| The movement of stock price appears to have some trends,which is very important for stock price prediction.There are two main types of stock price forecasting methods: fundamental analysis and technical analysis.This paper studies the trend of stock price movement in terms of technical analysis and transforms the stock price forecasting into the time series analysis.The two main tools of time series analysis are traditional econometric model and computational intelligence model.Traditional econometric models are good at interpreting but have poor prediction performance.Computational intelligence models,especially the artificial neural network models can be good at self-adaptive learning and have powerful nonlinear mapping ability,so this research uses artificial neural network for modeling.Preprocessing and network structure design are very important to establish neural network prediction model.The existing preprocessing methods are mostly automated,such as signal analysis,time series segmentation and principal component analysis,which lack of financial sense.In terms of network structure,the strong ability of recurrent neural network for time series analysis and the development of deep learning have brought new opportunities for financial forecasting.This study focuses on extracting features of more financial signal and comparing different recurrent neural networks.There are two main works in this study: firstly,the trend indicator is preprocessed in two steps.In the first stage,the trend indicator is extracted from five aspects: price,return,momentum,price-volume relationship and volatility.The features of price reflect the level of price in different time scales.The features of return reflect the speed of price movement.The characteristics of momentum reflect the accelerated speed of price movement.The feature of volume adds extra information which price feature do not have.The volatility feature includes the information brought by the interval data.The second stage is discretizing the features based on the internal relationship between each indicator and trend,mining more dynamic characteristics of the trend.The second work is constructing different recurrent neural network models,comparing three representations of the input and different network architectures,and making a primary exploration of deep hidden layers.The experimental results show that the preprocessing of trend factor significantly improves the prediction performance.In the three types of recurrent neural networks,GRU is the best performer,and the deep structure performs better than the shallow layer model under the condition that the total number of hidden neurons remains unchanged.In all hyper-parameter settings,the discretized trend factor performs best of GRU which has three hidden layers.Compared with random prediction method,our model improved the accuracy of stock returns prediction from 33.3% to 68%.According to the experimental results,this study obtains the following important conclusions: price,return,momentum,price-volume relationship and volatility contain stock trend information,which can effectively improve the performance of prediction model.In the three types of recurrent neural networks,GRU is more suitable for the low-frequency financial time series with small simple size.The results also show that deep learning is not omnipotent.Feature representation with more financial sense can get twice the result with half the effort,which is much more efficient than the deep structure.There are three main innovation in this research: the two-steps data representation based on the inner logic of technical analysis and trend prediction;the elaborative preprocessing of the output variables,using 5-days return to set class labels and adding the “concussion” label,which is helpful to set up trading strategy;the analysis of three types of recurrent neural network in short-term prediction of stock index trend and the preliminary exploration of deep learning. |
PDF Full Text Request