Eindhoven University of Technology MASTER Improving the promotion forecasting accuracy at Unilever Netherlands van der Poel, M.J. Award date: 2010 Link to publication Disclaimer This document contains a student thesis (bachelor's or master's), as authored by a student at Eindhoven University of Technology. Student theses are made available in the TU/e repository upon obtaining the required degree. The grade received is not published on the document as presented in the repository. The required complexity or quality of research of student theses may vary by program, and the required minimum study period may vary in duration. General rights Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain
96
Embed
Improving the promotion forecasting accuracy at Unilever ...
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Eindhoven University of Technology
MASTER
Improving the promotion forecasting accuracy at Unilever Netherlands
van der Poel, M.J.
Award date:2010
Link to publication
DisclaimerThis document contains a student thesis (bachelor's or master's), as authored by a student at Eindhoven University of Technology. Studenttheses are made available in the TU/e repository upon obtaining the required degree. The grade received is not published on the documentas presented in the repository. The required complexity or quality of research of student theses may vary by program, and the requiredminimum study period may vary in duration.
General rightsCopyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright ownersand it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
• Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain
n.e. = no effect on the promotional sales a The baseline group for the different product categories is the product group “Four_or_five_for” b The baseline group for the different retailers is the retailer “Albert Heijn” c The baseline group for the different product groups is the product group “Homecare”
To increase the usability of the model in practice, a model is constructed with the above most
important variables. The model fit of this model with a limited number of variables is still surprisingly
high and almost equal to the model fit of the full model. However, not all variables have data
availability at Unilever, since Unilever as a manufacturer is dependent on a retailer for information of
upcoming promotions. For two variables in the adapted model Unilever has no data availability.
These are the percentage of shops with a second placement and the extra number of shops where
the product is sold in promotion. To analyze what the effect is of the lack of data on the forecast
accuracy of the model a new model without these variables is tested. The model fit decreases to an
adjusted R-square of around 0.500, indicating that the exclusion of the two variables substantially
Improving the promotion forecasting accuracy at Unilever Netherlands
Page VII
worsens the performance of the forecasting model. Thus it is important for Unilever to gain data
availability on these variables.
Unilever not only wants to know how much a promotion sells on the shopping floor, but also wants
to know how much a retailer orders of a product. Therefore, the model results for the consumer
demand are adapted to retailer orders. The retailers included in the research order on average
between 39% and 85% more than is sold during the promotion. The forecasts for the consumer
demand are raised with this difference. The model performance decreases substantially because of
the extra variance in the retailer orders. The adjusted R-square for the Non-Food data set has
decreased to 0.103, meaning that the predictive power of the model is almost absent. For the Food
data set the R-square is 0.392. So, the variability in the retailer orders is a lot higher for the Non-
food products than Food products. Forecasting retailer orders for Non-Food products seems to have
little to no benefit, forecasting retailer orders for Food products has more practical value.
Implementation & conclusions
The different model adaptations in this research show that if the right information is available
Unilever is very well capable of accurately predicting the consumer demand. Unilever has an
advantage over the retailer because of their larger data pool of promotions over all retailers which
can be used to forecast upcoming promotions. Hence, with this skill Unilever is able to take the lead
in establishing a collaboration with retailers and increasing the forecast accuracy.
However, two aspects decrease the forecast accuracy of a manufacturer. First, a manufacturer has
less data availability than a retailer and thus important variables cannot be used to forecast the
promotional demand. Second, forecasting retailer orders has turned out to be far more difficult than
consumer demand, especially for Non-Food products. The bullwhip effect leads to a substantial
deviation between retailer orders and consumer demand. As a result, Unilever should first increase
their data availability on promotions by closer collaboration with retailers and better database
management. Thereafter, in order to be able to accurately forecast retailer orders, the disturbing
factors behind the bullwhip effect should be analyzed. Close collaboration with a retailer is needed to
successfully analyze these disturbances. When the disturbing factors are successfully analyzed, a
promotion forecasting model which forecasts the consumer demand and corrects for the disturbing
factors should be formulated and employed together with the retailer.
Concluding, close collaboration and information sharing is needed, where in the end Unilever and
the retailer together use one forecasting approach. Concepts like Vendor Managed Inventory (VMI),
Continuous Replenishment Program (CPR) and Collaborative Planning, Forecasting and
Replenishment (CPFR) can be used to increase the collaboration between Unilever and a retailer,
where VMI is the most basic concept and CPFR is the most advanced concept.
Improving the promotion forecasting accuracy at Unilever Netherlands
Page VIII
Preface
This master thesis is the result of the final part of my study Industrial Engineering and Management
at Eindhoven University of Technology. The master thesis project was executed Unilever Netherlands
in Rotterdam from the beginning of 2010 up to the end of the summer.
When I started my master thesis I just came back from an international semester in Hong Kong. Life
over there had been eye opening, and really interesting, but also relaxing and having a lot of fun in
one way or another. Therefore, starting my master thesis in Rotterdam really pushed me back into
normal hard working life. And I have to say that I still feel lucky that an opportunity for my master
thesis had presented itself at Unilever, since the working atmosphere is really good in the
headquarters in Unilever Rotterdam. Luckily the burden of the master thesis did not feel like that at
all, so I can look confidently in to the future where a real job is waiting for me.
I would like to grab the opportunity to express my gratitude towards a few people. First of all, I
would like to thank Patrick van Balkom, my supervisor at Unilever. His guidance and comments
provided very useful insights and shed light on my path the moments I needed it. I really enjoyed
working with him.
