Forecasting traffic flow time series with Vine-Transform ARMA Copula models

Abstract The prediction of traffic flows in urban areas is gaining importance in order to carry out urban planning for early warning systems and optimized logistics. Hence, there is a growing need for simple and high-performing statistical models. This study leverages the vine-transform autoregressive moving-average copula model to predict traffic data, with particular emphasis on evaluating forecasting performance. To do so, real-life data on origin-destination signals extracted from mobile phone data have been used. Performance evaluation was conducted using the rank-graduation box approach along with a moving window cross-validation strategy, incorporating rank-graduation accuracy for precision and rank-graduation explainability for component analysis. As a benchmark for comparison, the VARX-DHR model was used. Preliminary results reveal that the vine-transform ARMA copula approach performs well in terms of accuracy. Furthermore, we show that the copula component presents greater explainability compared to the autoregressive and moving average components. Residual diagnostics show significantly lower autocorrelation and partial autocorrelation with respect to the original data, and that they are approximately normally distributed. The developed method could provide valuable insights supporting urban planners and analysts in making informed decisions.