New product demand forecasting and distribution optimization : a case study at Zara

Abstract

The problem of optimally distributing new products is common to many companies and industries. This thesis describes how this challenge was addressed at Zara, a leading retailer in the "fast fashion" industry. The thesis discusses the development and evaluation of a modular system including distributional demand forecasting and dynamic programming distribution optimization. The demand forecasting module combined the practice of using similar products to predict the demand of a new product with a new store or customer cluster data aggregation scheme. Moreover, distributional forecasts were generated using a generic distribution of the expected relative forecast error constructed based on historical forecast performance. Finally, an empirical study of expert or qualitative forecasting within Zara was performed to evaluate the potential for forecast improvement. The distribution optimization module leveraged the distributional forecasts and dynamic programming to determine the optimal initial shipment quantities. The dynamic program directly accounted for the inventory constraints as well as the information dynamics that result from the improvement in forecast accuracy after initial sales are observed. The complete system was validated using extensive simulation. Overall, the new demand forecast reduced forecasting error by over 30% and the final simulation results showed that the overall system would be expected to improve initial sales by over 12%. Given Zara's scale, these results would translate to hundreds of millions in additional profit. Thus, a live pilot was approved and initiated by Zara with the goal of confirming the simulated impact of the system under real conditions. Assuming a successful pilot, full system implementation is expected in 2011.