مقاله انگلیسی رایگان در مورد راهبرد مبتنی بر شبکه عصبی برای پیش بینی روند بازار سهام – الزویر 2019

Stock market forecasting is a vital component of financial systems. However, the stock prices are highly noisy and non-stationary due to the fact that stock markets are affected by a variety of factors. Predicting stock market trend is usually subject to big challenges. The goal of this paper is to introduce a new hybrid, endto-end approach containing two stages, the Empirical Mode Decomposition and Factorization Machine based Neural Network (EMD2FNN), to predict the stock market trend. To illustrate the method, we apply EMD2FNN to predict the daily closing prices from the Shanghai Stock Exchange Composite (SSEC) index, the National Association of Securities Dealers Automated Quotations (NASDAQ) index and the Standard & Poor’s 500 Composite Stock Price Index (S&P 500), which respectively exhibit oscillatory, upward and downward patterns. The results are compared with predictions obtained by other methods, including the neural network (NN) model, the factorization machine based neural network (FNN) model, the empirical mode decomposition based neural network (EMD2NN) model and the wavelet de-noising-based back propagation (WDBP) neural network model. Under the same conditions, the experiments indicate that the proposed methods perform better than the other ones according to the metrics of Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). Furthermore, we compute the profitability with a simple long-short trading strategy to examine the trading performance of our models in the metrics of Average Annual Return (AAR), Maximum Drawdown (MD), Sharpe Ratio (SR) and AAR/MD. The performances in two different scenarios, when taking or not taking the transaction cost into consideration, are found economically significant.

Introduction

Stock market forecasting is always a remarkable topic and has attracted continuous attention in finance. Unfortunately, stock prices exhibit dynamic, non-linear, non-parametric and chaotic properties in nature (Oh & Kim, 2002; Wang, 2003). Many standard statistical and econometric models for forecasting must face significant 5 challenges, such as disobeying the statistical assumptions in dealing with non-stationary time series, or having unsatisfactory forecasting performance due to the requirement of observations to be distributed normally. There are numerous models and strategies proposed for the stock market predictions. Most of them can be classified into two categories: the ones based on statistical techniques and those using machine learning techniques. In the category of statistical approaches, there are autoregressive integrated moving average (ARIMA), 10 generalized autoregressive conditional heteroskedasticity (GARCH) volatility (Franses & Ghijsels, 1999), and the smooth transition autoregressive model (STAR) (Sarantis, 2001), just to name a few. These approaches are primarily based on the assumptions of stationarity in time series and linearity among normally distributed variables. However, the stationarity, linearity and normality assumptions are not satisfied in real stock markets. On the other side, machine learning models without these restrictive assumptions have been proposed in recent years, 15 and they can outperform the statistical methods (Hansen & Nelson, 2002; Zhang, 2003; Enke & Thawornwong, 2005; Ture & Kurt, 2006). Thus, machine learning approaches, such as support vector machine (SVM) (Kim, 2003; Qian & Gao, 2017), genetic algorithm (Kim & Han, 2000), fuzzy system (Wang, 2002; Shen & Han, 2004), neural network (NN) (Vellido et al., 1999; Chen et al., 2003; Rather et al., 2015) and hybrid methods (Armano et al., 2005; Wang et al., 2011; Patel et al., 2015), have been widely employed in forecasting stock prices. 20 Although the machine learning based models have achieved remarkable results, there are still limitations. Firstly, they do not have an explicitly mechanism to handle the non-stationarity of stock prices. Secondly, as far as we know, most models do not pay attention to the interactions between features at different scales. For example, for the neural network models or deep network models, which are also widely used in many other applications such as image processing (Krizhevsky et al., 2012; Szegedy et al., 2015; He et al., 2015a), mechanical 25 translation (Bahdanau et al., 2014; Luong et al., 2015), speech recognition (Hinton et al., 2012; Amodei et al., 2015) and so on, the non-linearities mainly handled by the activation functions, and there is few technique addressing the non-linear interactions among the inputs.