ASSOCIATIVE NETWORKS FOR THE DAILY PREDICTION OF OPENING AND CLOSING VALUES OF STOCK INDICES

Redes asociativas para la predicción diaria de los valores de apertura y cierre de los índices bursátiles

Simón Pedro J. Mejía Uribe 1, Henry Laniado Roda 2, Paula María

Almonacid Hurtado 3 y Jimmy Saravia Matus 4

Recibido: 1/5/2023 - Aceptado: 28/12/2023

ABSTRACT

This paper presents an associative neural network based on Long Short-Term Memory (LSTM) networks to predict the opening, minimum, maximum and closing prices of the Shanghai composite index, PetroChina Company Limited (PetroChina), and Zhongxing Telecommunications Equipment Corporation (ZTE). The data is transformed with time series techniques to render them stationary. Once good results are obtained in terms of the mean absolute percentage error (MAPE), the model is tested with the American Nasdaq Composite Index (IXIC). Similar works have been carried out, such as that of Ding & Qin (2020) where they predict the opening, minimum and maximum prices of an asset. This study goes a step further to predict the closing value following the proposed associative network methodology. Having the opening price and the closing price, it is possible to make investments to generate profitability based on the daily net change in value of the asset.

RESUMEN:

Este trabajo presenta una red neuronal asociativa basada en redes LSTM (Long Short-Term Memory) para predecir los precios de apertura, mínimo, máximo y cierre del índice compuesto de Shanghai, PetroChina y Zhongxing Telecommunications Equipment Corporation (ZTE). Los datos son transformados con técnicas de series temporales para hacerlos estacionarios. Una vez obtenidos buenos resultados en términos del error porcentual absoluto medio (MAPE), el modelo es probado con el American Nasdaq Composite Index (IXIC). Trabajos similares han sido realizados, como el de Ding & Qin (2020), donde predicen los precios de apertura, mínimos y máximos de un activo. Este studio va un paso más adelante en la predicción del valor de cierre siguiendo la metodología de redes asociativas propuesta. Teniendo el precio de apertura y el precio de cierre, es posible realizar inversiones para generar rentabilidad en base al cambio neto diario en valor del activo.

---

PALABRAS CLAVE: red neuronal asociativa, LSTM, series temporales, índice, precios

KEYWORDS: Associative neural network, LSTM, Time series, Index, prices

INTRODUCTIÓN

Predicting stock and financial indices prices is considered a complex problem given the amount of noise in the system, that is, it has small or large fluctuations that cannot be explained with the data available. This can be due to multiple factors that affect the value of financial assets such as news, a change in the laws of a country, a comment on social networks or it can simply be white noise which cannot be tracked. In most recent works there are different approaches to address this problem, some authors seek to build more complex models such as Ding & Qin (2020) with its associative network that predicts the opening, minimum and maximum values for an asset for the following day, others seek solve the problem by obtaining data that better explains the system as Mehta et al. (2021) who integrated data from public opinion, sentiment analysis, and historical data to train multiple models, some researchers prefer to attack the noise problem directly, such as Wu et al. (2021) or Bao et al. (2017) who used Wavelet transformations to reduce noise in the system.

The ability to correctly predict the behavior of the value of assets leads to profits for investors while reducing risk. Therefore, in this work a technique is designed that allows predicting the following day opening, minimum, maximum and closing values of a stock or financial index simultaneously, using associative LSTM neural networks which allow for predicting values simultaneously by taking advantage of the relationship that exists between asset prices. Similar works have been carried out, such as that of Ding & Qin (2020) where the opening, minimum and maximum values of an asset are predicted. This study goes a step further by predicting the closing value using the proposed associative network methodology. With the predictions of opening and closing values for the following day, it is possible to make investments that seek to profit from the daily net change in the value of the asset.

This study begins testing by using data from the Shanghai composite index, PetroChina and ZTE. These assets were selected in order to replicate the results obtained by Ding & Qin (2020). The characteristics used are money traded, net asset value change and the opening, minimum, maximum and closing values for each asset. The data is collected with a daily frequency.

As can be expected, the opening, minimum, maximum and closing values are found to be related to each other, therefore the LSTM associative network technique is an interesting approach to this problem. This document validates the efficiency of LSTM associative networks to predict related historical values and extend good performance to indices from other market sectors.

The rest of this paper is organized as follows. In the second section we present the state of the art and our theoretical framework, here we discuss related works and the different techniques used in the literature. In the third section, we introduce the methodology, present the data and explain analytical models for the construction of the network. In the fourth section we present our results. Finally, in the last section we present our conclusions.

