Markdown pricing using causal networks and regression trees

Early or large discounts of fashion items may cause early stock-outs with low revenue. Conversely, late or small discounts may result in low sales, low revenue and excessive stocks at the end of life-cycle of products. So, pricing strategy is very important for fashion retailers. This study introduces two causal estimation methods and an optimization method to find the optimal pricing strategy for regular and promotion periods of fashion items in the presence of stock availability data. The first estimation method is a graphical method where causal Bayesian network structure is learned and the network is quantified with parameter learning to estimate the average treatment effect (ATE) of price on sales. The second estimation method is a regression tree method which estimates different ATE of price on sales for different item groups that can be separated heterogeneously. With different approaches, these two methods estimate sales quantities for different pricing decisions. Comparing accuracy metrics of these two methods, it is observed that the regression tree method provided better results. Sales estimations from the regression tree method are used in an optimization algorithm, Exhaustive Feasible Space Simulation and Optimization (EFSSO), that compares revenue estimations of different scenarios and find the optimal scenario that yields the maximum revenue. Proposed method is tested on a dataset from a fashion retailer in Türkiye. Up to 27.84\% improvement is estimated.