1 Introduction

Money investment in a stock market has been one of the favorable investment strategies among a large number of people. Such investment demands knowledge of the market behavior as well as appropriate planning before trading through one or more stocks. The volatile stock market consists of nonlinear complex patterns; future stock price or trend prediction using such noisy data are a difficult task. The market experiences fluctuations due to several influential parameters such as political conditions, various events at national as well as international levels, global economy, to name a few [1]. Hence, having a considerably potential system to predict the stock market trend with good accuracy can be an important aspect. Various fundamental analysis and technical analysis are also performed before investing in a stock market. While fundamental analysis represents an analysis of the numeric data such as earnings, profits, and ratios to determine future predictions, technical analysis focuses on utilizing historical data and modeling techniques that help to identify trends [1,2,3].

To forecast the future stock price as well as price movement, various experiments and analyses have been carried out [4]. Several methods have also been developed using Artificial Neural Network (ANN), Support Vector Machine (SVM), Genetic Algorithm (GA), Linear Regression (LR), and other computational intelligence techniques. The pattern identification capabilities of such techniques are explored to predict the market trend and various models are developed to classify the nonlinear time-series stock data. One of the powerful machine learning techniques is ANN; it is used to learn the complex stock data and present the associations between input and output variables [5]. It is a widely used stock prediction model [6,7,8,9,10]; its capability to adapt the change in the data patterns is useful for deriving important information on the stock trend [11]. The widespread applicability of ANNs includes diverse domains such as intrusion detection systems [12], human action recognition [13], sentiment analysis [14, 15], recommender systems [16], personality trait identification [17], emotion recognition [18], face recognition [19], texture analysis [20], to name a few. Among the wide variety of applications, it can be observed that the prediction performance of a model can be largely dependent on its hyperparameters tuning as well as input features; while a stock market dataset may contain irrelevant features, selection of appropriate features plays a vital role to attain good performance for the given prediction model. It helps to eliminate the irrelevant features and hence, reduce the number of inputs given to a classifier. Such feature selection can enhance the overall prediction efficiency of the model [1, 21].

The stock price and stock trend prediction have attracted a large number of researchers from financial as well as computational fields. Various studies have been carried out using ANN models to forecast the stock market; along with an appropriate feature selection method, the weight initialization of an ANN model is a crucial task to improve the overall model performance. Here, an ANN can be considered as a combination of an input layer, one or more hidden layers, and an output layer with neuron(s) in each of the layers; the edges connect neurons of one layer to the next layer, and weights are provided over these edges. Various survey articles have reviewed the significance of artificial intelligence as well as machine learning techniques, including ANN for a variety of stock market applications [22, 23]. On the other hand, among a large number of diverse ANN applications, the stock market has also been remarkably considered [24]. One of the earlier work proposed solutions to initialize ANN weights which influenced the convergence ability of the model [25]. On the other hand, a statistical method was adopted to generate the weights of an ANN’s hidden layer neurons using vector quantization (VQ) [26]; the equivalence between VQ prototypes and circular backpropagation networks was also provided. Forecasting techniques presented in study [27] used the relevance of input variables and computed the predictive relationships of numerous financial input variables. A sensitivity-based linear learning method was proposed in Ref. [28] for providing an initial set of weights for ANN. Also, various studies adopted GA for optimizing the edge weights of ANNs [29,30,31] to further predict the stock market; GA was helpful in removing irrelevant features and hence, reducing the space complexity of the input features [32]. Results showed that the proposed model gave better results over the linear transformation with conventional ANN. On the other hand, a neuro-fuzzy system and scaled conjugate algorithm were used along with ANN to forecast the stock values in Ref. [33]. ANN was combined with the Genetic Fuzzy Systems in Ref. [34] to predict stock prices, wherein a regression analysis was performed to find the optimal feature space. Hence, integration of appropriate weight initialization and feature selection methods can benefit for the performance enhancement of ANN as well as other prediction models.