Second, I would like to thank my first supervisor at the TU/e, Karel van Donselaar. His thorough
knowledge on the subject led to some very good discussions. And without his efforts of finding an
internship I would not have had the opportunity at Unilever. Third, I would like to thank my second
supervisor at the TU/e, Jeroen Schepers. The feedback he gave on my work provided new insights
and improved the quality of my work.
Lastly, I would like to thank my girlfriend for supporting me during the project.
Thijs van der Poel
Rotterdam, August 2010
Improving the promotion forecasting accuracy at Unilever Netherlands
Page IX
Index
Abstract .................................................................................................................................... III
Management summary............................................................................................................... IV
Preface ................................................................................................................................... VIII
Index ........................................................................................................................................ IX
Part 1: Project definition .................................................................................................... 1
1 Introduction of research ....................................................................................................... 1
1.1 Structure of report ........................................................................................................ 1
1.2 Company description ..................................................................................................... 1
1.3 Problem introduction ..................................................................................................... 3
Preservability 0.001 0.291* 0.001 0.296* a a 0.001 0.168* 0.001 0.247*
log_size_of_product a a a a a a 0.112 0.057 0.362 0.180*
Frequency_of_purchase -0.043 -0.059 -0.098 -0.132* a a -0.053 -0.079* -0.115 -0.193*
Personalcare d a a a a 0.086 0.055 a a a a
Ice_and_beverages d -0.310 -0.133* -0.776 -0.440* a a -0.164 -0.069* -0.361 -0.236*
SCC_and_vitality_shots d 0.594 0.173* 0.454 0.185* a a 0.318 0.101* 0.396 0.204*
Savoury_and_dressings d 0.146 0.090** a a a a a a a a
winter_products_temp a a a a a a a a a a
summer_products_temp 0.042 0.306* 0.052 0.519* a a a a a a
* = significant at a 0.01 significance level ** = significant at a 0.05 significance level a The variable is not significant for this data set. b The baseline group for the different product categories is the product group “Four_or_five_for” c The baseline group for the different retailers is the retailer “Albert Heijn” d The baseline group for the different product groups is the product group “Homecare”
Table 7-3: Unstandardized and standardized Beta coefficients with significance level for all 5 data sets
Based on the Beta coefficients in Table 7-3 the hypotheses drawn in paragraph 6.3 will be discussed.
The promotional variables, Display, Folder and TV_support were expected to have a very positive
effect on the promotional sales. Display indeed has a very positive effect, Folder has a positive
effect; however, less than display. And TV_support only has an effect in the HPC dataset (H1 and
H2 confirmed and H3 rejected). Because of the unexpected result for the variable TV support an
extra analysis is performed. Since there is only data availability for promotions at the Albert Heijn for
the variable TV_support, it is worth to check if the variable is significant if loaded for all promotions
in the sample at Albert Heijn. Appendix 5 depicts the results for this single linear regression model.
The standardized Beta coefficient is significant at a 0.003 level with a value of 0.141, which is still
not very high. Since in a full model colinearity with other variables is likely to decrease this effect
size, it is concluded that the effect size is medium to small. Therefore, in this research TV_support is
not considered as an important variable for the model. However, when full information is available
for all retailers, a new analysis is needed to test this conclusion.
The promotion variables Holiday_products and Promo_length were suspected to be positively
correlated with the promotional sales. No significant effect is found at all for Holiday products.
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 31
Products like luxury ice cream do sell more in Holiday period; however, apparently this effect is lost
or hard to find for promotions in holiday period. Regarding the promo_length, this variable indeed
has a high standardized Beta coefficient, especially for the HPC models and data set where all
promotions are included. The effect in the Food data set is minor, since almost all promotions have a
duration of 1 week in this data set. (H4 rejected and H5 confirmed).
Regarding the discount on a promotion, both log_absolute_discount and percentual_discount are
expected to have a highly positive influence on promotional sales. However, only
percentual_discount confirms this hypothesis and log_absolute_discount has no influence or even a
significant negative influence in three of the data sets. The significant negative influence of the
variable log_absolute_discount in the data sets 1, 3 and 4 on the LF is contradictory to the
hypothesis. Since a higher absolute discount is very likely to result in higher promotional demand,
this result requires further investigation. When this variable is the only dependent variable in the
model the impact becomes positive with a Beta of 0.480 and a significance level of 0.000 (Appendix
6). Hence, correlation effects with other dependent variables are responsible for the negative effect
in the final model. The highest correlation in the correlation matrix (Appendix 7) of 0.759 between
the variable percentual_discount and log_absolute_discount is likely to be responsible for the
negative effect in the full model5. Therefore, the variables percentual discount and
log_absolute_discount should not be included in the same model. These results are consistent with
the results of Van Loo (2004) and Van den Heuvel (2006), where the absolute discount had no
impact or a negative impact on the promotional sales. However, so far the absolute discount for a
promotion is calculated per product. Another option is to calculate the absolute discount per offer,
since a consumer is probable sensitive for the total discount received on an offer. Appendix 6 depicts
individual linear regression analysis where the percentual discount, absolute discount per offer and
absolute discount per product6 are compared. Remarkably, both absolute discount measures result
in a higher model fit and have a higher standardized Beta value. However, when the absolute
discount per offer is included in the full model the effect reverses and becomes small (Appendix 6).
Furthermore, a threshold effect could occur at the absolute discount per offer, i.e. consumers are
only willing to especially go to a retailer for a promotion if the total discount per offer received is
high enough. Appendix 6 tests this effect as well for SPO promotions and all for all promotion
mechanisms together. However, no threshold effect is discovered. Altogether, the percentual
discount might be a better predictor because both absolute discount variables correlate too much
with other variables. This is reflected by the last table in 6 where the full models are fitted with the
5 All the other correlation heights in the correlation matrix in appendix 5 are below 0.8 as well (according to Field (2005) a
correlation of 0.8 or higher indicates a multicolinearity problem). 6 No transformation is applied on the absolute discount per offer and the absolute discount per product to simplify the
comparison, because the log transformation does only improve the results slightly.
