Direct Marketing Response Models Using Genetic Algorithms

Direct marketing response models seek to identify individuals most likely to respond to marketing solicitations. Such models are commonly evaluated on classification accuracy and some measure of fit-to-data. Given large customer files and budgetary limitations, only a fraction of the total file is typically selected for mailing promotional material. This desired mailing-depth presents potentially useful information that is not considered by conventional methods. This paper presents a genetic algorithm based approach for developing response models aimed at maximizing performance at the desired mailing depth. Here, depth of file information is explicitly taken into account during model development. Two modeling objectives, response maximization at selected mailing depth and fit-to-data, are considered and tradeoffs amongst these empirically explored. Resampling approaches, effective for controlling overfit to training data, are also investigated.