Macro-level accident fatality prediction using a combined model based on ARIMA and multivariable linear regression

Prediction is one of the key functions of traffic safety management system and fatality prediction is the priority. In previous studies, time series theory was widely used in traffic fatality predication. Researchers have applied Auto-Regressive Integrated Moving Average (ARIMA) model to the traffic safety data. To introduce relevant features, which influences the model's reliability and accuracy, Multivariate Linear Regression (MLR) model is combined with ARIMA model. And the number of variable message signs and carriageway area are picked out as features in this study. The combined model not only inherits the accurate trend prediction of ARIMA model, but also calculates comprehensive relevant features. Traffic crash fatalities and fatalities per 10,000 vehicles in Shanghai are selected as targets and two models are trained respectively for the two targets above - ARIMA(1,1,0) and ARIMA(1,2,0). Then MLR model is introduced to take the two selected features into consideration. Experiments show that both combined models perform better and more reasonably than simple ARIMA model where the error rate of prediction is under 1%.