Master’s thesis
Reducing out-of-stock and
food waste with segmented
promotions:
Exploring a new promotional strategy
12-11-2017
By G.I.S. van Grinsven
Dual Degree in Operations Management University of Groningen
Newcastle University
Abstract
We explore a new promotional strategy where the promotional peak in demand is spread over multiple periods and customer segments. By altering promotional variables between subsequent
periods, a retailer can control the promotional volume. We perform an analysis of promotional demand to estimate the effect of changes in the promotion variables. We simulate the both segmented and regular promotions in similar environments and compare them on out-of-stock and
food waste. We find that segmented promotion can reduce out-of-stock and food waste by 34 percent across different product categories. These finding show the potential of segmented
1
Table of Content
1. Introduction
2
2. Literature review
5
2.1.
Promotional strategies and promotional lift
5
2.2.
Deal effect curve
8
2.3.
Out-of-Stock and food waste
9
2.4.
Forecasting methods
12
3. Problem setting
14
3.1.
Case study
14
3.2.
Segmented promotions
15
3.3.
Incorporating promotional impact of variables
17
3.4.
Research Question
17
4. Methodology
18
4.1.
Regression analysis
18
4.2.
Case analysis
22
4.3.
Simulation
23
5. Results
26
5.1.
Results of regression analysis
26
5.2.
Results of case analysis
34
5.3.
Results of simulation
37
6. Discussion
40
7. Conclusion
42
8. Future research
43
2
1.
Introduction
Retail markets have been characterized by an increasing number of promotions in an effort to increase market share. Strong competition among retailers resulted in a sharp increase in the number, frequency and depth of promotion especially in grocery retailing (Ettouzani, Yates, & Mena, 2012; IGD Supply Chain Analysis, 2007; McKinnon et al., 2007; Mou, Robb, & Dehoratius, 2017). To attract customers retailers are using a variety of promotion techniques such as price promotions, deep discount deals, in-store displays and other feature advertising (store flyers). The Promotion Marketing Association estimated total promotion spending across all product categories of $429 billion (3,6 % of US GDP) in 2004 in the USA (Ailawadi, Beauchamp, Donthu, Gauri, & Shankar, 2009).
Meanwhile, there is an increasing awareness among retailers and manufactures of the losses incurred due high Out-of-stock rates and food waste. (Corsten & Gruen, 2003; McKinnon et al., 2007; van Woensel et al., 2007; Trautrims, Grant, Fernie, & Harrison, 2009). Ettouzani et al. (2012) conducted a research across seven major food retailers in UK, conducting
interviews with 110 practitioners. Their findings show that retailers struggle to maintain low Out-of-Stock levels without food waste during promotions. A study by ECR (2003) has revealed that the world-wide average OOS rate in grocery is 8.2%, ranging from 5% in
canned food to 18% in fresh food. During promotions the OOS rates increase significantly (up to 47%) as retailers struggle dealing with peaks and volatility in demand (Ettouzani et al., 2012). Careful planning and accurate forecasting are required to efficiently deal with promotions.
3
reduction do not always outweigh costs incurred for advanced techniques and knowledge (Ali et al., 2009). As a result, retailers are reaching to segmented promotions that aim to minimize the costs of OOS and product waste without investing in costly software and knowledge. Especially in online retailing many new promotional opportunities and strategies are emerging.
This research is focused on a new promotional strategy, segmented promotion, which is enabled by utilizing distinctive characteristics of online retailers and designed to spread the promotional peaks in demand evenly over multiple periods of time and control promotional volumes during these periods. The concept of segmented promotions is as follows: A retailer divides its total population of customers into n segments, where each segment is a
representative sample of the total population. The promotions is active for only one segment (1/n) of the population at a time during each period t. As such after a cycle of t=n periods the promotion has been available to each customer while the promotional peak has been 1/n times the original peak. During the segmented promotion cycle a retailer can alter the promo depth or mechanism to compensate for inaccuracies in demand. Furthermore these segmented promotional cycles can be successive or cycle length can be coordinated with the volume of the good sold.
The aim of this paper is to test the efficiency of segmented promotions in term of control in sales volume and reductions in OOS and food waste compared to regular promotions; moreover this paper aims to explore the characteristics of segmented promotions and
4
This paper is structured as follows. Section 2 will review literature current literature on this topic. Section 3 will elaborate on segmented promotions and corresponding problems. Also the case-study is described in this section . Section 4 will introduce the three-step
5
2.
Literature Review
This section will discuss relevant literature in the fields of promotional strategies and promotional lifts, deal effect curve, OOS and food waste and demand forecasting. This section will also illustrate the motivation for this research.
2.1
Promotional strategies and promotional lift
The focus of this research is on a new promotional strategy; segmented promotions. Here we will elaborate on the existing promotional strategies and the uptake in demand that result from these strategies.
2.1.1 Promotion strategies
Promotions in thefast-moving-consumer-good (FMCG) market can generally be grouped in
two types: trade promotions and consumer promotions. The former refers to marketing activities that occur at retailers in collaboration with manufacturers. The latter refers to promotions initiated at the retailer to attract consumers. Over the past decades there has been extensive research on trade promotions, both analytical and empirical models, yet research has been primarily on trade promotions from a manufactures perspective (Ailawadi et al., 2009). Literature on consumer promotions, from a retailer perspective, is much more scarce. Sales data, but especially data on costs of a retailer, is not publicly available making
6
Segmented promotions originate from price promotion but spread over multiple periods. Gilbert & Jackaria (2002) find that product trial (i.e. trying out a new product) and purchase acceleration (i.e. buying more quantity or shortening the time between repurchases) were the most influential factors to increase purchases among the respondents. Multibuy promotions are basically price promotion that require the customer to buy multiple units (i.e 1+1 free or
2nd half price). Although Gilbert & Jackaria (2002) find that multibuy promotions do not
significantly influence consumer’s buying behavior, the authors do find a significant relation of multibuy promotion on purchase acceleration and brand switching
7
2.1.2 Promotional lift
Articles that are promoted experience a substantial increase in demand. Promotional demand can be up to 90 times the regular non-promotional demand (Ali et al., 2009). The promotional lift is influence by a large set of variables but base volume (i.e. sales during regular period), price discount and type of display are the most significant (Ailawadi, Harlam, César, & Trounce, 2006; Donselaar et al., 2016). Bijmolt, Heerde, & Pieters (2005) perform meta-analysis on the price elasticity of promotional demand. The authors find that a decrease in price by 1 unit results in an uptake in promotional demand of 2.62. This increase in
promotional demand is not endless for every decrease in price. Gupta (1988) was one of the first authors that tried to decompose this promotional lift in demand. Gupta (1988) focused his research on the decomposition of promotion elasticity. Brand switching was believed to be the major driver of the promotional uptake in demand. From there van Heerde, Gupta, & Wittink (2003) moved to the decomposition of the unit sales. The authors find that influence of brand switching is only a fraction of what was previous believed. Gupta (1988) estimated that 80 percent of the of the uptake in demand resulted from brand switching, whereas van Heerde et al. (2003) estimate the effect to be 30-45%.
The components of the promotional lift in demand that benefit the retailer are not similar to the components that benefit the manufacturer (Ailawadi et al., 2006). Uptake in demand that results from store switching is a major benefit for the retailer, yet it does not help much for the manufacturer. On the other hand, the component of the lift in demand that results from brand switching is very beneficial for the manufacturer and less for the retailer (assuming equal margins). Much of this research takes the perspective of the manufacturer (Gupta, 1988; Heerde & Neslin, 2017; van Heerde et al., 2003) and research from the retailers perspective is scarce. Ailawadi et al. (2006) use data from a U.S. drug store chain to estimate the
components of the promotional lift from the retailer’s perspective. On average, 45 percent of the total lift results from incremental sales, 45 percent results from brand switching and 10 percent results from purchase accelerations (i.e. purchase moved forward in time).
