Long-term prediction in Tehran stock market using a new architecture of Deep neural networks

Financial markets play an important role in the economy of modern societies. Therefore, many researchers have investigated to forecast these markets using various statistical and soft computing methods. Financial time series are essentially complex, dynamic, nonlinear, noisy, nonparametric, and chaotic, so they cannot be described by analytical equations, because their dynamics are too complex or unknown. In recent years, deep learning methods have attracted lots of attention, due to their exceptional performance compared to other existing approaches in many learning problems. The objective of this paper is longterm prediction of price time series in Tehran Stock Exchange. For this purpose, a new architecture of two deep learning methods, LongShort Term Memory (LSTM) and Recurrent Neural Network (RNN), for tenstep ahead simultaneous prediction, are proposed. That is a multivariable structure with multi outputs. By using the output error feedbacks as internal inputs, the network can learn error dynamics during the training phase. Experimental results show the high capability of the proposed structure for both methods in multistep ahead stock price forecasting and the superiority of the LSTM network compared to RNN for longterm predictions.