Abstract: Most practical problems of forecasting time series are characterized by a high level of nonlinearity and nonstationarity, noise, the presence of irregular trends, jumps, and anomalous emissions. Under these conditions, statistical and mathematical assumptions limit the possibility of applying classical forecasting methods. The main disadvantage of statistical models is the difficulty of choosing the type of model and selecting its parameters. An alternative to these methods may be methods of computational intelligence, which include artificial neural networks, which can significantly improve the accuracy of time series prediction. A significant advantage of neural networks is that they are able to learn and generalize the accumulated knowledge, highlighting the hidden relationships between input and output data. At the moment, the most time series forecasting solutions based on this toolkit involve the use of feed-forward neural networks (perceptrons, convolutional neural networks, etc.). The article provides an overview of the architecture, principles of operation, and methods of teaching known models of recurrent neural networks. In the study, we built and compared the architectures of Elman and Jordan neural networks for solving the problem of forecasting prices for agricultural products. The corresponding statistical comparisons of the above models are also given. The experimental results show that such approach provides high accuracy in predicting the values from the price of agriculture products.