Marketing analytics for sustainable & efficient retail operations

Perishable products, such as food, clothes, and flight tickets, play an important role in major global industries. Their time-sensitive sales cycle and deteriorating value are particularly challenging in retail operations. Early insights from data are essential to inform decisions early in their lifetime, which are the most impactful. However, due to the natural trade-off between data availability and the product's value over time, managers face significant challenges in ordering, stocking, and pricing decisions. The present thesis investigates how marketing analytics can better inform management decisions for perishable products to allocate their resources efficiently & sustainably. The first chapter presents a model for size preferences of seasonal fashion products. The model overcomes data challenges by incorporating historical data and expert knowledge into a hierarchical Bayesian structure. Fine-tuning determines the optimal model structure and balances different data sources. Performance analyses show that the model is more accurate, less biased, and more profitable than a management heuristic. Overall, it shifts 15.6% of our industry partner's inventory from overstocked to understocked sizes to optimize their pre-season ordering and stocking decisions. The second chapter introduces a unique longitudinal online grocery marketing mix and expiration date-based dynamic pricing data-set. It uniquely contains residual shelf lives of dynamically discounted food with a lot of variety, which is expensive to collect observationally or experimentally. The data-set can be used to study consumers' price-freshness trade-off, expiration date-based pricing, and grocery inflation. The third chapter reviews deep generative models for synthetic data. Retailers face many data challenges when applying marketing analytics techniques, including scarcity, lack of variety, and privacy. Synthetic data generation can help with these challenges, e.g., by unlocking restricted data-sets and augmenting insufficient data.