Daily Collision Prediction with SARIMAX and Generalized Linear Models on the Basis of Temporal and Weather Variables

Abstract

Short-term collision prediction is a relatively new area of research in the field of traffic safety because of the high randomness of data and the methodological complexity. Motivated by requirements from frontline traffic operations and enforcement services, the authors conducted this study to develop models that predicted daily total collisions. The study started with decomposition analysis of time series data to determine trends, seasonality, and randomness of daily collisions before it proceeded with an investigation of potential collision contributors. Temporal factors (i.e., months, weekdays, and holidays) and weather forecasts (i.e., daily mean temperature, amount of rainfall, and amount of snowfall) were selected as predictive factors. Accordingly, the seasonal autoregressive integrated moving average model with external regressors (SARIMAX) was identified, and a series of SARIMAX models of different orders was estimated and diagnosed. A generalized linear model (GLM) was also developed and compared with the SARIMAX models by validation measures. Finally, a calibration mechanism was recommended to optimize predictions. Model validations provide evidence that both SARIMAX and GLM are adaptable; however, the SARIMAX models are a viable and preferable option because they can provide greater accuracy than GLM in the short-term prediction of collisions. The models developed in this paper are now being applied (a) to support scheduling of traffic operations, maintenance and enforcement, and dispatch of material and personnel resources and (b) to provide situation awareness for all road users and stakeholders.