THE EFFECTS OF LABOR PLANNING ON
PROFITABILITY IN RETAILING
Etienne Schreuder
Student no. 10326960
Supervisor:
Drs. Frank Slisser
MASTER THESIS
Executive Program in Management Studies – Marketing Track
Preface
The subject of this thesis was initiated by two different parties, which were Foot Locker Europe and myself. Nowadays, customer demand and customer traffic flows to stores fluctuate a lot, but there is a lack in research on flexible ratios in labor hours. Since Foot Locker Europe is the world biggest retailer is athletic footwear and apparel, the organization would like to get some insights in these variables. Additionally, for myself it was a great opportunity to analyze ‘Big Data’ files, and to further develop my knowledge in the retail industry.
During the master thesis several people supported me with information, knowledge and great recommendations. Some people earn a special remark. First, I would like to thank Foot Locker Europe and the project team members for the supervision of my research paper. A special remark for Paul Jansen for his support and the very interesting discussions we have had. Second, Andrei Grigoras for enriching SQL queries which I could use for combining data and data analyzing. Third, my supervisor drs. Frank Slisser for the stimulating Skype sessions, discussions and the useful feedback during the thesis development. Last but not least, I would like to thank my friends and family who supported me during the thesis project by giving critical comments and stimulating words or just having fun and forgetting the stress.
The Executive Program in Management Studies gave me the opportunity to enrich my knowledge on how things work out scientifically but even more important, this part time study at the University of Amsterdam gave me some wonderful years and new friends from multiple business industries.
Table of contents
Abstract ... 4
1.0 Introduction ... 5
2.0 Literature review ... 6
2.1 Effects on business performance ... 6
2.2 Labor planning based on sales / profit (Gap 1) ... 8
2.3 Labor hours and traffic (Gap 2) ... 11
2.4 Differences among geographical areas (Gap 3) ... 13
2.5 Differences among traffic ranges (Gap 4) ... 14
3.0 Hypotheses ... 15
3.1 Conceptual model ... 17
4.0 Research design ... 19
4.1 Research design and data collection ... 19
4.2 Measures and controls ... 21
5.0 Results ... 23
5.1 Descriptive statistics ... 24
5.2 Data validation ... 31
5.3 Outcomes ... 33
6.0 Discussion ... 37
6.1 Limitations and future research ... 38
7.0 Conclusion ... 40
8.0 References ... 42
Abstract
Nowadays, labor planning in retail business becomes more and more a trending topic for store managers. Most retail organizations use electronic planning tools to schedule their employees. These planning tools are often based on sales response modelling in order to anticipate on demand fluctuations. Since sales demand fluctuates a lot in the current retail business, this paper provides new planning ratios that are focused on profit maximization on an hourly basis. Therefore, it is analyzed whether a profit model is more effective in terms of profit rather than a sales model and if so, is there is an optimum in labor hours to maximize this profit? A large detailed dataset with records from 130 stores of a retailer in athletic footwear and apparel was analyzed to find an accurate model for the estimations of optimum labor hours to maximize profit. Herewith, an Ordinary Least Squares (OLS) regression model is used based on high traffic data (300+ visitors per hour) to find a saturation point in the profit formula based on hourly traffic and labor hours. Using an OLS methodology with a profit outcome is a new contribution to the labor planning literature and it sheds a new light on existing performance indicators for labor planning. Herein significant coefficients were found for the variable labor hours
traffic . This variable was missing in existing
literature and seemed to be a valuable improvement on existing sales models. The outcomes show that a decrease in labor hours leads to an optimum in profit, where the sales model predicted an increase of labor. With an average decrease of ~17.11% in labor hours during peak hours, an average increase of ~8.86% on profit was found at a 0.001 significance level. Based on managers’ insights in expected sales demand, the created ratios will help them to save on labor costs. Moreover, the outcomes of this research also prove that the saturation point in labor hours differs per geographical area and per traffic range, which should be taken into consideration. As expected, the required number of employees is different per country and should be lower in countries with a low profit
traffic ratio. However, the outcomes also indicate that
the potential profit improvements vary among these countries. These results suggest that every geographical area needs to have their own unique planning ratios in order to reach the optimum in labor productivity and profit.
1.0 Introduction
In today’s retailing, business organizations foresee different cost saving challenges. One of these challenges is to efficiently schedule store employees. Actually, a very interesting challenge in retailing is to predict future demand as accurate as possible. As Mani et al. (2012) stated: “The ability to meet shifting customer demand in a timely and cost-effective manner is an important driver of retail store performance. Retailers need to match in-coming traffic with labor to provide a consistent quality of service to their customers” (p. 2). Kabak et al. (2008) mentioned that achieving the best configuration of staff, considering time, demand, and cost constraints, can provide as an important competitive advantage. On the other hand, inappropriate workforce schedules can lead to a wasteful oversupply of staff or an undersupply that leads to loss of business. Ernst et al. (2004) said the following about staff scheduling: “It is an area that has become increasingly important as business becomes more service oriented and cost conscious in a global environment” (p. 3). This labor research arises from this cost consciousness thinking. The described labor planning problem does still need further research and significant evidence, so new findings can help managers to improve their labor planning. Since most of the current planning models (Kabak, 2008) are based on sales response, the main research question of this research is about a comparison between two dependent variables: sales and profit. Sales demand fluctuates a lot in the
retail business, due to this reason a new variable labor hours
traffic was measured as a new planning ratio for
short term decision making.
Main research question: To what extent would it be more profitable to use a planning model based on profit response rather than sales response, and what would be the added value of flexible labor planning ratio(s) on profit optimization?
Moreover, in this paper differences in traffic ranges and geographical areas were analysed, which are both part of the sub research questions. This research paper includes a case study of Foot Locker Europe (FLE). Since FLE implemented a new planning system in their stores, it is very interesting to do research on the subject labor planning in retail stores. At FLE, a store manager receives a certain amount of planning hours on a weekly basis. This number is based on forecasts of sales, units, and transactions
for every single week. When a store manager uses less hours than expected, he will be rewarded. The new system helps the managers to optimize their labor planning based on predicted sales demand; in this case FLE should take labor costs into consideration. Currently, the average store labor costs / sales ratio is approximately 14,9%. FLE uses this percentage as a budget and schedules their employees based on this budget, herewith the ratio is varying per geographical area. However, the question is whether this percentage is an optimal sales / labor costs ratio and whether it leads to higher profits. The literature review provides an overview of the most important prior literature regarding to this labor planning issue. What research has been done already in order to increase business performance? What is the role of labor planning? Where is new research required in order to optimize the current labor planning activities at FLE? There were of three main objectives that initiated this research. First, was to add empirical evidence to the existing literature by searching for the effects of workforce ratios on business performance. Secondly, there was the goal to create a more accurate theoretical framework in labor planning. Finally, new findings in labor planning research would help FLE in managing their employees by creating a better understanding of the costs-benefit ratios during daily operations.