In this paper, we adopt a feed-forward Vanilla Neural Network (VNN) consisting of an input layer, a single hidden layer, and an output layer for stock trend prediction. The use of VNN architecture is motivated by universal approximation theorem which states that “a feed-forward network with a single hidden layer containing a finite number of neurons can approximate continuous functions on compact subsets of \(R^n\), under mild assumptions on the activation function” [35]. Generally, the initial weights of VNN are randomly assigned and hence, it may not give the desired performance within the specified number of iterations. Therefore, we propose a new weight initialization technique for VNN using Pearson Correlation Coefficient (PCC) and predict the open price trend of the stock market; we also consider ten technical indicators as provided in [36] for this experiment. While PCC is mainly used to find how well two variables are associated with each other, we propose this novel application of PCC to derive initialization weights of a VNN model; experimentations are performed using feed-forward VNN using ten years of stock trading archival data from years 2008 to 2017 for the stocks of Reliance Industries, Infosys Ltd, HDFC Bank, and Dr. Reddy’s Laboratories. The considered stocks are traded in a good amount and belong to different sectors, viz. Technical, Banking, and Pharmaceutical.

The rest of the paper is organized as follows: an overview of the proposed approach along with the prediction model and proposed weight initialization techniques are provided in Sect. 2; Section 3 discusses the collection of data and computation of technical indicators; the experimental results, ablation study, and analysis are given in Sect. 4; concluding remarks and future scope are discussed in Sect. 5.

2 The proposed approach

In this study, we focus on predicting the direction of the stock price movement using VNN and by employing ten technical indicators as given by [36]. In VNN, initial weights are given on the edges of the input neurons that connect to the neurons of the hidden layer. As convergence tends to increase at each iteration, the weight on each of the edges becomes more accurate toward the prediction of the final output. Hence, it is safe to conclude that the efficiency of a VNN in terms of accuracy and f-measure depends on the edge weights.

To study the significance of weight initialization techniques on stock trend prediction using VNN, we initially consider the random initialization (RI) for the VNN model, wherein the edge weights of VNN are randomly initialized. As discussed earlier, we can improve the overall efficiency of the model by initializing edge weights with the appropriate initial values. Thus, we propose a PCC-based approach to initialize the edge weights connecting neurons between the input and the hidden layers of a VNN; the Pearson correlations are useful in providing the relationship of a particular parameter with the others. In our proposed approach, PCC of an input variable (feature) quantitatively shows its correlation with the output. The correlation values range from \(-\,1\) to 1; while a negative correlation shows that the input variables vary inversely with the output, a positive correlation shows that the input variables vary linearly to the output value. Thus, adopting PCC to derive the edge weights helps the underlined VNN to improve performance. As mentioned in [37], edge weights having non-negative weights tend to give better results than the ones with negative weights. Therefore, we introduce another technique to initialize VNN with Absolute Pearson Correlation Coefficient (APCC) values as the edge weights wherein all edge weights range between 0 and 1.

Using different weight initialization techniques, we demonstrate a VNN model to predict the stock open price movement; a structural overview of the proposed approach is given in Fig. 1. Here, the historical datasets are collected and ten technical indicators (features) are selected as given in Ref. [36]; also, the derived technical indicators are represented in continuous as well as discrete data. These features are combined into a feature set and Pearson correlations are derived; as it can be seen, three techniques, namely, RI (the standard approach), PCC values as weights, and APCC values as weights are considered for model training; the predicted trend is evaluated using accuracy and f-measure metrics. The formulas to compute the selected technical indicators are shown in Table 1; these features serve as inputs to the VNN model along with the three weight initialization techniques.

