Outlier detection is a subject of interest in data mining. With the evolution of data acquisition methods such as wireless sensor networks, there is a need to detect outliers in data streams. However, dealing with data streams is challenging due to the amount of data that grows infinitely and the nonstationarity of the distribution. On top of that, the detection has generally to be done in an unsupervised way. Some methods have been proposed to tackle this problem but none of them can be easily parameterized. This paper proposes a novel method based on the Christoffel-Darboux kernel borrowed from the theory of approximation and orthogonal polynomials. This method perfectly applies to data streams while being deployable with no tuning at all. It is compared to some state-of-the-art methods for outlier detection in data streams and applied to the data from an industrial luggage conveyor showing very convincing results.