In response to the problem that agricultural futures are affected by many factors so that nonlinear and nonsmooth features are difficult to extract and lead to low prediction accuracy, this paper proposes a forecasting method based on CEEMDAN and TransformerEncoder-TCN, which is from the idea of "decomposition-ensemble". First, the complementary ensemble empirical mode decomposition with adaptive noise (CEEMDAN) is used to decompose the time series into multiscale Intrinsic Mode Function (IMF) and residuals, reducing the complexity of series modeling. Second, each subseries is predicted using Temporal Convolutional Network (TCN) incorporating multi-stage self-attention unit (TransformerEncoder) which optimizes the modeling weights of the significant features of the sequence. Finally, the prediction results of each subseries are linearly summed and integrated to obtain the final prediction results. The soybean futures revenue index in the agricultural product index of South China Futures Company is used as the research object. The model is retrained by using time-series cross-validation and parameter transfer. The ablation and comparison experimental results show that the model proposed in this paper has superiority in RMSE, MAE and DS, verifying its effectiveness in predicting agricultural product futures.