Impact of chart image characteristics on stock price prediction with a convolutional neural network

Guangxun Jin; Ohbyung Kwon

doi:10.1371/journal.pone.0253121

. 2021 Jun 23;16(6):e0253121. doi: 10.1371/journal.pone.0253121

Impact of chart image characteristics on stock price prediction with a convolutional neural network

Guangxun Jin ¹, Ohbyung Kwon ^2,^*

Editor: Thippa Reddy Gadekallu³

PMCID: PMC8221485 PMID: 34161352

Abstract

Stock price prediction has long been the subject of research because of the importance of accuracy of prediction and the difficulty in forecasting. Traditionally, forecasting has involved linear models such as AR and MR or nonlinear models such as ANNs using standardized numerical data such as corporate financial data and stock price data. Due to the difficulty of securing a sufficient variety of data, researchers have recently begun using convolutional neural networks (CNNs) with stock price graph images only. However, we know little about which characteristics of stock charts affect the accuracy of predictions and to what extent. The purpose of this study is to analyze the effects of stock chart characteristics on stock price prediction via CNNs. To this end, we define the image characteristics of stock charts and identify significant differences in prediction performance for each characteristic. The results reveal that the accuracy of prediction is improved by utilizing solid lines, color, and a single image without axis marks. Based on these findings, we describe the implications of making predictions only with images, which are unstructured data, without using large amounts of standardized data. Finally, we identify issues for future research.

I. Introduction

Successful stock trading is highly important for investors. Considering multiple stocks when trading, they must buy or sell by selecting appropriate stocks with attention to the timing of the sale. Accordingly, stock price prediction is a long-standing research issue. Because stock prices are determined by a wide variety of variables [1], prediction seems to be a random walk, especially using past information [2].

Stock price prediction has traditionally been performed using linear models such as AR, ARMA, and ARIMA and its variations [3–5]. However, the assumption that stock price fluctuations are linear is an oversimplification of the factors affecting stock prices; other assumptions are that hidden dynamics are at work and only overall trends can be seen [6,7]. Nonlinear models such as artificial neural networks (ANNs) and convolutional neural networks (CNNs) can also be used to predict stock prices [6,8]. Deep neural networks (DNNs), which utilize deep learning algorithms for stock price prediction, may also been used. Although recurrent neural networks (RNNs) and the like are also being proposed [9], CNNs appear to be relatively superior for the purpose of stock price prediction [10].

Existing studies of stock price prediction mainly use numerical data. In recent years, attempts have begun to predict stock prices through image-based discrimination using the stock price graph itself as input data without using profile information or numerical data [2,11,12]. Even stock forecasting experts use the stock chart itself to make predictions. The reason for this is that the chart contains information that can form the basis for prediction; the extracted knowledge pertains highly to stock price prediction. In this study, we examine the possibility that this knowledge can be extracted using a deep learning model. Some previous studies involving deep learning-based stock price prediction have used additional charts as training data. However, most stock price chart types are fixed so that the characteristics of the chart image that can affect stock price prediction are not easily understood [2,12]. In deep learning algorithms such as CNNs, since the characteristics of the image affect predictive performance, the shape of the chart may be an important variable [11,13]. However, only a few studies examine how the characteristics of stock price charts affect the performance of deep learning algorithms (especially CNNs) for stock price prediction.

The purpose of this study is to analyze the effects of stock price chart characteristics on the stock price prediction performance of CNNs [14]. The present study focuses on improving the quality of stock price image data by implementing a rigorous preprocessing technique: selecting optimal image characteristics. To this end, various types of images were generated from actual stock price data and significant differences in CNN prediction performance were identified for each characteristic. CNN parameters included in this study are dropout, number of filters, and activation functions, as these can strongly affect inference performance. The results of this study reveal that stock prices can be predicted with images only using CNNs.

The findings of this experimental paper are as follows. First, even if no numerical information is prepared, stock price predictions can be made with a high level of accuracy using only image data. Second, when predicting using image data, the preprocessing step of selecting an optimal image shape that can increase accuracy is absolutely necessary. Our experiments identify which factors affect prediction accuracy and which characteristics of charts are useful to increase that accuracy. We hope that our results may be useful for deep learning practitioners to select optimal hyperparameters in a minimal amount of time for the purpose of stock price prediction [15].

II. Background

2.1 Stock price prediction

As previously mentioned, stock price prediction has traditionally been performed using linear models such as AR, ARMA, and ARIMA and its variations [3–5]. In these models, the stock price is used as in a time series model as an independent variable for forecasting. For multivariate models, the company profile (returns, sales, various financial ratios, etc.) [16], stock trading volume [17], stock price trends over a certain period of time, and the number of times a specific company’s name was mentioned in search engines [18] have been proposed (see Table 1).

Table 1. Literature on stock price prediction.

Method	Determinants	Data Frequency	References
SVM	Single words, bigram, polarity, noun phrase	Daily	[19]
ANN, SVM, linear regression	Economic info, financial ratios	Yearly	[16]
MLP, RNN, CNN	Close price	Daily	[10]
ANN, genetic algorithm	Close price	Daily	[20]
LSTM	Rate of return	Daily	[21]
CNN	Open and close prices	Daily	[7]
CNN	Close price	Monthly	[12]
ARIMA, LSTM, CNN	DJIA index closed	Daily	[22]
SVM, CNN	Close price	Every 30 minutes	[2]
LSTM, CNN	Close price	Daily	[23]
LSTM, CNN	Four commodity futures and two equity index closed	Every 5 minutes	[24]

Open in a new tab

Recently, attempts have been made to predict stock prices using nonlinear models. Various classification algorithms such as ANNs, naïve Bayes, support vector machines (SVMs), and random decision forests have been used. Nonlinear classifiers for stock price prediction show better performance than existing linear models [3–5,10]. Among various options, artificial neural networks (ANNs) have provided good predictive performance and become widely used financial forecasting tools [25–27]. For example, a comparison between ANN and SVM classifiers revealed that ANNs perform better than SVMs in forecasting on the Istanbul Stock Exchange [25]. In order to improve the predictive performance of ANNs, certain features need to be more easily extracted; therefore, feature engineering methods such as principal component analysis (PCA) are now being used to improve the performance of ANNs [28]. However, ANNs do not always perform well. In another study, ten technical indicators were used to classify the up and down movements of stocks using ANN, SVM, random forest, and naïve Bayes classifiers. The authors found that the random forest model outperformed the others [29]. Therefore, a hybrid method has been proposed that combines several prediction techniques such as finding the initial weight of the ANN using a genetic algorithm and a simulated annealing algorithm and then learning the network using a back propagation algorithm. This combinative approach had better results than the standard ANN method [20].

