Starting from the availability of Origin-Destination data from mobile phone signals for an hourly time series (1 year length), in this talk, I present the Vector Auto Regressive with exogenous variables I proposed for the modelling of time series of traffic flows, where the inherent complex seasonality is accounted by imposing a Dynamic Harmonic Regression in place of the exogenous term.
Since, for technical reasons, data are provided with some hours of delay, the challenging task here is that of identifying a performative model (in terms of nowcasting accuracy) which functioning does not rely on recent (i.e., latest 24 hours) values. Particular attention will be devoted to: i) data characterization, to circumscribe the role of measurement errors; ii) the description and the motivation of the “ad-hoc” cross validation strategy adopted; iii) the discussion on the role of residuals, which exhibit a leptokurtic distribution with heavy tails and a first-order serial correlation. The application of this model, if associated to early warning systems for the alert of people, may turn useful during (e.g) floods episodes.