Stock Index Forecasting using Deep Learning

Abstract

Stock index forecasting have always been an attractive field for researchers. Stock index prices are affected by some obvious factors such as supply and demand, investor sentiment etc and an uncertain noise. With increasing research to forecast stock index using deep learning techniques, many neural networks, particularly RNNs (recurrent neural networks) have been applied to forecast stock prices successfully. In this research, we implement direct input-to-output connections (DIOCs)in Elman neural networks (ElmanNN) , as proposed in previous research [17] and introduce DIOCs in the Long short term memory networks (LSTM). We show that Elman network with DIOCs (Elman- DIOC) and LSTM network with DIOCs (LSTM-DIOC) significantly outperforms their vanilla counterparts without such DIOCs. Four different global stock indices- the Shanghai Stock Exchange (SSE) Composite Index, the Korea Stock Price Index (KOSPI), the Nikkei 225 Index (N225), the Standard & Poor’s 500 Index (SPX) are used to show the efficacy of Elman-DIOC and LSTM-DIOC networks. To evaluate the effect of DIOCs, we train eight different model configurations depending on the presence/absence of hidden and output layer biases, direct input-to-output connections. We show that DIOCs always improves the forecasting accuracy as long as the problem has linear components.