As an international quick-service restaurant and one of the largest restaurant chains in the world, our sponsor has more than 18 thousand stores in over 80 countries. Their business model depends on its ability to provide high-quality products to customers quickly and efficiently. With the rapid change and the increased rivalry in the industry, they have re-strategize themselves in order to remain competitive and continue the growth. Their business development team is shifting their focus towards “Delco”, which represents delivery and takeaway services. Their aim is to open more Delco outlets in the future. Hence, we would require a systematic way of selecting potential sites and visualisation tools to analyse their store performance. Our proposed solution will be a visualisation dashboard which is built based on Tableau and a statistical and machine learning model for the potential site’s recommendation.
Organisations and corporate giants are always finding innovative ways to maximise business profits (Kumar, 2013). To create competitive advantage, many companies have practised various data mining techniques in areas such as customer relationship management (CRM) (Chen, D., Sain, S. & Guo, K., 2012), profit prediction (SAS Institute Inc., 2011) and etc. Less are aware of the value of adopting location intelligence. With the use of spatial analytic techniques, companies are able to automate the process of store location selection while optimizing the amount of cost invested in opening outlets in a region (Kantor, 2019).
Re-strategizing to Delco
Our sponsor in the Philippines has excelled in providing multiple customer services namely dine-in, takeaway and delivery services. However, the dine-in regulations placed during the pandemic phase this year have greatly affected the company’s business revenue which heavily depended on sales from dining services. On the other hand, the dine-in outlets are worn-out and lack maintenance to provide adequate dining services. Extra maintenance time and cost are required to elevate the customer experience. Considering both external and internal concerns, they are shifting their focus to expanding business scope in customer services other than dine-in, such as “Delco” which represents a combination of delivery and takeaway services.
Current challenges and goals
Our sponsor has diverted investments in opening up more Delco outlets. However, the entire process of outlet location selection is carried out manually by the business development team. The manual process has incurred a series of serious issues:
- As the entire process is assessed and evaluated totally based on personnel experience, it could lead to misevaluation and misjudgement of strategic decisions which could result in huge business losses.
- The workflow and evaluation steps which are based on personnel experience are inconsistent and hard to pass on to new staff should there be any change in team formation.
On the other hand, the regional operations team wants to further understand and monitor the opening of new Delco outlets. They are interested in identifying interesting sales performance trends over time to assess the success of their strategic decisions. However, they faced some difficulties during the process:
- There are no common metrics with standard definitions which the team can use to visualise and compare sales performance across time.
- There is a lack of systematic and professional way to visualise and analyse the data, which makes it difficult for the team to make meaningful conclusions during the process.
- The team is still at an exploring stage to experiment with different analytical tools, they are open to trying out new analytics tools to handle their demands.
As such, our sponsor is seeking a cost-effective and innovative solution which enables them to:
- Accurately choose the optimal location for opening up new Delco outlets by analysing data which determines the success of opening Delco outlets.
- Conveniently analyse sales performance data in a timely manner which allows the team to visualize interesting trends, identify potential weaknesses and business opportunities.
In order to tackle existing challenges faced by our sponsor, we proposed to use Tableau as the analytical tool to generate visualisations of sales performance at a deeper level. The dashboard generated would allow them to visually track, analyse and understand the key performance indicators of their sales.
On the other hand, we would like to build a model using machine learning to predict the optimal location for opening Delco outlets after analysing information such as trade area and outlet sales performance.
Potential business benefits
By using the Tableau dashboards, the regional operations team is able to:
- Combine and visualize data from multiple data sources in different formats such as CSV, shop, xlsx etc. in an easy and efficient way.
- Easily toggle with the dashboard which supports interactive and dynamic experience.
- Compare user experience of both Tableau and internal instituted visualisation software to choose the most suitable tool which suits their needs.
By using the machine learning model, the business development team is able to:
- Receive recommendations on locations to open Delco outlets with less consumption of human resources, time and effort.
- Access to a highly customised model that completely suit the needs of the team which is looking for an optimal solution in an urgent situation
Solution Concept and Implementation
Our solutions consist of a decision model which provides the recommendations on potential Delco sites and a dashboard which provides the visualization of store sales performance and analysis.
The model was built based on statistical learning and machine learning. There are three main steps needed for the model building process:
- Trade area analysis
Theoretical trade areas will be generated for each Delco store. The current stores’ performance within each store’s trade area such as average weekly sales performance which will be used as the performance indicator for model building. Additional data such as estimated population, road density and point of interest for each trade area are calculated which will be used as the variables.
2. Model building
Decision tree which is a machine learning algorithm that transforms the data into a tree representation is used for the final model building. The model takes in all the variables that were generated previously and creates a tree that indicates the decision rules for potential sites selection.
3. Potential sites selection
Trade area for each municipality is created with an estimated population, road density and point of interest calculated. Potential sites will be selected based on the decision rules generated previously. The output will be both new and in field sites suggestion. For each potential site, further analysis such as visual analysis using QGIS is used to check the validity. Redistricting is applied to the infield recommendations which is to ensure the population and demand are equally split for the current store and new store. This is to ensure that the current store and new store have different trade areas so that their performance will not be affected by each other.
