Cloud Hosting

Foundation theory

When discussing the stock market, with its inherent and complexity, the predictability of stock returns has always been a subject of debate that attracts much research. Fama (1970) postulates the efficient market hypothesis that determines that the current price of an asset always reflects all prior information available to it immediately. In addition, the random walk hypothesis states that a stocks price changes independently of its history, in other words, tomorrows price will depend only on tomorrows information regardless of todays price (Burton, 2018). These two hypotheses establish that there is no means of accurately predicting stock prices.

On the other hand, there are other authors who argue that, in fact, stock prices can be predicted at least to some extent. And a variety of methods for predicting and modeling stock behavior have been the subject of research in many different disciplines, such as economics, statistics, physics, and computer science (Lo and MacKinlay, 1999).

A popular method for modeling and predicting the stock market is technical analysis, which is a method based on historical data from the market, primarily price and volume. Quantity. Technical analysis follows several assumptions: (1) prices are determined exclusively by supply and demand relationships; (2) prices change with the trend; (3) changes in supply and demand cause the trend to reverse; (4) changes in supply and demand can be identified on the chart; And (5) the patterns on the chart tend to repeat. In other words, technical analysis does not take into account any external factors such as political, social or macroeconomic (Kirkpatrick & Dahlquist, 2010). Research by Biondo et al. (2013) shows that short-term trading strategies based on technical analysis indicators can work better than some traditional methods, such as the moving average convergence divergence (MACD) and the relative strength index (RSI).

Technical analysis is a well method of forecasting future market trends by generating buy or sell signals based on specific information obtained from those prices. The popularity and continued application of technical analysis has become widely recognized with techniques for uncovering any hidden pattern ranging from the very rudimentary analysis of the moving averages to the recognition of rather complex time series patterns. Brock et al. (1992) show that simple trading rules based on the movement of short-term and long-term moving average returns have significant predictive power with daily data for more than a century on the Dow Jones Industrial Average. Fifield et al. (2005) went on to investigate the predictive power of the filter rule and the moving average oscillator rule in 11 European stock markets, including covering the period from January 1991 to December 2000. Their key findings indicate that four emerging markets: Greece, Hungary, Portugal and Turkey, are information inefficient, compared with seven more advanced other markets. Past empirical results support technical analysis (Fifield et al. 2005); however, such evidence can be criticized because of data bias (Brock et al. 1992).

Elman (1990) proposed a Recurrent Neural Network (RNN). Basically, RNN solves the problem of processing sequence data, such as text, voice, and video. There is a sequential relationship between samples of this data type and each sample is associated with its previous sample. For example, in text, a word is related to the word that precedes it. In meteorological data, the temperature of one day is combined with the temperature of the previous few days. A set of observations is defined as a sequence from which multiple sequences can be observed. This feature of the RNN algorithm is very suitable for the properties of time series data in stock analysis as the Fig. 1:

Source: Lai et al. (2019).

Figure 1 shows the structure of an RNN, in which the output of the hidden layer is stored in memory. Memory can be thought of as another input. The main reason for the difficulty of RNN training is the passing of the hidden layer parameter . Since the error propagation on the RNN is not handled, the value of multiplies during both forward and reverse propagation. (1) The problem of Gradient Vanishing is that when the gradient is small, increasing exponentially, it has almost no effect on the output. (2) Gradient Exploding problem: conversely, if the gradient is large, multiplying exponentially leads to gradient explosion. Of course, this problem exists in any deep neural network, but it is especially evident due to the recursive structure of the RNN. Further, RNNs differ from traditional relay networks in that they not only have neural connections in one direction, in other words, neurons can transmit data to a previous layer or same class. Not storing information in a single direction, this is a practical feature of the existence of short-term memory, in addition to the long-term memory that neural networks have acquired through training.

The Long Short Term Memory (LSTM) algorithm introduced by the research of Hochreiter and Schmidhuber (1997) aims to provide better performance by solving the Gradient Vanishing problem that repeated networks will suffer when dealing with long strings of data. In LSTM, each neuron is a memory cell that connects previous information to the current task. An LSTM network is a special type of RNN. The LSTM can capture the error, so that it can be moved back through the layers over time. LSTM keeps the error at a certain maximum constant, so the LSTM network can take a long time to train, and opens the door to setting the correction of parameters in the algorithm (Liu et al. 2018). The LSTM is a special network topology with three gateway structures (shown in Fig. 2). Three ports are placed in an LSTM unit, which are called input, forget, and output ports. While the information enters the network of the LSTM, it can be selected according to the rules. Only information that matches the algorithm will be forwarded, and information that does not match will be forgotten through the forget gate.