In the literature review four major gaps were found. The hypotheses are based on these four major gaps and contribute the literature with new contributions of labor planning in retailing. In the fourth chapter the research design is described, followed by a results chapter with data distributions, descriptive statistics and outcomes of the hypotheses tests. Subsequently, the discussion and the conclusion were outlined.
2.0 Literature review
The described labor planning issue is an important factor that influences business performance, or more specific, retail store’s profit. How significant is this factor, and what will be other variables that influence a store’s total profit? In this literature review the most obvious factors are discussed, and it provides an overview of the most important research papers in labor planning as well.
2.1 Effects on business performance
Nowadays a lot of research has been done on factors that have significant effects on business performance, profit and customer satisfaction in retailing. Maxham et al. (2010) proved that an increase in
employee satisfaction has an indirect positive effect on store sales. On the other hand, a decline in employee satisfaction has been found to be linked to a decline in store sales. Information about these effects is very valuable for retail managers. What are the most important factors, in particular; how important is an optimal labor planning? Another example is the effect of employee behavior on customer behavior and customer purchase intentions in retailing. Goff et al. (1997) found that the relationship between salesperson and customer has a significant influence on sales. Subsequently, Reynolds et al. (1999) stated that relationship benefits among customers and salespersons are positively associated with satisfaction, loyalty, word of mouth and purchases. Also the role of store managers is researched in the literature; Netemeyer et al. (2010) showed that managers’ performance and satisfaction have a positive effect on Average Customer Value Transaction and so on business performance. Store managers get more and more understanding of customer needs, wants and customers’ expectations of services and products. Ton & Raman (2010) showed that increases in product variety and inventory levels are both associated with higher sales. Another interesting factor in retail stores is the store layout. Sulek et al. (1995) claim that stores can improve customer satisfaction by enhancing store design through the use of technology, innovative layout, and aesthetically appealing décor, therewith they have found significant evidence for their statement. At the end there are many factors that will have an effect on business performance and customer satisfaction and, as mentioned before, labor planning is definitively one of them. Chapados et al. (2013) indicate that “it stands to reason that effective sales staff scheduling should be of critical importance to the profitable operations of a store, since staffing costs typically represents the
second largest expense after the cost of goods sold” (p. 1). As described, profits and business
performance are influenced by many factors and independent variables. Labor planning is one of these independent variables and it has some major gaps in prior literature, which are interesting for both academic and managerial perspectives. These perspectives will be further discussed in the research design chapter. In Figure 1 the main theories of factors with an effect on business performances are displayed in a visualization of prior research, where labor planning has been separated from other influencing factors.
Figure 1: Effects on business performance in retailing
Relationships; Customers vs salesperson
Goff et al. (1997), Reynolds et al (1999).
Store manager’s job performance
Netemeyer et al. (2010)
Product variety , inventory levels
Ton & Raman (2010)
Store layout Sulek et al. (1995) Business performance Labor planning (figure 2)
2.2 Labor planning based on sales / profit (Gap 1)
Before the 21th century, there was not much research of labor planning. However, Hise et al (1983)
already indicated that the number of employees is one of the most important factors in retail stores. Many years later, Lam et al (1998) proposed a model which sets store sales potential as a function of store traffic volume, customer type, and customer response to sale force availability. This model is called the sales response model and it is used in several labor planning research papers. The use of electronically-tracked traffic data made it possible to create a large and reliable dataset. They found an optimal number of employees and simulated their model by creating predicted sales. The results show that understaffing could be a major problem, in order to maximize store profits. The model is also capable of identifying the time periods when service is most heavily demanded by customers. Lam (2000) created a formula for gross profit net of hourly staff costs, which is a derivative from results of the sales response model:
𝜋𝜋
𝑖𝑖𝑖𝑖= 𝑆𝑆
𝑖𝑖𝑖𝑖𝑔𝑔
𝑖𝑖𝑖𝑖− 𝑙𝑙
𝑖𝑖𝑖𝑖𝑤𝑤
𝑖𝑖(1)
Where π is gross profit net of staff costs in an hour at time t for store i, S is the overall store sales, g is the average gross margin, l is the number of sales people present, and w is the average wage rate per hour. In the last decennia labor planning became more and more an upcoming trending topic. Referring to Ton (2009): “While the costs of increasing labor are obvious and easy to measure, the benefits are often indirect and not immediately felt” (p. 1). Where Sulak et al. (1995) mentioned that the relationship
between service quality and profitability remains only partially understood, Ton (2009) proved that an increasing amount of labor hours has a positive effect on conformance quality and service quality. In their setting, conformance quality is an internal measure of quality and is defined as the degree to which stores conform to prescribed standards related to logistics activities. Only the conformance quality had a statistically significant effect on profit margins. Mani et al. (2011) reviewed the drivers and consequences of understaffing, and concluded that the ideal labor planning varies significantly, since the costs of labor vary significantly among different stores. They also mentioned that understaffing has been found to be negatively associated with store association satisfaction. So, empirical research is done on both, increased labor (Ton, 2008) and understaffing (Mani, 2012) and their effects on customer satisfaction / business performance.
As previously mentioned, there are definitively interesting planning tools for retail stores, but most of them are not universal and so too restricted for most industries. Ernst et al. (2004) reviewed different planning solutions: 1) demand modelling - similar to the sales response model by Lam et al. (1998) 2) artificial intelligence approaches 3) constraint programming 4) meta-heuristics 5) mathematical programming approaches. A numerous of existing planning tools are available, although the current literature is very limited in research about key performance indicators for labor planning and profit ratios on hourly basis. The distribution of customer demand over the week changed a lot in the last decennia. Due to this reason it is important to enhance research in this specific field. How could managers maximize their total profits nowadays? What is an optimal labor per hour ratio during peak periods? How should a general manager manage his sales managers in order to get an optimal planning, related to an increase in profit? In most retailing companies, labor planning is quite similar to the sales response model, so it is mainly is based on predicted sales demand.
The efficiency results of the sales response model depends a lot on sales planning accuracy. Arnold et al. (2009) already researched that sales planning has a significant positive effect on product sales, where the manager’s behavior is a mediator between these variables. Labor planning is not part of their research, but sales planning and sales efforts would be interesting additional mediators to a new conceptual framework. When managers receive / create better forecasts on sales demand, they will be able to create a better labor planning as well. Corominas et al. (2005) created an annual planning model for the service
industry, with a finite set of working hours. “Annualising working hours (AH) consists of hiring workers for a certain number of hours per year and distributing those hours irregularly over the course of the year to face demand fluctuations” (Corominas, 2005, p. 231). With their model they calculated ratios in terms of sales demand, capacity and shortages. Actually, the model will be useful for analysing labor planning with a finite number of hours in retailing as well. An extra facet in retailing is that working with a finite number of hours per store is not easy to determine and not always accurate enough, since sales demand is a prediction of the future. Moreover, they explicitly mentioned that planning a weekly number of working hours could become a problem. Finally, Corominas’ research is based on a service industry. How should the total number of working hours in retail stores be managed? It requires understanding and knowledge of labor planning by both general managers and store managers, which is still a challenge today. Netemeyer et al. (2010) already proved that manager’s performance does have a significant effect on both, employee satisfaction and customer satisfaction. As described before, Maxham et al. (2010) found that an increase in satisfaction has a positive effect on business performance. How can organizations positively influence the planning performance of their store manager? Herewith, insights in labor planning ratios might be one of the keys to success. Store managers can use the ratios to maximize their profits, and their managers can use these insights to manage the store managers and to improve their performance.
