Industrial topic with Delhaize : analysis and forecasting of clients' purchases using time series

The main objective of this thesis was to review and explore time series forecasting models in the context of Delhaize client purchasing behaviors. We had access to a four-year data base sample of client total amounts spent per visit at Delhaize. We have explored the subject from two different approaches: the statistical and the deep learning approaches. From the statistical point of view, we mathematically analyzed the time series made of the weekly client purchases. We first study the stationary of client time series and we search how to better stationarize them by either differencing or decomposition. A a next step, we try to fit the time series to an ARIMA model which is the most common and used technique for time series forecasting. In the second approach, the machine learning point of view, we used the Artificial Neural Networks (ANN) technique, one of the major tools in machine learning. We begin by introducing these particular architectures before presenting a specific ANN, the Recurrent Neural Network (RNN). We first implemented a simple RNN followed by more ingenious networks with the LSTM (Long Short-Term Memory) networks. Lastly, we checked whether departmental purchase history enabled to improve the forecasting.