2.1.3 The profit impact of the promotional lift
Once the components of the promotional lift were established, Ailawadi et al. (2006)
8
combination with decreased margins are the reason for this loss. Nijs, Dekimpe, Steenkamps, & Hanssens (2001) find similar results in their study across 460 product categories over four years. From all promotions, only 58 percent results in profit for the retailer. In their research, Srinivasan, Pauwels, Hanssens, & Dekimpe (2004) account for cross-category (i.e. switching brand/article in other category) and store-traffic effects for the retailer but still find that profits from promotions are mainly received by manufactures. The research aforementioned,
illustrates the importance of more knowledge on promotion from a retailers perspective. However, the lack of publicly available cost data limits empirical analysis on the costs effects of promotions, while this can be fairly different from the sales effect of promotions.
2.2
Deal effect curve
The effect of an increase in discount percentage is not equal for all discount percentages. This non-linear relation between the percentage discount and the promotional lift is described in literature as the deal-effect curve. The deal-effect curve was introduced by Blattberg, Briesch, & Fox (1995) and focused on perishable articles. The reason is the restriction to stockpile perishable items which results in a saturation level where an additional increase in discount will not increase sales. Little is known on the actual shape of the deal effect curve and only a few papers discuss this (Donselaar et al., 2016; Marshall & Leng, 2002; Martínez-Ruiz, Mollá-Descals, Gómez-Borja, & Rojo-Álvarez, 2006; Van Heerde, Leeflang, & Wittink, 2001).
Donselaar et al. (2016) analyze promotional demand for perishable products at a Dutch
9
2.3
Out-of-Stock and food waste
This section will elaborate on consequences of the promotional uptake in demand on out-of-stock (OOS) and food waste. We will discuss the importance of availability of products, food waste, drivers behind OOS and food waste and customer reactions to OOS.
2.3.1 The importance of availability of products
Over the last decade high service levels have become of vital importance in the fast moving consumer goods (FMCG) and retail sector (Corsten & Gruen, 2003; Mou et al., 2017; Moussaoui, Williams, Hofer, Aloysius, & Waller, 2016; Trautrims et al., 2009). Moreover, availability of products is described as major component of service level in literature (Chuang, Oliva, & Liu, 2016; Corsten & Gruen, 2003; Mou et al., 2017; Moussaoui et al., 2016). A study among American and European grocery shoppers ranked the most important components of customer service (ECR, 2003). Availability came third after shorter queues and more promotions. However, in online retail, queues are less relevant making promotions and the availability of products the key components for service levels in online grocery retailing (Chuang et al., 2016). Trautrims et al. (2009) describe availability of products as the customer service output of a successful supply chain but the authors also note that higher availability does not automatically lead to higher service levels. In their research a trial at Marks & Spencer supermarket showed that increasing availability can raise unexpected problems, such as perceived freshness and timing of store replenishments. This is especially true when dealing with perishable product freshness. Mou et al. (2017) find that perishability can influence customer-perceived utility in two ways. Product quality will decay over time and product attractiveness can also decrease over time. Both issues will increase waste and spoilage.
2.3.2 Food waste
10
supply chain in many studies (Ahumada & Villalobos, 2009; Kaipia et al., 2013; Mena et al., 2011; D. H. Taylor & Fearne, 2009). Authors seem to agree that very fresh product
categories, such as fruit & vegetables, are the main contributor to avoidable food waste. The mismatch originates from poor promotional management, forecasting difficulties, efficiency and availability, poor ordering and a lack of information sharing (D. H. Taylor & Fearne, 2009). The minimalization of this mismatch is therefore crucial to decrease avoidable food waste. Inventories with large buffers to account for demand volatility are no op option because of the perishability of the articles. Fresh food supply chains require flexibility instead of economies of scale (Ahumada & Villalobos, 2009). Donselaar et al. (2016) stress the increasing awareness for food waste and they argue that higher forecast accuracies will decrease food waste and improve the food quality perceived by customers. Segmented promotions attempt to minimize the mismatch by increasing the accuracy of the estimated mismatch and by enabling more flexibility during promotions.
2.3.3 Drivers behind OOS and food waste
One of the root causes for OOS and food waste at retailers is promotional activities (Corsten & Gruen, 2003; J. C. Taylor & Fawcett, 2001). Corsten & Gruen (2003) find OOS rate to be 100 percent higher for articles on promotion versus articles not on promotion. Taylor & Fawcett (2001) compared OOS rates at six retailers comparing promotional periods versus non-promotional periods. There findings show that OOS rates were 240 percent higher during promotional periods. The availability of products is influenced at all stages of the supply chain. However, Ettouzani et al. (2012) conducted an extensive literature review and find that two-thirds to three-quarters of OOS causes originate at store and retailer level.
Moussaoui et al. (2016) systematically review the drivers of availability of products in literature. The authors find a general consensus that most availability drivers are of
11
historic sales have not fully captured historic demand because part of demand could not be supplied. This principle will cause inaccurate predictions of future demand which leads to more unavailability. This vicious circle illustrates the effect of demand autocorrelation. Moreover, demand unpredictability can add another layer of complexity in the availability of products and demand is especially unpredictable during promotional activities (Ehrenthal & Stölzle, 2013; Fernie & Grant, 2008; Moussaoui et al., 2016).
From the operational drivers, demand fluctuation and poor forecasting are significant contributors to OOS and food waste (Corsten & Gruen, 2003; Ettouzani et al., 2012; Moussaoui et al., 2016). Ettouzani et al. (2012) examine availability of products during promotions. In their multiple-case study involving seven major grocery retailers in the UK, the authors find that all retailers suffer from OOS and food waste due to demand fluctuations and poor forecasting. The retailers indicate that causes are demand variability as a result of promotional lift in demand, weather changes, seasonality and different characteristics of products. All retailers indicate a difficulty in forecasting, especially during the introduction of a new product. One can expect that the major grocery retailers in the UK have experienced demand planners
2.3.4 Customer reactions to OOS
There are five basic reaction when customers face unavailability (Corsten & Gruen, 2003; Peinkofer, Esper, Smith, & Williams, 2015):
1. Consumers buy the item at another store (store switching)
2. Consumers delay ordering or purchasing the item (postpone purchase at the same store)
3. Consumers do not purchase the item (a lost sale)
4. Consumers substitute the same brand (different size or type) 5. Consumers substitute another brand (brand switching)
12
a competing brand than US customers. The author do not provide an explanation for this differences other than the conclusion that US customers are more brand loyal. Indirect losses can be the result of a loss in store loyalty as a result of unavailability. Store loyalty has become as important as brand loyalty, if not more, in grocery retail (Trautrims et al., 2009). Interestingly, Peinkofer et al. (2015) researched unavailability during promotion in online retail of consumer electronics and find that customers are actually less dissatified when promoted items are unavailable. These findings are contrary to prior beliefs that customers tend to have higher availability expectations when demand is stimulated by promotions (J. C. Taylor & Fawcett, 2001). Peinkofer et al. (2015) suggest that a promotion in online retail can signal consumers to have lower inventory availability expectations. However, there are several limitations to this research; the authors only tested in cases with large price
promotions and the authors only considered individual stockout occurences. Their research focused on short-term consumer behavour, in the long-term repeated unavailability can lead to loss in consumers and loss in store loyalty.