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 32
three different discount variables. The model with percentual discount clearly has a higher model fit
than the other two models (0.689 against 0.654 and 0.651). Concluding, in a full model the
percentual discount is a better predictor than the absolute discount per offer or per product and
both predictors do not function together in a model because the Beta coefficient of the absolute
discount variable turns negative, which decreases the understandability of the model (H6 rejected
and H7 confirmed).
For the next hypothesis, the different promotion mechanisms, the result is less conclusive. The
promotion mechanism Four_or_five_for_X functioned as the baseline variable and has the most
positive impact on the promotional sales. SPO has a very similar result as Four_or_five_for_X and
the mechanisms Two_for_X and Three_for_X have the most negative result. For the variables
Premiaat and Free_product no effect is found which could indicate that their effect size is similar as
the baseline group (four_or_five_for_X) or that their effect is insignificant. One would expect that
promotions with a price off should sell better than promotions which offer a premiaat or (unrelated)
Free_product. It could be that this effect is already inherited in the variable percentual_discount
since both free_product and premiaat have no percentual discount. This can be tested by running a
regression analysis on all promotions of 2009 where the promo mechanism variables are the only
included independent variables. The results for this analysis are depicted in Appendix 8, where the
variable four_or_five_for_X is maintained as the baseline variable. And indeed this confirms the
hypothesis that the promo mechanisms Free_product and Premiaat have the most negative/least
positive impact on promotional sales. However a SPO promotion still sells better than a two_for_X or
three_for_X promotion, which was not expected in advance (H8 partly confirmed). The result of
the variable number_of_products_in_promotions corresponds with the hypothesis that more
products in the same promotions negatively affect the promotional sales (H9 confirmed).
Regarding the retailer variables, similar promotions sell better at the C1000, average at Albert Heijn
and lower at Plus and Kruidvat. However, the retailer dummy variables are not significant for all data
sets. C1000 and Plus have a significant effect in four of the five data sets; Kruidvat has a strong
significant effect in all Non-Food data sets. No clear hypothesis was drawn on forehand for this
variable(s) (H10 not tested). Another variable which relates directly to the retailer is the number
of selling points at which a promotion is sold. If that number of selling points in a promotion is
higher than the usual number of selling points than the promotional sales is higher. The effect of the
extra number of selling points is very strong (H11 confirmed).
Next, the effect of the brand variables will be discussed. Both the percentage_of_repeat_buyers and
the promotion_pressure of a SKU have almost no impact on the promotional sales. An explanation
might be that the measures are not directly related to a promotion; therefore, clear effects decrease.
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 33
(H12 and H13 rejected). The LF_of_former_promotions does have a very strong positive effect
on the promotional sales. Meaning that when a SKU had high promotional sales in the past, it is
more likely to have high promotional sales in the future (H14 confirmed). For the variable
market_penetration no effect is found, meaning that the promotional sales is not affected by the
penetration a product has in the market (H15 rejected).
The variables preservability, size_of_product and frequency_of_purchase are inherited in the
research to describe the susceptibility of stockpiling. Products with a longer preservability, a smaller
size and a lower frequency of purchase are thought to be more susceptible for stockpiling and to
have higher promotional sales. Indeed a longer preservability and a lower frequency of purchase
result in higher promotional sales, especially in the food categories. This might be caused by a more
frequent shopping pattern for food categories, which makes the variable frequency of purchase
more important. Also, the fact that the preservability is of less importance in the HPC categories
explains the lack of effect in the HPC data set (H16 and H18 confirmed). The size of a product
positively affects the promotional sales. This is not in line with the hypothesis and could be the result
of the higher value of large products, which concurs with the absolute discount of a product (H17
rejected).
For the different categories (Homecare, Personalcare, Savoury & Dressings, Ice & Beverages and
Spreads & Cooking) no conclusive results are found over the different data sets for the direction and
magnitude of the categories. Only for the categories Ice_and_beverages and SCC_and_vitality_shots
a medium effect size is found, where Ice_and_beverages has a negative impact on promotional sales
and SCC_and_vitality_shots has a positive impact on promotional sales. Again as with the retailer it
was unclear in advance which effects should be expected (H19 not tested).
For the variables winter_products_temp and summer_products_temp the effect of the temperature
is tested on temperature sensitive products. Temperature is expected to have a positive effect on
summer products since sales is expected to be higher at higher temperatures and a negative effect
on winter products, since sales is expected to be higher at lower temperatures. However, no effects
are found for both variables. A reason could be that seasonality effects are already taken into
account and the extra temperature differences per week are not significant enough to be found. A
more detailed research on temperatures for temperature sensitive products would most likely find an
effect. However, because of the inclusion of all products, temperature does not produce a better
forecast (H20 and H21 rejected).
7.4 Validation full model
The results in the previous paragraph sketched the performance of the full model fitted on the
promotional datasets of 2009. To test the robustness of the model, the promotions of 2010 are
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 34
forecasted with the same variables and coefficients as in the model of 2009. Continuously, the
sample size and number of predictors in the validation period are equal to the calibration period. The
results of this robustness check are depicted in Table 7-4. The R-square7 of the data sets is slightly
higher than the R-square in the calibration period. Hence, it can be concluded that the model fitted
on the calibration data sets is robust and generalizable for other data periods. Furthermore, the data
sets without the Magnum products have a higher R-square and a lower MAPE (in line with H22).
