Artificial Intelligence in Retail Site Selection Building Smart Retail Performance Models to Increase Forecast Accuracy By Richard M. Fenker, Ph.D. Abstract The term Artificial Intelligence or AI has been used since the 1960s to describe the programmed intelligence contained in machines that think. While many of the early attempts to build AI systems were unsuccessful, today s combination of fast computers, unlimited memory and statistical methods that complement the logic in AI systems has led to many successes. Today, AI programs can land a passenger jet without help, track weather systems around the world, beat human opponents in chess, recognize speech or handwriting, determine battle strategies, and even perform various types of surgery. At Experian Business Strategies, AI methods are also being used to build simulation systems that help restaurant and retail chains make better decisions about their locations, marketing and operations.
What AI Adds to Conventional Statistical and GIS Models Companies greatly increased their use of statistical and GIS models in the 1990s, when personal computers and the data to do modeling (such as demographics) became readily available. These models typically looked back at the history of a company s performance and then attempted to forecast this history using various types of statistical methods. While often useful, these models lacked the ability to see through the statistical noise created by the differences in operations, marketing, building style or age, name recognition, and the many other factors that can influence sales. The Prediction approach to this problem was to: Study the behavior of customers to create a logic or set of AI rules to describe how the world works for a particular concept. Combine AI rules with knowledge from statistical or GIS-based analysis to create predictive models. Integrate these models into computer systems that simulate the retail world of a particular concept and can be used for business functions such as market planning, site selection and sales forecasting. The biggest advantage of this approach is that it makes intelligent use of statistics and GIS while AI-based logic keeps the models from being exposed to the statistical noise that is always present in real-world situations. In a conventional approach, the researcher would need to see that good visibility correlates with high sales in order to include this variable in a predictive model. Experian Business Strategies uses Artificial Intelligence to overcome the limitations that are inherent in conventional modeling approaches. Leveraging AI, we begin with a rule-based logic that knows visibility is important. But we also know that this importance differs in mall locations, urban locations, suburban locations and so forth. The rules or logic, similar to those of a real estate expert, guide the decision-making based on these nuances helped by the statistics but not determined by them. Artificial Intelligence Helps Filter the Noise How Statistics and AI Combine to Create the Best Models For nearly 30 years, members of the Prediction team developed Retail Performance Models, using information obtained from nearly 3 million consumers about how they use the retail world. Prediction leveraged the following key findings from this research to substantially increase model accuracy and reliability: 1. There is a lot of consistency in how people use the retail world for all concepts. If you divide concepts into three classes destination, convenience, and in-between you already know a great deal about the rules of the world. Page 2 of 8
2. When you get to a particular type of concept such as a casual theme restaurant, convenience store or bigbox discount store, there is a great deal of similarity in the logical rules that describe what matters as you plan markets or attempt to forecast sales. 3. In building modeling solutions for different categories of retail, Prediction begins with the logic foundation already in place. The points of uniqueness tend to be details around the customers, what competitors those customers use and how they use a specific concept. In the case of a casual theme restaurant, for example, the presence or absence of a bar changes usage patterns significantly. 4. Each retail situation has a somewhat different set of rules. We use retail situation to describe the setting for a store, such as a mall, strip center, highway location, downtown or urban location, and so forth. A large concept such as Blockbuster or McDonald s may have 16 different retail situations, as generally every chain has five to 10 varying situations. Because the samples for most concepts are not large enough, depending on statistics to study the rules that operate in every situation is impossible. If you had to depend on statistics to study the rules that operate in every situation, you cannot do it well because you just don t have large enough samples for most concepts. However, Prediction starts with detailed, logical knowledge about each type of setting based on its research. 5. Finally, there is no such thing as a simple sales prediction, just as there is no simple weather forecast. A model doesn t predict sales, it evaluates the conditions for sales and then assumes that, if your market development strategy, marketing and advertising, operations, competitive positioning and all of the other non-location factors are consistent with what is happening in your historical locations, then the location factors will probably give us an expected sales of $N. If any of these other conditions changes, as you know, sales can fluctuate. Prediction s philosophy is to use all the sophistication possible to make an informed sales prediction, and surround this prediction with logic-based information so you can put the sales forecast into the appropriate context for making the best decision. Incorporating AI through CHARMS Chaotic Heuristic Automatic Recursive Modeling System CHARMS is an AI-based system that is designed to increase the accuracy of retail sales forecasts. Not only can CHARMS evaluate hundreds of sales prediction equations simultaneously, it is also a dynamic, adaptive model. Unlike most other retail real estate models, CHARMS learns from information you add for new sites and adjusts its output accordingly. The choice of equations also changes with each site, giving you the best sales forecast for the unique circumstances of that site. When CHARMS is added to the collection of models within Prediction s Retail Performance Modeling ( RPM ) solutions, it automatically selects the best models to use for each sales forecast. It is a little like the sophisticated combustion systems in cars today that adjust the valves, the fuel mixture, the spark and the timing to compensate for any driving situation. CHARMS is designed not just to use the multiple models in the RPM platform, but to fine tune these models in real time for every prediction. Page 3 of 8
First, data mining techniques are used to identify site and demographic variables that classify stores in terms of sales. Using neural networking analyses, a sales prediction equation is created for each level of each classification variable as well as for all crosses of the levels of those variables. As a very simple example, let s consider store size and type of retail setting as classification variables. Within the CHARMS logic, an individual sales prediction equation would be constructed for each level of store size, for each type of retail situation, and for each combination of size and retail setting. The AI logic associated with each equation gives it permission to recursively consider, by reviewing the whole history of the company and all of its stores, how relevant each combination might be for the current store being evaluated. This is similar to how your mind might work as you quickly review an entire history of times you were in a conflict with your spouse and mentioned her mother before you decided to bring up this touchy subject in the current argument. That is how CHARMS processes a detailed and complex history of any retailer to bring that knowledge into the moment for better decisions. Page 4 of 8
