Demand Forecasting for Size Curve Optimisation in Luxury Retail

Retail firms face considerable issues as a result of product overstock and under stocking, especially in the context of fundamental tourism dynamics and seasonal bias. These problems can lead to missed sales opportunities, lower client satisfaction, and increased business expenditures. Intrinsic tourism characteristics, such as changes in visitor numbers and spending habits, can make it difficult for shops to precisely forecast demand. Seasonal biases, such as higher demand for certain items at various periods of the year, affect inventory management even more. To effectively address these difficulties, retailers must establish strategic inventory management procedures and use data analytics to estimate demand and improve product offers. An effective size distribution is a crucial element in the management of inventory levels, the reduction of waste, and the improvement of customer satisfaction. That's why this thesis tackles this issue by introducing a data driven technique to select the best size mix in replenishment phase. The methodology applies an analytical approach to reveal size mix trends from historical sales and inventory data. Specifically, the technique examines sales data to detect trends in customer demand for different product sizes. It also takes into account stock data to assess the availability of different sizes of inventory as well as the proportion of inventory sold to consumers. Initially, key stakeholders were engaged to acquire a thorough grasp of the methods involved in the buy sales campaign, as well as the challenges faced during size allocation. This qualitative insight was useful in determining the objectives and methods of the study. The exploratory data analysis (EDA) technique was used to look at historical sales data and uncover patterns and trends that can help forecast demand. A sophisticated algorithm was used to estimate the weekly demand for each size in numerous functions and locations, allowing stockouts to assure accuracy. The dataset was supplemented with exogenous factors, such as climatic conditions and public holidays, that were appropriate to individual places and nations. This innovation sought to incorporate extrinsic elements that impact purchase behaviour, hence improving prediction models. For both single time series and aggregated data, several forecasting models were created, including ARIMA, SARIMA, Random Forest, and LightGBM. The performance of these models was evaluated using the Symmetric Mean Absolute Percentage Error (SMAPE) metric, which facilitated the identification of the most accurate predictor. The Global Random Forest model demonstrated the lowest median SMAPE and was selected for the final demand predictions. The size curve was then computed with the same granularity as the anticipated demand. This detailed size distribution enables for more accurate inventory planning and greater alignment with actual customer demand. Finally, future research topics are described, stressing the potential for combining sophisticated machine learning approaches with real time data analytics to improve demand forecasting accuracy and inventory management.