And the more specific HPC and Food data sets do not generate in both data sets (H23 not
confirmed).
validation period (Q1 2010)
All
(1)
Food
(2)
HPC
(3)
All w/o
Magnum (4)
Food w/o
Magnum (5)
sample size 246 94 152 243 91
number of predictors 21 16 14 19 18
R-square 0.676 0.574 0.742 0.713 0.711
MAPE (actuals) 33.9% 34.9% 31.8% 31.4% 28.8%
MAPE (Unilever) 34.6% 42.3% 29.1% 31.6% 32.9%
Table 7-4: Model summary validation period for the full model
8 Generalizability of model results
In this chapter the generalizability of the model results will be discussed. First, the generalizability of
the sample size taken within Unilever will be discussed. Second, the research is compared with other
research in the field. The goal of this chapter is to check if the results are generalizable within
Unilever and if the results are consistent with other research in the field. If not, further investigation
will be done.
8.1 Generalizability of sample size
The sample size within Unilever is defined on the dimensions retailer, time, region and products. Of
these the choice for region and time are assumed not to disturb the sample size, since the region is
the whole of the Netherlands and the time horizon is longer than 1 year. The retailers in the sample
size are all among the larger retailers in the Netherlands. Furthermore, the retailers included in the
sample are the most important retailers for Unilever in terms of volume and thus also for the
Unilever wide forecast accuracy. Because of their high sales, the impact on safety stock levels of
7 No adjusted R-square is reported for the validation period, because the adjuste R-square only makes sense in the
calibration period.
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 35
Unilever is higher than that of smaller retailer. Hence, it is concluded that the retailers in this
research form a solid representative base for the sample size. Lastly, the selection of SKU’s included
in the sample size is analyzed. In paragraph 3.4 the criteria for including SKU’s are stated. The
sample size selection has been taken over all different product categories of Unilever. However, it
would still be possible that the sample size selection is not representative for all products of
Unilever. Especially the 3rd criteria in paragraph 3.4, that more high volume SKU’s should be
included, could cause an unrepresentative sample size. One way of checking the effect of this
assumption is to analyze the sample size on ABC classification. The ABC classification is a method
used within Unilever to rank SKU’s on their importance. In this classification A SKU’s are high
volume, high turnover and high gross profit SKU’s, and C SKU’s are low volume, low turnover and
low gross profit SKU’s8. And the ABC classification is not made over all products of Unilever at once,
but over the five different categories named in paragraph 1.2. Figure 8-1 shows the normal
deviation within Unilever and the deviation in the sample size. Within the sample size the A SKU’s
are overrepresented, the B SKU’s and the C SKU’s are underrepresented.
Figure 8-1: ABC partition for all products of Unilever and for the sample size (based on volume)
The next step is to analyze what the effect of this deviation of the normal situation is on the
performance. Figure 8-2 depicts the full model MAPE values of data set 4 (all promotions without
Magnum products) for the A, B and C SKU’s. The C SKU’s perform the worst, the A SKU’s are in the
middle and the B SKU’s perform best. One would expect that the MAPE values decrease for A SKU’s
because of the higher sales volumes of these SKU’s. Normally, higher volumes should result in a
decrease of variance. Table 8-1 shows the average and the variance of the LF’s for the SKU
classification. Interestingly, the variance for A SKU’s is higher than the variance of the other SKU’s,
which explains the difference in MAPE values. The difference in variance is very large meaning that
8 The ABC classification is based on these three criteria. However, the sales departments have the final call over the ABC
classification.
A - products
20%
B - products
60%
C - products
20%
ABC partition Unilever (based on volume)
A - products
51%B - products
42%
C - products
7%
ABC partition sample size (based on volume)
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 36
A SKU’s contain a lot more variance than B or C SKU’s. A closer look to the data suggest that the
very large LF’s of a SKU have a large contribution to the total variance of that SKU. Table 8-1 indeed
depicts that A SKU’s contain more very high LF’s (20 or higher) than B and C SKU’s. This could be
caused by the fact that A SKU’s are more often severely promoted and the forecasting model might
not be able to adequately forecast such heavy promotions. Another remarkable issue in Figure 8-2 is
that the MAPE (actuals) value is higher for A SKU’s than the MAPE (Unilever) value. At the C SKU’s
this is the other way around. This arises from the fact that in the calculation of the MAPE (actuals)
overforecasting is heavier punished and in the calculation of the MAPE (Unilever) underforecasting is
punished more severely. Meaning that, A SKU’s tend to be overforecasted and C SKU’s tend to be
underforecasted in the model.
Figure 8-2: Comparison of MAPE values data set 4 over the ABC classification
SKU type Average Variance Average number of
promotions per SKU Number of promotions with a LF higher than
20 per SKU
A 7.51 34.95 19.31 0.938
B 6.37 12.87 13.81 0.269
C 6.15 16.82 12.80 0.200
Table 8-1: The average, variance, number of promotions and number of high LF’s on the ABC classification
Concluding, the sample size does deviate from the total Unilever product portfolio on the ABC
classification. However, this has no clear implication on the performance and generalizability of the
model. Furthermore, it has been reasoned that the choice of time horizon, region and retailer are
done in such a way that the sample size is generalizable. One other aspect which could disturb the
sample size is the exclusion of SKU’s which are sold less than a year. Newer products tend to be
more difficult to forecast, because of the lack of stable base line sales and the lack of historical
0.0%
5.0%
10.0%
15.0%
20.0%
25.0%
30.0%
35.0%
40.0%
All A B C
MAPE (actuals)
MAPE (Unilever)
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 37
comparable promotions. This will hold for the model as well as the current promotion forecasting
process of Unilever. It is difficult to determine what the impact is on new products. This research is
focussed on more stable products, since the effect size of dependent variables is easier determined
for these promotions.