2.4
Forecasting methods
In promotional forecasting, the widespread industry approach, is to calculate the promotional lift and multiply this with the baseline sales (Cooper, Baron, Levy, Swisher, & Gogos, 1999). The promotional lfit is calculated by dividing the promotional sales by the regular sales in the period prior to the promotion. The amount of historic promotions per article is often limited. Promotional lifts are therefore calculated at a higher aggregation level (i.e. per product
category) to more accurately predict (Huber, Gossmann, & Stuckenschmidt, 2017; Thomassey & Fiordaliso, 2006; Zotteri, Kalchschmidt, & Caniato, 2005). Zotteri, Kalchschmidt, &
Caniato (2005) research the impact of aggregation level on forecasting performance using data from a food retailer. The authors specify three aggregations levels that should be clear when forecasting demand. The time bucket (i.e. the period over which demand is aggregated), the forecasting horizon (i.e. the set of items the demand refers to) and the locations (i.e. the set of stores/customers the demand refers to). The author show that there is not one best
aggregation level as this is dependent on the three factors aforementioned. Zotteri et al. (2005) propose a clustering method where data is aggregated on their degree of similarity rather than on the basis of other features such as region or brand. The authors show that the right
13
complexity of such a clustering approach is very high. It is questionable if similar
improvements can be achieved without the advanced knowledge and software used in the paper.
Ma, Fildes, & Huang (2016) increase forecast accuracy by incorporating intra and inter category sales data in the model. The authors show that forecasting accuracy can improve by 12.6 percent over a set of product categories. 95 percent of this improvement is the result of incorporating intra-category information. Again, it is questionable if the increase in accuracy outweighs the complexity of the method. Their findings are especially interesting for category managers that make the selection of articles to be promoted. The right combination of articles in each product category will enable the highest possible lift. Huang, Fildes, & Soopramanien (2014) propose a similar increase in forecast accuracy by including competitive information (i.e. prices and articles in promotion). They find that indeed the forecast accuracy can improve, yet the authors also stress the abundant amount of information. The selection of relevant competitive information is a key challenge. Besides, competitive information is often not available at time of promotional planning (i.e. several weeks in advance).
Generally we see that improvements in forecast accuracy go hand in hand with increases in complexity. This is confirmed by Ali et al. (2009) who find that, from 30 candidate models, only five models are on the efficiency frontier in terms of accuracy and complexity. The use of regression threes with explicit features yields an increase in accuracy of 65 percent. The complexity of such a model is outweighed by this large increase in accuracy.
14
3.
Problem setting
The literature review has shown that promotional strategies are of vital importance in
retailing. Besides, there is a general consensus in literature that OOS and food waste are two major issues for retailers. Problems in OOS and food waste are getting worse as a result of promotional activity and retailers reach out to complex forecasting techniques, yet the impact of promotions on demand remains difficult to model. Numerical analysis on the impact of promotion is scarce and authors call for more research (Ali et al., 2009; Donselaar et al., 2016; Ettouzani et al., 2012). In an effort to reduce OOS and food waste that result from promotional activity, retailers are using new promotional strategies. This research will focus on a new promotional strategy implemented by the case company; segmented promotions. This section will elaborate on the case study used and the characteristics and issues related to segmented promotions.
3.1
Case study
We propose a case study conducted at a Dutch pure online grocery retailer. A case study will enable data and knowledge to validate theoretical theory in a practical environment. Unlike the main competitors in the Netherlands (i.e. Albert Heijn and Jumbo), the company does not own any physical stores. Customers can only shop via an mobile application after which the groceries are delivered to their home address at a chosen time. The lack of physical stores enables the company to run operation very efficiently since they are only tailored to home delivery. Besides all customers order via an mobile application which provides the company with an unique dataset (i.e. viewing time per product, unavailability in seconds and a rich order history). In an effort to reduce demand peaks and the OOS and food waste that is
associated with this peak, the company has started experimenting with segmented promotions. The case study has taken place over a period of four months in which we analyzed the
15
up to two weeks ahead of the promotion week but this is not possible for other perishable products (i.e. dairy products). The company is obligated to purchase the registered volume during the promotion period. As a result of the increased promotional volumes, the articles on promotion need to be slotted at a different and larger shelf in the warehouse. The warehouse management system and the customer application are synchronized thus when there is no inventory to sell the article becomes unavailable in the application. When the actual
promotional volume is lower than the registered volume, the articles are sold in the following period. Yet, for many perishable product, the best-before date is binding constraint that prevents articles being sold the next week. Once articles have passed the best-before date they are thrown away. In addition, left over volumes of large articles (i.e. toilet paper) or very slow moving non-perishable articles can take up take up valuable space in very crowded
warehouses.
3.2
Segmented promotions
Segmented promotions originate from an incentive to decrease the demand peaks and control promotional sales volume. Aforementioned, demand of an article during promotion can be up to 90 times the regular demand. Promotional demand is characterized by uncertainty and volatility, therefore it cannot be planned accurately in advance. Inventory policies with high safety levels can provide a buffer to control for fluctuation but this is very inefficient and, on top of this, it can lead to a waste for perishable products. Segmented promotions can pose a solution to this challenges while maintaining efficient operations that minimize costs. For segmented promotions, a retailer divides its total population of customers into segments. Each segment should contain a set of unique customers that are not part of other segments. A critical constraint for efficient segmented promotion is therefore the ability to differentiate between unique customers. In online retailing, this is possible through accounts that require log-in. The number of customer segments depends on the length of the promotion cycle and driven by the need to spread demand peaks. A longer promotion cycle enables to spread the total promotional uptake further thus minimizing the promotional lift per period.
16
segmented promotion. The regular promotion shows a major uptake in period seven, resulting in a lift of six times the base volume. In the segmented example the promotion is active in period six, seven, eight and nine for segments A, B, C and D respectively. Thus in period eight, segments A, B and D have no active promotion (i.e. customers in these segments pay the regular price) while segment C has an active promotion (i.e. customers in this segment pay the discounted price). After the segmented promotion cycle is complete (i.e. four periods have elapsed) all customers have had access to the discounted price yet the peak in demand has decreased to 2.25 times the regular volume for four consecutive weeks. For different articles, multiple segmented promotion cycles can run paralleled with a different order in customer segment per period. This enables that customers in each segment have access to a promotion in each periods.
Figure 1 "Illustration of segmented promotion"
The ability to differentiate between unique customers has evolved since the introduction of online retailing, where customers are often required to log-in to their specific account. Therefor this strategy is not easily applicable on traditional in-store retailing. Additionally, when attracting new customers through wide-spread promotional advertisement (i.e. radio/tv commercial) a retailer is not able to differentiate between customers. This makes segmented promotions a less suitable strategy for this cause. Targeted email advertisement is an
exception and can work in combination with segmented promotions.
17
3.3
Incorporating promotional impact of variables
From the case study we learn that retailers are submitted to the registered volumes, without proper control of the promotional volume in regular promotions. Segmented promotions enable control over the promotional volume during the segmented promotional cycle. After the first period in a segmented promotion the retailer can update their forecast. An
extrapolation of the first period sales across the remaining periods of the cycle can provide a much more accurate forecast. A retailer is now aware of the mismatch between the original registered volume and the more accurate predicted volume of the remaining periods. The promotional lift is dependent on many variables where some are article specific (i.e. weight, volume or perishability of an article) and other are time specific (i.e. holidays or season at time of promotion). Few variables can be determined by the retailer namely: percentage discount, multibuy or single buy and type/rank of display. In the course of this paper we will refer to the retailer determined variables as promotional variables. During the sequential periods in a segmented promotion cycle, the retailer can control the promotion lift by tweaking these variables. The goal, hereby, is to minimize the mismatch between the registered volume and the predicted volume. Consequently it is essential that retailers are aware of the impact of changes in promotional variables.
