The invention discloses an improved GM (1, 1) model prediction method based on a trapezoidal formula, comprising the following steps: 1, selecting a raw data sequence adopted by a prediction model according to a prediction target, wherein the original data sequence is a set of non-negative data sequences; 2, performing an accumulation process on the original data sequence to generate a cumulativesequence; 3, constructing a quadratic interpolation polynomial, building a trapezoidal formula construction background value based on the quadratic interpolation, and using the least-square method tofind parameters a, u; 4, based on the solved parameters a, u, establishing a time response sequence X<^(1)>[(k+1)] to restore the predicted value X<^(0)>[(k+1)] of an initial point, wherein the predicted value is a predicted value sequence of the original data sequence; and 5, according to the predicted values of the original data sequence solved in the previous step, performing error check to determine the prediction accuracy of a GM (1,1) model. The method of the invention can effectively adapt to various types of data changes and effectively improve the prediction accuracy of the model.