Most chain store modeling experts will tell you that a “good” sales forecasting model will estimate sales +/- 20% in 80-90% of the cases.
Most chain store real estate dealmakers believe that they need a model with no more than +/- 15% error 85% of the time.
Most people don’t agree on how this error is measured or what the role of human judgment should be in determining the “official” sales estimate used in calculating the projected return on investment.
Everyone wants to “Keep It Simple Stupid” because it’s hard to make decisions when you are confused about the facts or their implications. This definitely applies to real estate planning and site selection for chain store operators. However, most sales forecasting models are anything but simple and are often intimidating to those without strong backgrounds in statistics (which includes the CEO, CFO, and VP Real Estate).
There is a huge push these days to use technology and mathematical models to increase the quality of business decisions. From the rigorous discipline of “Six Sigma” in the late 80’s to the recent business analytics wins of companies such as Capital One and Harrah’s, there seems to be an unbridled confidence in the application of computers and statistics to financial analysis.
The problem is that some situations cannot be modeled with enough precision to be useful.
Chain store sales forecasting is one example.
The reason is simple: historical data about the retail marketplace are not static and therefore cannot be used to reliably estimate future sales.
A Framework for Complexity
Let’s consider some different decisions that face chain store operators ranging from simple to complex.
A simple problem is one that can be reduced to an equation and applied repeatedly with very similar results. An example would be the selection of the size of a steel beam to support a roof in a building. The force of gravity is consistent and can be used to compute the load requirements of structural steel. Even if the equation is complicated (to those who are not structural engineers), it is simple, straightforward, and reliable.
A complicated problem is one in which the relationship between cause and effect requires analysis and expertise. Many business problems fall into this category such as staffing for checkout lines to minimize wait times for customers, logistics for deliveries in the supply chain, and inventory management based on seasonality of demand. In these cases, historical data provide a reasonable basis for predictive models and can provide a solid foundation for planning and investment decisions.
A complex problem consists of a situation where the relationship between cause and effect can only be determined in retrospect, not in advance. This is due to large number of variables that influence the outcomes, the changing values of these variables, and the non-linear interactions among the variables.
Chain store sales forecasting is a complex problem.
Although our use of statistical models in sales forecasting has outstripped its usefulness, it would be a mistake to simply revert to “gut feel.” The chain store industry has a great opportunity to build upon the advances in technology, data, and analytical methods and create a new approach that uses the best of “art and science.”
The best tool for integrating art and science in real estate decisions is the oldest tool: analogs.
Analogs allow decision-makers to look at a new opportunity, find similar situations from past experience, and use them as a guide to estimating the future performance of trade areas and sites. Computers and market data can be used to present the “patterns” for comparison and the human brain can be used to assess the similarity of the analogs and adapt them to the new situation.
In chain store sales forecasting, the analog method was first formalized by William Applebaum in the 1930’s. Since then a vast array of methods have been used to create classification schemes for markets, trade areas, stores, competitors, and customers. The frustration of this effort is that no two entities are exactly alike, and any attempt to fit them into a scheme will result in a large number of cases near the boundaries of the categories. For example, let’s say that we define “urban” stores as those with a a population density of 5,000 people per square mile within a 2 mile radius. Does that mean that a store with 4,999 people per square mile is not urban?
The computer can easily compute population density for any location in a second; a human being can’t do this in a year. However, a human can look at a map of an area and instantly classify it based on a variety of attributes: its density, proximity to major highways, the presence of retail activity, traffic congestion, and relationship to surrounding cities and towns; a task that a computer program would find daunting, generating comical results in many cases.
Over the next few months we will further explore some new ways to integrate art and science for better real estate planning and site selection.
Interesting take, and to a large degree, I agree 🙂
One reason computers find looking “at a map of an area and instantly classify(ing) it based on a variety of attributes: its density, proximity to major highways, the presence of retail activity, traffic congestion, and relationship to surrounding cities and towns;” a daunting task is because the researcher/programmer often fails in translating those attributes into numeric data.
If we correctly coded this information, I am confident that clustering and neural net algorithms would outperform human projections. One of the reasons that is a big “If” is because, as you mention, the retail environment is constantly changing. Not only would one have to code the information correctly, but then the data would need to be routinely updated (with more precision than the US census for example).
I came across an interesting paper that compares/contrasts analog methods with other statistical techniques: A New Perspective on Forecasting Store Sales: Applying Statistical Models and Techniques in the Analog Approach. Again, I agree with you that such techniques will only be as useful as your data is accurate.
apologies if this remark is too long,
The problem is one that is truely complex but not insurmountable with the proper use of a combination of different tools. The suggestion by the other commentary is that data accuracy is paramount, and I would completely agree. Models are only as good as the data that is put into them, but I find it rather interesting that your article discards the use of model for a single (albeit a good tool), analogs. I would argue that the combination of the two tools along with solid data and experience will bridge the gaps. At the very least, a model should never be used in a “push button” fashion. There is a lot of knowledge and understanding that is required to run these models successfully. There isn’t a single tool out there that should be considered the best; it is a combination of tools; an intergrated approach that will move the needle. There is no push button solution out there and no one tool is the best solution.
Brian, I completely agree with your points about using more than one tool and the perils of “pushbutton” solutions. I was actually referring to “pushbutton” models when I said that they have outlived their usefulness. I’m sure that you know that a lot of companies want a “pushbutton” solution and consider anything less to be a “workaround” rather than a blend of art and science. The best decision processes take advantage of mathematical models, analog comparison, analyst judgment and synthesis of the results of tools and field work, and senior executives asking tough questions that assess risk and return in investment committee meetings. And the best possible data, of course.