Addressing the issue that existing slope deformation predictions rarely consider the spatial variability of soil property parameters, this paper proposes a stochastic deformation prediction method for high fill slopes. By analyzing the one-dimensional consolidation creep characteristics of loess, a semi-theoretical semi-empirical creep model describing its deformation behavior is developed. Based on the variability of loess consolidation creep observed in repetitive tests, a cross-correlated Gaussian random field model for the creep model parameters is established using the Karhunen-Loève series expansion method. Combined with ABAQUS secondary development, MCS-SFEM is employed to perform deformation predictions. The method is applied to a specific high fill slope, where 89.06% of the measured deformations at monitoring points fall within the predicted deformation range. The RMSE, MAPE, and R2 of the predicted mean curve are comparable to those of the hyperbolic method and not significantly different from the BP neural network model. Compared with the hyperbolic method, the proposed prediction method can account for nonlinear deformation caused by gravity changes during construction, incorporates post-construction nonlinear deformation through deformation limits, and its extrapolation prediction better conforms to the primary deformation trend, thereby demonstrating the rationality and reliability of the prediction method.