The invention provides a vehicle driving risk prediction method based on time varying state transition probability markov chain. Firstly, an offline vehicle driving risk prediction model training: based on samples of accidents and near accidents, real-time vehicle driving risk states are divided by clustering time window characteristics parameters and regarded as countable states of the markov chain, and a multiterm logistic model of vehicle driving risk states transition in different vehicle driving risk states is built. Secondly, an online vehicle driving risk model real-time prediction: under the circumstance of car networking, the variable parameters required by a prediction model are collected in real time, through a risk state clustering center position and markov property, an original state probability distribution vector and a markov chain n steps transition probability at any time in the future are calculated, and the prediction result of the vehicle risk states in the futureis obtained. According to the invention, by means of a recurrence algorithm, the estimation of markov chain n steps time varying state transition probability is achieved, which can reflect the characteristics of the vehicle driving risk states changing with the characteristics of the transportation system, and can meet the requirement of early warning in real time.