Data-driven models for e-commerce sales predictions

Abstract: Future predictions have various applications, including stock prices, house market prices, and company sales. For sales, predictions can guide future expectations and suggest ways to cut costs. Due to the value of predictions, researchers have developed plenitude of prediction algorithms. Still, many companies use simplistic prediction algorithms that fail to provide accurate results. Also, the vast number of existing algorithms makes it difficult to find the best algorithm for a specific data set. In this thesis, I predicted an e-commerce company’s future sales based on its historical trans-action data with three different models: the machine learning algorithm LSTM, a forecasting library released by Facebook called Prophet, and a model that I developed inspired by Prophet, called the Average Sales Prediction (ASP) model. I compared these models to each other and a benchmark model. The benchmark model I used is one of the simplistic algorithms that some companies currently use. It takes the mean value of the past month’s sales to predict the upcoming day. Using the mean absolute percentage error (MAPE), I found that LSTM had the best overall performance on these data, with a MAPE of 18%. The second-best performing model was ASP, which resulted in a MAPE of 26%. Finally, Prophet resulted in a MAPE of 31%. The results gauge the company’s future performance and will help them improve its sales through streamlined workloads and better warehouse and transportation planning. The models can be enhanced further for sales that, for example, depend on the weather or other external factors.