Recently, research was conducted to predict stock prices using deep learning algorithms. For example, DNNs and numerical data have been used [30,31] and long short-term memory (LSTM) or RNN classifiers have been used with unstructured data [32,33]. In general, deep learning-based stock price prediction outperforms conventional stock price prediction in the sense that deep neural network architectures are capable of capturing hidden dynamics and making more accurate predictions [6]. Inference is also possible with a large number of input features and a convolution technique unique to deep learning [7]. However, when applying a nonlinear model such as those involving deep learning algorithms, there is a concern about overfitting. Among deep learning algorithms for stock price prediction, DNNs and RNNs have been proposed in addition to CNNs [9], but CNNs appear to be relatively superior in terms of performance [10]. However, this method has the disadvantage that a large amount of data must be prepared, which requires considerable cost and time. Data availability and data preparation costs are also important when corporate information is not well shared. Despite these difficulties, the input data and processing methods involved in CNNs play an important role in identifying the quality of the extracted features and predicting stock prices. For example, when performing stock price prediction with CNNs using 10-day data from 100 companies listed in Borsa Istanbul, the prediction performance improved to some extent by combining similar features, generating technical indicators, and identifying time-lagged features [34].

While most existing studies of stock price prediction as described above were based on structural data, big data research has recently progressed to stock price prediction using informal data such as text and images. In particular, information from social media on corporate activities is unstructured text information. Social media provides information that can be determined as favorable or unfavorable to companies through sentiment analysis [35]. However, it is difficult to explain the rise and fall of stock prices using only data based on sentiment analysis. For this reason, big data researchers combine this data with other variables to make predictions, but even in this situation, the problem of data availability reappears.

2.2 Stock price prediction based on image-based deep learning

In recent years, attempts to predict stock prices through image-based discrimination have begun using the stock price graph itself as input data without using profile information or numerical data related to stock prices [2,11,12]. This is due to the assumption that future stock price predictions can be made more accurately by learning trends of the past and the premise that a pattern similar to those in the past will occur repeatedly [12]. When deep learning algorithms such as CNNs are used for stock price prediction, they predict the stock price only, not using an image as a dataset [6,8]. The fact that stock prediction analysis experts are also making predictions based on the image characteristics shown in stock charts implies that the chart image contains information that can aid prediction. Image characteristics can provide clues; the knowledge these images provide can be extracted using deep learning models such as CNNs.

III. Methods

This study examines the effects of the features of images from various stock charts on the accuracy of stock price predictions, focusing in particular on the role of filter and dropout, the main hyperparameters of CNNs, as mediating variables between image features and prediction performance (see Fig 1). To verify this role, we propose the following hypotheses:

[Hypothesis 1] Image characteristics of stock charts affect the accuracy of stock price prediction by CNNs.

[Hypothesis 2] The filter parameter plays a mediating role between image composition variables and prediction accuracy.

[Hypothesis 3] The dropout parameter plays a mediating role between image composition variables and prediction accuracy.

3.1 Image characteristics

Table 2 shows the image characteristics of the stock price chart to be considered in this study. Several types of charts were targeted: plots, barplots, and histograms, which are the most commonly used stock price charts. Second, we determined whether the direction of the chart (horizontal or vertical) also affects discrimination and prediction performance. Third, an axis heading displayed can also affect discrimination and prediction performance. Explanations or scale marks on the x- or y-axis provide helpful information for understanding the chart for human analysts, but in the case of CNNs, if text information is not interpreted and is viewed only as part of the image, it may be considered noise. Therefore we included it in the analysis. Fourth, we compare performances between two different forecasting dates: the forecast for the next day of the period shown in the stock price chart, or the forecast 5 days later (i.e., one week later, excluding the weekend). In the literature, there are studies that examine predictions after 1 day and studies that do so after 5 days [2,12]. Fifth, whether data on the chart is displayed in the form of a bar or a line has a similar meaning to the presence or absence of color on the surface (see below). If it is displayed as a bar, the thickness and shape of the bar are reflected in learning, and if it is expressed as a line rather than as a bar, it corresponds to the narrowing technique in image processing, meaning that only direction and length information is provided. Sixth, whether the bars and lines are solid or dotted can affect performance.

Table 2. Image characteristics.

Image characteristics	Variable name	Values
Series	X1	time series = 1, frequency = 2
Graph type	X2	plot = 1, hist = 2, barplot = 3, stripchart = 4
Graph direction	X3	horizontal = 1, vertical = 2
Appearance of axis	X4	no = 0, yes = 1
Forecast date	X5	after 1 day = 1, after 5 days = 2
Bar/line	X6	bar = 1, line = 2
Color	X7	no = 0, yes = 1
Color tone	X8	monotone = 0, colorful = 1
Dotted/solid line	X9	solid line = 0, dotted = 1 (density)

Open in a new tab

Finally, the presence or absence of color on the surface of the chart can also affect discrimination and performance. If there is color on the surface, it may be recognized as an area, and if not, it may be recognized as an outline. In image processing, there is a skeletonization process [36]. Combining the skeletonized algorithm with the CNN as the recognition algorithm reduces the impact of the shooting angle and environment on the recognition effect and improves the accuracy of gesture recognition in complex environments [37]. In the context of stock price charts, a color on the side provides information indicating the share of the stock price in the overall image, and no color means it is seen as a line, which means that the degree of change in the stock price is emphasized.

Also, depending on the problem area to be identified, the presence or absence of color may be advantageous for discrimination. Even black, white, and gray color may be advantageous, especially in fields such as medical imaging diagnosis [38] or security search [39]. In some cases, preprocessing, or graying, in a color image is performed in order to remove unnecessary features that interfere with interpretation. Thus, both black and white and color images affect discrimination and prediction performance. This is especially true when it is difficult to give a special meaning to a color.

3.2 CNN characteristics

CNNs are feed forward neural networks that have a unique effect on graphic image processing. A convolutional layer and a pooling layer are included in the network structure. CNNs are particularly widely used as deep learning networks in many recent studies. Representative CNNs include LeNet-5, VGG, and AlexNet. In a fully connected neural network, all neurons in adjacent layers are connected together and the number of outputs can be arbitrarily determined. However, the problem with fully connected neural networks is that the form of data is not represented. For example, input data in the form of an image is composed of the height, length, and channel of the three-dimensional shape (H, W, C). However, the input data, which is a fully connected neural network, must be compressed from 3D data to 1D data. With a MNIST data set, for example, the data image must be input in 784 (28*28) format with 3D data consisting of 1 channel, 28 pixels high and 28 pixels long, or (28,28,1).

Since image information is compressed into one dimension, some spatial information is lost during this process. On the other hand, since CNNs input 3D data and the output value is also 3D, they are more appropriate for classifying images, as the form can be maintained and the data type need not be changed. Therefore, in this paper, we selected a CNN algorithm for the prediction of stock prices using graphic data (see Fig 2).

In addition to the image characteristics considered after image formation, the shape of the dropout and filter can also affect CNN performance. In fact, there are thousands of network characteristics that affect CNNs. Therefore, in this study, only the numbers of dropouts and filters were considered (see Table 3).

Table 3. CNN characteristics.