3.4
Research question
From theoretical analysis we understand that the ability to control the promotional volumes during consecutive periods in a segmented promotions cycle can minimize the mismatch between forecasted and actual volumes and therefore reduce OOS and food waste. We aim to test the efficiency of this new promotional strategy in term of reductions in OOS and food waste compared to regular promotions. This has resulted in the following research question:
To what extend can segmented promotion reduce OOS and food waste compared to regular promotions in online grocery retail?
To answer this question, first we have to estimate the effect of changes in each promotion variable on the promotional lift (i.e What happens to the promotional lift if the percentage discount changes from 10% to 15%?). After we have to find the correct parameters to
18
4.
Methodology
In this paper we propose a three-step methodology that enables comparison between regular and segmented promotions in terms of OOS and food waste. Figure 2 shows the sequential steps and the associated goal for each analysis. This section will elaborate on method used in each step and illustrate its functionality.
Figure 2 "Three-step methodology”
4.1
Regression analysis.
The goal of this regression analysis is to estimate the impact of changes in the promotion variables on the promotional lift. Regression analysis is well known statistical method to estimate the relation between an dependent variable and one or more independent variables (or predictors). Regression analysis are widely used in the field of economics and business and is frequently found in forecasting literature (Ali et al., 2009; Johnson Ferreira, Hong, & Simchi-Levi, 2016; Ma et al., 2016; J. W. Taylor, 2007; Trapero, Kourentzes, & Fildes, 2015; Zotteri et al., 2005). For this step we adopt the methodology from Donselaar, Peters, Jong, & Broekmeulen (2016). The authors analyze demand during promotional activity at a traditional grocery retailer using a regression model with promotional lift as depend variable.
4.1.1 Promotional lift
19
standard (Cooper et al., 1999). The sales quantity can be decomposed by the number of customers, the penetration (i.e. the percentage of total customers that buy an article) and the items bought per buying customer (Equation II).
(I) 𝑃𝑟𝑜𝑚𝑜𝑡𝑖𝑜𝑛𝑎𝑙 𝑙𝑖𝑓𝑡 (𝑃𝐿)𝑖𝑡 = 1 𝑆𝑎𝑙𝑒𝑠 𝑞𝑡𝑦𝑖𝑡
3(∑−1𝑡=−3𝑆𝑎𝑙𝑒𝑠𝑞𝑡𝑦𝑖𝑡)
(II) 𝑆𝑎𝑙𝑒𝑠 𝑞𝑡𝑦𝑖𝑡 = 𝐶𝑢𝑠𝑡𝑜𝑚𝑒𝑟𝑠𝑡∗ 𝑃𝑒𝑛𝑒𝑡𝑟𝑎𝑡𝑖𝑜𝑛𝑖𝑡∗ 𝑖𝑡𝑒𝑚𝑠 𝑝𝑒𝑟 𝑏𝑢𝑦𝑖𝑛𝑔 𝑐𝑢𝑠𝑡𝑖𝑡
From the literature review we learn that the promotional uptake in sales is the cause of customers switching store (i.e. lift in number of customers), brand or article type (i.e. lift in penetration) and/or customers buying more (i.e. lift in items per buying customer). The penetration is calculated by dividing the number of unique customers that buy article i in period t by the total number of customers in period t. The first effect is described as the substitution effect while the latter is described as the stockpiling effect. Many authors speculate on the effect of promotions on penetration and stockpiling (Ailawadi et al., 2009; Ali et al., 2009; Donselaar et al., 2016; Ettouzani et al., 2012; Mou et al., 2017),yet no empirical analysis has been done to estimate the effects because no accurate point-of-sales (POS) data was available. Traditional retailers often lack exact data on the total number of customers in store and cannot identify repetitive purchases by a customer. Detailed POS data provided by the case company enables us to separately estimate the substitution and
stockpiling effect.. Similarly to the promotional lift, the penetration lift and items per buying customer lift are calculated by dividing the values during the promotional period by the average value in the three week prior to the promotional period.
20
4.1.2 The model
From the literature we have selected the most influential independent variables that are
available in the dataset in line with Donselaar et al. (2016). Variables like front page on flyers, in-store displays and coupon savings are reoccurring in the promotional lift literature.
However, these types of promotion strategies are not used by the company in this study. Some variables are excluded to prevent multicollinearity, since they are highly correlated with other variables (i.e. the absolute price discount and the price per unit of weight). To make
comparison between articles more relevant we take the natural logarithm of the promotional lift. The relative promotional lift opposed to the absolute value as the dependent variable results in standardized values for all promotions (Donselaar et al., 2016). The primary focus for this regression analysis is on the promotional variables. Other variables obtained from the literature review on the promotional lift are included to control for the effects of these
variables.
(III) 𝐿𝑛(𝑃𝐿𝑖𝑡) = 𝛽0+ 𝛽1𝐷𝑖𝑡+ 𝛽2𝐿𝑛(𝑃𝑟𝑖𝑐𝑒𝑖𝑡) + 𝛽3𝐿𝑛(𝐵𝑎𝑠𝑒𝑖𝑡) + 𝛽4𝑅𝑎𝑛𝑘𝑖𝑡+
𝛽5𝐿𝑛(𝐹𝑟𝑒𝑠ℎ𝑖) + 𝛽6𝐶ℎ𝑖𝑙𝑙𝑒𝑑𝑖+ 𝛽7𝑃𝑟𝑒𝑚𝑖𝑢𝑚 𝑏𝑟𝑎𝑛𝑑𝑖+
𝛽8𝑆𝑡𝑎𝑐𝑘𝑎𝑏𝑙𝑒𝑖 + 𝛽9𝑀𝑢𝑙𝑡𝑖𝑏𝑢𝑦𝑖𝑡+ 𝛽10𝐿𝑛(𝑊𝑒𝑖𝑔ℎ𝑡𝑖) + 𝛽11𝐻𝑜𝑙𝑖𝑑𝑎𝑦𝑡+ 𝜖
Where
𝑷𝑳𝒊𝒕 Promotional lift factor for item i in period t
𝑫𝒊𝒕𝒋 The percentage discount for item i in period t
𝑭𝒓𝒆𝒔𝒉𝒊 The minimal days of guaranteed preservability after sale for item i
𝑷𝒓𝒊𝒄𝒆𝒊𝒕 The price of item i at period t without discount
𝑩𝒂𝒔𝒆𝒊𝒕 The base volume of item i in the three weeks prior to period t
𝑹𝒂𝒏𝒌𝒊𝒕 The rank on promo page in application for item i in period t (range
from 1 to 25 where 1 is at the top of the page and 25 is at the bottom of the page)
𝑪𝒉𝒊𝒍𝒍𝒆𝒅𝒊 Indicator variable if item i is stored in chilled (1) or ambient
temperature (0)
𝑷𝒓𝒆𝒎𝒊𝒖𝒎 𝒃𝒓𝒂𝒏𝒅𝒊 Indicator variable if item i is a premium brand (1) or private label (0)
𝑺𝒕𝒂𝒄𝒌𝒂𝒃𝒍𝒆𝒊 Indicator variable if item i is able to be stacked (1) or not (0)
𝑴𝒖𝒍𝒕𝒊𝒃𝒖𝒚𝒊𝒕 Indicator variable if item i was part of a multibuy (1) or single-buy
(0) promotion in period t
𝑾𝒆𝒊𝒈𝒉𝒕𝒊 The weight of item i in kilograms
21
4.1.3 Analysis
To estimate the effect of changes in promotional variables on the promotional lift we run this
regression model (Equation III). The resulting coefficients (𝛽1 𝑡𝑜 𝛽11) show the effect of a
unit change, in the corresponding variable, on the promotional lift (i.e. if the rank changes
from 10 to 11, the natural logarithm of the promotional lift changes by 𝛽4). The corresponding
significance level for each coefficient reports if we can reject the null hypothesis that the coefficient equal to zero (i.e. the variable has no effect on dependent variable). A significance level with p-value ≤ 0.10 indicates that we can reject the null hypothesis. We focus primarily
on coefficients for the promotional variables; Discount (𝛽1), Rank (𝛽4) and Multibuy (𝛽9) and
the corresponding significance levels. These variables can be altered by the retailer during the course of a segmented promotion cycle while the other variables are article or period specific. If the coefficient is significant we use this value to calculate the effect of a change in
promotion variables, this will be explained in more detail in section 4.3. The model in Equation III is also estimated using the penetration lift and items per buying customer lift as dependent variables instead of the promotional lift. Results from these two regression analysis can provide insights on the substitution effect (Penetration lift) and stockpiling effect (Items per buying customer lift). Analysis on the difference between both effects is useful to validate theory on this matter, which in turn can be useful to control promotional sales volume. From the literature review we learn that there is a non-linear relation between percentage discount and the promotional lift (Bijmolt et al., 2005; Donselaar et al., 2016; van Heerde et al., 2003). The effect of an increase from 5 to 10 percent discount is not equal to the effect of an increase from 20 to 25 percent discount. The non-linear form of this relation can cause threshold or saturation levels in the percentage discount. We introduce a dummy variable for the percentage discount in the regression model to capture these non-linear effects. The alternative regression model now becomes (Equation IV).
