Handwriting is an everyday life human activity. It can be collected off-line by scanning sheets of paper. The resulting images can then be processed by a computer-based system. Thanks to digitizing tablets, handwriting can also be collected on-line. From the collected raw signals (pen position, pressure over time), the dynamics of the writing can be recovered. Since handwriting is unique for each individual, it can be considered as a biometric modality. Biometric systems predicting gender from off-line handwriting, have been recently proposed. However we observe that, in contrast to other modalities such as speech, it is not straightforward for a human being (even expert) to predict gender. In this study we explore the limits of automatic gender prediction from on-line handwriting collected from a young adults population, homogeneous in terms of age and education. In our previous work [1], a statistical analysis of on-line dynamic features has shown differences between male and female groups. In the present study, we provide these features to a classifier, based on a machine learning approach (SVMs). Since datasets are relatively small (240 subjects), several evaluation frameworks are explored: cross validation (CV), bootstrap, and fixed train/test partitions. Accuracies obtained from fixed partitions range from 37% to 79%, while those estimated by CV and bootstrap are around 60%. This shows to our opinion the limits of the gender recognition task from on-line handwriting, for our observed young adult population.