Hyperparameter	Values
dropout	(0.25, 0.25), (0.5,0.5)
filter	(2,4,8),(3,6,12),(4,8,16),(5,10,20),(6,12,24),(7,14,28),(8,16,32),(9,18,36)(10,20,40),(11,22,44),(12,24,48),(13,26,52),(14,28,56),(15,30,60),(16,32,64)

Open in a new tab

IV. Experiment

4.1 Data

To create a stock price chart, we first collected daily closing prices for the 5 years from 2015 to 2019 for all 789 companies listed in the KOSPI Index from Dataguide. Note that companies closed during this period were excluded, and non-business days were excluded for other companies. Next, a chart image was created for each company at the closing price for each period of 30 days, with attention to parameters such as vertical/horizontal, line/bar, and colored/colorless for three types of charts: barplot, plot, and histogram. In total, 30 types of chart images were created by making various changes in chart characteristics. Examples of the generated chart images are shown in Fig 3.

In total, 45,407 images were created for each chart, and 1,424,065 images were collected to make up the image dataset. Charts were created by collecting data for each company every 30 days throughout the 5-year period. Labels were automatically generated for each image, and each image was compared to the closing price of stocks 5 days later. Fig 4 provides an example.

Changes in the closing price on the forecast date reveal how much the price rose or fell (after 1 day or 5 days) based on the last day of the 30 days after the image was created (see Eqs (1), (2)). If the image name rate of return was greater than 0, 1 was automatically added to the end of the previously created label; otherwise, 0 was automatically added. If the result was greater than 0, the value was added to the stock price increase dataset; if it was less than or equal to 0, the value was added to the stock price decline dataset. Training and verification data were randomly arranged in a ratio of 9:1.

Y i e l d a f t e r 5 d a y s = \frac{v a l u e 35 - v a l u e 30}{v a l u e 30} * 100

(1)

Y i e l d a f t e r 1 d a y = \frac{v a l u e 31 - v a l u e 30}{v a l u e 30} * 100

(2)

4.2 CNN model

In this paper, the CNN model structure was constructed using Keras. Keras is an advanced ANN API that enables stock price prediction with the simplest possible code[40]. Keras can also shift computing from CPU to GPU acceleration without code changes.

In the learning phase, a CNN was constructed to train and test the above-described data. The constructed CNN (see Fig 5) consisted of an input layer (28x28), 3 convolution layers, 3 max pooling layers, 2 dropout layers (0.25, 0.50), fully connected layers (128), and an output layer. By reducing the filter size, more detailed image features can be captured. In this study, this parameter was optimized to a 3x3 filter size. Since convolutional and pooling layers are composed of basic units, the number of convolutional layers and the size and number of each convolutional layer filter can affect the performance of the CNN.

A neural network structure as shown in Eq (3) was formed for the convolution operation in the CNN structure (W is the weight, x is the input, and b is the bias). In the middle step of the network, the ReLU function as in Eq (4) was used, and in the last step, the output was obtained using softmax.

e i = \sum_{j} W i, j X j + b i

(3)

r e l u f (x) = {\begin{matrix} 0, x < 0 \\ x, x \geq 0 \end{matrix}

(4)

4.3 Evaluation methodology

Next, 10-fold cross validation was performed to determine the inference performance for each image with the image dataset secured for the test and the configured CNN model. At this point, evaluation metrics are needed to compare the results of our method with other methods. Accuracy is a common metric used for this purpose [2,11,12]. Accuracy was used in this paper because there is no imbalance issue in the data used in this study; thus, it is unnecessary to use the F-measure [7].

4.4 Experimental environment

To establish the experimental environment, we followed the method of Eapen [41]. We developed the model using the Python programming language (Python 3.7 in a Windows 64-bit system environment). In addition, development tools included PyCharm and Anaconda3; Keras was used to build the network model structure, and the TensorFlow framework was used at the bottom. The relevant parameters are outlined in Table 4.

Table 4. Experimental parameter settings.

Hyperparameter	Values
Filter size	(3,3)
Size of max-pooling	(2,2)
Optimizer	Adam
Activation function	Activate function combinations between layers: {ReLU, Elu, Selu, Softsign, Softplus, RRelu, Gelu} Output: softmax
Learning rare	0.001
Epochs	200
Batch size	200
Padding	Same
Dropout rate	(0.5,0.25),(0.5,0.5)
Number of layers	4

Open in a new tab

We used RMSE as a measure of loss function and activation function. The image size was 28*28. Moreover, rather than using a standard CNN such as VGG and AlexNet, we tested the validity of our results on benchmark datasets.

We used the following libraries for various functionalities:

Keras [40], a popular open-source library that enables researchers and software engineers to define and train many deep learning models in a short amount of time. It is used for creating and training neural networks. It provides a simple interface to existing libraries like Tensorflow [42], which allows use of GPUs for faster training and prediction. With Keras, we can achieve fast experimentation, which is key to gaining insightful feedback and improving the accuracy of deep learning models for stock price prediction. Note that in Keras version 2.0, metrics such as recall, precision, and the F-measure have been removed to promote the use of accuracy as the main metric for CNNs that are trained on balanced datasets [40].
Tensorflow [42], the backend for Keras. It facilitates the processing of low-level operations such as tensor products and convolutions. It is developed and maintained by Google.
Scikit-learn [43] for performing the grid search of the model to find the best parameters using 10-fold cross-validation.
Matplotlib [44] for plotting the graphs for the actual time series as well as predicted trends.
Pandas [45] for reading values from csv files as DataFrames.
Numpy [46] to perform matrix operations like flip and reshape and to create random matrices. Although there are various optimization algorithms such as Adam, AdaDelta, and RMSProp, Adam optimization was selected in our study following the protocol in a previous study [2]. The Adam optimization algorithm can be used instead of the classical SGD procedure to update network weights iteratively based on training data [47].

Hyperparameter optimization also affects the performance of CNNs [48,49]. A preliminary experiment was therefore conducted to select the hyperparameters for the CNNs that influence stock price prediction. In accordance with previous studies [48,50,51], among the possible convolution layers, pooling layers, dropouts, and filters, we focused on the number of filters and dropouts, examining the activation function for each layer as the hyperparameter to improve prediction. To determine whether the shape of dropouts and filters affects the accuracy of CNN inferences, the dropout variable was set to (0.25,0.5) and (0.5,0.5), and the number of 3-layer convolution kernels was set to (1,2,4), (2,4,8), (3,6,12), (4,8,16), (5,10,20), (6,12,24), (7,14,28), (8,16,32), (9,18,36), (10,20,40), (11,22,44), (12,24,48), (13,26,52), (14,28,56), (15,30,60), and (16,32,64). By designating the number of filters in this format, we obtained 960 sample cases per image. The activation functions reviewed in the pretest to optimize the activation function in the CNN were ReLU, Elu, Selu, Softsign, Softplus, RRelu, and Gelu. As a result of experimenting with various combinations, we obtained the following sequence: [ELU, ELU, ELU, SELU]. We then applied 10-fold cross-validation for performance evaluation. Algorithm 1 summarizes the above experimental sequence.

