1 week
e20 squared forecast error initial value 130
D̂0 average forecasted demand initial value 39 consumer orders
n number of observation periods 7 – 22
ρ average utilization of the supplier (manufacturer) 65%
R review period (time between successive orders) 1 week
SF safety factor 2.33
T number of time steps (product class life time) 76 weeks
L0 replenishment lead time initial value 2 weeks
δt autocorrelation parameter in period t model estimation
βit model parameters in period t model estimation
ct constant in period t model estimation
Dt product demand in period t see fig. 1
pt product price in period t see fig. 1
Results
We use a subset of 76 periods (2002/10/28 to 2004/04/05) of all data for out of sample
forecasting to evaluate the forecast quality with respect to our evaluation model. Performance
is measured in terms of the bullwhip effect (eq. 24), service level (eq. 21) and inventory measures
(eq. 22).
The analyzed models result in the following e2 for the forcast D̂t :
P
9
1
0.99
0.98
0.97
Service Level SL
0.96
0.95
0.94
0.93
0.92
Model 0
Model 1
0.91 Model 2
Model 3
Model 4
0.9
13 14 15 16 17 18 19 20 21 22
periods
4.5
Bullwhip Effect
3.5
2.5
1.5
13 14 15 16 17 18 19 20 21 22
periods
10
500
Model 0
Model 1
Model 2
Model 3
450 Model 4
400
350
Inventory
300
250
200
150
13 14 15 16 17 18 19 20 21 22
periods
The identification of the ideal forecast methods and inventory policy parameters (i. e. the
number of calculation periods) is simple. We assume that all forecast methods cause the
same handling cost. Therefore, an alternative with a simulated service level of 100% and
the lowest average inventory should be selected. In particular, most models show their ideal
performance with 20 observation periods, only model 1 shows best performance with 21 periods.
Furthermore we are not able to identify any significant difference between the forecast methods
under consideration of this criterion.
Let n∗ denote the range of observation periods n with a simulated service level of 100%. This
evaluation criterion is used to evaluate the robustness of the selected improvement alternatives.
Model 2 presents the best performance with n∗ = 6 (in observation periods n = 15 . . . 20),
followed by model 1 with n∗ = 4 (n = 18 . . . 21). Model 3 and model 4 result in n∗ = 1 for the
period n = 20, while model 0 results in n∗ = 2 for the periods n = 19 . . . 20.
The analyzed product is characterized by a very long product life cycle. Nevertheless, the
robustness criterion should occupy a major position by the selection of the forecast method,
due to the fact that the optimal number of observation periods could only be determined ex
post. If the number of observation periods is limited by a short life cycle this criterion will be
an important factor.
The performance of the different forecast models is also determined by the assumptions
of our illustration example, e. g. utilization, maximum allowed lead time and initial values.
There is no significant difference in terms of the forecast error between the models including
and not including reference prices (see table 2) but the caused lead time variability is higher
for reference price models. Therefore flexibility (utilization, etc.) of the manufacturer has an
important influence on the overall performance of the analyzed supply chain process.
Our research is based on the research work of Chen et al. (2000). They show how the
bullwhip effect is influenced by the number of observation periods for the standard deviation
11
of the forecast error as well as the calculation of the forecasted demand. We considered
additionally the lead time variability caused by the base stock inventory policy for different
forecast methods. Our illustration example (see Figure 1) confirms this analytical result but
under consideration of the service level it is obvious that the number of periods influences also
the service level performance. We confirm also the restriction that an improvement of the
bullwhip effect by increasing the number of observation periods is limited. In our illustration
example period 21 (22 for model 1) leads to decreasing results, caused by the characteristic of
the empirical data set.
Second, we investigate the performance of the analyzed forecast methods under consideration
of extended model with additional dampening of the order variability. The results are very
similar to the basic model. On the one hand a little decrease of the bullwhip effect measures
can be observed and on the other hand the average inventory can be decreased. The problem
is that the service level will be also decreased. In particular the robustness criterion n∗ shows
the following results: n∗ = 2 for model 1 and 2 (n = 20 . . . 21) and n∗ = 0 for all other models.
These results confirm that the bullwhip effect is an important but not the only performance
measure that should be used to evaluate process improvements.
Conclusion
Based on analytical research work about the bullwhip effect (see Section ) we developed a
dynamic model that can be used to evaluate supply chain process improvements under con-
sideration of different forecast methods. In particular we used for evaluation a bullwhip effect
measure, the service level (fill rate) and the average on-hand inventory. We define and apply a
robustness criterion to enable the comparison of different process alternatives. This criterion
can help managers to reduce variability by applying robust process improvements because the
optimal number of observation periods can be only determined ex post.