Based on the literature review, the main conclusion is that there is a major gap in planning the number of employees in retail stores on an hourly basis as a function of a profit curve. Lots of organizations (FLE, 2014) use the conversion and basket value ratios by Perdikaki et al. (2012) to stimulate their store sales. Perdikaki et al. (2012) explicitly mentioned that this conversion rate is different in certain circumstances, for example for different location types, or per higher per capita income. There are schedule methods that take care of the employee planning based on predicted sales demand, the sales response model for example, but effective ratios between labor hours and traffic are missing. Empirical research before the
21st century has already shown that more employees in the store results in more sales (Lam, 1998). On
the other hand, understaffing leads to a decrease in sales. The impact of understaffing on profitability is much higher than that of overstaffing (Mani, 2012). Their paper is focused on periods when retails need to improve their staffing decisions, and they admitted that: “there is a long line of work in operations
management that has employed linear programming or queuing theory for staffing decisions” (p. 5). Their focused approach limited them in the data analysis.
Research gap 1: A sales response model is basically about forecasting sales by traffic demand, but with the Gross Profit Net by staff costs (Lam, 1998), planning labor directly to profit expectations would be a valuable addition to the literature. Moreover, standard ratios based on historical data could be added.
2.3 Labor hours and traffic (Gap 2)
Normally, more employees in a store will generate more sales, but at some point the sales per employee will go down. It is generally known that there will be a point where the average sales per employee is decreasing, but where the business performance would still be increasing. In this case the extra sales by adding an extra employee will be higher than the additional salary costs, in other words; it generates more profit. In most retail companies, sales is the driver to success of their stores, but eventually it is the profit that makes a store successful or not. The question that arises is: At what point will profit go down and saturate? A few years ago Kabak et al. (2008) concluded that standard scheduling problems do not recognize the effect of staff availability on customer sales. In their paper a two-stage model is proposed to handle this staff availability issue. The aim of their optimization model was to satisfy the hourly determined staff requirements with a minimum number of store employees. A mixed integer optimization (MIP) provided them outcomes of which the hourly staff requirements for maximizing profit can be obtained. Their main limitation was that the model was only applicable within certain staff schedule restrictions. As suggested before, Kabak et al. (2008) did also recognize that following staff requirements specified on hourly basis may not always result in the maximum profit on a weekly basis, due to fixed staff costs and legal obligations. Limitations in Kabak’s (2008) research are restrictions in schedule patterns, fulltime and part-time employees, and sample size. Also follow-up research by Chapados (2013) is characterized by these limitations, they admitted; “In the retail store context, finding the optimal number of salespersons for each time period (the optimal staffing) is not sufficient, because there is generally no way to match these optimal numbers exactly given the various constraints on the admissible shifts of employees” (p. 5). However, since multinational organizations have a lot of daily variation in peak hours,
the proportion of temporary employees has increased rapidly. Such a shift in peak periods and employee planning patterns makes it more attractive to research hourly staffing optimums for the future. The research by Chapados et al. (2013) is based on a developed stochastic model, that includes the number of employees, sold units, estimated store traffic, and certain explanatory variables that account for seasonal effects and special events.
Recently, Perdikaki et al. (2012) researched the relationships between store traffic characteristics, labor, and sales performance. Their study was the first in operations management that analysed traffic data in the retail context. This is measured by counting the number of transactions and sales volumes, supported by collected hourly traffic data from a retail chain in women’s apparel over a one-year period. The effects on business performance were measured, but also the effects of the conversion rate, which was defined as the ratio of transactions to traffic. They researched how labor influences the impact of traffic on store sales performance. Subsequently, they showed that 1% increase in labor results in a 0.5% increase in conversion rate, for the women’s apparel stores in their sample size. Ton (2009) found that an increase of one standard deviation in labor at a store is associated with a 10% increase of yearly profitability. Their research objectives are similar to the thoughts of Mani et al. (2012); they already mentioned that retailers have to match traffic and labor. As suggested by Perdikaki et al. (2012), an increase in the conversion rate is positively associated with an increase in customer loyalty. A valuable finding was a decreasing non-linear relationship between traffic and conversion rate. One of the findings by Perdikaki et al. (2012) was that labor moderates the impact of traffic on the number of transactions. They suggested that the development of analytical models would enable better staffing decisions in order to find better traffic / labor ratios. An improvement to the traffic variable in the sales response model was a factor of x between the hourly sales / traffic combination, by Kabak et al. (2008). It is not said that the sales and traffic by a certain customer take place in the same time range. The addition was based on: “the consideration of customers who visit the store and only purchase items after spending a considerable time there” (p. 79). Netessine et al. (2010) also mentioned that it is not argued that labor planning should be based on sales rather than traffic to determine staffing levels in stores. They measured the quality of store labor planning using month-to-month deviations (mismatches) between forecasts of store traffic and planned labor hours: the higher the deviation, the lower the quality of planning. In their research the actual costs of
adding an extra employee, based on traffic and potential sales is missing and would be valuable additional information. Several years later, Tan and Netessine (2013) researched an optimal labor planning ratio for a specific restaurant industry. It appears that reducing the staffing level may improve sales and save labor costs. The result was an optimal number of tables per server.
Unfortunately, there is no empirical research done in an optimal labor hours
traffic ratio in retailing for different
timeframes yet, so organizations do not have accurate control tools on understaffing and overstaffing. Moreover, it would be very interesting to research the effects of labor, as a moderator between traffic and profit. The actual ratios of this research gap would probably be different for several store types by sales volume and within geographical areas, since variable costs are fluctuating by customer demand. The net profit formula by Lam et al. (1998) could be modified by adding the conversion rate and average basket value as variables. This makes sales a dependent variable of traffic and transactions:
𝜋𝜋
𝑖𝑖𝑖𝑖= (𝑆𝑆 ∗ (𝛼𝛼 ∗ 𝐵𝐵) + (𝛽𝛽 ∗ 𝐶𝐶))
𝑖𝑖𝑖𝑖𝑔𝑔
𝑖𝑖𝑖𝑖− 𝑙𝑙
𝑖𝑖𝑖𝑖𝑤𝑤
𝑖𝑖(2)
Where π is gross profit net of staff costs in an hour at time t for store i, g is the average gross margin, S is the overall store sales, l is the number of sales people present, and w is the average wage rate per hour. Where C is the conversion rate and B is the average basket value.