Algorithm 1: Proposed Model Procedure

1: procedure StockPrediction()

2: Phase Data Preparation:

3: company = read(companyList, stockPrice)

4: trainingDataset = dataset.split(dates = 2015 − 2019)

5: testDataset = dataset.split(dates = 2015 − 2019)

6: Phase Labelling Data:

7: slopeRef[1..n] = calculateSlopeReferences(farFutureValue = 5, nearFutureValue = 1)

8: calculate distribution of the class(Up, Down) to ind separation values

9: firstSepPoint, secondSepPoint = find the separation values(slopeRef[1..n])

10: for(all dataset):

11: slopeCurrent = calculateEachImageSlope(farFutureValue = 5, nearFutureValue = 1)

12: if(slopeCurrent > secondSepPoint):

13: label = 1 (”Up”)

14: elif(slopeCurrent < = secondSepPoint):

15: label = 0 (”Down”)

16: merge labels and images

17: Phase Chart Generation:

18: for(all chartNum):

19: for(all company):

20: for(all period):

21: chartProperty = getChartProperty(chartNum)

22: chart = generateChart(chartProperty, imagefile)

23: putImage(label, chart)

24: Phase Predicting Label:

25: model = CNN(epochs = 200, learningrate = 0.001, activation = (‘Activation function combination’, ‘softmax’))

26: model.train(trainingDataset)

27: model.test(testDataset)

V. Results

5.1 Differences in prediction accuracy according to image characteristics

In this study, stock price data from January 1, 2015 to December 31, 2019 were used in a CNN, and learning and prediction were performed by setting the closing price as 1 or 0 after 1 or 5 days. Table 5 shows the results of an ANOVA analysis comparing differences between factors according to the accuracy of the stock price prediction. The results of the analysis showed a significant difference in prediction accuracy for X1, X2, X4, X6, X7, X8, and X9. Putting together the results for overall accuracy and the ANOVA analysis, we see that when a stock price chart is used as a training dataset, the accuracy of the prediction can be increased by drawing a solid line, coloring the lower area, and using an image without axis marks. Therefore, Hypothesis 1 was partially supported.

Table 5. Results of ANOVA analysis of the difference in accuracy between stock price predictions using image characteristics.

Variables	N	Accuracy	SD	F-value
X1	Time series (308)	0.639	0.048	6.744***
X1	Frequency (112)	0.626	0.043	6.744***
X2	Plot (112)	0.643	0.050	5.263***
	Histogram (112)	0.630	0.041
	Bar plot (168)	0.640	0.049
	Strip chart (28)	0.608	0.038
X3	Horizontal (238)	0.635	0.0501	0.028
X3	Vertical (182)	0.636	0.0431	0.028
X4	No (252)	0.641	0.044	7.506***
X4	Yes (168)	0.628	0.050	7.506***
X5	After 1 day (210)	0.636	0.048	0.019
X5	After 5 days (210)	0.635	0.046	0.019
X6	Bar (308)	0.633	0.046	3.388***
X6	Line (112)	0.643	0.050	3.388***
X7	No (238)	0.629	0.049	10.295***
X7	Yes (182)	0.644	0.043	10.295***
X8	Monotone (252)	0.629	0.048	12.077***
X8	Color (168)	0.645	0.044	12.077***
X9	No (252)	0.640	0.042	5.632**
X9	Yes (168)	0.629	0.053	5.632**

Open in a new tab

X1_Series, X2_Graph type, X3_Graph direction, X4_Appearance of axis, X5_Forecast date, X6_Bar/line, X7_Color, X8_Color tone, X9_Dotted/solid line.

5.2 Mediating role of filter

To test Hypothesis 2, which states that filter characteristics play a mediating role between image characteristics and stock price prediction accuracy, Chow’s verification was performed by classifying the filter variable into two groups: large (high filter) and small (low filter). The Chow’s F(10,940) value was statistically significant at 5.223; therefore, Hypothesis 2 was supported. According to Table 6, which shows the results of the Chow verification, variables X1 and X5 had a positive effect on accuracy in the low-filter population, but a negative effect in the high-filter population. In addition, the X7 variable negatively affected accuracy in the low-filter population, but had a positive effect on the high-filter population. Both X4 and X6 positively affected accuracy, but the results show that the high filter had a greater effect than the low filter. Also, X2 negatively affected accuracy, but the high filter had a greater effect than the low filter. X1 and X2 had no effect in the low-filter condition, but had a negative effect in the high-filter condition. The X9 variable did not affect the accuracy, meaning that the filter size was not adjusted.

Table 6. Results of Chow verification for filter.

	Variables	B	SE	t	p
Low filter	(const)	3.708	0.44	8.432	0	0.044
	X1	-0.114	0.073	-1.574	0.116
	X2	-0.005	0.085	-0.064	0.949
	X3	0.039	0.071	0.546	0.049**
	X4	0.101	0.083	1.217	0.022**
	X5	0.175	0.053	3.328	0.001***
	X6	0.064	0.196	0.326	0.074*
	X7	-0.032	0.14	-0.231	0.008***
	X8	-0.144	0.077	-1.858	0.064*
	X9	-0.153	0.134	-1.147	0.252
High filter	(const)	4.104	0.363	11.309	0	0.142
	X1	-0.165	0.06	-2.751	0.006***
	X2	-0.136	0.071	-1.932	0.044**
	X3	-0.136	0.059	-2.32	0.021**
	X4	0.514	0.069	7.476	0.000***
	X5	-0.024	0.044	-0.55	0.006***
	X6	0.278	0.162	1.713	0.087*
	X7	0.402	0.116	3.475	0.001***
	X8	-0.198	0.064	-3.099	0.002***
	X9	-0.072	0.11	-0.649	0.517

Open in a new tab

Chow’s F(10,940): 5.223; p value: 0.0001.

X1_Series, X2_Graph type, X3_Graph direction, X4_Appearance of axis, X5_Forecast date, X6_Bar/line, X7_Color, X8_Color tone, X9_ Dotted/solid line.

5.3 Mediating effect of dropout

To verify Hypothesis 3, which states that dropout acts as a mediator between image characteristics and stock price prediction accuracy, the dropout variable was divided into two groups: (0.25, 0.5) and (0.5, 0.5), and Chow verification was performed. The results show that Chow’s F(10,940) = 2.3627 was statistically significant; therefore, Hypothesis 3 was supported (see Table 7). According to Table 7, variables X2, X3, and X6 did not affect either Dropout1 or Dropout2. It was found that X1, X4, X7 had more influence in Dropout1, X8 had a significant effect only on Dropout1, and X5 and X9 had a significant effect only on Dropout2.

Table 7. Results of Chow verification for dropout.