8.2 Comparison with other research in the field
In van der Poel (2010a) the available research on promotion forecasting was split in two parts. The
first paragraph was theoretical research papers and the second part was about more practical
master theses. To judge if the results of this research are comparable, on what aspects the research
differs and what implications these differences have for the result of the forecasting model, a
comparison will be made in this paragraph. Both the theoretical research papers and the practical
master theses will be included in this comparison. The advantage of the research papers is that the
approach is more scientific and the advantage of the master theses is that the model and model
performance have been described more extensively. The following research will be included in the
comparison:
• Cooper et al: PromoCast ™: A New Forecasting Method for Promotion Planning.
• Wittink et al: SCAN*PRO: the estimation, validation and use of promotional effects based on
scanner data (internal paper).
• Van Loo: Out-of-Stock reductie van actieartikelen, Model voor vraagvoorspelling en
logistieke aansturing van actieartikelen bij Schuitema/C1000.
• Van den Heuvel: Action products at Jan Linders Supermarkets.
Table 8-2 makes a comparison between the different research papers on promotion forecasting. All
four papers have been performed from a retailer point of view. Furthermore, the SCAN*PRO and
Promocast models are directed at the store level of a retailer instead of the supply chain level. All
methods use linear regression. Regarding the performance of the models, the paper of the
Promocast model does not contain any comparable performance measures, since the authors
measure the number of case packs missed. For the other models the performance measures differ
considerably. The adjusted R-square of the models of Van Loo en Van den Heuvel is similar, while
the adjusted R-square of this research is substantially higher. Regarding the MAPE, the model of Van
der Poel and Van Loo perform similar. However, the MAPE calculation of Van Loo is not based on the
absolute sales number but on a transformation of the LF. Since this transformation brings the values
of the dependent variable closer together, this measure understates the real MAPE values (based on
absolute sales).
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 38
Van der
Poel
SCAN*PRO
- model
Promo-
cast
Van Loo Van den
Heuvel
Year 2010 1988 1999 2006 2009
Point of view Manufacturer Retailer Retailer Retailer Retailer
Commercial use No Yes Yes No No
Aggregation level Supply chain Store level Store
level
Supply
chain
Supply
chain
Method Linear
regression
Linear
regression
Linear
regression
Linear
regression
Linear
regression
Ln LF as dependent var. Yes Yes Yes No No
Sample size 1238 20801 n.a. 1556 n.a.
Average LF 6.08 n.a. n.a. 9.04 4.59
Variance 26.95 n.a. n.a. 29.93 7.84
Standard deviation 5.19 n.a. n.a. 5.47 2.80
Minimum 1.00 n.a. n.a. 1.13 1.00
Maximum 49.38 n.a. n.a. 34.00 14.28
Adjusted R-square 0.691 a 0.507 b n.a. 0.45 0.44
MAPE validation period (full model) 31.3% 37.1% n.a. 31.1%c n.a.
a : The adjusted R-square of the full model of data set 4 is taken here. b : MAPE value of SCAN*PRO model in research Van Loo (2006). c : The MAPE calculation in the research of Van Loo seems to be based on the ln of the LF. This calculation understates the MAPE based on absolute promotional demand.
Table 8-2: Comparison research on promotions forecasting
All models in Table 8-2 differ substantially in performance9. To investigate where this difference in
performance originates from Table 8-3 shows an overview of the most important variables included
in the research. The current research is taken as the frame of reference. The SCAN*PRO-model is a
concise model, where only a few important variables are taken into account. The Promocast model is
by far the most elaborate model with 67 independent variables. This model makes extensive use of
LF’s of former promotions and since the model is directed at the store level, the promotion database
is a lot larger. The model of Van Loo does not include the important variables display and folder,
which are included in all other models and are among the most important variables. The model of
Van den Heuvel includes the most important variables and contains some interesting research on the
effect of other actions in the same product category and the effect of Out of Stocks.
9 The performance of the Promocast model is not available for the R-square and MAPE measures. The paper on that model
only states the performance in case pack size difference on retailer store level.
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 39
The adjusted R-square model performance is known for four of the five models. The performance of
the model build in this research compared to the other models is considerably higher. Here, the
underlying factors for this difference will be discussed. The adjusted R-square of the model in this
research might be higher, because Van Loo did not include the critical variables display and folder.
Furthermore, Van Loo, the SCAN*PRO-model and Van den Heuvel did not include all of the following
variables: promo mechanism, the average LF of former promotions, the number of products in
promotion, the growth of the number of selling points, the size of a product, preservability and TV-
support. Finally, Van Loo did not transform the LF as dependent variable and thus the dependent
variable is not normally distributed. This has a very negative impact on the performance of the
model. Altogether, the model of Promocast is the most sophisticated model regarding the included
variables. However, the results of this model cannot be compared and the model is directed at the
store level of a retailer.
Van der
Poel
SCAN
*PRO model
Promo-
cast
Van Loo Van den
Heuvel
Retailer x n.a. n.a. n.a. n.a.
Product category x x x
LF former promotions SKU x x
Display x x x x
Folder x x x x
Promo-length x x All 1 week All 1 week
Promo mechanism x x x
tv-support x x
Number of products in promotion x
Growth # of selling points x n.a. n.a.
