People Do Not Know What They Want Until You Show it to Them. But When

Recommendation systems are critical tools for online retailers in their pursuit of enhanced operational performance and improved shopper experience. As such, firms invest heavily to improve their algorithms. Notwithstanding these efforts, there is usually a serious omission of demand-timing element in prevalent online recommendation systems. As a result, recommendations are often presented out of synchronization with the next demand cycle, leading to squandered marketing opportunities and customer dissatisfaction. In this research, we propose a novel demand-driven recommendation system that factors in predicted demand timing. The core of our novel design consists of a predictive model that forecasts product-level repurchase cycles for the online retail environment. We propose a new approach to incorporate the predicted repurchase cycles into three key recommendation generating stages: (i) \textit{retrieval}, (ii) \textit{ranking}, and (iii) \textit{re-ranking}. Using large-scale online experiments, we demonstrate that this novel demand-driven approach outperforms the prevalent recency-based recommendation system in two of the three stages, resulting in higher recommendation click value rate (CVR), higher revenue per mille (RPM), and improved customer satisfaction. Our study contributes to the growing literature on recommendation system design in general and recommendation timing research in particular. Further, our predictive model of the product-level repurchase cycle is novel for the online retail environment and can serve as the basis for improvements and planning of many business decisions. Our research findings are actionable and impactful to online retailers in their pursuits of revenue growth via efficient recommendation-system design. Due to its proven effectiveness, our recommendation system has been implemented by a large online retailer in its commercial platform.