Forecasting Booking Behaviour in the Freight Ferry Industry: An Artificial Neural Network Approach

Abstract

The freight ferry industry operates in a stochastic environment where the arrival of vehicles on the departure day is uncertain. This thesis aimed to explore the application of artificial neural networks (ANNs) to predict whether active bookings at a given time before departure are going to become transferred, cancelled, or no-show. The sum of these bookings at a given time before departure was assumed to follow a Poisson binomial random variable. A comparative analysis was conducted between ANNs and Gradient Boosting Trees (LightGBM) trained on structured booking data, such as customer number and check-in status, and recurrent neural networks (RNNs) trained on a sequence of previous cumulative transfer and cancellation rates. The findings revealed that LightGBM was the better option to predict the mean of the distribution, while an ensemble of ANNs demonstrated promising potential in predicting the distribution’s variance and error estimates.