Percentual discount x x x x x
Size of product x
Preservability x x
number of actions in same product group No data x
More specific data on display location No data x
More specific data on size and place folder advertisement
No data x
LF former promotions SKU with matching advertisement and display
Not enough data
x
n.a.: Not applicable in this model because model is build at a single retailer or model is build on store level
Table 8-3: Comparison of the variables included in the different promotion forecasting research
Lastly, Van Loo (2006) used a different dependent variable than the other research. In his research
Van Loo fitted a log normal distribution on the LF’s and then used the cumulative lognormal
distribution of each LF as dependent variable (P(LF)). In the research Van Loo concluded that this
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 40
measure gave superior results against the LF of a promotion; however, no comparison was made
with another widely used dependent variable in literature, the ln transformation of a LF. Appendix 9
shows the results if the P(LF) is used in the model of this research instead of the ln(LF). The new
dependent variable is tested on the full model on data set 4. The results indicate that indeed the
P(LF) gives superior results against the LF. But the P(LF) has a lower model fit than the ln(LF). An
explanation might be that the ln(LF) meets the requirements of a normal distribution better than the
P(LF) as shown in appendix 9.
Concluding, this chapter provided insight in the generalizability of the model results by checking the
assumptions underlying linear regression, analyzing the generalizability of the sample size and
comparing the research with other relevant research in the field. The sample size is regarded to be
generalizable over the other Unilever SKU’s. Only the introduction of a new SKU will cause deviation
from the current sample size and quite likely lower the performance. But, forecasting new SKU’s has
always been difficult. Regarding the comparison against other research, the results of the model
constructed in this paper are quite high. The inclusion of important variables in this research, which
were not included in the comparable research, is very likely to be responsible for the good model fit.
Conclusion part 3: In this part the results for the full model were depicted. Both the model fit and
forecast accuracy values are quite high for the full model. However, the full model on consumer
demand level only provides the first part of the total picture. Because of the functional requirements
stated in paragraph 2.4, the full model needs to be adapted to the retailer demand level. This will be
done in the next part, so the model becomes useful in practice.
Improving the promotion
Part 4: Model adaptation
The results of the full model provided det
of the different variables and the performance of
the different data sets. In order to translate
these results into a model that can be used in
practice, this part decreases the number of
variables in the models. Then the
evaluated based on their model results. This gives
insight in the performance of the reduced model
against the full model and thus in the practical usefulness of the forecasting model.
There are three main reasons to adapt the full model o
1. To increase the usability
2. To correct for data availability
3. Adapt the model of consumer demand to retailer orders
The first adaptation will result in a model with 5 to 10 variables, since this nu
useful in practice (interviews Unilever). The number of variables has to be limited because an
employee of Unilever should be able to quickly work with the model. The variables will be selected
on their effect size and direction.
first adaptation on data availability. The variables which are normally not known within Unilever will
be deleted from the set of variables. Hence, adaptation two includes the same variables
adaptation one without the variables with low data availability. In the third adaptation, the consumer
demand will be adjusted to the retailer orders. The consumer demand serves as the basis for the
discussion with the retailer and for the On Shelf Ava
orders form the real demand that should be met within Unilever.
The adapted models will be tested on data sets 3, 4 and 5, because the disturbing effect of the
Magnum products was too large to include t
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 41
Part 4: Model adaptation
The results of the full model provided detailed insights in the
of the different variables and the performance of
the different data sets. In order to translate
these results into a model that can be used in
practice, this part decreases the number of
variables in the models. Then the models will be
evaluated based on their model results. This gives
insight in the performance of the reduced model
against the full model and thus in the practical usefulness of the forecasting model.
There are three main reasons to adapt the full model of the previous part:
usability of the model in practice.
data availability in practice.
consumer demand to retailer orders.
The first adaptation will result in a model with 5 to 10 variables, since this number of variables is still
useful in practice (interviews Unilever). The number of variables has to be limited because an
employee of Unilever should be able to quickly work with the model. The variables will be selected
on their effect size and direction. The second adaptation will inspect the variables included in the
first adaptation on data availability. The variables which are normally not known within Unilever will
be deleted from the set of variables. Hence, adaptation two includes the same variables
adaptation one without the variables with low data availability. In the third adaptation, the consumer
demand will be adjusted to the retailer orders. The consumer demand serves as the basis for the
discussion with the retailer and for the On Shelf Availability of a product, but in the end the retailer
orders form the real demand that should be met within Unilever.
The adapted models will be tested on data sets 3, 4 and 5, because the disturbing effect of the
Magnum products was too large to include these products in the further analysis.
forecasting accuracy at Unilever Netherlands
effect size
mber of variables is still
useful in practice (interviews Unilever). The number of variables has to be limited because an
employee of Unilever should be able to quickly work with the model. The variables will be selected
The second adaptation will inspect the variables included in the
first adaptation on data availability. The variables which are normally not known within Unilever will
be deleted from the set of variables. Hence, adaptation two includes the same variables of
adaptation one without the variables with low data availability. In the third adaptation, the consumer
demand will be adjusted to the retailer orders. The consumer demand serves as the basis for the
ilability of a product, but in the end the retailer
The adapted models will be tested on data sets 3, 4 and 5, because the disturbing effect of the
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 42
9 Adaptations to increase the usability and check for data availability
9.1 Adaptation 1: Increase the usability by reducing the number of variables
To reduce the number of variables in the full model some criteria are needed. The goal is to reduce
the number of variables to less than ten variables and analyze the impact of the reduction of
variables on the performance of the model. The criteria to select the variables are:
1. A strong effect in the three data sets for the full model, i.e. an average standardized Beta of
0.150 or higher over the three data sets.
2. A persistent effect in the three data sets for the full model, i.e. the standardized Beta does
not have an opposing direction in the three data sets.