Fig. 1
figure 1

A structural overview of the proposed approach

Full size image
Table 1 Selected technical indicators and their formulas [36]
Full size table

2.1 Feed-forward VNN: an ANN model

Various machine learning models are used for financial predictive research. ANN is one of the most frequently used methods for the same; it works well with more number of input features and a larger dataset. It has a capability to analyze nonlinear relationship between input and output to derive unknown patterns [38]. The stock market prices are highly volatile in nature; also, various factors can affect these prices. In order to derive the historical data patterns, the stock market dataset of many years can be used for training such ANN models.

In this study, we adopt feed-forward VNN with an input layer, a single hidden layer, and an output layer. Motivated by the universal approximation theorem [35], we provide a feature set comprising of ten technical indicators as an input to the VNN model. For the hyperparameter tuning, neurons in the hidden layer are varied from 10 to 100, whereas the output layer contains one neuron that provides the value, 0 or 1, as an output, i.e., trend down or up, respectively. Every neuron of a particular layer is connected to every other neuron in the next layer via edges that create a dense neural network. These edges have some weights that are used to perform computation on a particular input feature. With the help of these weights, the features that contribute more to predicting the output as compared to others can be assigned with appropriate weights. Thus, finding a correlation between input parameters and the output and further, using this correlation as weight for respective input neuron can result in improvement in stock trend prediction accuracy.

2.2 PCC-based weight initialization of feed-forward VNN: a novel application of PCC

Correlation between the input and output data is a measure of how well they are related to each other. The most common measure of correlation is the Pearson Correlation Coefficient (PCC). PCC shows a linear relationship between the two sets of data.

In this article, we compute PCC between the input features and output feature, i.e., open price; we initialize weights on the edges connecting neurons from the input layer to the hidden layer with respective PCC values. Hence, all the edges emerging out from an input feature are initialized with the PCC value of that input feature and the open price. Here, PCC values range between \(-\,1\) and 1. If two parameters are positively correlated then the coefficient is positive, i.e., between 0 and 1, otherwise, the coefficient is negative, i.e., between \(-\,1\) and 0 which indicates a negative correlation. Using these coefficients as weights increases the accuracy because the highly correlated features receive positive weights and negatively affecting features receive negative weights as per their correlation with the output. The PCC (r) for variable x and y can be calculated using Eq. 1 [39].

$$\begin{aligned} r = \frac{ N \sum xy - \sum x \sum y }{ \sqrt{[N \sum x^{2} - (\sum x)^{2}] [N \sum y^{2} - (\sum y)^{2}]} } \end{aligned}$$
(1)

where, N denotes the number of pairs of scores, \(\sum xy\) is the sum of products of paired scores, \(\sum x\) and \(\sum y\) indicate the sum of x scores and y scores, respectively; similarly, \(\sum x^2\) and \(\sum y^2\) give the sum of squared x scores and y scores, respectively.

After initializing the input neurons with respective PCC values, the weight update algorithm changes weights using Stochastic Gradient Descent (SGD) at every iteration during the training. The loss is measured in terms of binary cross-entropy. Here, the hidden layer has a Rectified Linear Unit (ReLU) activation function and the output layer has a sigmoid activation function. Learning rate (lr), number of neurons in the hidden layer (n), number of iterations, i.e., epochs (ep), and momentum constant (mc) are the parameters of VNN that must be determined accurately to get good results. In this study, we consider continuous (as given by Ref. [36]) as well as discrete (as given by Ref. [40]) data representations of the considered technical indicators. For each representation, n is varied from 10 to 100 with an increment of 10, mc is varied from 0.1 to 0.9 with an increment of 0.1, ep is ranged from 1000 to 10,000 with an increment of 1000, and lr is set to 0.1; these parameter levels for feed-forward VNN are similar to Ref. [40]. The parameter combinations give \(10\times 10\times 9=900\) different parameter settings for a specific company with a particular weight initialization technique using which VNN is trained. Thus, for a single company and for each data representation, we have performed 2700 iteration treatments in total which includes 900 training runs for each of the three weight initialization techniques, i.e., RI, PCC, and APCC.