Research gap 2: The variable 𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥 𝐡𝐡𝐥𝐥𝐡𝐡𝐥𝐥𝐡𝐡
𝐭𝐭𝐥𝐥𝐥𝐥𝐭𝐭𝐭𝐭𝐭𝐭𝐭𝐭 is missing in existing planning models and would probably be
an accurate predictor of profit, since this variable is useful for flexible work hours. The effects of this variable should be measured in a retail context and it would be interesting to research the effects of the
𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥 𝐡𝐡𝐥𝐥𝐡𝐡𝐥𝐥𝐡𝐡
𝐭𝐭𝐥𝐥𝐥𝐥𝐭𝐭𝐭𝐭𝐭𝐭𝐭𝐭 ratio on both, profit and sales to find out whether the new variable would be a valuable
addition for sales and profit models.
2.4 Differences among geographical areas (Gap 3)
As suggested by Kabak et al. (2008) there could be substantial differences within geographical areas / locations. Kabak et al. (2008) indicated that each industry and organization has its own characteristics and constraints on workforce scheduling: “A specific mathematical model must thus be developed for each area of application, which cannot be easily transferred to another industry”. For multinational
organizations as FLE, research in differences among geographical areas would be very valuable. An important finding by Mani et al. (2011) is that: "the imputed cost of labor varies significantly among different stores, even though they belong to the same retail chain” (p. 3). Actually, yearly Eurostat statistics by Laursam (2014) show that there are substantial differences among labor costs and productivities within all countries of Europe. Herewith, the average labor costs in France are substantial higher than the average labor costs in the Netherlands and Italy. This would suggest that stores in France would need a higher labor productivity to reach a similar profit? On the other hand, as suggested by Kabak (2008) and Mani (2011), each industry has its own characteristics and ratios should be calculated on an individual basis, which would be per country in this case. Labor costs in France, Italy and the Netherlands were checked because they belong to the dataset of this research (FLE) as well.
Research gap 3: Research in differences among geographical areas is not present in the current literature.
2.5 Differences among traffic ranges (Gap 4)
Research in differences among traffic ranges is very limited as well, which is also suggested by Mani et al. (2011). He admitted that parameters of a sales response function could be different across timeframes. In their model different clusters within the week are present to distinguish differences in traffic volume in these different timeframes. Perdikaki et al. (2009) showed that the conversion rate and the basket value are both influenced by the amount of traffic in an apparel retail business. Surprisingly, there are several suggestions / ideas about differences among traffic ranges, but there a no academic analyses on this topic. For future planning models it would be interesting to anticipate quickly on demand fluctuations, which makes research in traffic ranges very valuable.
Research gap 4: Differences among time ranges should be measured in a profit model. As mentioned
by Mani et al. (2011), there is a lack of information about an optimal 𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥 𝐡𝐡𝐥𝐥𝐡𝐡𝐥𝐥𝐡𝐡
𝐭𝐭𝐥𝐥𝐥𝐥𝐭𝐭𝐭𝐭𝐭𝐭𝐭𝐭 ratio during extreme
The four research gaps that were defined are the first steps to accurate labor planning models based on profit. Since research in these gaps will provide managers with short term planning ratios it is required to first research these major gaps before getting into more detail, as Netessine (2010) already did for a service industry.
3.0 Hypotheses
In many research papers planning models are measured by sales response behaviour. Since labor costs are included in the profit variable and not in the sales variable, the profit model is expected to be more accurate in terms of optimizing the number of store employees to maximize profit. In this research a new variable labor hours
traffic is measured on both outcome variables: profit and sales. The main research question
is directly related to research gap 1 and research gap 2 from the literature review, and therefore it would provide new contributions to existing planning models in retail.
Main research question: To what extent would it be more profitable to use a planning model based on profit response rather than sales response, and what would be the added value of flexible labor planning ratio(s) on profit optimization?
Hypotheses 1 and 2 are based on the main research question, where the first hypothesis includes a comparison between the dependent variables profit and sales.
Hypothesis 1: A profit model based on hourly sales records is a better indicator of an optimum in labor hours in order to reach optimum profit than a sales model based on the same data records.
As discussed in papers by Perdikaki et al. (2012) and Kabak et al. (2012), it would be very valuable to find a tipping point in the profit curve per number of labor hours. Due to the current fluctuations in retail of traffic and demand, a short term decision ratio would be a valuable addition to existing planning methods.
Therefore a hypothesis with an additional independent variable labor hours
Hypothesis 2: A profit model with an additional variable 𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥 𝐡𝐡𝐥𝐥𝐡𝐡𝐥𝐥𝐡𝐡
𝐭𝐭𝐥𝐥𝐥𝐥𝐭𝐭𝐭𝐭𝐭𝐭𝐭𝐭 gives a more reliable and optimized
planning model in terms of profit than a profit model without this variable 𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥 𝐡𝐡𝐥𝐥𝐡𝐡𝐥𝐥𝐡𝐡
𝐭𝐭𝐥𝐥𝐥𝐥𝐭𝐭𝐭𝐭𝐭𝐭𝐭𝐭 .
Since most store managers are mainly focused on stores’ turnover (Ton, 2008), i.e. the sales, the suggestion is that the actual number of labor hours is rather too high than too low. Due to this reason, for optimizing profit a decrease in labor hours (on the x-axis) is expected. The third hypothesis suggests that profit will increase up till a certain labor hours
traffic optimum and after reaching this it will decrease to zero labor
hours and a negative profit amount.
Hypothesis 3: The new variable 𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥 𝐡𝐡𝐥𝐥𝐡𝐡𝐥𝐥𝐡𝐡
𝐭𝐭𝐥𝐥𝐥𝐥𝐭𝐭𝐭𝐭𝐭𝐭𝐭𝐭 in the profit model would decrease compared to actual
labor hours in order to find an optimum in profit.
Due to a lack of analyses in the current literature about differences among countries an extra
sub-question arose: To what extent does the optimum of the variable labor hours
traffic differ per geographical area?
As explained in the literature review, there are probably substantial differences in labor costs and consumer behaviour within countries. The only issue is that there are no elaborations of these differences within countries in existing literature. What is the effect of an optimal number of employees on the profit per visitor? Therefore, in the last hypothesis differences among countries for the effect of the variable
labor hours
traffic , which is the labor productivity, on an additional variable profit
traffic was checked. In many sales
models outcomes are measured by the conversion rate, which is sales per visitor (Perdikaki, 2012). In a profit approach this would be replaced by the profit per visitor. Logically, the expectation is that stores with a lower profit
trafficvalue need to decrease more in labor to reach their optimum number of hours.
