Resumo: Digital markets have altered how economic agents interact and have changed the behaviour of tourists. In addition, the COVID-19 pandemic has shown that it is necessary to constantly monitor the evolution of digital consumer behaviour and the factors that influence it, as they are dynamic elements that evolve over time. This paper analyses digital inequalities and validates the main factors influencing tourists to book online tourism services. This research uses a set of microdata with 69,752 and 23,779 observations to analyse the booking mode of accommodation and transportation services, respectively, obtained from the Resident Travel Survey of the National Statistics Institute of Spain during the period 2016-2021. The article confirms variations in the online consumer profile and in the trip's characteristics. One of the most relevant findings is the narrowing of the generational gap in the online contracting of tourist services. However, there are remaining digital inequalities, such as regional inequalities and others based on the education level and income of tourists. It is also highlighted that different types of trips, depending on the destination, the type of accommodation or transport have a different propensity to be booked through digital purchase channels. The accessibility to big data sources and recent advances in machine learning models have also made the methodologies for analysing digital consumer behaviour evolve and must be incorporated into tourism studies. This study compares the predictive performance of different methodologies in the context of e Tourism. In particular, we evaluate the potential predictive power that could be obtained using machine learning techniques to explain consumer behaviour in e-Tourism and use it as a benchmark to compare it with the results obtained using traditional statistical methods. The selected predictive evaluation metrics show that the logistic regression statistical model outperforms the predictive power of the Multilayer Perceptron neural network and presents values very close to the maximum predictive power achieved by the Random Forest algorithm.