(IV) 𝐿𝑛(𝑃𝐿𝑖𝑡) = 𝛽0+ ∑5𝑗=1𝛽1𝑗𝐷𝑖𝑡𝑗+ 𝛽2𝐿𝑛(𝑃𝑟𝑖𝑐𝑒𝑖𝑡) + 𝛽3𝐿𝑛(𝐵𝑎𝑠𝑒𝑖𝑡) + 𝛽4𝑅𝑎𝑛𝑘𝑖𝑡+
𝛽5𝐿𝑛(𝐹𝑟𝑒𝑠ℎ𝑖) + 𝛽6𝐶ℎ𝑖𝑙𝑙𝑒𝑑𝑖+ 𝛽7𝑃𝑟𝑒𝑚𝑖𝑢𝑚 𝑏𝑟𝑎𝑛𝑑𝑖+
22
Where 𝐷𝑖𝑡𝑗 is a dummy variable for percentage discount for item i in period t. 𝐷𝑖𝑡1=1
(0-10%), 𝐷𝑖𝑡2=1 (10-20%), 𝐷𝑖𝑡3=1 (30-40%), 𝐷𝑖𝑡4=1 (40-50%) or 𝐷𝑖𝑡5=1 (50-60%). Discretizing
continuous data into buckets has its shortcomings, yet it does enable comparison between one discount class to the next. To identify threshold and saturation levels we run the alternative regression models with the promotional lift as dependent variables and the only difference being the default dummy class as proposed by Donselaar et al. (2016). For threshold levels, we start with 0-10 percent class as the default dummy class and check if there is a significant positive difference between the 0-10 percent discount class and higher. If the difference is equal to zero or negative and significant we can conclude that a threshold is present. A similar methodology is used to identify saturation levels but in this case we start with the highest discount class as default and compare this to lower classes.
4.2
Case analysis
The goal of the case analysis is to find values of the parameters to realistically simulate segmented and regular promotions. The case company has experimented with segmented promotions in three intervals. Direct comparison between results from historic regular promotions and the results from the test intervals will be influenced by the demand volatility (Donselaar et al., 2016) and forecast accuracy (Ali et al., 2009) thus making such a
comparison unmeaningful. Instead we propose to simulate regular and segmented promotions in equal time periods with similar forecasting inaccuracies. As a result we must estimate the volatility in demand, to simulate differences between periods in a segmented promotion cycle. Secondly, the accuracy of the registered volume is estimated (i.e. the prior forecasted volume compared to the actual promotional volume). Other input values like base volume and average promotional lift are provided by the case company.
23
a frequently used method in the industry that works simple yet effective to simulate normal random variables.
Ali et al. (2009) showed that there are opportunities to improve the five weeks ahead forecast of the registered volume by as much as 65 percent. However, this is not the focus of this study and therefore during the course of this paper we will take the registered volume as
predetermined. We estimate the difference of historic registered volumes versus the actual promotional demand in that period. Again we will use the distribution of this historic difference (forecast inaccuracy) with its mean and standard deviation to simulate random forecast inaccuracies. To account for the possible improvements in forecasting accuracy, suggested by Ali et al. (2009), we simulate both a case of low forecast accuracy, at the current level, and a case of high forecast accuracy, with a 65 percent improvement.
4.3
Simulation
The final step in the methodology is to simulate regular and segmented promotions in similar environments and compare the difference in OOS and food waste. Here we will elaborate on the scenarios and decisions that occur in the simulation.
During a segmented promotion cycle, actual sales data from the first to the current period can be used to more accurately predict the volume sold in the sequential period. Again, to avoid misunderstandings, we will refer to this volume as the predicted volume (PV). In the
simulation study, this PV is calculated using a simple, yet effective, moving average of the available sales data in the previous periods (Equation V) (Ali et al., 2009). Using the base volume, we can calculate a predicted lift (PL).
(V) 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑉𝑜𝑙𝑢𝑚𝑒 (𝑃𝑉𝑖𝑡) = 𝑛1 ∑𝑛𝑡=1𝑆𝑎𝑙𝑒𝑠 𝑞𝑡𝑦𝑖𝑡 |𝑤ℎ𝑒𝑟𝑒 𝑡 = 1,2, . . , 𝑛
Basically, there are three different scenarios to account for per period in a segmented
24
decision variables) to minimize the mismatch (Equation IX and X). The effect of these
alterations in the decision variables is, per product category, obtained from regression analysis
(i.e. βi). When a the coefficient for a promotional variable is insignificant, that variable will
not be included (i.e. βi = 0). The coefficients denote the change of the natural logarithm of the
promotional lift that result of a change in the corresponding variable. As a result we write the objective function for the mismatch in terms on the natural logarithm (Equation VI). The promotional variables (or decision variables) are bounden to following constraints; The maximum difference in discount between original set value at t=0 and the focal period is five percent (Equation XI). The rank on promo page is an integer value ranging from 1 to 25 (Equation XII). Multibuy is an indicator variable where 1 denotes multibuy promotion and 0 denotes a single buy promotion (Equation XIII). We will use the coefficients and its standard deviation with randomly generated probabilities to simulate the actual effect of changes in the decision variables compared to the expected effect used to solve the optimization problem.