We could demonstrate with our illustration example that the evaluation of forecast methods
should not be restricted to the classical performance indicators like MAD, MSE, MAPE, SAE,
etc. but using our proposed robustness criterion provides further insights. These results
confirm that the bullwhip effect is an important but not the only performance measure that
should be used to evaluate process improvement. It can be recommended to apply also further
measures like average on hold inventory and service level (e. g. fill rate).
Furthermore, dampening of the order variability in our illustration example decreases the
bullwhip effect and the average on-hand inventory but with the problem of a decreasing service
level.
Further research activities will deal with the dynamic effects in more detail, e. g. to identify
the ideal (in terms of the overall supply chain process performance) number of observation pe-
riods ex-ante by a given demand characteristic forecast method, inventory policy and bullwhip
dampening procedure.
References
Boute, R. N., Disney, S. M., Lambrecht, M. R., Van Houdt, B., April 2007. An integrated
production and inventory model to dampen upstream demand variability in the supply
chain. European Journal of Operational Research 178 (1), 121–142.
12
Briesch, R. A., Krishnamurthi, L., Mazumdar, T., Raj, S. P., 1997. A comparative analysis of
reference price models. Journal of Consumer Research 24, 202–214.
Cachon, G. P., Fisher, M., 2000. Supply chain inventory management and the value of shared
information. Management Science 46 (8), 1032–1048.
Chen, F., Drezner, Z., Ryan, J. K., Simchi-Levi, D., 2000. Quantifying the bullwhip effect in a
simple supply chain: The impact of forecasting, lead times, and information. Management
Science 46 (3), 436–443.
de Kok, T., Janssen, F., van Doremalen, J., van Wachem, E., Clerkx, M., Peeters, W., 2005.
Philips electronics synchronizes its supply chain to end the bullwhip effect. Interfaces 35 (1),
37–48.
Fibich, G., Gavious, A., Lowengart, O., 2003. Explicit solutions of optimization models and
differential games with nonsmooth (asymmetric) reference-price effects. Operations Research
51, 721–734.
Fransoo, J. C., Wouters, M. J., 2000. Measuring the bullwhip effect in the supply chain. Supply
Chain Management: An International Journal 5 (2), 78–89.
Greenleaf, E., 1995. The impact of reference price effects on the profitability of price promo-
tions. Marketing Science 14 (1), 82–104.
Ho, T.-H., Zhang, J., 2004. Does format of pricing contract matter? Working paper, University
of California at Berkeley.
Hopp, W. J., Spearman, M. L., 1996. Factory Physics: Foundations of Manufacturing Man-
agement. Irvin Inc., Chicago.
Hosoda, T., Disney, S. M., 2006. On variance amplification in a three-echelon supply chain
with minimum mean square error forecasting. Omega 34, 344–358.
Jedidi, K., Zhang, Z. J., 2002. Augmenting conjoint analysis to estimate consumer reservation
price. Management Science 48, 1350–1368.
Klassen, R. D., Menor, L. J., 2007. The process management triangle: An empirical investiga-
tion of process trade-offs. Journal of Operations Management 25 (5), 1015–1034.
Kopalle, P. K., Rao, A. G., Assuncao, J. L., 1996. Asymmetric reference price effects and
dynamic pricing policies. Marketing Science 15 (1), 60–85.
Lee, H. L., Padmanabhan, V., Whang, S., 2004. Information distortion in a supply chain: The
bullwhip effect. Management Science 50 (12), 1875–1886.
Mazumdar, T., Raj, S., Sinha, I., 2005. Reference price research: Review and propositions.
Journal of Marketing 69, 84–102.
13
Natter, M., Reutterer, T., Mild, A., Taudes, A., 2007. Practice prize report – an assortment-
wide decision-support system for dynamic pricing and promotion planning in DIY retailing.
Marketing Science 26 (4), 576–583.
Neale, J., Tomlin, B., Willems, S., 2003. The role of inventory in superior supply chain per-
formance. In: Harrison, T. P., Lee, H. L., Neale, J. J. (Eds.), The Practice of Supply
Chain Management: Where Theory and Application Converge. Kluwer Academic Publish-
ers, Boston Dordrecht London.
Popescu, I., Wu, Y., 2007. Dynamic pricing strategies with reference effects. Operations Re-
search 55 (3), 413 – 429.
Winer, R. S., 1986. A reference price model of brand choice for frequently purchased products.
Journal of Consumer Research 13, 250–256.
14