Source: Ding et al. (2015).

This gate-based architecture allows information to be selectively forwarded to the next unit based on the principle of the activation function of the LSTM network. LSTM networks are widely used and achieved some positive results when compared with other methods (Graves, 2012), especially in terms of Natural Language Processing, and especially for handwriting recognition (Graves et al. 2008). The LSTM algorithm has branched out into a number of variations, but when compared to the original they do not seem to have made any significant improvements to date (Greff et al. 2016).

Data on the stock market is very large and non-linear in nature. To model this type of data, it is necessary to use models that can analyze the patterns on the chart. Deep learning algorithms are capable of identifying and exploiting information hidden within data through the process of self-learning. Unlike other algorithms, deep learning models can model this type of data efficiently (Agrawal et al. 2019).

The research studies analyzing financial time series data using neural network models using many different types of input variables to predict stock returns. In some studies, the input data used to build the model includes only a single time series (Jia, 2016). Some other studies include both indicators showing market information and macroeconomic variables (White, 1988). In addition, there are many different variations in the application of neural network models to time series data analysis: Ding et al. (2015) combine financial time series analysis and processing natural language data, Roman and Jameel (1996) and Heaton et al. (2016) use deep learning architecture to model multivariable financial time series. The study of Chan et al. (2000) introduces a neural network model using technical analysis variables that has been performed to predict the Shanghai stock market, compared the performance of two algorithms and two different weight initialization methods. The results show that the efficiency of back-propagation can be increased by learning the conjugate gradient with multiple linear regression weight initializations.

With the suitable and high-performance nature of the regression neural network (RNN) model, a lot of research has been done on the application of RNN in the field of stock analysis and forecasting. Roman and Jameel (1996) used back-to-back models and RNNs to predict stock indexes for five different stock markets. Saad, Prokhorov, and Wunsch (1998) apply delay time, recurrence, and probability neural network models to predict stock data by day. Hegazy et al. (2014) applied machine learning algorithms such as PSO and LS-SVM to forecast the S&P 500 stock market. With the advent of LSTM, data analysis became dependent on time becomes more efficient. The LSTM algorithm has the ability to store historical information and is widely used in stock price prediction (Heaton et al. 2016).

For stock price prediction, LSTM network performance has been greatly appreciated when combined with NLP, which uses news text data as input to predict price trends. In addition, there are also a number of studies that use price data to predict price movements (Chen et al. 2015), using historical price data in addition to stock indices to predict whether stock prices will increase, decrease or stay the same during the day (Di Persio and Honchar, 2016), or compare the performance of the LSTM with its own proposed method based on a combination of different algorithms (Pahwa et al. 2017).

Zhuge et al. (2017) combine LSTM with Naiev Bayes method to extract market emotional factors to improve predictive performance. This method can be used to predict financial markets on completely different time scales from other variables. The sentiment analysis model is integrated with the LSTM time series model to predict the stocks opening price and the results show that this model can improve the prediction accuracy.

Jia (2016) discussed the effectiveness of LSTM in stock price prediction research and showed that LSTM is an effective method to predict stock returns. The real-time wavelet transform was combined with the LSTM network to predict the East Asian stock index, which corrected some logic defects in previous studies. Compared with the model using only LSTM, the combined model can greatly improve the prediction degree and the regression error is small. In addition, Glmez (2023) believed that the LSTM model is suitable for time series data on financial markets in the context of stock prices established on supply and demand relationships. Researching on the Down Jones stock index, which is a market for stocks, bonds and other securities in USA, the authors also did the stock forecasts for the period 2019 to 2023. Another research by Usmani Shamsi (2023) on Pakistan stock market research on general market, industry and stock related news categories and its influence on stock price forecast. This confirms that the LSTM model is being used more widely in stock price forecasting recently.

See original here:
Applying machine learning algorithms to predict the stock price trend in the stock market The case of Vietnam ... - Nature.com