(VI) min 𝑀𝑖𝑠𝑚𝑎𝑡𝑐ℎ = {ln(𝑅𝐿) − ln(𝑃𝐿ln(𝑃𝐿 𝑡+1) 𝑖𝑓 ln(𝑃𝐿𝑡) < ln(𝑅𝐿) 𝑡+1) − ln(𝑅𝐿) 𝑖𝑓 ln(𝑃𝐿𝑡) > ln(𝑅𝐿) Where; (VII) 𝑃𝐿𝑡 = 𝐵𝑎𝑠𝑒 𝑣𝑜𝑙.𝑃𝑉𝑡 (VIII) 𝑅𝐿 = 𝐵𝑎𝑠𝑒 𝑣𝑜𝑙.𝑅𝑉 (IX) ln(𝑃𝐿𝑡+1) = ln(𝑃𝐿𝑡) + ∆𝐷𝑉 (X) ∆𝐷𝑉 = 𝛽1 (𝐷𝑡+1− 𝐷𝑡) + 𝛽4 (𝑅𝑎𝑛𝑘𝑡+1− 𝑅𝑎𝑛𝑘𝑡) + 𝛽9 (𝑀𝑢𝑙𝑡𝑖𝑏𝑢𝑦𝑡+1− 𝑀𝑢𝑙𝑡𝑖𝑏𝑢𝑦𝑡) Subject to; (XI) 𝐷𝑡+1= 𝐷0± 0.05 (XII) 𝑅𝑎𝑛𝑘𝑡+1= 1,2,3, . .25 (XIII) 𝑀𝑢𝑙𝑡𝑖𝑏𝑢𝑦 ∈ {0,1}
25
promotion cycle we calculate the PV for that period. The sales in the first period are 150 units
(due to forecast inaccuracy of -0.25) this results in PLt=1.5. We notice that PL<RL thus we
are selling then forecasted. For the subsequent period we want to minimize this difference by changing the decision variables. The optimization problem is solved using Microsoft Excel solver. In this example we expect an increase in discount (to increase volume in subsequent period), a decrease in rank (to move promoted article higher on promo page) and possibly a
change to multibuy promotion dependent on the coefficient (from 25% discount to 2nd article
at half the price). Now we simulate sales for the second period, we incorporate the actual effect of changes in decision variables and volatility of demand. Sales in second period are
190 units this results in PLt=1.7 (average period 1 and period 2). Again we minimize the
26
5.
Results
This section will report the results obtained in each step of the three-step methodology.
5.1
Results from regression analysis
First we will report on the dataset used in the regression model and check if all the assumption regression model are met. After we will discuss the regression results with respect to the promotional variables. Lastly we will discuss the result from the analysis on thresholds and saturation levels.
5.1.1 Data and regression assumptions
The dataset is provided by the case company includes data on 3976 promotions over the course of 91 weeks. In total there are 1855 unique articles that are all assigned to a specific
product category by the case company. Table 22 provides an overview of the promotion and
articles in each product category. Additionally, we have described the average values for the promotional variables in each product category. There are four product categories that exist out of non-perishable items; Candy & Snacks, Housekeeping, Pantry and Pharmacy. The other product categories include only perishable items, with a very limited shelf life.
Table 2 "Data description per product category"
Prior to running the regression analysis we test if all necessary assumption are met. Violation of the assumptions would bias the results and make them untrustworthy. The adjusted R-squared value ranges from 0 to 1 and reflects the amount of the variation in dependent variable that is described by the independent variables. The higher the adjusted R-squared
Candy & Snacks
Dairy products
Drinks Fruit & Vegetables House-keeping Meat, Fish, Vega
Pantry Pharmacy Ready made Sandwich filling # of articles 143 167 15 222 78 79 321 229 47 189 # of promotion 251 360 46 540 195 139 786 101 69 399 # of weeks 91 91 91 91 91 91 91 91 91 91 # of multibuy promotions 162 150 25 8 137 6 362 184 28 78 Average % discount 22% 22% 26% 13% 35% 15% 22% 30% 22% 22% Average rank 7 6 4 9 9 10 9 11 9 6
Average promotional lift 14.0 8.9 12.5 3.8 19.6 6.5 9.2 13.0 7.2 6.9
27
value, the better the model fits the data provided. There is no benchmark for a minimal adjusted R-squared value since this depend heavily on the predictability of the dependent variable. Therefore we compare our adjusted R-squared values, that range from 0.42 to 0.87 to prior research in grocery promotional demand. Our results compare favorably (higher adjusted R-squared values) to the models estimated by Martínez-Ruiz et al. (2006) but are slightly lower than the models estimated by Donselaar et al. (2016). Nonetheless a great part of variation of the promotional lift is indeed explained by these models. Variance Inflation Factor (VIF) statistics are generally used to check the presence of multicollinearity between the variables in the model. The maximum VIF statistic values for perishable category models; Sandwich fillings (1.84), Dairy products (2.45), Fruit & Vegetables (1.42), Meat, Fish & Vegetarian (2.11), Ready-made (6.79), Drinks (6.88). The maximum VIF statistic values for non-perishable category models; Candy & Snacks (1.37), Housekeeping (2.34), Pantry (1.87), Pharmacy (1.75). VIF statistic values that exceed 10 indicate that the least squares estimates are excessively influenced by multicollinearity (Breusch & Pagan, 1979). Yet, in our models the maximum VIF statistic does not exceed 6.88 and most values are below 3. This shows that the assumption of no multicollinearity is met.
Next the Breusch-Pegan test was conducted to check the assumption of constant variance of the error term (homoscedasticity). However the test results indicated heteroskedasticity. The residuals vs fitted it values plot can visualize the severity of heteroscedasticity. Figure shows the residuals vs fitted values for the dairy products model, it shows clearly that the variance in the residuals is not constant but increasing for the fitted values thus clearly indicating
heteroscedasticity.
28
Heteroscedasticity violates the assumptions of an Ordinary Least Squares regression and thus the estimates are no longer the best linear unbiased estimators (BLUE). Heteroscedasticity does not necessarily mean the coefficients are biased but the significance of the coefficients can be incorrect. The standard approach to this is to run the regression with robust standard errors (Breusch & Pagan, 1979). Robust standard errors relax the assumption of independent and identical distributed errors. Next we can run a Generalized Least Squares (GLS)
regression, which, in case of heteroscedasticity, gives a smaller weight to observation expected to have error terms with large variances compared to observation that have error terms with small variances (Breusch & Pagan, 1979). The Durbin-Watson statistic is generally used to indicate autocorrelation, in our models the Durbin-Watson statistics range from 2.01 to 2.14 which indicates no serial/autocorrelation. Results from the GLS regression with robust standard errors are reported. Next we look at the coefficients for the promotional variables in model with promotional lift as dependent variable. To gain more insights in these finding we compare these coefficients with the coefficients obtained from the models with penetration lift and items per buying customer lift as dependent variables.
5.1.2 Percentage discount
Table 3 reports the values of the coefficient β0 to β11 for each product categories obtained
from the GLS regression with robust standard errors. The corresponding number of asterixis represents the significance level of that coefficient. For some independent variables a value for the coefficient per product category is missing. Reason for this is that all promotions in that product category have the similar or no value for this independent variable and thus no variation.
In table 3 we notice that the coefficient for discount percentage (β1) is positively significant
for every product category but the drinks and housekeeping categories. From this we can conclude that, with an exception of promotions in the drinks category, an increase in
percentage discount will increase the promotional demand for that item. The product category of drinks comprise fully out juices and has a very limited number of observations.
Furthermore, the percentage discount is very similar for all promotions in the drinks category which explains the insignificance of the coefficient.