	Variables	B	SE	t	p
Dropout1	(const)	3.883	0.439	8.854	0	0.235
	X1	-0.149	0.072	-2.051	0.041**
	X2	-0.06	0.085	-0.703	0.482
	X3	-0.069	0.071	-0.981	0.327
	X4	0.322	0.083	3.878	0.000***
	X5	0.041	0.053	0.783	0.434
	X6	0.134	0.196	0.685	0.494
	X7	0.312	0.14	2.233	0.026**
	X8	-0.241	0.077	-3.127	0.002***
	X9	0.06	0.133	0.451	0.652
Dropout2	(const)	3.929	0.377	10.413	0	0.286
	X1	-0.131	0.062	-2.096	0.037**
	X2	-0.082	0.073	-1.115	0.265
	X3	-0.028	0.061	-0.456	0.649
	X4	0.293	0.071	4.101	0.000***
	X5	0.11	0.045	2.439	0.015**
	X6	0.207	0.168	1.231	0.219
	X7	0.058	0.12	0.478	0.0633*
	X8	-0.1	0.066	-1.512	0.131
	X9	-0.285	0.115	-2.485	0.013**

Open in a new tab

Chow’s F(10,940): 2.363764; p value: 0.0092.

X1_Series, X2_Graph type, X3_Graph direction, X4_Appearance of axis, X5_Forecast date, X6_Bar/line, X7_Color, X8_Color tone, X9_ Dotted/solid line.

5.4 Performance comparison

Using the results of hypothesis testing in this study, we compared the prediction performance of our method with those in other similar studies in terms of the chart image, filter, and dropout characteristics optimized for CNN-based stock price prediction. As shown in Table 8, the accuracy of the proposed method in this study was 64.3%, which was significantly higher than that of previous studies (52.1%–57.5%).

Table 8. Comparative performance between previous and present results.

Method	Data	Methods	Accuracy
Gunduz et al. [34]	Numeric	CNN	56.0%
Nelson et al. [32]	Numeric	LSTM	55.9%
Zhong & Enke [28]	Text	PCA	57.5%
Gunduz et al. [34]	Numeric	CNN	56.0%
Nelson et al. [32]	Numeric	LSTM	55.9%
Zhong & Enke [28]	Text	PCA	57.5%
Di Persio & Honchar [10]	Numeric	MLP	52.1%
Di Persio & Honchar [10]	Numeric	LSTM	52.2%
Di Persio & Honchar [10]	Numeric	CNN	53.6%
Proposed method	Chart Image	Proposed	64.3%

Open in a new tab

VI. Discussion

6.1 Implications

As ways are developed to understand and analyze images through deep learning, research is being conducted to predict the ups and downs of stock prices using only the image of a stock price chart, not using an enormous amount of numerical information [3–5,10,30–33] that takes a lot of time and effort to collect and process [2,11,12]. In this study, we identify the specific characteristics of images that have a significant influence on the prediction accuracy of deep learning algorithms.

The greatest contribution of this study is its focus on the CNN algorithm to investigate the effects of the characteristics of stock price chart images on prediction performance. In fact, performing the CNN algorithm using image datasets or multimodal datasets has been actively attempted in various domains such as medicine [14]. In addition, prediction performance has been improved by selecting an appropriate preprocessing method according to the domain [52,53]. However, only a few studies have mentioned a preprocessing method that selects the optimal image characteristics in advance for stock price prediction [54].

In this study, we experimented with basic algorithms for CNNs under various conditions using a relatively large amount of image data (charts for training: 1,281,173, charts for testing: 142,352). The results revealed several significant variables: the type of chart, the use of dotted or solid lines, indications of the area, and presence or absence of axis information. Based on the results of this analysis, we plan to develop an optimal method for determining the shape of the chart for providing to CNNs for the purpose of stock price prediction.

In addition, utilizing a CNN with selected chart images and specifying optimal filters and dropouts resulted in better prediction accuracy than in other, similar studies. This is the first study to show that stock price prediction is possible simply by observing the image characteristics of a chart, not relying on numerical information.

In the past, in addition to developing algorithms to improve stock price prediction, efforts have been made to improve feature selection on the premise that the quality of input data characterization plays an important role in prediction performance [20]. This study supports the suggestion that considering the features of the image contributes to improving prediction performance.

6.2 Limitations

First, this study did not consider the patterns of stock price fluctuations. In future, patterns of stock price fluctuations will be meaningfully grouped for the purpose of pinpointing which image or CNN network characteristics contribute to improving prediction accuracy in which groups. Second, this study focused only on dropout and filter among the possible network characteristics of CNNs. However, there are various other points of comparison to study such as algorithm optimization, the activation function, and the number of layers. To focus on image characteristics in this study, we preoptimized some data in advance using Adam as the optimization algorithm, ReLU as the activation function, and including 5 layers (3 hidden layers). Future research will elucidate the role of these other parameters in prediction performance. Third, this study considered only CNNs among the possible deep learning algorithms. Although CNNs are the most commonly used image-based discrimination algorithms, and hence our choice was reasonable for this first study, it will be necessary to expand and study other image discrimination algorithms such as RNNs and LSTMs in future research. Last, only the CNN algorithm was used in our experiment, but in future work, we will also use algorithms such as BI-GRU and LSTM among RNNs. However, the main significance of this paper is its proof that there is a difference in the performance of deep learning algorithms according to image characteristics. We hope that tour results will inspire further research.

6.3 Conclusion

Stock price chart images are an important source of information for stock price forecasting. In the past, researchers had no choice but to use traditional classification algorithms that rely on numeric or textual data because there was no suitable preprocessing method to use image data. Recent development of deep learning algorithms such as CNNs has made it possible to discriminate using images and to make predictions using only stock price charts. This study reveals that there is an optimal fit between image characteristics and algorithm characteristics. The result (Hypothesis 1) suggests that we need to optimize the image characteristics for better stock price prediction. We identify a causal relationship between image characteristics and discriminant performance in order to improve the reliability and performance of image-based prediction. Moreover, the results of testing of Hypotheses 2 and 3 suggest that when selecting the characteristics of the stock chart, the filter parameter and dropout must also be changed according to the image characteristics.

Our method outperforms previous methods in terms of prediction accuracy. Through this study, image data may become widely used as an important resource for business intelligence. The results also suggest that there may be no need to prepare hybrid data (combining images and numerical data) for price prediction. Stock price prediction can be easy even in various formats of information systems (mobile apps, video-based SNS, image databases, etc.)–wherever chart images only are available. Although stock market prediction appears to be a random walk, a stock price expert looks at a company’s stock price chart and predicts whether a company’s stock price will rise or fall based on his knowledge and experience. We believe that the results of the performance test in this paper can be evaluated as showing the possibility that CNN can acquire knowledge and, to some extent, imitate the implicit knowledge of stock price prediction experts using charts. This paper assumes that people who predict stock prices using only charts without numerical data on stock prices have image-based forecasting knowledge. In this study, the authors tried to augment their predictive ability through deep learning. It would be appreciated if you could understand the significance of this study in this way.

Data Availability

The data (data set and meta data) underlying the results presented in the study are available from: [1] Variable description and experimental data Link https://figshare.com/articles/dataset/Variable_description_and_experimental_data/14074496 [2] Company-dataset Link: https://figshare.com/articles/dataset/dataset/14074502 [3] Image-dataset Link: https://figshare.com/articles/dataset/cnn-dataset/14074292 The authors confirm that the authors of the present study had no special privileges in accessing these datasets which other interested researchers would not have.

Funding Statement

This work was supported by the Ministry of Education of the Republic of Korea and the National Research Foundation of Korea (NRF-2020S1A5B8103855).

