Recently promising new methods have been developed to optimize e-commerce A/B testing using dynamic allocation. They provide faster results to determine the best variation and reduce test costs. However, dynamic allocation by traditional methods of reinforcement learning is restrictive on the type of data used : time series cannot be considered as a context. In this paper, we present two new methods, based on a common approach, that allow time series to describe visitors. Our numerical results on data from real tests leads to an improvement on dynamic allocation apply to A/B testing.