Next we look at the coefficients for discount (β1) in the models with penetration (Table 4) and
29
reasons for the positive effect. We notice that generally the coefficients for percentage
discount (β1) are much higher in the model with the penetration lift (table 4) compared to the
items lift (table 5). From this we conclude that the increase in sales that result from increases in percentage discount are mainly because of the substitution effect and much less because of the stockpiling effect. For the non-perishable product categories, we notice relatively high and significant results on items per buying customers. This indicates that higher percentage
discount will indeed invite customers to buy more of a single product. The fact that these articles can easily be stored for later use can explain these higher and significant coefficients.
5.1.3 Rank on promotion page
Overall, the impact of rank (β4) on the promotional lift is very limited. The coefficients for
rank in table 3 are close to 0 and most are insignificant. This does not compare to results found in other studies that analyze the promotional lift (Donselaar et al., 2016; Nijs et al., 2001; van Heerde, Leeflang, & Wittink, 2004). Reason can be that prior research focused on offline promotion flyers compared to the online promotion page in this research. The online promotion page is the landing page in the application, thus the first thing you see and it easy to scroll down. All historic promotions used in the dataset were displayed on this page thus explaining the minor influence.
Again we look at the tables 4 and 5 to gain more insight on this effect, since there can still be relevance for the retailer. There can be an incentive to differentiate between selling more to customers that regularly buy the article (increase in items lift) or increase the number of customers that substitute or try out the promoted article (increase in penetration lift). We notice negative significant coefficient for rank on the penetration lift (table 4). A lower position on the promo page will thus decrease the number of customers that substitute a product for the article in promotion. The effect of rank on items per buying customer lift is very low and insignificant (table 5). This indicates that the amount of units of that are purchased per article per customer is independent of the position on the promo page. Thus a retailer face with limited inventory can choose to position an article low on the promo page and enable customers that buy the article regularly to stockpile the item. The other option is to position the article on the top of the promotion page which will lead to more customer
30
5.1.4 Multibuy versus Single buy
The difference between a multibuy and a single buy promotion is captured in the dummy
variable for multibuy (β9). In table 4 we see that most product categories show a positive and
significant coefficient. This means that the promotional lift is higher when the article is
promoted in multibuy (i.e. 2nd halve price or 1+1 free). Exception to this are the categories
fruit & vegetables (-0.027), meat & fish & vegetarian (-0.371) and ready-made (-0.086). The articles in these categories are often very perishable and thus the restriction to stockpile might be the explanation for this negative effect. A restriction to stockpile of very perishable articles
would show through lower coefficients for multibuy (β9) in the items per buying customer lift.
However, when we look at table 5, we notice that the coefficients for multibuy for these product categories are actually larger than other perishable categories. The restriction to stockpile is therefore not the reason for the negative coefficients in table 3.
We look at table 4 to further examine the effects of multibuy promotions on very perishable articles. The coefficient for multibuy for the product categories fruit & vegetables, meat & fish & vegetarian and ready-made are all significant and strongly negative compared to the other perishable categories. From this we can conclude that customers are less like to substitute for or try out the promoted articles when they are required to buy multiple units. Combining both findings, we conclude that the restriction to stockpile very perishable articles is not the reason for the overall negative effect of a multibuy promotion but the lack of the substitution effect is.
5.1.5 Other variables
In line with the research by Donselaar et al. (2016), the base volume is an important predictor for promotional lift. For all product categories we see a negative significant relation of base volume on promotional lift indicating that routinely purchased articles show lower
31
Promotional lift Candy & Snacks
Dairy Products
Drinks Fruit & Vegetables
House- keeping
Meat, Fish & Vega
Pantry Pharmacy Ready-made Sandwich
filling β β β β β β β β β β β1 Discount 1.684 *** 2.466 *** 0.247 2.327 *** 0.436 2.677 *** 1.641 *** 3.927 *** 3.392 *** 1.462 *** β2 Ln(Price) -0.320 *** -0.071 -0.561 *** 0.074 0.015 0.037 -0.155 ** -0.421 *** -0.296 0.179 β3 Ln(Base) -0.152 *** -0.144 *** -0.210 *** -0.114 *** -0.193 *** -0.130 *** -0.147 *** -0.177 *** -0.150 *** -0.118 *** β4 Rank 0.007 -0.011 * 0.011 0.002 -0.013 0.007 -0.012 *** 0.009 * 0.001 -0.001 β5 Ln(Fresh) -0.274 *** 0.000 -0.153 *** 0.025 -0.008 0.221 * β6 Chilled -0.121 0.000 -0.009 0.000 0.000 0.000 * β7 Prem. brand -0.040 0.122 0.239 ** 0.153 -0.204 -0.342 *** -0.028 0.903 *** -0.093 0.163 *** β8 Holiday -0.055 0.008 -0.15 *** -0.081 ** -0.063 0.086 -0.088 ** -0.067 0.394 ** 0.049 β9 Multibuy 0.227 *** 0.087 ** 0.570 *** -0.027 0.554 *** -0.371 ** 0.009 0.114 -0.086 0.140 ** β10 Ln(Weight) 0.083 -0.126 ** -0.444 *** 0.063 ** -0.255 ** -0.041 0.060 0.001 -0.479 -0.214 ** β0 _cons 4.477 *** 3.164 *** 5.956 *** 1.928 *** 2.906 ** 2.051 *** 3.315 *** 2.955 *** 2.888 0.419 Observations 268 320 46 539 105 139 487 163 61 382 R-squared 0.45 0.70 0.87 0.59 0.18 0.57 0.49 0.42 0.48 0.56
Penetration lift Candy & Snacks
Dairy Products
Drinks Fruit & Vegetables
House- keeping
Meat, Fish & Vega
Pantry Pharmacy Ready-made Sandwich
filling β β β β β β β β β β β1 Discount 1,151 *** 0,943 ** 2,180 * 1,560 *** -0,092 1,145 * 1,220 *** 2,608 *** 1,512 ** 0,615 ** β2 Ln(Price) -0,424 *** -0,035 0,156 0,041 0,022 -0,076 -0,130 * -0,547 *** -0,539 * -0,056 β3 Ln(Base) -0,604 *** -0,654 *** -0,715 *** -0,405 *** -0,409 *** -0,424 *** -0,504 *** -0,480 *** -0,641 *** -0,420 *** β4 Rank -0,031 *** -0,032 *** -0,097 *** -0,028 *** -0,049 ** -0,025 *** -0,041 *** 0,000 -0,024 ** -0,011 ** β5 Ln(Fresh) -0,180 ** -0,162 *** 0,010 -0,062 0,179 β6 Chilled -0,601 *** -0,025 β7 Prem. brand -0,011 0,019 -0,159 0,097 -0,184 -0,394 *** -0,049 0,853 *** -0,047 0,064 β8 Holiday 0,002 -0,030 0,053 -0,070 ** -0,043 0,088 ** -0,035 -0,175 ** 0,505 *** 0,049 β9 Multibuy -0,171 ** -0,022 -0,143 * -0,337 * -0,210 -0,351 * -0,415 *** -0,586 *** -0,314 *** -0,061 β10 Ln(Weight) 0,193 ** -0,019 -0,127 0,082 *** -0,275 *** 0,010 0,029 0,088 -0,228 0,074 β0 _cons 1,325 ** -0,453 -2,633 * -0,276 -0,055 0,148 -0,076 0,820 0,540 -0,441 Observations 268 320 46 539 105 139 487 163 61 382 R-squared 0,51 0,70 0,80 0,63 0,16 0,57 0,43 0,37 0,57 0,47
Items lift Candy & Snacks
Dairy Products
Drinks Fruit & Vegetables
House- keeping
Meat, Fish & Vega
Pantry Pharmacy Ready-made Sandwich