Hypothesis 4: The new variable 𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥𝐥 𝐡𝐡𝐥𝐥𝐡𝐡𝐥𝐥𝐡𝐡
𝐭𝐭𝐥𝐥𝐥𝐥𝐭𝐭𝐭𝐭𝐭𝐭𝐭𝐭 in the profit model would decrease less for stores in
countries with a higher 𝐩𝐩𝐥𝐥𝐥𝐥𝐭𝐭𝐭𝐭𝐭𝐭
A second sub-question that came up from the literature review is: To what extent does the optimum of the variable 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 ℎ𝑙𝑙𝑜𝑜𝑙𝑙𝑜𝑜
𝑖𝑖𝑙𝑙𝑙𝑙𝑡𝑡𝑡𝑡𝑖𝑖𝑡𝑡 vary in different timeframes? Since every store has peak hours, moderate hours, and
low traffic hours, it would be interesting to keep the total profit in all timeframes as high as possible. Herewith, it is important to categorize the hours and to determine what makes an hour a peak hour, which could be different per store type. As discussed in the literature review, the sales response model by Lam et al. (1998) is capable of identifying these time periods as well, but is limited to sales forecasts instead of profit maximization. There are no separations in traffic ranges in existing research papers available. Therefore, it is chosen to not create a hypothesis about this research gap, but to include an analysis of
labor hours
sales within different timeframes based on FLE standards. At FLE a peak hour is defined as at least
300 visitors per hour and an extreme peak hour contains at least 400 visitors per hour, which is based on traffic analyses over the last couple of years. Besides these separations in timeframes, there could also be seasonal differences in optimal planning ratios in retailing (Netessine et al., 2010), but this topic will be out of the scope of this research and could be researched in the future.
Answers to the research questions will provide the current literature with knowledge about the effects of optimal planning ratios on profit, and correlations among different timeframes and geographical areas. From a managerial perspective it will help managers to create key performance indicators in retail stores in order to optimize profit. All research questions will be answered within a quantitative research design.
3.1 Conceptual model
This research is focused on labor planning, as in number of employees in retail stores. The basic model in Figure 2a includes a simple interrelationship where the new variable 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 ℎ𝑙𝑙𝑜𝑜𝑙𝑙𝑜𝑜
𝑖𝑖𝑙𝑙𝑙𝑙𝑡𝑡𝑡𝑡𝑖𝑖𝑡𝑡 (=AHTraffic) moderates
between Traffic and Profit / Sales. If this would be proved the model would be useful for coefficient estimation of optimum profit. Subsequently, two valuable ratios in the retail business were added to the initial model. As discussed in the literature review, these variables would probably make the model stronger and more reliable. The two ratios, both defined by Pedikaki et al. (2010), are present in the bottom of the conceptual model in Figure 2b. The first addition is the interrelationship is between traffic and transactions, which is defined as the conversion rate. The second one, the basket value, is the
interrelationship between transactions and sales volume. For both the basket value and the conversion rate is analysed to what extent these variables influence the outcomes profit and / or sales, and how the variables improve the basic model. Traffic is defined as the number of store visitors per hour. The conceptual model, which is displayed in Figure 2b, suggests that the planned number of employees influences all variables that are related to Profit and / or Sales. In the following chapters these interrelationships were further analysed, and with these results the most reliable model is defined. At the end, the model is used to find an optimum of the AHTraffic variable.
Figure 2a: Conceptual model
Traffic
AHTraffic
Profit / Sales
Figure 2b: Conceptual model with Basket Value and Conversion
Traffic
AHTraffic
Profit / Sales
Basket Value
Conversion
The labor hours
traffic variable is a moderator for the described interrelationships in the conceptual model. In
existing planning models, profit is restricted by daily profit and a planned number of employees per day. The new conceptual model is completely focused on labor planning variables, but other influencing
factors, as displayed in the literature overview by Figure 1, should be taken into consideration. The hourly measure of number of employees, and the hourly labor costs with their effects on conversion rates and profit are unique contributions to the existing literature. As Mani et al. (2012) admitted: “store profit data, especially at the individual hourly level, is rarely collected. Further, store profit data are often considered to be of high strategic value, so retailers tend to be reluctant in disclosing this information” (p. 10). Not only the hourly profit and the addition of a new variable are unique additional research, but also the use of data records from 130 retail stores on hourly basis are unique academic deliverables. Herein the actual hours were not measured by employee schedules as in previous research papers, but by an accurate time and attendance system that even exclude employee break punches.
4.0 Research design
4.1 Research design and data collection
To research daily flows in FLE’s retail industry data is obtained for 130 FLE stores. The outcomes of validity tests made it to decide to only include hours where the traffic was high, which is defined as traffic with at least 300 visitors per hour. This part of the dataset was used in order to examine the four hypotheses. The reason to choose for only high traffic explained in the data validation chapter (5.2). FLE is a large retail organization in athletic footwear, apparel and accessories. The organization operates over 600 stores in Europe, mainly in the big cities of fifteen different countries. The stores are located primarily in shopping malls or downtown. FLE implemented a system that collects traffic and sales data from more than 200 stores on hourly basis. It provides the opportunity to research new hourly parameters on labor planning at FLE. The internal planning system and the traffic counters at FLE were gradually implemented, so the sample size has its limitations.
The influence of the variable labor hours
traffic on the interrelationship between traffic and profit was measured in
a linear regression analysis, to find out whether this variable is a functional moderator. Subsequently, the effects of labor hours
traffic were analyzed by hourly data records, and within an Ordinary Least Squares (OLS)
regression. For example, the research by Netessine et al. (2010) is based on monthly data, where they mentioned this as one of their limitations. The data is collected by an internal labor planning system, SQL
sales databases, and store traffic counters. Currently, this data is used for key performance indicators as in daily conversion measures. By extracting these raw data files into Excel formats, it was possible to combine different variables on hourly basis. Since different files with raw data were combined, only hourly measures that included all required variables were part of the dataset. Moreover, measuring with the internal planning system and the traffic counters is an accurate method to describe the influence of
several continuous predictor variables, where the most important one is the labor hours
traffic
variable
. Herebythe outcome variable would be the profit and / or sales. Furthermore, the OLS model could be useful in analyzing the effects of several predictor variables, for example the BasketValue and the Conversion, on the optimal number of labor hours. At the end differences within groups were analyzed, so the OLS model includes all requirements for the total elaboration of this research. As mentioned before, the actual sample size was 130 stores, due to the data limitations and validity tests, which are discussed in chapter 5.2; the number of hourly records is 18.111 records. The dataset is based on 170 days and 10 working hours a day. These 18.111 records include 106 retail stores and that is enough to get a normal distribution, which is required for the use of the OLS model. The events took place in the period August 2013 to January 2014, which are the third and fourth quarter in FLE’s financial year. The dataset includes only hours where the traffic number was at least one customer. The research dataset has the highest number of hourly store data in retail compared to research papers by prior literature, which is required to find any new significant ratios in labor planning. Due to differences in closing times of the stores it is decided to include only regular daily hours that are similar for all stores. All working days of the second semester of 2013 were included within an 8:00 AM – 6:00 PM time range. Besides that, there are considerable differences in store opening times in the dataset, in particular the Sunday is an interesting exception for further research. These kinds of differences are part of the data analysis as well. Any unrealistic outliers, due to system errors, were excluded from the dataset. In this case examples of outliers are high traffic numbers without any sales, or hours with a high number of transactions but hardly any labor hours in the store. The number of records in the dataset was more than enough to find significant results about the effects of store planning behavior on profit. Hence, the 106 FLE stores are mainly divided by five primary countries (Germany, Italy, Netherlands, United Kingdom, France), and all highest sales stores per country were included in the dataset. Generally, the highest sales stores are
accommodated with traffic counters, and probably these stores would have the highest profit potential as well, which makes them attractive to analyze.