References

1.Carpenter GA, Grossberg S, Markuzon N, Reynolds JH, Rosen DB, et al. Fuzzy ARTMAP: A neural network architecture for incremental supervised learning of analog multidimensional maps. IEEE Transactions on Neural Networks 1992;3(5):698–713. doi: 10.1109/72.159059 [DOI] [PubMed] [Google Scholar]
2.Sim HS, Kim HI, Ahn JJ. Is Deep Learning For Image Recognition Applicable to Stock Market Prediction?. Complexity 2019:1–10. 10.1155/2019/4324878. [DOI] [Google Scholar]
3.De Gooijer JG, Hyndman RJ. 25 Years of Time Series Forecasting. International Journal of Forecasting 2006; 22(3):443–473. 10.1016/j.ijforecast.2006.01.001. [DOI] [Google Scholar]
4.Menon VK, Vasireddy NC, Jami SA, Pedamallu VTN, Sureshkumar V, Soman. Bulk Price Forecasting Using Spark Over Use Data Set. In International Conference on Data Mining and Big Data 2016; 9714:137–146. 10.1007/978-3-319-40973-3_13. [DOI] [Google Scholar]
5.Box GEP, Jenkins GM, Reinsel GC, Ljung GM, et al. Time Series Analysis: Forecasting and Control. John Wiley and Sons Inc 2015; 37(5):709–712. 10.1111/jtsa.12194. [DOI] [Google Scholar]
6.Selvin S, Vinayakumar R, Gopalakrishnan EA, Menon VK, Soman KP. Stock Price Prediction Using LSTM, RNN and CNN-sliding Window Model. In International Conference on Advances in Computing, Communications and Informatics 2017; 23(1):1643–1647. 10.1109/ICACCI.2017.8126078. [DOI]
7.Hoseinzade E, Saman H. CNNpred: CNN-based Stock Market Prediction Using A Diverse Set of Variables. Expert Systems with Applications 2019; 129:273–285. 10.1016/j.eswa.2019.03.029. [DOI] [Google Scholar]
8.Cao J, Wang J. Stock Price Forecasting Model Based on Modified Convolution Neural Network and Financial Time Series Analysis. International Journal of Communication Systems 2019; 32(12):e3987. 10.1002/dac.3987. [DOI] [Google Scholar]
9.Jain S, Gupta R, Moghe A. Stock Price Prediction on Daily Stock Data Using Deep Neural Networks. In International Conference on Advanced Computation and Telecommunication 2018; 10:1–13. 10.1109/ICACAT.2018.8933791. [DOI]
10.Di Persio L, Honchar O. Artificial Neural Networrs Architectures for Stock Price Prediction: Comparisons and Applications. International journal of circuits, systems and signal processing 2016; 10:403–413. Available from:https://www.semanticscholar.org/paper/41487776364a6dddee91f1e2d2b78f46d0b93499. [Google Scholar]
11.Guo S, Yan Z, Zhang K, Zuo W, Zhang L. Toward Convolutional Blind Denoising of Real Photographs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2019; 49(1):1712–1722. 10.1109/CVPR.2019.00181. [DOI]
12.Sezer OB, Ozbayoglu AM. Financial Trading Model with Stock Bar Chart Image Time Series with Deep Convolutional Neural Networks. In Intelligent Automation and Soft Computing 2019; 26(2):2007–2012. 10.31209/2018.100000065. [DOI] [Google Scholar]
13.Lan R, Zou H, Pang C, Zhong Y, Liu Z, Luo X. Image Denoising Via Deep Residual Convolutional Neural Networks. Signal, Image and Video Processing 2019;15(10):1–8. 10.1007/s11760-019-01537-x. [DOI] [Google Scholar]
14.Lakhe S, Mariwalla R, Reddy C. Regression Analysis Based Linear Model for Predicting Stock Prices. Industrial Engineering Journal 2017; 10:154–157. 10.26488/IEJ.10.1.9. [DOI] [Google Scholar]
15.Reddy T, Bhattacharya S, Maddikunta PKR, Hakak S, Khan WZ, Bashir AK, et al. Antlion re-sampling based deep neural network model for classification of imbalanced multimodal stroke dataset. Multimedia Tools and Applications 2020:1–25. 10.1007/s11042-020-09988-y. [DOI] [Google Scholar]
16.Ballings M, Van den Poel D, Hespeels N, Gryp R. Evaluating Multiple Classifiers for Stock Price Direction Prediction. Expert systems with Applications 2015; 42(20):7046–7056. [Google Scholar]
17.Bessembinder H, Kaufman HM. A Comparison of Trade Execution Costs for NYSE and NASDAQ-listed Stocks. Journal of Financial and Quantitative Analysis 1997; 287–310. [Google Scholar]
18.Preis T, Kenett DY, Stanley HE, Helbing D, Ben-Jacob E. Quantifying the behavior of stock correlations under market stress. Scientific reports 2012; 2(1):1–5. doi: 10.1038/srep00752 [DOI] [PMC free article] [PubMed] [Google Scholar]
19.Hagenau M, Liebmann M, Neumann D. Automated News Reading: Stock Price Prediction Based On Financial News Using Context-capturing Features. Decision Support Systems 2013; 55(3):685–697. 10.1016/j.dss.2013.02.006. [DOI] [Google Scholar]
20.Qiu M, Song Y. Predicting the Direction of Stock Market Index Movement Using an Optimized Artificial Neural Network Model. PloS One 2016;11(5):e0155133. doi: 10.1371/journal.pone.0155133 . [DOI] [PMC free article] [PubMed] [Google Scholar]
21.Fischer T, Krauss C. Deep Learning with Long Short-term Memory Networks for Financial Market Predictions. European Journal of Operational Research 2018; 270(2):654–669. https://ojs.aaai.org/index.php/AAAI/article/view/9354. [Google Scholar]
22.Deng S, Zhang N, Zhang W, Chen J, Pan JZ, Chen H. Knowledge-driven stock trend prediction and explanation via temporal convolutional network. InCompanion Proceedings of The 2019 World Wide Web Conference 2019 May 13 (pp. 678–685).
23.Jin X, Chen Z, Yang X. Economic Policy Uncertainty And Stock Price Crash Risk. Accounting Finance 2019;58(5):1291–1318. 10.1111/acfi.12455. [DOI] [Google Scholar]
24.Wang J, Wang J. Forecasting Stock Market Indexes using Principle Component Analysis and Stochastic Time Effective Neural Networks. Neurocomputing 2015;156:68–78. 10.1016/j.neucom.2014.12.084. [DOI] [Google Scholar]
25.Kara Y, Boyacioglu MA, Baykan ÖK. Predicting Direction of Stock Price Index Movement Using Artificial Neural Networks and Support Vector Machines: The Sample of the Istanbul Stock Exchange. Expert Systems with Applications 2011; 38(5):5311–5319. 10.1016/j.eswa.2010.10.027. [DOI] [Google Scholar]
26.Guresen E, Kayakutlu G, Daim TU. Using Artificial Neural Network Models in Stock Market Index Prediction. Expert Systems with Applications 2011; 38(8):10389–10397. 10.1016/j.eswa.2011.02.068. [DOI] [Google Scholar]
27.Shastri M, Roy S, Mittal M. Stock Price Prediction Using Artificial Neural Model: An Application of Big Data. EAI Endorsed Transactions on Scalable Information Systems 2019; 6:156085. 10.4108/eai.19-12-2018.156085. [DOI] [Google Scholar]
