Article Preview
TopIntroduction
The application of time series prediction in financial fields is generally considered one of the most complicated issues in time series analysis due to the financial data noise and vast factors. Time series forecasting is a research program that has some useful applications in numerous other research fields, especially in the financial fields. Forecasting by time series can help people make wise decisions to reduce the risks of investments. In a time series, time is usually an important variable used for making decisions and predictions. When the authors use time series to predict the trend of the future financial market, for example, the stock indices, the authors need to introduce some detailed historical data over some time to train a model. Researchers usually use historical data to predict various future events, such as the forecast of stock prices and changes in product sales.
In this study, the authors focus mainly on the multi-feature stock selection and the deep learning models in the financial market’s trend prediction. Specifically, like most of the applications of classic time series models, the authors require strict assumptions regarding the distributions and stationarity of time series. Meanwhile, all units of the input vectors are independent of each other. The authors select 13 typical stock technique features, which include the result of autoregressive integrated moving average (ARIMA). After being denoised by Fourier transforms, they will be input into LSTM for future predictions. The model in the researcher’s study is named 13f-LSTM. The flowchart of the 13f-LSTM model for the financial time series is shown in Figure 1. Then the authors evaluate and compare the predictive accuracy of both 12 features LSTM, called 12f-LSTM in this paper (without ARIMA) and 13f-LSTM for a long-term stock time series prediction. Moreover, the authors will compare 13f-BLSTM and 13-LSTM in model performance and predictive accuracy.
Figure 1. The flowchart of the 13f-LSTM model for financial time series
TopThe organization of this paper is as follows. In the next section, the authors discuss the existing works related to the research and the advantages of the proposed approach is over the existing works. In the following section, the authors proposed a hybrid model, named 13f-LSTM, with introductions to Fourier transform, ARIMA, LSTM, and its variants. The next sections present the inputs and data resources of 13f-LSTM, followed by the design and analysis of our experiment and its results, and finally, the authors present conclusions from the study.
In the last decade, several approaches, including statistical models and machine learning models, have been proposed to predict financial time series. Most of the traditional statistical models assume that the time series under study is generated from a linear process (Kumar & Murugan, 2013). However, financial time series are essentially dynamic, chaotic, nonlinear, complicated, and highly noisy (Si & Yin, 2013). Machine learning models, such as the support vector regression (SVR; Wikipedia, n.d.a) and especially deep learning models, i.e., artificial neural networks (ANNs; Guo et al., 2014), have been applied successfully in modeling, optimizing, and predicting the financial time SERIES (Lee, 2009). Most of them can capture linear or nonlinear relationships between key features with no prior knowledge about the input data (Atsalakis et al., 2009).
Throughout the traditional statistical techniques, for the quantitative stock market, multi-factor or multi-feature stock selection strategy is the most widely used stock selection model in quantitative stock investment (Zhang et al., 2018). The basic principle is based on mathematical and statistical methods, testing the validity of a series of factors related to stock prices, combining multiple valid features to establish a quantitative model to model stocks, and selecting the best performance according to the corresponding principles for an excess return of the stock portfolio (Peng et al., 2022).