Predicting Retail Location Performance


A nonprofit organization that operates over 630 thrift stores across the U.S. is always searching for ways to grow its retail presence and expand support in local communities.

With limited budgets, tight margins, and declining foot traffic, it became increasingly difficult to understand why some locations significantly outperformed others and where the organization should open new stores.


Intuition had historically been the leading driver in store-location decisions — an approach that often paid off. But market and economic pressures have increasingly complicated the retail landscape, and with this has come more scrutiny of existing and potential store locations.

We recognized that the organization needed a data-backed framework for validating location decisions. Our solution needed to account for the thousands of factors that impact sales as well as in-kind donations that serve as the pipeline of goods sold in the thrift stores. In other words, we needed to include data to inform both the supply and demand sides of the equation not only to understand past performance but also to predict future successes and failures.


We had several hypotheses going into this process. Fortunately, our approach involved the following simple but effective framework that identifies a clear objective and allows for collaboration and an iterative modeling process:
  • Discovery
    The process started with stakeholder interviews, alignment on the definition of a successful location, and identifying the data assets needed for the analysis.
  • Exploration
    After aggregating sales and donation data from each territory across the country, we added extensive third-party geographic, demographic, and psychographic data. Our next step was to set the market area around each location with a dynamic drive-time analysis. We then identified the top predictors of store performance and developed the Income Economic Theory, which established a statistically significant correlation between profitable locations and the balance between likely donors and buyers.
  • Model development
    After evaluating several modeling techniques, we selected a machine-learning algorithm known as a random forest model, which best predicted the location’s success. The model generated an index, which served as the primary metric to guide expansion decisions. We then created a report that combined the index with several descriptive variables about the area surrounding a location.
  • Delivery
    Initially, we developed a report for each existing retail location and ran ad hoc analyses for several potential locations. As demand for reports grew, we decided to create a custom self-service portal that allowed for location evaluation and real-time report generation.
  • Enhancement
    The model is regularly refreshed with updated retail location performance data, marketplace data, and consumer demographics.


The site evaluation tool revolutionized how the organization evaluates potential thrift store locations as well as the performance of existing stores.

Within the first month, over 300 reports were generated. Moreover, the tool aided in identifying underperforming stores and facilitated lease agreement negotiations with property managers for existing locations.