The financial market is characterized by being a dynamic, complex and non-linear system, characterized by data intensity, noise, non-stationary nature, unstructured and with a high degree of uncertainty [1] . As so many factors interact simultaneously, such as political events, macro and microeconomic conditions and investor expectations, predicting these movements is a very challenging task.

The growing role that the stock market plays in the world economy stimulates the development of research aimed at building theories, involving the topic of stock price prediction, and accurate methods are crucial for the management of investor portfolios.

Assessing expected returns relative to total exposure assumes that portfolio managers understand the distribution of the portfolio. Specialists can model the influence of tangible assets in relation to market value, but not of intangible assets such as rights, experiences or brand equity.

An important ally in the quest to minimize risk in relation to exposure, artificial intelligence and machine learning with their neural networks have provided a great balance of quality in recent decades, improving detection, diagnosis, prediction and problem solving [2] ; this is because in this market, future events are at least partially dependent on past events and data [3] , and not entirely random.

The aim of this study is to map and analyze the published academic literature on stock price prediction using artificial neural networks Long-Term Memory Artificial Neural Networks—RNA LSTM. To this end, a bibliometric analysis and a systematic review of the literature on the subject are carried out, during the period from January 1, 2000 to March 31, 2022, with a final sample of 99 articles. Bibliometric analysis refers to quantitative analysis, which is developed by counting frequencies and co-citations. The systematic review, a qualitative analysis, considers the correlation between the most significant themes, but still little studied by the academy. The research base used comes from the Web of Science—WoS database, and both the bibliometric analysis and the systematic review do not dispense with the use of R, RStudio, Biblioshiny and VOSViewer software. In the bibliometric analysis, the verification of the main laws is adopted: Lotka [4] and Bradford [5] .

The literature review is presented in item 2, with the identification of theories and methods of forecasting stock prices with multilayer perceptron artificial neural networks mentioned in the articles of the final sample. The bibliometric analysis and systematic review methodologies are described in item 3, and in item 4 the results of both methodologies are reported, with descriptive statistics of the most relevant characteristics of the articles in the final sample and the knowledge gaps on the topic. Item 5 presents the conclusions, paths for future studies and limitations of this research.

The current stock price of a publicly traded company reflects the company’s past operation, current timing and future profitability prospects. To obtain a more accurate projection of the stock price, several types of studies have already been carried out. The most classic ones focus on the financial data of the target companies, added to micro and macroeconomic aspects. However, the non-linear and non-stationary features of financial data sequences make predictions more challenging [6] . In 1969, Akaike [7] used the autoregressive (AR) model for prediction. Then, the combination of the autoregressive model (AR) with the moving average (MA) model was proposed, forming the ARMA model. In 1970, Box and Pierce [8] proposed the Autoregressive Integrated Moving Average System (ARIMA) model, which remedy some drawbacks of the ARMA model for dealing with non-stationary sequences. In 1982, Engle [9] proposed the autoregressive conditional heteroskedasticity (ARCH) model to process time series volatility, in 1986 Bollerslev [10] proposed the generalized autoregressive conditional heteroscedasticity (GARCH) model. In 1987, Hull and White [11] proposed a model to solve the stochastic problem of time series volatility (SV). All these methods lay the groundwork for the development of time series forecasts.

With the development of computer technology, new prediction methods have been proposed, such as artificial neural networks [12] , whose objective is to replicate the way the human brain works. Recurrent neural networks [13] are a powerful set of artificial neural network algorithms, especially useful for processing sequential data such as sound, temporal or language. Some recurrent neural networks—RNN performed better in predicting financial data and became popular, such as Long Short-Term Memory [14] .

Thus, LSTM network is a specific type of RNN that has been widely applied to solve supervised learning issues [15] . It has non-linear memory cells and gate units [16] , capable of processing non-stationary long-term sequences. In addition, it can extract the characteristics of financial data and reflect the characteristics of the network [17] . It offers good performance in predicting prices in the stock market, as it is an algorithm capable of identifying non-linear and hidden relationships in the data, that is, it is a supervised learning algorithm, capable of learning from a set of data in training (given a dataset, LSTM can learn a nonlinear function for regression).

Maknickiene and Maknickas [18] improved the performance of measurements in the foreign exchange market or using TM; Chen, Zhou, and Dai [19] used LSTM for Chinese design market returns and performed well. After that, several experiments with the modern LSTM network in a literary way in isolation, hybrid or combined methods with classics and financial series studies were completed, A indicates that LSTM is quite suitable for time series financial models [20] .

Concretely, the LSTM network consists of three parts, including an input layer, an output layer and several hidden layers between them. The hidden layers have memory modules. The core of the memory module is the self-connecting memory cell with three ports, input, output, and forgetting. The value of each of these ports controls the flow of information in the memory module.

Information is retained by cells and memory manipulations are done by gates. Gateway: where useful information is added. Oblivion Gate: where information that is no longer useful is removed. Output Port: the task of extracting useful information from the current cell state to be displayed as a result. A vector is generated, and the information is regulated using the function that filters the values to be remembered. Vector values and regulated values are multiplied to be sent as output and input to the next cell [21] .

