Stroke patients often exhibit large imbalances in strength between the two upper limbs (ULs) and between the different muscle groups of the paretic UL. The aim of the study was to compare the ability of hemiparetic and normal subjects to produce symmetrical forces with both ULs and to determine whether the differences between force vectors can be predicted by a model accounting for the weakness of the different muscle groups or from clinical characteristics. Sixteen hemiparetic and 16 control subjects were assessed using static dynamometers in which both ULs were fixed. They were asked to produce symmetrical sub-maximal forces (in terms of direction and magnitude) with both ULs in four directions in the sagittal plane, which covers all possible combinations of flexion and extension at the shoulder and the elbow. Fifteen trials were performed, and the subjects were asked to gradually increase the forces produced from one trial to another. In addition, the maximal voluntary force (MVF) was measured unilaterally on both sides under two conditions: single-joint MVF (isolated flexion or extension at the elbow or the shoulder) and multi-joint MVF (same directions as the bilateral task). During the bilateral task, the stroke subjects generally showed larger errors between limbs than the control subjects, although they were able to produce sufficient force in the required directions during multi-joint MVF. The difference between groups was statistically significant in the four target directions for the errors in the magnitude, and in two target directions for the errors in direction (p<0.01). Differences predicted by a model based on the relative single-joint MVFs show a moderate association with errors observed for the magnitude (p=0.001-0.11, R2=0.17-0.54), but not for the direction of forces (p>0.1, R2<0.16). Several clinical features were examined as potential predictors using stepwise multiple regressions. The mean multi-joint MVF and proprioceptive impairments were the best predictors of mean errors in the magnitude of force (p<0.001, R2=0.82) whereas the mean angular errors were best explained by proprioceptive impairments (p=0.01, R2=0.38). Inaccurate internal representations of the paretic UL might explain why asymmetries were observed even though motor capacities were sufficient to perform successfully.