4.2 Measures and controls
As displayed in the conceptual model in Figure 2b, the main variables that were used in this study are; store traffic, labor hours (AHTraffic), conversion, basket value, sales, and profit. All variables were collected on hourly basis, and therefore the Date d and Hour h for Store i were created, where h is the first hour number in the measured hourly range. As discussed, the model is based on the principles of the sales response model created by Lam et al (2000), which was eight years later extended by Kabak et al (2008). In contrary to Chapados et al. (2013), who used traffic estimations, this sales response model defines potential store sales in terms of measured store traffic and customer response to workforce availability on weekly basis. Lam’s (2000) profit formula for hourly staff scheduling (1) is the main variable in this research. The hourly sales data records are subtracted by an average costs of goods sold, which leads to the profit margin; subsequently this number is subtracted by the actual hourly costs. The remaining amount is the dependent variable in the OLS model, which is the GrossProfitNet by staff costs. For accurate modelling a variable Traffic(t-1) is added as an extra control variable, since a part of the sales could be obtained in the follow-up hour after customers’ arrival. The Traffic(t-1) indicates the traffic number of the previous hour. The dataset has no sales data for the 7:00 AM - 8:00 AM hour range, so for the variable Traffic(t-1) the 8:00 AM - 9:00 AM hour range is excluded. As described in the literature review, with new additional variables as in ActualHours, Traffic and GrossProfitNet by staff costs, the new conceptual model will be capable of scheduling on hourly basis as well. The ActualHours variable has been deducted by all break hours, which could be easily done because all employees punch their breaks as well. As already mentioned in the previous chapter, WeekDay is another additional variable to the dataset. Normally, days in the weekends have more traffic per hour than weekdays, so WeekDay is included as categorical control of the model. In finding the proposed planning ratios all variables of Table 1 were measured on hourly basis for the 130 stores of the original sample size. Subsequently, the effects of some existing ratios could be determined as well; the hourly BasketValue and the hourly Conversion. In this way, new accurate planning ratios could be derived and used for strategic decision making and
academic planning research. The results of this analysis will provide the literature with information about which number of employees per hour generates the highest profit. In the GrossProfitNet by staff costs, the labor costs were calculated by an average labor costs per hour on store level in 2013. Since all individual stores have their own rate, the actual labor hours per shopping hour were calculated over all unique records. However, a separation in timeframe was made in order to find an optimal number of employees at certain traffic numbers, where these traffic ranges are continuous interval variables. Due to FLE standards, the low / moderate / high traffic hours were respectively 1-150 / 151-300 / 301+ visitors per hour, since there were substantial differences found by internal research in FLE’s planning behavior among the distribution patterns of these three ranges. The variable is called TrafficRange. Actually, the categorization of traffic ranges would be varying per store, but it is assumed that these ranges fitting the data very accurate, and so give reliable results. In the descriptive statistics the traffic ranges were analyzed in more detail. A separation in time frames makes it possible to compare profit patterns and create the desired ratios per traffic range. The peak hours and moderate traffic hours are of most interesting, because during these hours relatively all employees will be working in a sales function, and could be booked as direct costs on the selling units. An extra variable TrafficRangeHigh was added to create an extra cluster within the high traffic range. This extra variable was created because the high traffic data was separately analyzed in the hypotheses testing. Regarding to traffic in the current literature; by using optimum labor hours per traffic range as in planning ratios, the effects of traffic forecasts are important as well. Perdikaki et al. (2012) found that traffic uncertainty affects store performance, but traffic forecasting will be out of the scope of this research, because it certainly is another discipline. The last hypothesis is about the effects on profit by geographical areas and store type. The Country, which is an unordered categorical variable, was added to separate the data into groups. Possibly a store in Italy needs more / less employees at a certain traffic range in order to find an optimal profit peak, than a store in Spain. Since store sales volumes increased rapidly, and the traffic numbers during peak hours increased as well, the StoreSaleVolume variable was added. The ranges were created by FLE standards based on sales volumes in 2013 and separated into three groups; low, moderate and high sales volumes. These insights could directly improve the existing planning methods in order to save money (less employees) or make more profit (more employees), and adds valuable planning ratios to the
current literature of labor planning. Different store types will be defined in terms of average store sales volume, as in Mani et al (2012).
In the OLS model, which is a replicate from Netessine’s (2010), an hourly independent variable AHTraffic is defined as the ActualHours for store i at DateHour t divided by the hourly Traffic for store i at DateHour t, multiplied by 100 visitors. E.g. AHTraffic = 𝐿𝐿𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝐿𝐿𝑙𝑙𝑜𝑜𝑙𝑙𝑜𝑜𝑖𝑖𝑖𝑖
𝑇𝑇𝑙𝑙𝑙𝑙𝑡𝑡𝑡𝑡𝑖𝑖𝑡𝑡𝑖𝑖𝑖𝑖 x 100. This potential moderator is the main contribution within this research and is checked for both profit and sales outcomes. Table 1 includes a list with all variable definitions used for the data analysis.
Table 1: Variable definitions
Variable Definition
Date Categorical variable: the date d for store i.
E.g., the first day is 08-04-2013 and the last day is 02-01-2014
Hour Categorical variable: hour h for store i
WeekDay Categorical variable: the weekday w for store i on date d.
Sales Sales of FLE retail units measured in dollars for store i at date hour t
SalesVolume Categorical variable based on Sales
Transactions Number of transactions for store i at date hour t
Traffic Number of visitors for store i at date hour t
TrafficRange Categorical variable based on Traffic
TrafficRangeHigh Categorical variable based on Traffic, only where traffic is high.
Traffic(t-1) Number of visitors for store i at date hour t-1.
ActualHours Actual labor hours for store i at date hour t
AHTraffic ActualHours / Traffic ratio for store i at date hour t
GrossProfitNet by staff costs
GPN, which is the gross profit net by staff costs as defined by Lam et al. (1998) for store i at date hour t.
Conversion Proportion of customers who made a transaction for store i at date hour t
BasketValue Value in dollars of customers’ shopping basket for store i at date hour t
Country Categorical variable: country c for store i at date hour t
The next step is to set up the new OLS model for the coefficients estimation. The AHTraffic is centred to the mean for interpretation purposes. WeekDay, Hour, Traffic and SalesVolume are added as control variables in the formula to increase the model validity. Table 4 illustrated why WeekDay, Traffic and Hour are valuable controls in the model. SalesVolume is added because of the fact that peak hours in high volume stores require different labor planning compared to peak hours in a low volume store with less floor space.