Analysis of the standardized Beta coefficients of the full model in Table 7-3 leaves nine variables
which meet these criteria: Display, Folder, Promo_length, Percentual_discount, Two_for_X,
Three_for_X, log_growth_number_selling_points, ln_LF_former_promotions_EAN and Kruidvat.
Since the SPO is significant as well and falls under the same variable (promo mechanism) as
Two_for_X and Three_for_X this variable is included as well. This argument also holds for the
retailers C1000 and Plus, which fall under the same variable as Kruidvat, namely retailer. Table 9-1
shows the model results of the calibration and validation period. The number of variables is larger
than the functional requirement of 10 variables; but, the variables SPO, Two_for_X and Three_for_X
as well as the variables Kruidvat, C1000 and Plus are dummy variables for the variable promo
mechanism and retailer and can be regarded as one variable in practice, since an employee only
needs to complete one data field. Hence, the number of variables comes to 8.
Table 9-3: Summary of the reduced model corrected for data availability (adaptation 2)
Again the adjusted R-square decreases, since fewer variables are used to fit the data. As a result the
MAPE values in the calibration period decrease as well. The MAPE values in the validation period
show similar results. The different data sets have a very similar model fit in the calibration period
with the Food model performing slightly better. However, in the validation period the Food model
performs quite a lot better than the HPC model for the MAPE (actuals), but not for the MAPE
(Unilever). When the MAPE (actuals) value is higher than the MAPE (Unilever) value this indicates
underforecasting. When the MAPE (actuals) value is lower than the MAPE (Unilever) value this
indicates overforecasting. In this case the promotions for the HPC data set are slightly
underforecasted and the promotions for the Food data set are slightly overforecasted. Overall, the 10 The variables excluded in some of the data sets are dummy variables which fall under retailer or promo mechanism.
Therefore, the number of variables can be 6 for each data set.
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 45
performance of the model decreased considerably because of the exclusion of the variables with
limited data availability (Display and Log_growth_number_of_selling_points). And the impact is more
severe on the HPC data set. The B-coefficients and Standardized Beta coefficients of the data sets
are depicted in Table 9-4. The variables Folder, Promo_length, Percentual_discount and Kruidvat
have the largest influence in the model (standardized Beta of 0.150 or higher in data set 4).
n.e. = no effect on the promotional sales a The baseline group for the different product categories is the product group “Four_or_five_for” b The baseline group for the different retailers is the retailer “Albert Heijn” c The baseline group for the different product groups is the product group “Homecare”
Table 12-1: Overview of the effect size and direction of the variables on the promotional sales
The ideal model shows that Unilever has the ability to forecast consumer demand. With the right
information Unilever is able to forecast the consumer demand at least as good as a retailer. Hence,
with this capability Unilever is able to take the lead in establishing a collaboration with retailers and
increasing the forecast accuracy. However, because of the practical requirements of a forecasting
model the ideal model formulated cannot be used in practice within Unilever. First, the model should
have a high ease of use, second the variables used should have data availability and third the
retailer orders need to be forecasted. Hence, some adaptations are needed on the full model, which
are discussed hereafter.
12.2 Adaptations needed on ideal model
To increase the usability of the forecasting model the most important variables are included in an
adapted model. The model fit of this model with a limited number of variables is still surprisingly
high and almost equal to the model fit of the full model. However, not all variables have data
availability at Unilever, since Unilever as a manufacturer is dependent on the retailers for information
of upcoming promotions. For two variables in the limited model Unilever has no data availability.
These are the percentage of shops with a second placement and the extra number of shops where
the product is sold in promotion. To analyze what the effect of the lack of data is on the forecast
accuracy of the model a new model without these variables is tested. The model fit decreases to an
adjusted R-square of around 0.500, indicating that the exclusion of the two variables substantially
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 61
worsens the performance of the forecasting model.
Moreover, Unilever needs to forecast retailer orders instead of consumer demand. Therefore, the
model results for the consumer demand are adapted to retailer orders. The retailers included in the
research order on average between 39% and 85% more than is sold during promotion. The
forecasts for the consumer demand are raised with this difference. The model performance
decreases substantially because of the extra variance in the retailer orders. The R-square for the
HPC data set has decreased to 0.138 in the calibration period, meaning that the predictive power of
the model is very low. For the Food data set the R-square is 0.411 in the calibration period. So, the
variability in the retailer orders is a lot higher for the HPC products than Food products. Forecasting
retailer orders for HPC products seems to have little to no benefit, Food orders can be forecasted
with a higher accuracy. The difference is partly caused by the height of the sales of a promotion.
Because Food promotions sell 4 to 5 times more than HPC promotions the variability in de retailer
orders decreases. This reasoning also holds for the A, B and C categorisation where the A SKU’s are
the more important high volume products. And indeed A SKU’s have a substantial higher forecasting
accuracy than C SKU’s.
The adaptations indicated that the reduction in the number of variables in the model does not lead
to a lower model performance. But when two of the most important variables are excluded, because
of a lack of data at Unilever, the model performance decreases substantially. Furthermore, the
transition from consumer demand to retailer orders leads to a high loss of predictive power. To
overcome these problems further steps need to be taken.
12.3 Future steps to increase the forecast accuracy
Since a direct forecast of the retailer orders turned out to be inaccurate and not all variables had
data availability, future steps need to be taken to deal with the problems which diminish the forecast
accuracy (implementation plan in paragraph 11.2). The first implementation step focuses on the
enhancement of the data usage within Unilever. Quite some promotion data is available somewhere
in the organization; however, the available data of historical and upcoming promotions should be
recorded more centrally and accessible. Then the data can actually be used by the logistic employee
to forecast promotions. The second step is to ensure that the important information of upcoming
promotions is provided by the retailers in advance. Retailers are afraid to do so because of the
sensitivity of the data. Unilever should win their trust to get hold of the important promotion data.