28.Zhong X, Enke D. Forecasting Daily Stock Market Return using Dimensionality Reduction. Expert Systems with Applications 2017; 67:126–139. 10.1016/j.eswa.2016.09.027. [DOI] [Google Scholar]
29.Patel J, Shah S, Thakkar P, Kotecha K. Predicting Stock Market Index Using Fusion of Machine Learning Techniques. Expert Systems with Applications 2015; 42(4):2162–2172. 10.1016/j.eswa.2014.10.031. [DOI] [Google Scholar]
30.Naik N, Mohan B. Intraday Stock Prediction Based on Deep Neural Network. National Academy Science Letters 2020; 43:241–246. 10.1007/s40009-019-00859-1. [DOI] [Google Scholar]
31.Chung H, Shin KS. Genetic Algorithm-optimized Multi-channel Convolutional Neural Network for Stock Market Prediction. Neural Computing & Applications 2020; 32(12): 7897–914. 10.1007/s00521-019-04236-3. [DOI] [Google Scholar]
32.Nelson DM, Pereira AC, de Oliveira RA. Stock Market’s Price Movement Prediction with LSTM Neural Networks. In Neural Networks (IJCNN) IEEE 2017:1419–1426. doi: 10.1109/ICORR.2017.8009447 [DOI] [PubMed] [Google Scholar]
33.Pang X, Zhou Y, Wang P, Lin W, Chang V. An Innovative Neural Network Approach for Stock Market Prediction. The Journal of Supercomputing 2020; 76:2098–2118. 10.1007/s11227-017-2228-y. [DOI] [Google Scholar]
34.Gunduz H, Yaslan Y, Cataltepe Z. Intraday prediction of borsa istanbul using convolutional neural networks and feature correlations. Knowledge-Based Systems 2017; 137:138–148. 10.1016/j.knosys.2017.09.023. [DOI] [Google Scholar]
35.Schumaker RP, Chen H. Textual Analysis of Stock Market Prediction Using Breaking Financial News. ACM Transactions on Information Systems 2009; 27(2):1–19. 10.1145/1462198.1462204. [DOI] [Google Scholar]
36.Saha PK, Borgefors G, di Baja GS. A Survey on Skeletonization Algorithms and Their Applications. Pattern Recognition Letters 2016; 76:3–12. 10.1016/j.patrec.2015.04.006. [DOI] [Google Scholar]
37.Jerripothula KR, Cai J, Lu J, Yuan J. Object Co-skeletonization With Co-segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition 2017; 33(2):3881–3889. 10.1109/CVPR.2017.413. [DOI]
38.Paracchini M, Marcon M, Villa F, Tubaro S. Deep Skin Detection on Low Resolution Grayscale Images. Pattern Recognition Letters 2020; 131:322–328. [Google Scholar]
39.Chen Q, Chen Q, Liao Q, Jiang ZL, Fang J, Yiu S, Xi G, Li R, Yi Z, Wang X, Hui LC, Liu D. File Fragment Classification Using Grayscale Image Conversion and Deep Learning in Digital Forensics. IEEE Security and Privacy Workshops 2018: 140–147. 10.1109/SPW.2018.00029. [DOI]
40.Keras Team, Keras 2.0 release notes 2020 [Online]. [Accessed 29 April. 2020]. Available from: https://github.com/keras-team/keras/wiki/Keras-2.0-release-notes.
41.Eapen J, Bein D, Verma V. Novel Deep Learning Model with CNN and Bi-Directional LSTM for Improved Stock Market Index Prediction. IEEE 9th Annual Computing and Communication Workshop and Conference 2019; 9:0264–0270. 10.1109/CCWC.2019.8666592. [DOI]
42.TensorFlow: An Open Source Machine Learning Framework for Everyone [Online]. [Accessed 29 April. 2020]. Available from: https://www.tensorflow.org/.
43.Scikit-Learn: Machine Learning in Python [Online]. [Accessed 29 April. 2020]. Available from: https://scikit-learn.org/stable/.
44.The Matplotlib Development Team (2018, Nov. 28). Matplotlib Version 3.0.2 [Online]. [Accessed 29 April. 2020]. Available from: https://matplotlib.org/.
45.Mester, T. Pandas Tutorial 1: Pandas Basics [Internet]. [Cited 2020 April 29]. Available from: https://data36.com/pandas-tutorial-1-basics-reading-datafiles-dataframes-data-selection/.
46.NumPy 2020 [Online]. [Accessed 29 April. 2020]. Available from: http://www.numpy.org/. [Google Scholar]
47.Kingma DP, Ba J. Adam: A Method For Stochastic Optimization, Machine Learning; 2013 [Cited 2020 April 20]. Available from: https://arxiv.org/abs/1412.6980.
48.Feurer M, Springenberg JT, Hutter F. Initializing Bayesian Optimization via Meta-Learning. In Twenty-Ninth AAAI Conference on Artificial Intelligence 2015; 14(2):135–147. https://ojs.aaai.org/index.php/AAAI/article/view/9354.
49.Feurer M, Hutter F. Hyperparameter Optimization. In Automated Machine Learning 2019; 3–33. Available from: 10.1007/978-3-030-05318-5_1. [DOI] [Google Scholar]
50.Wang Y, Liu M, Yang J. Data-driven Deep Learning for Automatic Modulation Recognition in Cognitive radios. IEEE Transactions on Vehicular Technology 2019; 68(4): 4074–4077. 10.1109/TVT.2019.2900460. [DOI] [Google Scholar]
51.Loni M, Sinaei S, Zoljodi A, Daneshtalab M, Sjödin M. DeepMaker: A Multi-objective Optimization Framework For Deep Neural Networks In Embedded Systems. Microprocessors and Microsystems 2020; 73:89–93. 10.1016/j.micpro.2020.102989. [DOI] [Google Scholar]
52.Gadekallu TR, Khare N, Bhattacharya S, et al. Deep neural networks to predict diabetic retinopathy. J Ambient Intell Human Comput 2020: 1–14. Available from: 10.1007/s12652-020-01963-7. [DOI] [Google Scholar]
53.Reddy GT, Reddy MPK, Lakshmanna K, Kaluri R, Rajput DS, Srivastava G, Baker T, et al. Analysis of dimensionality reduction techniques on big data. IEEE Access 2020; 8:54776–54788. 10.1109/ACCESS.2020.2980942. [DOI] [Google Scholar]
54.Song Y, Lee J. Performance Evaluation of Deep Learning Stock Price by Chart Type for Buying Policy Verification. In Fuzzy Systems and Data Mining IV 2018:646–652. 10.3233/978-1-61499-927-0-646. [Google Scholar]

PLoS One. doi: 10.1371/journal.pone.0253121.r001

Decision Letter 0

Ruxandra Stoean

22 Dec 2020

PONE-D-20-29152

Impact of Chart Image Characteristics on Stock Price Prediction of a Convolutional Neural Network

PLOS ONE

Dear Dr. Kwon,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

The paper needs substantial improvements and explanations on multiple levels: the image data set generation and choice of features, use of technical indicators along image feature in tandem or in comparison, reformulation of the hypotheses, a more transparent model validation, a statistical comparison with state-of-the-art deep learning methods for the same data set, inclusion of more performance metrics.