The initial model (Figure 2a), based on research by Netessine & Tan (2012) was measured for the different traffic ranges, and is specified as follows:
log (𝐺𝐺𝐺𝐺𝐺𝐺𝑖𝑖𝑖𝑖) = 𝛼𝛼0+ 𝛼𝛼1𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑖𝑖𝑖𝑖 + 𝛼𝛼2𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴2𝑖𝑖𝑖𝑖+ 𝛼𝛼3𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑖𝑖𝑖𝑖 + 𝛼𝛼4𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑖𝑖(𝑖𝑖−1)+ 𝛼𝛼5Controls
The basic model is used to compare a profit model with a sales model, and to determine the optimum AHTraffic ratio. As shown in the second conceptual model (Figure 2b) BasketValue and Conversion for store i at Date d and Hour h were added to the basic model. The analysis would confirm whether or not these variable are valuable additions to the model.
log (𝐺𝐺𝐺𝐺𝐺𝐺𝑖𝑖𝑖𝑖) = 𝛼𝛼0+ 𝛼𝛼1𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑖𝑖𝑖𝑖 + 𝛼𝛼2𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴2𝑖𝑖𝑖𝑖+ 𝛼𝛼3𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑖𝑖𝑖𝑖+ 𝛼𝛼4𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑖𝑖(𝑖𝑖−1)+ ( 𝜶𝜶𝟓𝟓𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩 ) + ( 𝜶𝜶𝟓𝟓𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑩𝑩𝑪𝑪𝑩𝑩𝑪𝑪𝑪𝑪𝑪𝑪 ) + 𝛼𝛼7Controls
5.0 Results
This chapter starts with the descriptive statistics, followed by the data validation, because the correlations and data patterns were required to optimize the model. At the end, the results of the hypotheses testing were displayed.
5.1 Descriptive statistics
As explained in the previous chapter, this research used the GrossProfitNet by staff costs as the dependent variable, which is the profit margin minus the labor costs at Date d and Hour h. The profit formula ignores operational fixed costs, as in rent, utilities and yearly investments, so it is important to be sure that there is a high relationship between profit and profit margin. For this purpose the correlation between these two variables is measured for full year totals of all 130 stores. Analyzing the effect of profit margin on profit, it seemed that the R2 value is 0.68 (correlation = 0.825). Moreover if the sample size would be subtracted by a certain outlier (one single high volume store in the UK) the R2 value is even 0.78. This outlier is a store with an exceptional high cost center for operational fixed costs, and in
particular for rent. With the considerable high R2 measure and correlation the GPN, which is based on the
profit margin, may be used as an indicator of Profit. Another relationship that needs to be checked is the one between Sales and GPN. Since the variables from the sales response model would be compared to
a profit model, it is necessary to know what percentage of the profit could be explained by the sales outcomes. The result was a R2 value of 0.831 (correlation = 0.708), which says that the interrelationship is considerably high.
Table 2 presents the summary statistics of the main basic variables. On average each store hour generates $546.88 on sales with 4.47 available labor hours for an average cost rate of $113.22 per hour. The average hourly GrossProfitNet by staff costs is $169.35. The hourly sales is generated by 16.47 sold units on average, based on 10.45 transactions generated by almost 134 visitors on average.
Table 2: Summary statistics of variables
Sales ($) Units
Trans-actions Traffic Actual hours Labor costs Gross Profit Net (staff) N 182234 182234 182234 182,234 182234 182234 182234 Mean 546.88 16.47 10.45 133.75 4.47 113.22 169.65 Std. Deviation 676.07 18.48 11.632 133.89 3.83 104.22 288.66 Minimum -317 -15 0 1 0 0 -850.38 Maximum 15837 274 197 1708 47.80 1375 7721.2 1st Quarter 145 5 3 45 2 53 7.18 2nd Quarter 342 11 7 90 3.08 80 88.19 3th Quarter 698 22 14 177 5.48 135 231.07
Table 3 shows the correlations among the main variables. Compared to the summary statistics in Table 2, Conversion, BasketValue and AHTraffic were added. In order to linearize the GPN in the model the GPN was transformed into natural logarithms. The GPN has a large standard deviation relative to the mean. With this logarithmic transformation the normality increased, and the skewness is acceptable (skewness log(GPN) = -0.54). As you can observe, log(GPN) is positively associated with ActualHours (correlation = 0.459), Conversion (correlation = 0.116) and BasketValue (correlation = 0.193). The correlations among the initial predictors, which are hourly Conversion rate, BasketValue, and ActualHours, are relatively low. This is suggesting that “the predictors should not cause the multicollinearity issue in the model estimation” (Netessine, 2010).
Table 3: Correlation matrix
Variable Sales ($) Units
Trans-actions Traffic Actual Hours Con-version Basket-Value AH-Traffic Log(GPN) Sales ($) 1
Units .950** 1 Transactions .932** .966** 1 Traffic .781** .822** .863** 1 ActualHours .721** .739** .755** .737** 1 Conversion .196** .192** .180** -.102** .002** 1 BasketValue .301** .207** .125** .072** .083** .217** 1 AHTraffic -.112* -,125** -.128** -.155** -.016** .014** -.134** 1 Log(GPN) .708** .698** .674** .592** .459** .116** .193** -.304** 1
**. Correlation is significant at the 0.01 level (2-tailed).
Regarding to Table 3, an interesting correlation is the one between Conversion and Traffic, which is negative (correlation = -0.102). Surprisingly, on average the stores are not able to raise their Conversion when the Traffic is increasing, which is also displayed in the scatterplot in Figure 3a (average Conversion = 7.8%). Moreover, there is no correlation between Conversion and ActualHours on hourly basis (correlation = 0.002), and that makes it even more interesting to find the optimal ActualHours in order to increase the GrossProfitNet by staff costs. The Conversion rate does have a very low influence on Log(GPN), AHTraffic and the BasketValue. Reviewing these correlations suggests that since the influence of Conversion on Log(GPN) is very low, this variable shouldn’t be the main control parameter to manage in practice. The surprisingly low correlations between Conversion on Log(GPN) and Conversion on ActualHours made it that this variable was excluded from the hypotheses testing (this is measured in Chapter 5.2 Data Validation as well). Conversion has also a low positive influence on Sales, Units and GPN, moreover the relationships with Traffic and ActualHours are totally different than expected. Based on this decision several other variables were added as control variables in the OLS model, in order to set up a reliable estimation of coefficients of AHTraffic. AHTraffic has the highest negative correlation with Traffic, which confirms the expectation that when Traffic increases there are less employees working per store visitor. For testing differences in peak hours a separation in traffic ranges is required. In Figure 3b the relation between Traffic and ActualHours is displayed. The pattern confirms that several planning skills/strategies for store managers were included. Testing the hypotheses allows them to better schedule
their hours and that will probably lead to a more linear function and a higher correlation, but managers have to keep in mind that this correlation varies by different stores / countries.
Figure 3a & 3b: Conversion vs Traffic and ActualHours vs Traffic
The correlation between Conversion and ActualHours is very low as well (correlation = 0.002) and the pattern (Figure 4a) looks like the Conversion / Traffic pattern. Due to this reason the Conversion, as in the adjusted formula for GrossProfitNet by staff costs (2), is excluded from the OLS model, because the expectation is that it wouldn’t have any significant effects on profit. The relation between Traffic and Traffic(t-1) is very high (Figure 4b), which suggests that Traffic(t-1) could probably be excluded from the model as well. Nevertheless this variable is included as an extra control, since it was recommended by prior research (Kabak, 2008).
