Models for tourism flash statistics integrating MNO and survey data

Mobile Network Operators (MNO) data are increasingly investigated by National Statistical Institutes as complementary data sources to be integrated into traditional statistical production processes, due to their timeliness and high spatial granularity. In this work, we produce flash estimates of tourist overnight stays in Emilia-Romagna at the municipal level by integrating MNO and survey data through augmented and transfer learning approaches. Prior to the application of the learning methods, an adjustment procedure is applied to the MNO data in order to obtain a more reliable proxy variable. For each approach, several model variants are tested and compared within a random forest framework. Learning-based estimates are also compared with predictions obtained from a traditional univariate time series model (TRAMO) for selected municipalities. Despite relying on a more limited amount of data, learning-based approaches yield satisfactory results compared to time series models, even outperforming them in specific settings and municipalities. A scalability analysis is conducted by progressively increasing the number of municipalities whose survey observations are replaced by proxies in the learning process. The results show that both augmented and transfer learning scale well and are effective in reducing the structural bias of MNO data with respect to the survey data they aim to approximate. These findings suggest promising directions for fully transductive learning solutions, which could bypass the need for early respondent data during the training phase.