Patterns with “unusual” frequencies are new functional candidate patterns. Their identification is usually achieved by considering an homogeneous m-order Markov model (m≥ 1) of the sequence, allowing the computation of p-values. For practical reasons, stationarity of the model is often assumed. This approximation can result in some artifacts especially when a large set of small sequences is considered. In this work, an exact method, able to take into account both nonstationarity and fragmentary structure of sequences, is applied on a simulated and a real set of sequences. This illustrates that pattern statistics can be very sensitive to the stationary assumption. Keywords and phrases: stationary distribution - pattern Markov chain - biological patterns - finite Markov chain embedding