Aiming at reducing the influence caused by uneven fields to the working quality of precision agricultural equipment, lots of agricultural machinery were designed with special mechanisms to profile the farm surface, while the working efficiency is still limited to the response rate of servo and control system. For above reasons, a kind of ARMA (Autoregressive and moving average) algorithm was put forward and then be applied to predict vehicle attitudes so as to realize compensation control for tractor implement. The theory and applicability of ARMA were analyzed in detail before modelling using roll angle data collected from a working rice transplanter by the Inertia and GNSS measurement system. Finally, a 90slength data set was acquired of which the first 60s data was used to model, while the last 30s data was designed to make comparison with predicted values. In order to reduce the amount of calculation, the 100Hz raw data at frequency of 2Hz, 5Hz and 10Hz was output and then predicted the future roll angle of the three sets of data in 1~2 s respectively. First of all, the different method was deployed to make angle series stationary; secondly, a combining method of box ordersearch solution and the MLE (Maximum likelihood estimation) was adopted to build models; at last, with the aid of AIC (Akaike information criterion), the final best model was identified after confirming the AIC. For better following of the dynamic trend of rice transplanter attitude and eliminating the negative influence brought by the increasing amount of sample data, the online modelling method that older sample data was substituted constantly by new observational values in the modeling process was needed. By using three groups of sample data, prediction of rice transplanter attitude in future 1~2s was conducted utilizing Matlab, meanwhile, the whole predicting time was 30s. The results showed that: the validity of ARMA model in predicting attitude trend of farm vehicles was proven; for all the three sets of angle samples, 2s predicting error shows larger than 1s predicting error; the 5Hz roll angle series showed the best predicting effect, with 1s RMSE (root mean square error) and std (Standard deviation) of 0.6567°and 0.6565°, and 2s RMSE and std of 0.6712° and 0.6769° respectively.