Single-period dynamic assortment planning involves the retailer's decision of selecting a set of products to offer and determining their initial inventory levels, considering stochastic demand and dynamic substitution. The objective is to maximize the expected revenue, subject to a capacity constraint on the total number of stocked products. This problem is notoriously difficult to solve, and existing literature offers effective heuristics. However, these heuristics are primarily designed for small-scale problems faced by brick-and-mortar retailers, where the capacity is generally limited due to shelf-space constraints. In contrast, online retailers often have much larger capacity levels, necessitating novel approaches. In this study, we introduce the first heuristic capable of handling large-scale dynamic assortment planning problems encountered by online retailers. Our approach is able to tackle scenarios with hundreds of thousands of customer arrivals, a capacity of thousands of units, and a wide range of product varieties. Moreover, our algorithm is relatively simple, making it more implementable in practice. To evaluate the effectiveness of our algorithm, we conduct extensive numerical experiments covering various customer types and scenarios. For small-scale problems (maximum of 200 customer arrivals), our approach achieves comparable revenue performance to the state-of-the-art approach while providing a significant speed advantage, being 30 times faster on average. As the problem size increases, the benefits of our algorithm become even more evident. For medium-sized problems (maximum of 1000 customer arrivals), our algorithm outperforms in both revenue and computational speed. While the state-of-the-art approach requires an average of 86 seconds, our algorithm produces superior revenues in just 0.3 seconds, on average. Real-world problems faced by online retailers may involve hundreds of thousands of customer arrivals, further highlighting the need for efficient and effective solutions. To illustrate the practical applicability of our approach, we apply it to a case study based on the online home goods retailer Wayfair. On average, our algorithm returns a solution in less than 80 seconds, while we had to stop the state-of-the-art algorithm after 24 hours. The revenue performance of our approach is on average 2.57 times better than that of the state-of-the-art approach in this case study
