Estimating Patronage at Dunkin’
How can we predict patron demand origins to Dunkin’ Donut shops in Cambridge and Somerville, MA?
Thanks to the kindness of a Boston-based Dunkin’, this project utilized hourly transaction counts for all Dunkin’ chains in Cambridge and Somerville, MA. The goal of this study was to speculate where Dunkin’ customers might be originating from based on transaction counts and census data. All work was completed by a cross disciplinary MIT-Harvard team consisting of: Firas Suqi, Marco Miotti, Carmelo Ignaccolo, and Adham Kalila.
Average Hourly Patronage of Dunkin’ Shops in Cambridge & Somerville
The project started by visualizing all hourly transaction data, and narrowing in on the 8am rush hour values to calibrate the model. After utilizing the Find Patronage tool of the UNA Toolbox for Rhino 3d, the team made assumptions about customer origin points and simulated these customers on routes to Dunkin’. Our simulation model was further calibrated by taking into account other demand points that customers might have stopped at along the way to Dunkin’.
Since the results of our pedestrian model were a bit underwhelming, we decided to factor car trips into account when considering patronage. To do so, we created a proxy for car trips that was based on data scraped from Google Maps. The Road Access Measure (based on congestion, number of lanes, a Large Road Index (LRI), and whether each chain had dedicated parking and a drive-thru) was factored in as a destination weight for each Dunkin’ Donuts location. The results then began to tell us something about where Dunkin’s customers are coming from.
Road Access Measure
Calibrating Destination Weights
Conclusion
In our findings, we discovered that calibrating our model based on origin weights alone cannot fully determine patronage at Dunkin’ Donuts locations in Cambridge and Somerville. Being bedroom suburbs of Boston with higher rates of automobile ownership, we found that that indicators of road access proved to increase the reliability of our patronage model significantly. Our results could be further validated by differentiating the data we received from Dunkin’ by drive-thru transactions.
In sum, this project was an interesting exploration combining spatial analytics with corporate data to better understand how location impacts retail patronage. In taking this study one step further, we can use our findings to propose a location that will optimize patronage counts given our model’s parameters, should Dunkin’ choose to expand in the area. For now, the model returns an r-squared of .89, meaning we did a pretty good job determining where Dunkin’s customers are originating.
Correlation Results
Individual Correlations of O+D Pairs
