Abstract: 

We present a Bayesian state-space model implemented in the probabilistic programming language (Stan) for sales forecasting in the automotive sector. The developed model follows a structural time series modelling approach. We utilise hierarchical modelling principles to share information among different markets and product lines through shared hyper-priors.
In this presentation we describe the problem context and the challenges of sales forecasting we face in the automotive industry. We will then present a summary of the model and an overview of the implementation details while motivating our modelling choices from the constraints of the real-world application we are working in. We also evaluate the predictive performance of the model in comparison to an existing corporate forecasting model that uses a combination of ARIMA and hierarchical linear regression and argue why this type hierarchical Bayesian models is preferred in forecasting under uncertainty.
The model’s state transition is described using a combination of “Local linear trend”, a linear regression augmentation of annual seasonality as well as a linear regression component to incorporate macro-economic explanatory time series. To reduce computation time the model implements a multidimensional Kalman filter in Stan to marginalise out the state transitions.

Bio: 

Dr. Mohamed Sayed is a Data Science Lead at Jaguar Land Rover’s Corporate Analytics Centre of Excellence where he leads business transformation activities utilising machine learning and statistical modelling techniques. He has led projects ranging from demand forecasting and assortment optimisation to warranty and quality management. Prior to that he was a Senior Research Associate with Loughborough University where he led a major European Union funded collaborative research project developing machine learning and decision support tools in the manufacturing domain.