In this paper, auto regression between neighboring observed variables is added to Dynamic Bayesian Network (DBN), forming the Auto Regressive Dynamic Bayesian Network (AR-DBN). The detailed mechanism of AR-DBN is specified and inference method is proposed. We take stock market index inference as example and demonstrate the strength of AR-DBN in latent variable inference tasks. Comprehensive experiments are performed on S&P 500 index. The results show the AR-DBN model is capable to infer the market index and aid the prediction of stock price fluctuation.