Understanding and predicting demand through visual characteristics and neural networks

Every season, online fashion retailers are confronted with the questions, which products to order and in what quantity. These questions are particularly important, because replenishment is difficult due to the short selling season and fashion products are perishable. Therefore, knowledge about typical product sales patterns and precise predictions can prevent costly over- or understocking for retailers. From a theoretical perspective, fashion products are special in the sense that they possess a social function. They are visible, inherently conspicuous, and can communicate identity and belonging. Thus, I propose that for fashion products, appearance is a key driver of consumers’ purchase decision. However, we lack knowledge about a detailed set of visual characteristics and their relation to demand (i. e., sales and returns), which could help us understand consumers’ purchase decisions better. To fill this void, I determine a product’s design typicality and brand prominence based on the theories of processing fluency and conspicuous consumption. Then, I relate these visual characteristics to typical product sales patterns to understand how product appearance and sales patterns are related. Existing measurement approaches of visual characteristics mostly rely on manual procedures restraining research to small datasets. I therefore make use of state-of-the-art neural networks to extract visual characteristics. Thereby, I propose an automated measurement approach to determine the brand prominence of products, which was previously measured by human raters. My approach makes questions concerned with the conspicuousness of brands applicable to large-scale datatsets. Further, I analyze whether including a product’s appearance in sales or return predictions can make predictions more precise. To this end, I operationalize appearance by extracting image features from convolutional neural networks on top of a product’s design typicality and brand prominence. I draw on regression and various machine learning algorithms to predict sales and returns and find that visual characteristics have predictive power for sales but cannot help explain or predict product returns. Building on studies that analyze the importance of vision and touch in the purchase process, I propose that touch is the key driver to explain returns and not vision.