The intelligence in CHARMS was strongly influenced by recent developments in nanotechnology, in which it is possible to create a number of smart, tiny nano-units, each of which are intelligent in some specific way. And, much as human cells combine to form muscles, bone androgens, the nano-units in CHARMS combine to form successively more intelligent and larger views of the retail world until eventually they can make a prediction about sales or an optimum market strategy. Unlike simple statistical models that operate with fixed equations and rules, the nano-units in CHARMS organize themselves in real time when a problem (a site to be evaluated) is presented. This creates a unique prediction model designed to solve that specific problem. The next site that is evaluated will most likely have a completely different model. Putting this in the context of retail forecasting, each nano-unit creates its own forecast. The next-level units in CHARMS evaluate the reliability and validity of each sales forecast. CHARMS then weights each forecast based on how similar it is to the current situation described by the nano-unit. Situations with more recursive history receive more weight. Eventually, another set of higher-level nano-units will evaluate all of the prediction results and select the combination of predictions most likely to work for the site being evaluated. An oversimplification of this process would be to say that CHARMS uses a reliability/validity analysis to weight each sales prediction, and makes a final sales prediction based on the overall weighted average. Actually, the recursive properties of CHARMS, described above, mimic human thinking in that, in the moment, CHARMS has access to previous decisions made in similar situations and how well they worked. Essentially, CHARMS uses rules that simulate how an expert in real estate and statistics, who had a complete understanding of your concept and all of your sites, would evaluate them. Case Study AI Delivers a Significant Increase in Sales Forecasting Accuracy Traditional modeling approaches are not able to capture the complexity of the retail world any more than measuring temperature, atmospheric pressure and wind speed are able to produce accurate weather forecasts. Both the retail environment and weather forecasting require complex, AI-based simulation models that generate a variety of forecasts, then use special logic or different starting assumptions to narrow these many predictions into a single best forecast. CHARMS uses AI agents to simulate a variety of retail conditions for any site being evaluated, then selects the most likely set of conditions for the site and uses these to generate a final prediction. Different modeling approaches, markets, retail settings, store types, competitive positions, product offerings or other complex factors are handled seamlessly and automatically within the CHARMS system to produce the most accurate forecast. A large, global retail client spent more than $1 million using conventional modeling approaches to build a sales forecasting system. Working with outside consultants, this company measured over 100 demographic factors for each site and then added a variety of site features such as visibility and access. All of these measures were combined within a regression analysis to forecast sales. Page 5 of 8
The client s approach resulted in an R2 of 0.34 and an average absolute sales forecasting error rate of 21% (Figure 1a). Scatter plots show the relationship between predicted and actual sales in a graphical format. ( Outliers have been removed for purposes of this illustration.) The better the prediction, the closer the scatter plot will represent a straight line. Figure 1a Regression Model The client then turned to Experian Business Strategies for our expertise in data aggregation and modeling techniques, to help them improve the accuracy of the company s forecasting approach. Initially, the Prediction team added a gravity-based model and built different demand models for each retail setting. This resulted in an R2 of 0.53 and an absolute average error rate of 12.6% (Figure 1b). Figure 1b Gravity Model Page 6 of 8
Finally, the team applied CHARMS to the forecasting system, achieving an R2 of 0.89 and an absolute average error rate of 5% (Figure 1c). Figure 1c CHARMS Model This example illustrates how an AI-based approach is able to clean-up much of the statistical noise in conventional modeling approaches by separating the different prediction components, building multiple models to explain the behavior of each component, and then combining the results to deliver a much more accurate result. On average, when CHARMS is added to conventional modeling approaches such as a regression analysis or gravity models, the improvement in prediction capability is 20% to 40%. Benefits Leveraging AI helps untangle the complexity of the many variables influencing sales to give the best possible assessment of potential store performance. Companies that have made the shift toward combining statistics, GIS and AI for location decisions not just financial estimates have seen a tremendous benefit on many dimensions, not the least of which is their bottom line. One of the simplest statistics that illustrates the value of this approach is sales volume. In an analysis of more than 7,000 locations for 13 different concepts evaluated using CHARMS, locations that had a Site Quality rating of 65 or higher (the average rating is 50 on a 100-point scale) also had 17% higher sales than the average store. Reducing the risk of opening an underperforming store is another key benefit associated with this approach. Clients have estimated that the total cost to close a store can be as much as three times the cost of opening a store. When you combine the hard-dollar costs left on the books with the opportunity costs and negative impact on the brand, the risk/reward equation favors an investment in an AI-based modeling approach to increase the probability of accurately predicting the future. Page 7 of 8
To find out more At Experian Business Strategies, we believe real estate investments are one of the most costly decisions a retailer can make, and making the best site selections for your business is a critical factor that will drive profit and success. To find out more about how we can help, contact Debbie Diot, Vice President of Sales at 972-874-5858. To access our full series of white papers, visit the Resource page of our website at www.experianstrategies.com. About the author Richard M. Fenker is Global Product Architect for Experian Business Strategies. Dr. Fenker holds a B.S. in Astronomy/Mathematics and a PhD in Mathematical Psychology. Dr. Fenker blends his knowledge and experience in programming and applied mathematics, demographics and GIS, research methodology, and over 25 years of modeling the behavior of consumers with his love for solving complex problems. He is also the author of The Site Book, the best selling guide to retail site selection. Most of his work today involves a mixture of consulting, the design and implementation of real-world modeling systems and writing. Page 8 of 8