This　paper　considers　variable　selection　for　spatial　autoregressive　models　based　on　the　minimum　prediction　error　criterion.　Firstly,　based　on　an　initial　consistent　estimator,　a　new　loss　function　is　constructed　from　the　perspective　of　prediction,　and　then　we　proposed　a　novel　variable　selection　method.　This　method　can　efficiently　select　the　significant　variables　via　penalizing　the　loss　function　proposed.　Under　mild　conditions,　the　large　sample　properties　of　the　resulting　method　are　established.　The　finite　sample　performances　are　investigated　via　the　extensive　Monte　Carlo　simulations.　Finally,　this　resulting　method　is　applied　to　the　Boston　housing　price　data,　further　validating　the　practicability　of　the　proposed　method.