However, designing a good network architecture for the problem studied is not a simplistic task. The model’s architecture directly interferes with its performance. The refinement process can be time consuming, as there is no formal method to perform this classification task, this is necessarily through the performance of iterative tests with several parameters, in which only the structure of greater assertiveness is maintained.

The quality of the information that the network is fed, as much or more, is reflected in the accuracy of the output layer’s response. The selection of information that will be provided to the input layer is a key factor in the design of an intelligent decision system, because even if the model is the best, it will perform poorly if the features are not well chosen. Specific methods must be used in the selection of relevant information.

The aim of this study is to answer the question—using LSTM artificial neural networks, can we get reliable predictions of stock prices? For this, the 7 steps described in Figure 1 , detailed below, are implemented.

Step 1—Choosing the database. Sample articles come from WoS, the world’s leading citation database.

Step 2—Using WoS Initial Search Parameters for the period from January 1, 2000 to March 31, 2022. Initially, 276 articles are identified based on variations of the keywords stock, market, LSTM, forecast, stock, predictive, regression, supervised, learn, backpropagation, supervised and backpropagation. Subsequently, exclusions are performed by applying filters in WoS itself, resulting in an intermediate sample of 127 articles, as shown in Table 1 .

Step 3—Exclusion of 19 of the 127 articles for not being available in the researched sources (Google Scholar, Science Direct and Web of Science) and another 09 for not being directly related to the topic of our research, namely: 01 e-commerce, 02 cryptocurrencies, 01 real estate price bubble detection, 01 hierarchical temporal memory, 01 neuromorphic vision datasets, 01 Gray Wolf-Elman optimization, 01 stock movement during the Covid-19 pandemic and 01 stock price prediction based on in morphological similarity clustering and hierarchical temporal memory. Thus, the final sample is composed of 99 articles [15] [20] - [122] .

Step 4—Creation of a database and collection of articles. The 99 articles in the final sample are obtained from the following academic research databases: Web of Science, Science Direct, and Google Scholar. From its analysis, the following information is collected to capture the general data of the article: title, author name, affiliated institution and country of origin of authors/researchers, journal name, volume and issue number, homepage and page final, year of publication, country of origin of data and number of years of sample data, keywords, Digital Object Identifier (DOI), Journal of Economic Literature (JEL) and number of citations of articles in the WoS database.

Step 5—Bibliometric analysis. Through the R, RStudio, Biblioshiny and VOSviewer software, objective data from the articles are analyzed—countries, authors, keywords, institutions, etc., for the preparation and analysis of relationship/co-citation tables and maps. The analyzes carried out by both tools are complemented by the verification of the main laws of bibliometrics, namely: 1) Bradford’s Law [5] —verification of journals that produce many articles in contrast to those that produce few on a given topic, and 2) Law de Lotka [4] —identification of researchers with a higher frequency of production in a given area of knowledge.

Step 6—Reading and coding the articles. Identification of the objectives, sample, methods and contributions of the articles. In addition, they are classified and coded into categories and subcategories structured according to Table 2 . Each of the 08 categories has non-exclusive subcategories. This means that the same article can be classified in more than one subcategory. Thus, the sum of the frequency count of the subcategories—for each category—is what adds up to 100%. In the coding process, as many subcategories as necessary per article are assigned.

Step 7—Systematic review. After coding the (sub)categorization matrix in Table 2 —for the final sample—a frequency count of the subcategories is performed to enable the identification of knowledge gaps. Such gaps are then compared with the subcategories of category 08—paths for future studies, in order to obtain aspects that can be the object of further studies on the subject.

Item 4.1 presents the results of the bibliometric analysis, mentioned in Step 5 of the Methodology. In turn, item 5.2 contains the results of the systematic review, whose steps are described in Steps 6 and 7 of item 3 of this study.

Publications on this topic are concentrated from 2020 onwards. The keywords most associated with these studies are model, neural networks, prediction and time series. 333 authors wrote on the subject between 2018 and March 2022; 43 of the 99 articles published in this period are associated with Chinese institutions. The journals that publish the most significant articles on the topic are Expert Systems with Applications, IEEE Access, Big Data Journal and Neural Computing and Applications. The most cited article is by Kim and Won [20] , Stock Price Index Volatility Prediction: A Hybrid Model Integrating LSTM with Various GARCH-Type Models, which studies the volatility of Kospi 200 stock index returns and capitalization of the stock market in South Korea, cited 235 times. The daily closing price of shares is the most analyzed type of data, and studies are still concentrated on American (21%) and Chinese (20%) stock exchanges. 57% of the studies present adjustments to already tested neural network models and 42% present new theories or new projection models. In 27% of the articles, the authors suggest future studies with news sources, other stock data, or the use of different time series. 22% suggest studies with hybrid models of LSTM not yet tested, associated or not with other neural networks, other data, other sources and periods.