Inertial and magnetic sensors based PDR approaches are particularly interesting for pedestrian location since they don't require any specific infrastructure. Estimating the walking direction, which is essential for PDR strategy, remains difficult with handheld sensors. WAISS is a new method that integrates hand movement and is independent of the misalignment between the walking direction and the pointing direction that estimates the walking direction. It uses statistical models of the hand accelerations in the horizontal plane. The paper studies how to create the best possible models. Among the features under study are the number of strides used to learn the models, different acquisition contexts and walking directions. Finally, the complexity of models needed for a given person is discussed. 100 strides over curved and straight line walks combined with a bi-modal Gaussian Mixture Model gives the best walking direction estimate with a 15° mean error over a 325 m indoor/outdoor walk performed by four subjects.