Geographic Demand Projection Model from Pilot Campaign Results

Profitability and effectiveness of a company are highly affected by accurate demand prediction, which is one of the most demanded tasks among new and developing businesses. The lack of user-friendly instruments for demand estimation motivated us to provide the study based on creating machine learning models which predict the potential demand on particular goods and services based on different geographical features, income, weather and industrial factors. The research is focused on creating the Shiny web-application to provide users with the data related to the most profitable potential locations (zip-codes) for products or services placement. The data triangulation approach is applied for including a higher number of potential significant predictors and generating a combined dataset, which allows us to generate machine learning models with good predictive power. The empirical part of the study is the Gutter Guards data case. The user data of gutter guard sales were used to test the application. The model inside the application is based on machine learning algorithms, variable importance analysis and ROC AUC estimation for accessing the quality of predictive algorithms. Results of estimated models of Gutter Guards dataset showed that the most important predictors for Gutter Guards sales are amount of advertising circulation, level of potential forest fire danger, population density, household size, age, and google search requests. The performance of the constructed predictive model is estimated by ROC AUC curve, accuracy and confusion matrix. Relationships related to demand prediction and the implication in the form of Shiny application would help companies to increase the profitability and the effectiveness of distribution based on estimated demand volumes in various geographical locations. The research we provide might be potentially adapted to different countries and regions based on the corresponding census data in the public access.