3 Research data collection

This section discusses the process to build the dataset and computation of technical indicators used for the experiments. Ten years of daily stock details for the years 2008 to 2017 are collected for Reliance Industries, Infosys Ltd, HDFC Bank, and Dr. Reddy’s Laboratories from the National Stock Exchange (NSE) of India [41]. The increase and decrease information of the datasets are important to find out whether the given data are balanced toward the ups and downs of the market. To provide an overview of the selected datasets, we elaborate the case study of Reliance Industries; the dataset covers a total of 2479 trading days and its increase and decrease cases of the distribution for each year are given in Table 2. It indicates that the total number of days with increase direction is 1253, while that for decrease direction is 1226, i.e., 50.54% cases for the increase direction and 49.46% cases for the decrease direction. For the experimentation, a similar distribution is obtained for other datasets.

Table 2 The count of increase and decrease cases in the entire Reliance Industries dataset
Full size table

VNN architecture contains various parameters such as connection weights, biases, learning rate, and a number of nodes in a hidden layer; each one of them affects the results based on the value assigned to them. Therefore, parameter tuning is required to select appropriate parameter values. Hence, we use a subset from the dataset which is \(20\%\) of the entire dataset. This subset is called the parameter selection dataset. This dataset is used in preliminary experiments to check for different parameter combinations to find a suitable parameter set that can be further used for the final experimentation using the entire dataset. For the preparation of the parameter selection dataset, we use two equally distributed \(10\%\) of the dataset for training and the rest \(10\%\) for the holdout. Here, an equal distribution indicates that \(10\%\) of increase cases, as well as decrease cases, are considered for training, as well as holdout, to ensure an unbiased distribution. For the final experimentation, the entire dataset is equally divided into training and holdout cases. From each year’s dataset, \(50\%\) of increase cases are assigned for training and the rest \(50\%\) of increase cases are assigned to the holdout; the same procedure is carried out for decrease cases. Thus, equally distributed increase and decrease cases in training and holdout datasets are used to represent the entire dataset in equal parts. For each year, the count of increase and decrease cases for parameter selection of the Reliance Industries dataset is shown in Table 3 for parameter selection and for the final experimentation; these experimental settings are the same as [40].

Table 3 The count of increase and decrease cases for training and holdout in the Reliance Industries dataset
Full size table
Table 4 Statistics for the selected indicators for the Reliance Industries dataset
Full size table

The statistics of the selected indicators (Table 1) for the Reliance Industries dataset are given in Table 4; these technical indicators are used as a feature set. All indicators are further scaled down between \([-1,1]\) so that each indicator is linearly normalized and thus, the large valued technical indicator is on the same scale with a small valued technical indicator. The output dataset consists of the value, 0 or 1; here, if the price of a stock at day T is higher than the price of that stock at day \(T-1\), then the output is assigned with 1, otherwise 0. In this study, the feature set of ten technical indicators is provided as the input to VNN which is further trained with three different weight initialization techniques, namely, RI, PCC, and APCC.

4 Experimental results and analysis

Table 5 Simulation parameters
Full size table

As explained in Sect. 3, VNN model is trained and tested on each of the four datasets using all the three weight initialization techniques, RI, PCC, and APCC. The simulation parameters considered for the experiments are given in Table 5. Here, we carry out the experimentation on two different data representations as given by Refs. [36] and [40]; while the former approach considers continuous data and further provides the same to VNN model, the latter approach introduces an opinion layer, wherein each technical indicator undergoes the trend identifying calculation and the trading data of each technical indicator is further represented using \(+\,1\) or \(-\,1\). Such data representation has displayed performance improvement in Ref. [40]. To extend the applicability of our proposed approach, we consider the continuous (Ref. [36]), as well as discrete (Ref. [40]), data representations and compare results of companies, Reliance Industries, Infosys Ltd, HDFC Bank, and Dr. Reddy’s Laboratories, with respect to all the three weight initialization techniques as given in Tables 6, 7, 8 and 9, respectively. These experiments have been performed on Python using Keras framework [42]. For the analysis purpose, we consider accuracy and f-measure as the performance metrics. As we have carried out a series of operations with a different number of neurons, epochs, as well as mc values, the comparison tables present the top five results derived among the given combinations for each weight initialization technique and both of the considered data representations.