filling β β β β β β β β β β β1 Discount 0,388 *** 0,219 ** 0,047 0,035 0,391 * 0,240 ** 0,141 ** 0,264 *** 0,424 *** 0,026 β2 Ln(Price) 0,036 -0,165 *** -0,344 *** -0,032 *** 0,055 -0,055 ** -0,092 *** 0,022 -0,199 *** 0,015 β3 Ln(Base) -0,821 *** -0,314 *** -0,596 *** -0,161 *** -0,829 *** -0,372 *** -0,577 *** -0,782 *** -0,532 *** -0,473 *** β4 Rank 0,002 0,001 0,004 -0,001 * 0,002 -0,003 ** 0,000 -0,005 ** 0,002 0,002 ** β5 Ln(Fresh) 0,064 *** -0,004 -0,119 *** -0,154 0,064 *** β6 Chilled -0,138 *** -0,002 β7 Prem. brand -0,071 ** 0,061 *** -0,072 0,014 0,011 -0,016 0,049 *** 0,061 -0,025 0,005 β8 Holiday -0,069 *** -0,055 *** 0,043 *** -0,013 ** -0,004 0,009 -0,006 0,046 -0,026 -0,026 *** β9 Multibuy 0,177 *** 0,111 *** 0,126 *** 0,336 *** 0,433 *** 0,224 *** 0,242 *** 0,385 *** 0,149 *** 0,181 *** β10 Ln(Weight) -0,093 *** -0,023 0,116 ** 0,002 -0,021 -0,031 0,042 *** -0,094 *** 0,048 -0,015 β0 _cons 0,550 ** 1,092 *** 2,509 *** 0,343 *** 0,722 *** 0,905 *** 1,101 *** 0,577 ** 1,873 *** 0,297 ** Observations 268 320 46 539 105 139 487 163 61 382 R-squared 0,59 0,35 0,62 0,43 0,57 0,49 0,60 0,47 0,81 0,52
Table 3 "Regression model with promotional lift"
Table 4 "Regression model with penetration lift"
32
5.1.6 Threshold and Saturation levels
The prior analysis showed the significant influence that percentage discount has on
promotional lift. We decomposed the promotional lift in the penetration lift and items lift, yet we do not know what the boundaries are in terms discount percentages. We distinguish between perishable and non-perishable product categories. For this analysis we look at the alternative model (equation IV) with dummy variables for percentage discount, in particular the difference between the default dummy class and the subsequent discount class. The coefficients for the difference between discount classes are reported in Error! Reference
source not found. (perishable) and Error! Reference source not found. (non-perishable).
Table 6 and 7 report the statistical significance of these coefficients for perishable and non-perishable product categories respectively.
The absence of any threshold or saturation level in the percentage discount of a promoted article would express in positive significant coefficients between subsequent discount classes. When combining all perishable items into a single regression model we see indeed all that all differences between increasing discount classes are positive (Error! Reference source not
found.). The corresponding significant levels in table 6 are smaller than 0.10 indicating that
33
interesting though, we notice a clear and significant negative difference between 30-40% and 40-50% thus showing an intermediate saturation level. Nonetheless, the difference between 40-50% and 50-60% is very strong positive and significant. This shows that increasing
discount from 40% onwards has only positive effect on the promotional lift if increased above 50%.
Furthermore, a noteworthy observation is the absence of any linear relation per product category between changes in discount classes and the promotional lift. Therefore we must be careful when implementing the findings on the effects of discount percentage obtained in the regression analysis.
34
Figure 4 "differences for perishable categories"
Table 6 "Significance levels for perishable product categories"
Table 7 "Significance levels for non-perishable product categories"
5.2
Results from case analysis
The literature review showed that systemic errors, i.e. demand autocorrelation, are an important driver of forecast inaccuracies and thus OOS and food waste. When historic sales data is used that includes OOS articles, not all demand is captured, which can trigger under estimations of future demand. This vicious cycle shows the continuous effect of OOS and demand autocorrelation (Moussaoui et al., 2016). This paper avoids this effect by using only sales data from the second and third day of the historic promotion week to calculate
All perishable
Dairy products
Drinks Fruit &
Vegetables Meat, fish &Vega Ready-made Sandwich filling 0-10% vs 10-20% 0.015 0.007 0.409 0.000 0.180 0.777 0.400 10-20% vs 20-30% 0.005 0.005 0.293 0.764 0.984 0.088 0.186 20-30% vs 30-40% 0.043 0.121 0.411 0.000 0.396 0.054 0.030 30-40% vs 40-50% 0.006 0.000 0.016 0.000 0.064 0.000 40-50% vs 50-60% 0.008 All non-perishable Candy & Snacks
Housekeeping Pantry Pharmacy
0-10% vs 10-20% 0.012 0.248 0.000
10-20% vs 20-30% 0.331 0.146 0.303 0.534 0.945
20-30% vs 30-40% 0.000 0.023 0.001 0.000 0.000
30-40% vs 40-50% 0.010 0.044 0.065 0.075 0.000
35
promotional demand. The percentage of orders on these days compared to the total orders in the week is very stable. Besides, inventory managers from the case company find that on the first day of the promotion week there can be issues with deliveries while OOS situations very rarely occur before the fourth day of the promotion week.
5.2.1 Demand Volatility
We have analyzed the difference in demand between periods in the segmented promotion test phases. One thing we notice is that there is no clear difference between the volatility of demand for different product categories. Hence we aggregate the demand for all articles per testing phase. During phase 1 the mean difference between subsequent weeks was 1.0059 with a standard deviation of 0.4512. In phase two we measured a mean of 1.0120 with a standard deviation of 0.2983. In the final testing phase, the mean difference was 0.9969 with a standard deviation of 0.3781. The three test phases occurred at different periods in time yet the mean difference between periods is roughly equal for all test phases. Difference in the standard deviations per test phase can be the cause of a lower number of observations in the second and third testing phase. Interviews with demand planners at the company revealed no clear cause for this difference. As a result we use the mean difference and standard deviation for all articles over the three periods in the simulation. This results in a mean difference between periods of 1.0049 and a standard deviation of 0.3810.
36
Figure 6 "histogram of distribution"
5.2.2 Forecast inaccuracies
Companies with limited historic sales data are challenged to accurately predict promotional volumes. The company in this research is no exception. Error! Reference source not found. shows the promotional demand as a percentage of the RV per product category. The large differences in the percentage of RV sold highlight need for a solution. Each data point relates to a historic article promotion. We calculate the inaccuracy as the percentage difference from the mean in figure 2. The corresponding standard deviations per product category are
37
Figure 7 "Forecast accuracy"
Table 8 "Standard deviation of accuracy"
5.3
Results from simulation
The analysis of historic promotional demand has illustrated the effects that discount, rank and multibuy have on the promotional lift per product category. In case of insignificance of the coefficient for these variables, their effect of a unit change is not taken into account. From the analysis on threshold and saturation levels, we have learned that the relationship between discount and promotional lift is non-linear. Therefor we should be careful to model the effect of a percentage change. Additionally from a moral aspect, it might not be deemed desirable to have large differences in discount percentage between customers. Interviews with the
Category Mean Standard Deviation Sandwich filling 1 0,535714 Freezer 1 0,716667 Drinks 1 0,56044 Pharmacy 1 0,713483 Fruit & Vegetables 1 0,54955 Housekeeping 1 0,735849 Ready-made 1 0,463636
Candy & Snacks 1 0,677966
Meat/Fish/Vega 1 0,598214
Pantry 1 0,610738