In Table 4 the traffic / hour distribution was displayed for average hourly traffic, including the percentage of peak hours of all hours in that specific category. Obviously, the hours in the afternoon had more peak hours than the hours in the morning. Moreover, the weekends were more crowded than the other weekdays. This corresponds with the suggestions made in the previous chapter and it makes WeekDay a valuable control variable as well. On Saturdays and Sundays in the afternoon, the percentage of peak hours is even close to 35-40% on average.
Table 4: Traffic / hour distribution
Day/hour Variable 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00
Monday Average traffic 25 42 73 94 108 126 149 159 150 126
% high peak hours 0% 0% 1% 3% 5% 8% 11% 12% 12% 9%
Tuesday Average traffic 24 36 61 82 96 109 128 137 128 106
% high peak hours 0% 0% 1% 2% 3% 5% 7% 9% 8% 6%
Wednesday Average traffic 33 39 62 81 100 124 151 161 148 122
% high peak hours 1% 1% 1% 2% 3% 6% 11% 13% 11% 8%
Thursday Average traffic 35 39 63 84 99 114 135 145 142 126
% high peak hours 2% 1% 1% 2% 4% 5% 8% 9% 9% 7%
Friday Average traffic 27 39 68 91 109 130 156 171 165 141
% high peak hours 0% 0% 1% 2% 4% 7% 11% 14% 13% 11%
Saturday Average traffic 33 62 122 165 194 244 301 318 281 241
% high peak hours 0% 1% 5% 12% 19% 28% 38% 41% 36% 30%
Sunday Average traffic 61 89 124 164 196 235 276 276 218 179
% high peak hours 1% 3% 9% 12% 20% 27% 34% 36% 30% 24%
Number of peak hours <10% Number of peak hours 10-20% Number of peak hours >20%
In order to see the differences within the distribution of the variable AHTraffic, Table 5 displays the AHTraffic by TrafficRange. In the FLE sample size, the AHTraffic ratio clearly becomes lower when the Traffic increases. During medium traffic hours, FLE used 3.0 employees for every 100 visitors on average, within the hourly data records. For peak hours, this ratio decreased to 2.5 employees / 100 visitors, which makes the main research question even more interesting. Because of this decrease of 0.5 employees per 100 visitors and due to validity tests, the high traffic hours were used for the OLS analysis to find an optimum in actuals hours per traffic range. The averages suggested that the optimum number of ActualHours during medium traffic hours would probably be lower than the current average, in order to increase the GrossProfitNet by staff costs. Finally, the conclusion of Figure 3a and 3b was that an
analysis of high traffic hours will give us the most profitable and reliable results. For these hours the correlation was higher than for low / medium traffic hours, and the number of records was high enough for using a linear regression model. This conclusion was derived from the data validation as well, and is further discussed in the data validation chapter.
The high traffic category included 9.9% of all hourly records and a considerable 32.4% of the total sales. Since a logarithm of GPN is used, 96.1% of the high traffic hours are included and all negative GPN scores were excluded. Fortunately, the part that is excluded from the regression includes most of the “system errors”. For example; days with high traffic but no sales count, which is hardly possible, are excluded.
Table 5: AHTraffic summary statistics by traffic range
AHTRaffic
TrafficRange All Low Medium High
N 182234 125954 38169 18111
Mean 6.656 8.358 3.006 2.509
Std dev. 21.46 25.612 3.932 4.364
GPN 169.66 71.69 266.80 646.25
Figure 5 illustrates that the average GPN varies a lot for different AHTraffic ranges. The optimum profit would be reached by an AHTraffic value of approximately 1–1.25 labor hours per 100 visitors, but most actual hours are in the 2-3 ranges. Since all these data records are analysed by the OLS model, the expected AHTraffic is close to ~2 labor hours per 100 traffic. These detailed illustrations were initially analysed for different traffic ranges, low / medium / high, as well. In Appendix 1 the AHTraffic vs GPN graphs is displayed for the different traffic ranges. Low traffic gives a similar pattern, the medium and high traffic ranges are flat GPN flows. As already derived from the mean analysis of AHTraffic, the majority of the hours of this variable are mainly present in the lower ranges on the x-axis. Herewith, in the high traffic cluster the GPN gradually increased by higher AHTraffic ranges.
Figure 5: AHTraffic summary statistics by traffic range
Regarding to the fourth hypothesis, in Table 6a differences in AHTraffic among three major countries were displayed. For the high traffic ranges there are no substantial differences. Observing France and Italy, there is a difference of 0.219 hours per 100 visitors, where the average of the Netherlands this is slightly 0.057 hours per 100 visitors lower than France. The major difference between France / Italy and the Netherlands is the AHTraffic for low traffic (differences are respectively 3.838 and 2.881 labor hours per 100 visitors). Moreover, on average the Netherlands has a small GPN, where the GPN is more than twice as high for France and Italy.
Table 6a: AHTraffic summary statistics within different countries
AHTRaffic
Country France NL Italy
TrafficRange Low Medium High Low Medium High Low Medium High
N 48490 14906 8276 16350 7147 3541 10296 4734 1706
Mean 7.889 3.075 2.563 11.727 2.899 2.506 8.846 3.389 2.344
Std dev. 24.535 1.698 1.448 28.834 1.466 0.893 22.807 1.272 0.837
Ave GPN 88.27 312.12 774.02 40.77 268.34 648.25 119.21 295.97 673.23
Table 6b shows that stores in the Netherlands used the highest percentage of labor on average (based on the sales), where the average Traffic was lower than in the stores in France and Italy (for high traffic). The result is a lower profit per visitor, profit (GPN)
traffic ratio. Due to these statistics the expectation is that the
decrease for stores in Netherlands (low profit
traffic ) would be larger than for the other countries to reach
optimum of profit.
Table 6b: Profits and labor costs of different countries
France NL Italy
GrossMargin (%) 51.6% 50.1% 56.3%
GPN (%) 31.4% 27.8% 35.7%
Labor costs (% of Sales) 20.2% 22.3% 20.6%
GPN / Traffic 1.68 1.52 1.63
For the analysis of the last research gap the variable TrafficRangeHigh was added to find differences among traffic ranges in the high traffic cluster (by FLE standards). With this variable the high traffic dataset is divided into a 301-400 traffic range (high – low) and a 401+ traffic range (high – high). Table 7 shows the number of records and the record distributions of both clusters.
Table 7: AHTraffic summary statistics by TrafficRange
AHTRaffic TrafficRange High High – Low (301-400) High – High (401+) N 18111 8823 9288 Mean 2.509 2.640 2.385 Std dev. 4.364 1.318 1.166 GPN 646.25 467.79 815.78 Ave Traffic 449.42 345.04 548.57
5.2 Data validation
Before the outcomes of the OLS model for profit may be interpreted, the high traffic data records need to be validated within the basic conceptual model. The low / medium traffic ranges were validated as well, but these results were not robust and therefore excluded from the hypotheses tests. Possible reasons for