LITERATURE REVIEW

There are many related works that use different techniques and approaches to predict the future value of stocks. A general 3 by 3 forecasting model has been used to classify stocks into 3 categories: overvalued stocks, undervalued stocks, and reasonably priced stocks (Harel & Harpaz, 2021). Using this technique investors can build strategies around the current state of stocks. The classification of price behavior for the following day into “High” or “Low” categories, corresponding to a rise or fall in price, required the use of a decision tree which was trained using articles published worldwide in order to identify the words that most affect the market. The model was tested following this strategy and proved to have superior results to other recently studied techniques (Carta et al., 2021). The efficiency of different models when trained with continuous or discrete data was compared, better results were obtained using the RNN and LSTM models when trained with continuous data (Nabipour et al., 2020).

Using data from the COVID-19 pandemic from social networks where cases of infections and deaths are reported, measurements of sentiment and events that affect the stock market have been extracted. This information has been used as input for SVM and Random Forest Classification models to predict market movements for the next day as shown in Almehmadi, (2021). The latter study managed to obtain superior results by including data from the pandemic. Combining financial data from stocks together with data extracted from Twitter and testing different combinations of these, the random forest model, despite requiring 37 characteristics, offered better results than other models such as the K-Nearest Neighbors that uses 9 characteristics to give its best result (Evans et al., 2021).

It has been shown that it is possible to make more abstract representations of the data using deep learning tools, deep canonical correlation analysis, and deep canonically correlated autoencoders. Results show that these models surpass traditional statistical and machine learning techniques (Huang et al., 2021). The relevance of the variables is identified by combining the results of Logistic Regression, SVM and Random Forest. The variables identified as relevant are used as input in a deep generative model which shows to surpass other approaches in the state of the art (Haq et al., 2021).

Wu et al., (2021) show that Wavelet Transformation (DWT) for noise removal and Extreme Learning Machine as a learning algorithm permits them to obtain better results than other Machine Learning methods. Bao et al., 2017 build a model that reduces data noise and improves prediction accuracy and gains when put into practice by combining Wavelet Transformations with Stacked Autoencoders and LSTM.

Mehta et al., 2021 use SVM models, MNB Classifier, Linear Regression, Naive Bayes and LSTM Networks models, to analyze data from public opinions, sentiment analysis and historical stock data to make predictions. The results show that considering these sources improves predictions. Nti et al., (2021) extract data from multiple sources and randomly select characteristics to train a hybrid CNN-LSTM network model and present high accuracy results, validating the efficiency of taking data from multiple sources. Demi̇rel et al., (2020), compare Multilayer Perceptron (MLP), SVM and LSTM models as to their ability to predict daily opening and closing prices, their results show that the MLP and LSTM models are superior to the SVM model. Ding & Qin, (2020), build an associative LSTM network which simultaneously predicts the opening, maximum and minimum values of with an accuracy better than 95%.

Javed Awan et al., (2021) use big data techniques and structured and unstructured data from social networks to compare linear regression, random forest and decision tree models. Using large volumes of data, they identify that decision trees underperform compared to other models.

Zhang & Bai, (2019) build and compare several models to identify stocks that have the highest probability of rising in price. Random Forest obtained the best results in prediction accuracy and in generating higher annual returns. Khattak et al., (2019) use KNN to classify stock trends and in this way identify future price increases or decreases. Their results show that their proposed model outperforms other simpler models.

Wang & Bai, (2018) generate weak ANN models and demonstrate that it is possible to improve their predictions by creating assemblies of these models using boosting techniques. Wu et al., (2018) show that it is possible to effectively filter the noise in market characteristics by assembling weak classifiers using AdaBoost, thereby obtaining better prediction results. Kohli et al., (2019) integrate historical stock data along with historical prices of commodities such as gold, oil, silver among others, and show that an AdaBoost ensemble model provides better results than other techniques used.

Nti et al., (2020) compare multiple assembly techniques for different models in order to predict share prices. Their results show that Stacking and Blending assemblies demonstrated superior performance compared that of Bagging and Boosting assemblies.

After reviewing related works in the literature, we decided to use a deep learning approach. Deep learning is a subset of artificial intelligence which uses neural networks that learn from large volumes of data. For this study, we use LSTM associative neural networks which, due to their ability to store information over time, are capable of relating events and finding trends that other types of networks cannot.

As a basis for the development of this paper we use the article "Study on the prediction of stock price based on the associated network model of LSTM" by Ding and Qin (2020). In their paper Ding and Qin build an associative network of great predictive performance, and use it to predict the opening, minimum and maximum prices of different stocks and financial indices.

METHODOLOGY

Data

We collect data for the Shanghai composite index (Code 000001.SS), PetroChina (601857.SS) and ZTE (000063SZ). The codes correspond to the identification number used in the Shanghai stock exchange. We construct one dataset for each asset. The datasets contain daily data for the following variables:

• Open: Opening price.

• Low: Minimum price.

• High: Maximum price.

• Close: Closing Price.

• Change: change in the price of the asset.

• Money: Money traded during the day.

Table 1 shows an example of the data for the Shanghai composite index.