As shown in Table 6, the top five results for Reliance Industries dataset indicate that the proposed weight initialization techniques, i.e., PCC and APCC, outperform the conventional weight initialization technique for the VNN model in terms of accuracy for a continuous data representation. On the other hand, for discrete data representation, PCC and APCC have attained near-comparable results with respect to the standard weight initialization technique (RI). Also, it can be observed that the discrete data representation has provided higher results as compared to the continuous data for each of the three weight initialization techniques. Thus, the results show comparable or higher performance improvement for the Reliance Industries dataset.

In a similar way, Table 7 presents the top five results of Infosys Ltd dataset. Here, VNN model with APCC as weights outperforms while comparing the prediction accuracies for continuous data, whereas VNN model with PCC-based weights provides better or comparable results. Subsequently, VNN model with PCC-based weights outperforms for the discrete data representation, whereas VNN model with APCC-based weights indicates comparable results or an overall performance improvement. As in the case of Reliance Industries dataset, each of the weight initialization techniques for the discrete data representation approach has considerably higher prediction performance than that of the continuous data.

Table 8 depicts the top five results of the HDFC bank wherein the VNN model with PCC as weights perform slightly lower than the VNN model with RI for continuous data; on the other hand, VNN model with APCC values as weights outperforms the conventional technique. As it can be observed, the average accuracy of using RI with VNN for discrete data is slightly higher than the proposed weight initialization techniques here, however, the top most accuracy of PCC-based weights for VNN is equivalent to that of RI and f-measure indicates higher value. These results are significant improvements to that of the continuous data for HDFC Bank dataset as well.

On the other hand, Table 9 describes the top five results of Dr. Reddy’s Laboratories. While both the proposed weight initialization techniques, i.e., PCC and APCC, outperform the standard weight initialization technique, RI, for continuous data, a notable performance improvement is noticed for the discrete data for PCC-based VNN weights. Also, the higher results with discrete data representation are observed here. The above analysis is empirically carried out based on the accuracy metric.

Table 6 Comparison of VNN model with three weight initialization techniques for Reliance Industries dataset
Full size table
Table 7 Comparison of VNN model with three weight initialization techniques for Infosys Ltd dataset
Full size table
Table 8 Comparison of VNN model with three weight initialization techniques for HDFC Bank dataset
Full size table
Table 9 Comparison of VNN model with three weight initialization techniques for Dr. Reddy’s Laboratories dataset
Full size table

In order to verify the statistical significance of the proposed approaches, PCC and APCC as weight initialization techniques for the VNN model, we compare the results with the RI-based VNN using the Wilcoxon signed-rank test. The results show that the proposed approach, APCC, as a weight initialization technique for VNN outperforms randomly initialized VNN for all the four companies for the continuous data wherein the obtained p-value is less than .05 [43]. On the other hand, PCC-based weight initialization for VNN model using discrete data representation indicate comparable or higher performance accuracy; for Dr. Reddy’s Laboratories dataset, it attains p-value less than .05 using Wilcoxon signed-rank test.