The third step should bring insight in the factors causing the gap between retailer orders and
consumer demand. Because of the bullwhip effect a lot of extra variance is added to the retailer
orders, especially for HPC products. Unilever should focus on understanding the source of the extra
variance together with the retailer. The fourth step has to bring the insights of the third step into
action and take this insights one step further. During this process the alignment with the retailer
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 62
becomes more important as the collaboration becomes more intensive. In the end this will result in
a supply chain forecasting model where both the retailer and Unilever make use of and new
technologies like RFID can be used to evolve the forecasting model.
The first two implementation steps will solve the poor database usage within Unilever, the main
scope of this research paper as stated in paragraph 2.2. The problem areas customer (retailer) team
deviation and retailer dependency will be influenced by the implementation steps as well. Because of
a standard way of working is proposed over all retailers, the promotion forecasting process will
become more alike for the different retailers. Furthermore, retailers who are not as far as others in
the implementation steps can learn from the forecasting process of Unilever for the more developed
retailers. Regarding the retailer dependency, it has become clearer which variables are needed from
a retailer to accurately forecast a promotion. And the implementation steps will convert the
dependency on a retailer to collaboration with a retailer.
Altogether, this research showed that if the right information is available Unilever is very well
capable of accurately predicting the consumer demand. Unilever has an advantage over the retailers
because of their larger data pool of promotions over all retailers which can be used to forecast
upcoming promotions. However, forecasting retailer orders has turned out to be far more difficult
than consumer demand, especially for HPC products. The bullwhip effect leads to a substantial
deviation between retailer orders and consumer demand. As a result, in order to be able to
accurately forecast retailer orders, the disturbing factors behind the bullwhip effect should be
analyzed. In order to successfully analyze these factors close collaboration with the retailer is
needed. When the disturbing factors are successfully analyzed, a promotion forecasting model which
forecasts the consumer demand and corrects for the disturbing factors should be formulated and
employed together with the retailer. Close collaboration and information sharing is needed, where in
the end Unilever and the retailer together use one forecasting approach and the retailer orders can
be predicted accurately.
12.4 Contribution to literature
In paragraph 1.5 three gaps in the literature were discussed. The gaps are (1) the choice of the
dependent variable to predict the promotional sales, (2) the development of a forecasting model for
a manufacturer and (3) if it is an advantage or disadvantage to be a manufacturer.
The first gap exists because there is no clarity in the promotion forecasting literature which measure
should be used as dependent variable. Different research uses different dependent variables namely,
the LF of the promotional sales, the ln of the LF and the cumulative lognormal distribution of the LF
(P(LF)). This research concluded that the LF of the promotional sales is not an adequate measure
because of clear signs of non normality. Both other measures correct for this non normality, only the
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 63
cumulative lognormal distribution does that in a lesser extent. The performance of a promotion
forecasting model substantially improves for both the P(LF) and ln LF measure, where the model fit
of the ln LF was slightly higher. Concluding the natural logarithm of the LF matches the normal
distribution best and has the highest model fit.
Regarding the second gap, the main difference between a retailer and manufacturer is that a retailer
needs to forecast the demand of his shoppers (consumer demand) and a manufacturer has to
forecast the orders placed by his customers (retailer orders). This research both developed a model
which directly predicts the retailer orders and a model which predicts the consumer demand after
which this prediction is adapted to a forecast for the retailer orders. Retailer orders do in fact differ
remarkably from consumer demand, between the 39% and 86% for the retailers in this research:
therefore, a model which predicts consumer demand cannot be used at a manufacturer without an
adaptation.
Third, it is not clear if being a manufacturer is an advantage or disadvantage in producing an
accurate promotion forecast. This research built a model which has a high promotion forecast
accuracy on consumer demand level. The research made use of promotional data of SKU’s for
multiple retailers. The fact that the model performance based on this data is quite good, indicates
that the larger promotional database can act as an advantage for a manufacturer. However, not all
variables for the consumer demand forecasting model are available for upcoming promotions at
Unilever, because retailers are not willing to share some of the important promotion characteristics
with Unilever. This is a major disadvantage of which this research indicated that the model
performance evidently drops. The second disadvantage of being a manufacturer is that a
manufacturer has to deliver retailer orders instead of consumer demand. This research showed that
the variability of retailer orders is higher than that of consumer demand and that the model
performance decreases substantially when forecasting retailer orders. Therefore, overall it is
concluded that a manufacturer has a disadvantage compared to a retailer.
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 64
Improving the promotion forecasting accuracy at Unilever Netherlands
Page 65
References
Anderson, E., & Coughlan, A. T. (2002). Channel Management: Structure, Governance and
Relationship Management. In B. A. Weitz & R. Wensley (Eds.), Handbook of Marketing (pg.
223-247). London: Sage.
Barratt, M. and Oliveira, A. (2001). Exploring the experience of collaborative planning initiatives.
International Journal of Physical Distribution & Logistics Management, Vol. 31, No. 4, pg. 266-89.
Blattberg, R.C., Briesch, R., Fox, E.J. (1995). How Promotions Work. Marketing Science, Vol. 14, No.
3, pg. 122-132.
Buckers, J. (2010). The ordering process of dry food groceries under promotion: a study of order
commitment timing and ordering methods. Master Thesis. Eindhoven University of Technology,
Eindhoven.
Cooper, L.G., Baron, P., Levy, W., Swisher, M., Gogos, P. (1999). PromoCast ™: A New Forecasting