Aiming to find a more accurate method to assess the effect of agricultural machinery compaction on soybean yield, data of soil physical properties and soybean yield in different compaction environments were obtained by carrying out different numbers of compaction walks with different types of tractors. Soybean yield forecast models were developed from mechanical factors, soil factors, and composite factors which affected soybean growth, respectively. To find out the differences of models built by different types of machine learning algorithms, multiple linear regression (MLR), random forest (RF), adaptive boosting (AdaBoost), and artificial neural network (ANN) were used in modeling. In addition, the importance of model features was comprehensively analyzed. The results showed that the relationship between mechanical operation and crop yield was complex, and the models built by integrated learning algorithms (AdaBoost and RF) had a better fit and higher coefficient of determination. Among the machine learning algorithms used, the best performance of the models built was AdaBoost, followed by random forest, artificial neural network and multiple linear regression. The model built using composite factors for soybean yield had the best fit, followed by mechanical and soil factors. The AdaBoost-based composite factor for soybean yield forecast model had the optimal fit with R2 of 0.92, MAE of 1.33% and RMSE of 1.86%. Mechanical factors and soil factors all had an effect on the variation of soybean yield. The number of mechanical compaction, soil penetration resistance in the surface and subsurface layers were the important factors affecting soybean yield. Therefore, the effects from mechanical compaction can be relieved by reducing the number of mechanical operation and loosening soil penetration resistance of the surface and subsurface soils.