In this article, we select the ten technical indicators for a fair comparison based on Refs. [36] and [40]; the effectiveness of different weight initialization techniques can be further analyzed using the selected set of features, i.e., technical indicators, through an ablation study. An ablation study can be an important approach to understand the significance of individual indicators as well as a group of the selected indicators; such a study can be carried out to eliminate the least correlating input features from the considered indicators. In this article, we evaluate the impact of dropping a particular feature through an ablation study; the experiments are performed on discrete data representations due to their significant improvements observed through Tables 6, 7, 8 and 9. Here, for each dropped feature under the ablation study, there are a total of 2700 combinations of the number of neurons, epochs, mc values, and weight initialization techniques for each dataset. Hence, by dropping one of the ten features for all four datasets, a total of \(2700 \times 4 \times 10 = 108,000\) experimentations are carried out for the ablation study. As expressed earlier, comparisons of the mean and median accuracies of the top five results attained for each dataset are represented for three weight initialization techniques in Figs. 2, 3, 4 and 5. The graphical comparisons are further enhanced with coefficient of variation (CV) metric to indicate dispersion of the results [44]. Here, F0 indicates that all ten features are considered, whereas F1 to F10 denote the removal of SMA, WMA, STCK%, STCD%, MACD, RSI, WILLR%, CCI, MOM, and A/D Osc feature, respectively, and the rest nine features are considered for the specific set of experimentations.

Fig. 2
figure 2

Comparative analysis using ablation study for Reliance Industries dataset

Full size image
Fig. 3
figure 3

Comparative analysis using ablation study for Infosys Ltd dataset

Full size image

It can be observed that PCC and APCC-based weight initialization techniques attain higher accuracies as compared to RI-based weight initialization in a large number of cases for an individual set of considered features. As compared to F0 with all ten features, the removal of one of the features (F1 − F10) has considerable impacts on the stock trend prediction; while dropping F2, i.e., WMA, significantly reduces the prediction accuracy, removing F10, i.e., A/D Osc denotes performance improvement; on the other hand, removal of certain features have a minimum impact on the resultant accuracies. The effectiveness of our proposed PCC and APCC-based weight initialization techniques can be viewed through individual ablation studies. We further analyze the same using CV metric; the smaller values indicate smaller variability and thus, higher stability [44]. It is shown in Figs. 2 and 4 that CV values based on PCC, as well as APCC, are smaller than that of RI in the majority of the cases for Reliance Industries and HDFC Bank datasets; while close CV values are attained for Infosys Ltd dataset using all three weight initialization techniques (Fig. 3), overall smaller CV values can be observed using APCC for Dr. Reddy’s Laboratories dataset (Fig. 5). Thus, the ablation study indicates the potential effectiveness of PCC and APCC-based weight initialization techniques over RI for a different set of features; it also helps to evaluate the impact of individual technical indicators in deriving the stock trend prediction. The same can be further utilized to engineer network weights related to the feature importance.

Fig. 4
figure 4

Comparative analysis using ablation study for HDFC Bank dataset

Full size image
Fig. 5
figure 5

Comparative analysis using ablation study for Dr. Reddy’s Laboratories dataset

Full size image

5 Concluding remarks and future scope

The paper aimed to predict the direction of the stock movement. We have accomplished the task by using the VNN model based on ten years (2008–2017) of stock trading archival data of Reliance Industries, Infosys Ltd, HDFC Bank, and Dr. Reddy’s Laboratories. For this, we used ten technical indicators that were given as inputs to the VNN to predict the direction of the stock price.

In this paper, we have proposed a potential solution that computes the correlation coefficient between each input variable and the output variable and further uses these correlation values to initialize their respective edge weights that connect input neurons to the hidden layer neurons in the given VNN model. In the prior study, the aforementioned ten technical indicators were directly used to predict the stock movement; in contrast to that, we propose to initially compute the PCC values of these indicators and assign them as weights to their respective VNN edges. Subsequently, we also propose to extend the same approach using APCC-based weights for the VNN model. For a detailed understanding of the proposed approach, we carry out a series of experimentations using four datasets and three weight initialization techniques, i.e., RI, PCC, and APCC. The hyperparameter tuning can be a critical challenge and therefore, we consider to evaluate our proposed approaches with different combinations of the number of neurons, number of epochs, as well as variable mc values. Hence, the results obtained using various combinations of these hyperparameters are compared using accuracy as well as f-measure metrics. We further extend the applicability of our proposed approach and adopt an opinion layer as proposed in Ref. [40] which presents the technical indicator data using \(+\,1\) and \(-\,1\), i.e., discrete representation. We further carry out the experimentations with the discrete data representation and evaluate the performance for each of the four companies using each of the three weight initialization techniques. We also compared the prediction performance between the continuous and the discrete data representations as given in Tables 6, 7, 8 and 9. The results indicate that the proposed approaches deliver better or comparable results in predicting the movement of stock prices than that of RI. Subsequently, it has been observed that a significant performance enhancement is achieved with the discrete data representation. We further evaluate the statistical significance of the results that demonstrate that the proposed weight initialization techniques, PCC and APCC, achieve better or comparable results than the conventional weight initialization technique, RI.

