In the article we present a technique for time series modeling that joins concepts based Fuzzy Cognitive Map design with moving window approach. Proposed method first extracts concepts that generalize the underlying time series. Next, we form a map that consists of several layers representing consecutive time points. In each layer we place concepts obtained in the previous step. Fuzzified time series is passed to the map according to the moving window scheme. We investigate two most important aspects of this procedure: division into concepts and window size and their influence on model’s accuracy. Firstly, we show that extraction of concepts plays a big role. Fitted models have low errors. Unfortunately, it is not always possible to extract appropriate number of concepts. The choice of the number of concepts is a compromise between model size and accuracy. Secondly, we show that increasing window size improves modeling accuracy.