It is largely unknown how sensory information represented by a neuronal population is transformed to generate motor responses. Ocular following is a useful model system for studying these sensory-motor transformations as it is a eye movement reflexively triggered by motion of a visual stimulus within less than 100 ms. Previous electrophysiological studies have shown that neurons in the middle temporal area (MT) represent sensory information critical for generating ocular following responses. However, single-neuron studies were not able to assess how neuronal variability correlates with behavioural variability on the level of single trials. We used Neuropixels probes to simultaneously record neurons (n = 29-139) in area MT of marmosets  viewing visual motion stimuli that evoked tracking eye movements. We explored how well the population activity correlated with the timing and velocity of eye movements. We used partial-least-square regression (PLSR) to predict horizontal and vertical eye velocities from the population spiking rates on a trial-by-trial basis. For each trial, we extracted the mean firing rates of each neuron and mean eye velocities in 20 ms time windows relative to motion onset. The model started to predict eye velocities well 60 ms after stimulus motion onset (r = 0.58), when the eyes just started moving, and stayed high (r>0.72) in subsequent time windows (80-160 ms). Eye velocities at 60-160 ms were also well predicted by neural activity at 20-40 ms (r = 0.52). In conclusion, single-trial motor variability can be predicted from variations in sensory activity in the MT population activity.