The effectiveness of the proposed approach can be evaluated using an ablation study wherein one feature is dropped and the experiments are carried out with the remaining set of features to study the impact of the dropped feature. We considered to apply an ablation study on our set of features such that one of the ten technical indicators was dropped for each experiment. The removal of a feature can indicate the impact of that feature on the prediction performance. The exhaustive experimentations indicated how certain features could influence the prediction; while removal of WMA technical indicator showed to reduce the prediction accuracy, removal of A/D Osc technical indicator resulted in accuracy improvement. Thus, a thorough analysis can be beneficial for deriving the correlation, as well as significance, of individual technical indicators. We extended our study by incorporating CV metric to analyze the variability in derived prediction results. The smaller CV values indicated higher stability; among a different set of experimentations, an overall smaller or comparable CV value could be attained using PCC, as well as APCC, as compared to RI-based weight initialization. Such consistencies were observed for individual ablation studies as well. The results indicated the importance of PCC and APCC-based weight initialization techniques and their potential for further acceptance as weight initialization techniques for other models.

An empirical analysis of the proposed approach has been presented using accuracy and f-measure. However, along with accuracy and f-measure, several other metrics can be incorporated for further analysis. While the technical indicators have been derived using 10-day period, finding a suitable period for technical indicators is worth investigating. The proposed approaches can be extended for other domains that use VNN or ANN models for the given application. Initializing VNN with proper initial weights using PCC or APCC suggests that the model gets good initial weight values for the input parameters according to their contribution toward the output variable. Thus, we can state that these weights play an important role to determine the performance of VNN. As a part of the future work, other macroeconomic variables such as inflation, currency exchange rates, or an average stock volume, can also be used along with the technical indicators to understand the dominance of such factors on the stock market as well as to enhance the prediction performance; the variety of data can be further fused to enhance the forecasting capability of the model [45]. On the other hand, the impact of such indicators on the prediction performance can be studied through ablation study for possible improvement in accuracy and reduction in training time; also, customized weights can be initialized based on the feature importance. Some of the recent works have indicated that companies having been listed on different stock exchanges can provide useful information [3, 46]; such factors can be potentially explored with improved weight initialization such as that proposed in our approach. The effectiveness of using PCC and APCC can also be evaluated for a variety of deep learning-based models for stock trend prediction [47].

The concept of transfer learning can be useful to forecast the trend of the stock movement for indices with incomplete data. The neural network model can be trained on the dataset of any other company and further, the knowledge gained by the model can be used to forecast the trend of the stock movement for the identified company. While we present PCC and APCC-based weight initialization techniques, other feature weight techniques may be compared and/or potentially combined to evaluate different ways to derive useful features [48]. The focus of this study is to predict the stock direction for short terms; this work can also be extended for the long-term prediction which may require other parameters such as analysis of stock’s overall profit returns, the impact of the company throughout the financial year, revenue, to name a few. One of the considerable aspects is hyperparameter tuning; while, we have considered a series of experimentations to derive the top five combinations, a metaheuristic approach such as particle swarm optimization [1] can be further utilized. Other domains can also be explored using the proposed methods that may help in improving the performance of the model. The proposed approaches can also be integrated with deep learning techniques and can be considered as another future redirection.