Comparing Artificial Neural Network Architectures for Brazilian Stock Market Prediction

Annals of Data Science https://doi.org/10.1007/s40745-020-00305-w Comparing Artificial Neural Network Architectures for Brazilian Stock Market Prediction Suellen Teixeira Zavadzki de Pauli1 · Mariana Kleina1 · Wagner Hugo Bonat1 Received: 27 January 2020 / Revised: 20 June 2020 / Accepted: 25 June 2020 © Springer-Verlag GmbH Germany, part of Springer Nature 2020 Abstract Prediction of financial time series is a great challenge for statistical models. In general, the stock market times series present high volatility due to its sensitivity to economic and political factors. Furthermore, recently, the covid-19 pandemic has caused a drastic change in the stock exchange times series. In this challenging context, several computational techniques have been proposed to improve the performance of predicting such times series. The main goal of this article is to compare the prediction performance of five neural network architectures in predicting the six most traded stocks of the official Brazilian stock exchange B3 from March 2019 to April 2020. We trained the models to predict the closing price of the next day using as inputs its own previous values. We compared the predictive performance of multiple linear regression, Elman, Jordan, radial basis function, and multilayer perceptron architectures based on the root of the mean square error. We trained all models using the training set while hyper-parameters such as the number of input variables and hidden layers were selected using the testing set. Moreover, we used the trimmed average of 100 bootstrap samples as our prediction. Thus, our approach allows us to measure the uncertainty associate with the predicted values. The results showed that for all times series, considered all architectures, except the radial basis function, the networks tunning provide suitable fit, reasonable predictions, and confidence intervals. Keywords Artificial neural network · Forecasting · Time series · Stock market * Suellen Teixeira Zavadzki de Pauli 1 Federal University of Paraná (UFPR), Cel. Francisco H. dos Santos Avenue, 210, Curitiba, PR 81530‑000, Brazil 13 Vol.:(0123456789) Annals of Data Science 1 Introduction The stock market is a popular investment option for investors because of its expected high returns. In contrast, stock market prediction is a complex task to achieve with the help of artificial intelligence. It is due to stock prices depend on many factors, including trends and news in the market [1]. The prediction of financial market times series is challenging. Such information presents high volatility in the observed data, together with the uncertainty inherent in any forecast. Therefore, wrong decisions based on such predictions can have economically catastrophic consequences for individuals, institutions, and nations, as observed in the financial crises [2]. According to [3], financial institutions and banks are among those industries that have relatively complete and accurate data. Typical cases include stock investment, loan payment prediction, credit approval, bankruptcy prediction, and fraud detection. Techniques such as discriminant analysis, linear and logistic regression, entire programming, decision trees, expert systems, neural networks and dynamic models are commonly used in financial and banking applications. It is not surprising that academia and the financial industry have invested a lot of time in proposing and comparing computational methods for analyzing available data, in order to obtain reliable predictions. Such analyses are used in the prevention of macroeconomic crises, prediction of stock market movement as well as for composing investment portfolios [2]. Since the 1950s, as computing technology was gradually used in commercial applications, many companies have developed databases to store and analyze collected data. Artificial intelligence methods were used to deal with data sets, including neural networks and decision trees. In the 90s, the term data mining came to be used, which crosses the human intervention, machine learning, mathematical modeling and databases. The investigation of the theoretical components of big data, or data science, requires interdisciplinary efforts in mathematics, sociology, economics, computer science and management science [4]. Big data has been one of the most popular topics since last several years and how to effectively conduct big data analysis is a big challenge for every field. According to [5], the process of big data analysis can be described by a general data analysis, which consists of several steps, including data acquisition and management, data access and processing, data mining and interpretation, and data applications. The modeling methods can include many techniques and the challenge reflects when and which method is appropriate. In this paper we will limit ourselves to the study of data interpretation, that still according to authors, corresponds to how to utilize rules and principles of mathematics and statistics in big data mining and interpretation. [5, 6]. According to [4], three problems are urgent to solve in order to benefit from using Big Data in science, engineering, and business applications. The first is transformation of semi- and unstructured data to structured data. The second is about complexity, uncertainty, and systematic modeling. The third, understanding of the relationship of data heterogeneity, knowledge heterogeneity, and decision 13 Annals of Data Science heterogeneity. In this paper we focus on modeling the technique of artificial neural networks (ANN) as well as the inherent uncertainty of the model. ANN is an artificial intelligence technique for classifying, image processing and predicting the data [7]. Tkáč and Verner [8] reviewed publications where neural networks were applied in the business field, for the period from 1994 to 2015. According to the authors, in the two decades of study, artificial neural networks were widely used in several business applications. They classified the articles in 15 areas. The stocks and bonds area had the second-largest volume of publications and showed a significant increase during the study period. It was also observed that a variety of neural network models were presented in an attempt to overcome standard statistical techniques and thus obtain more assertiveness. Consequently, the popularity of new hybrid methods increased and overcame various difficulties of conventional neural networks in addition to achieving more precise results. The type of network most used, according to the authors, considering all the articles evaluated, is the feedforward multilayer and few studies have used other networks. The most common learning algorithm was the backpropagation performed by a descendent gradient search. Artificial neural networks (ANNs) are considered flexible computational structures and universal function approximators. In general, ANNs have a high degree of assertiveness on the (...truncated)