There isn’t any systematic way of new store selection for our sponsor. It is not feasible to copy existing models available on the market because their demand is specifically focused on Delco. Hence, it requires a high level of customization. Our model is better than their existing method because it provides recommendations with support from statistical evidence and tailor to their demands. Also, it can be shared or passed down to others in their team.
Tableau Sales Analysis Dashboard
The sales data given by our sponsor consists of store name, store address, type of facilities, type of assets and past five years of sales data. And the tableau dashboard uses the sales data given to do analysis in the following perspectives:
- Overall Sales Performance
This page is built to analyse and compare overall sales performance. It has split the data into sales performance by facility and sales performance by asset types. To further analyse the sales performance, apart from the annual sales, this page also includes the growth of different types of sales from the previous year.
2. Sales Performance by Location
This page aims for further analyse sales performance in different location levels, such as legal class, regions and cities. The dashboard will automatically calculate the average weekly sales performance by selected location level and store count in such location level from 2015 to 2019. The map also reflects the store location at each location level.
3. Each Store Performance
This page is created for a further breakdown of the sales performance of different stores. It includes the overall sales performance and average weekly sales performance. For better understanding, the sales performance of the store, the reference line of overall average sales performance and KPI has been added for comparison purposes.
Our sponsor originally planned to engage an external vendor to build the dashboard using Power BI. And we shared the benefits of using Tableau, including Tableau able to handle huge volumes of data compared to Power BI which only able to handle the limited volume. And Tableau has access to more data sources and servers than Power BI.
Challenges, Insights and Lessons Learnt
The decision tree model we built has achieved 96% accuracy and 95% F1 score. And it also retrieved 33 out of 35 true positives and 39 out of 40 true negatives. According to this, we believe that our suggestion of potential sites for new Delco outlets is trustworthy. And our sponsor could build their decision of new Delco outlet locations based on our model as well.
Apart from that, our dashboard contains sales analysis in three different perspectives which allows our sponsor to easily analyse the different levels of the sales performance.
Our sponsor has commended that using Tableau dashboard is a good initiative. They were also extremely impressed with our visualisation of trends, patterns and insights of sales performance analysis. They considered the dashboard as a great way for them to compare against other visualizations tools they used.
Other than visualization, they also found the enquiry on the feasibility of the sales target excellent. This helps them to discover that many stores were not able to reach the original sales target.
In addition, they have declared that we put in a great effort in deciphering complicated real-life problems with imperfect data as we managed to find alternative solutions.
Throughout the project, we have encountered many challenges. The trade area data provided to us is in image form which requires additional steps to digitise it before we can use it for further analysis. This is out of our expectation as digitizing takes up a huge portion of our time. Besides that, there are key data missing in the dataset such as incomplete data trade area boundaries and trade zone boundaries. There are also multiple stores within the same trade area. Due to COVID-19, it is impossible for our sponsor to update the data within a short period of time. Hence, we have to think about an alternative plan which is to use the theoretical trade area instead of the practical trade area. Nonetheless, we still carried on with the digitizing of the Delco trade area for the visualization in Tableau.
Moreover, this is our first time building the model, we have explored a few models in order to discover the most suitable model for this project. We build a logistic model at the start, but the results are not ideal and we proceed to try out other models such as decision tree models. During exploration, all of us have to learn new software and programming languages. For example, we built the logistic regression via BlueSky Statistics and decision tree via R language. Majority of us have no experience in them and have to spend additional time and efforts in learning.
Although we have experiences in Tableau, most of us did not take courses about visual analysis. Hence, we have difficulties in building the dashboard in the beginning as we need to create many dashboard versions for the comparison.
The sustainability of our dashboard could be one of the limitations. Since the dashboard is not linked to real-time data, this will require users to re-import updated sales data if further analysis is required.
Besides, since all our analysis and model building are based on theoretical trade areas, there may be differences between the model outputs and reality. This is because the theoretical trade area did not include local contexts such as people’s lifestyle and preference. Besides that, the trade areas represent all the asset types such as restaurants and express. Moreover, the model only focused on Delco stores which may hinder its accuracy as the trade area is used for factor analysis. These limitations can be mitigated by replacing the theoretical trade areas with practical trade areas in the future, especially for trade areas that are only for delivery and takeaway services.
Chen, D., Sain, S. & Guo, K. (2012, August 27). Data mining for the online retail industry: A case study of RFM model-based customer segmentation using data mining. https://link.springer.com/article/10.1057/dbm.2012.17
Kantor, M. (2019, January 30). How Companies Use Location Intelligence To Delight Customers, Fuel Growth. https://www.forbes.com/sites/esri/2019/01/30/how-companies-use-location-intelligence-to-delight-customers-fuel-growth/?sh=a451cae637e6
Kumar, T. P. (2013, October 7). How businesses use real-time location-based data to build competitive advantage? https://www.geoawesomeness.com/how-businesses-use-real-time-location-based-data-to-build-competitive-advantage/
SAS Institute Inc. (2011). Increase Customer Profitability Using Data Mining and Advanced Analytics. https://www.slideshare.net/chrandall/increase-profitability-using-data-mining