In　view　of　the　nonlinear　fluctuation　of　agricultural　futures　price　and　the　linkage　characteristics　of　domestic　and　foreign　futures　products,　considering　that　the　traditional　neural　network　prediction　model　can　not　quantitatively　characterize　the　causal　relationship　between　multi-source　input　variables,　this　paper　constructs　a　graph　neural　network　prediction　model　with　transfer　entropy.　Firstly,　the　adjacency　matrix　between　nodes　is　represented　by　calculating　the　transfer　entropy,　which　is　used　as　a　priori　information　to　identify　the　causal　relationship　between　variables.　At　the　same　time,　the　temporal　convolution　module　of　multi-scale　filter　is　set　to　extract　the　node　features　to　identify　the　time　dependence　of　the　sequence.　Secondly,　the　graph　convolution　module　is　set　to　realize　the　propagation　and　feature　selection　of　node　information　and　its　neighborhood　information.　Finally,　the　final　prediction　result　is　output　by　connecting　parameters.　The　empirical　study　on　soybean　futures　data　shows　that　compared　with　the　existing　general　forecasting　model,　this　model　can　achieve　the　best　forecasting　effect.　©　2024　Chinese　Medical　Journals　Publishing　House　Co.Ltd.　All　rights　reserved.