Please submit your revised manuscript by Feb 05 2021 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.
A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.
An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

We look forward to receiving your revised manuscript.

Kind regards,

Ruxandra Stoean

Academic Editor

PLOS ONE

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. Please provide links to all data sources used in the Methods section.

3. We note that you have stated that you will provide repository information for your data at acceptance. Should your manuscript be accepted for publication, we will hold it until you provide the relevant accession numbers or DOIs necessary to access your data. If you wish to make changes to your Data Availability statement, please describe these changes in your cover letter and we will update your Data Availability statement to reflect the information you provide.

4. Please upload a new copy of Figure 3 as the detail is not clear. Please follow the link for more information: https://blogs.plos.org/plos/2019/06/looking-good-tips-for-creating-your-plos-figures-graphics/

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Partly

Reviewer #2: Yes

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: No

Reviewer #2: Yes

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: In my view, the paper in its current form is in frontiers between major revision and direct rejection. The authors did not discuss the implementation of CNN in sufficient detail, lacking the parameters such as loss function, learning rates, and numbers of epochs, etc. There is no cohesion between the hypothesis and conclusion presented in the paper. However, the content of the paper shows a lot of work correctly oriented towards a goal. Hence, Major revision is suggested before it can be considered as competent for publication. The main concern that needs to be addressed are listed below:

1. I cannot see any novelty in this work. There are many publications in this regard except the dataset used.

2. Since dropout and filter-size are obvious parameters that affect CNN learning- so hypothesis 2 needs reformulation (hypothesis formulation adds no more value to this work).

3. There is no cohesion between the research hypothesis and conclusion.

4. Stock market prediction is a random walk and includes non-linear dynamics, authors are unable to address these issues in their experiments as well as conclusions.

5. Authors need to shed light on how chart-based stock prediction is feasible over technical indicator-based prediction.

6. The Paper lacks the details about image feature representation by CNN.

7. The information provided about image characteristic is ambiguous.

8. Model validation is weak. The author needs to specify the size of the increase dataset, decrease dataset, and Size of training, validation, and testing dataset explicitly.

9. Detail of CNN network training parameters is missing such as loss function.

10. Since the proposed CNN has very limited layers, the network overfitting might be the issue, the author needs to verify this with network generalization capability.

11. The comparison of the presented result with the previous result is not comprehensive. The reference cited in Table 7 (Namely Gunduz et.al -2017, Nelson et.al-2017) are missing in the reference list.

12. The validity of result presented in the paper need to be tested on benchmark datasets as author proposed their own CNN, instead of standard CNN such as VGG, and AlexNet, there are many CNN parameters still need to be optimized in the network such as numbers of layers, optimizers, activation functions, and normalizations.

13. The image dataset generation process needs to be discussed in detail.

14. The only accuracy is used as performance metrics which might be biased towards a single class either increases or decrease so the author needs to add a few more performance metrics.

15. There are many grammatical and syntactic errors inside. A native speaker can fix it.

Reviewer #2: PLOS ONE Review of the manuscript PONE-D-20-29152

The manuscript titled “Impact of Chart Image Characteristics on Stock Price Prediction of a Convolutional Neural Network” uses CNN on image features to forecast stock price. The research work presented in the manuscript is good by using image features for prediction. However, there are some weak points that should be addressed.

1. Scope of the manuscript is limited. The research work uses only image features for price prediction.

2. The manuscript uses only CNN on the image data for prediction. There are state-of-the-art methods that should also be used and compare with CNN.

3. The main idea of this research work is to use image features for prediction. But there is not enough information, how the images and there features are generated. So in section 4.1, please provide some more information how the chart images were created.

After the changes, the manuscript may be recommended for publication.

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

PLoS One. 2021 Jun 23;16(6):e0253121. doi: 10.1371/journal.pone.0253121.r002

Author response to Decision Letter 0

23 Feb 2021

Comments Response

[Reviewr 2]

[Q1] I cannot see any novelty in this work. There are many publications in this regard except the dataset used.

[A1] As you pointed out, the study of predicting stock prices with images has begun. Many example studies have been cited in the paper. However, there is still no research on ‘which’ characteristics of a chart image significantly affect the performance of deep learning-based stock price predictions. This is the novelty of this paper. This paper has been researched from this angle and its features are presented in the conclusion section.

[Q2] Since dropout and filter-size are obvious parameters that affect CNN learning- so hypothesis 2 needs reformulation (hypothesis formulation adds no more value to this work).

[A2] As the reviewer mentioned, filter size and dropout are factors that influence the CNN model. In addition, in this paper, a hypothesis was created based on the idea that there will be a difference in the degree to which filter size and dropout affect each image feature. The result of Hypothesis 2 and 3 means that when selecting the characteristics of the stock chart, the filter parameter and dropout must also be changed according to the image characteristics. Therefore, the authors treated these parameters as mediating rather than independent variables.

[Q3] There is no cohesion between the research hypothesis and conclusion.

[A3] The authors modified the conclusion as advised, adding information about the hypotheses to the conclusion.

[Q4] Stock market prediction is a random walk and includes non-linear dynamics, authors are unable to address these issues in their experiments [A4] as well as conclusions.

Although stock market prediction appears to be a random walk, a stock price expert looks at a company's stock price chart and predicts whether a company's stock price will rise or fall based on his knowledge and experience. The results of the performance test in this paper can be evaluated as showing the possibility that CNN can acquire knowledge to some extent imitate the implicit knowledge of stock price prediction experts using charts.

[Q5] This paper assumes that people who predict stock prices with only charts without numerical data on stock prices have image-based forecasting knowledge.

[A5] In this study, the authors tried to augment their predictive ability through deep learning. It would be appreciated if you could understand the significance of this study in this way. These are newly added in conclusion.

[Q6] Authors need to shed light on how chart-based stock prediction is feasible over technical indicator-based prediction.

[A6] Table 8 has been revised to show that using chart images results in more accurate stock price prediction than numeric and text-based prediction.

The authors are considering a method of combining image-based prediction and technical indicator-based prediction for future research. Information to this effect has been provided in the discussion section of the revised version.

[Q7] The Paper lacks the details about image feature representation by CNN.

[A7] The process of creating an image is as follows.

1. Daily data for 5 years (2015-2019) was collected on 789 companies included in the KOSPI Index calculation in the Dataguide program.

2. Images were composed of data collected on a monthly basis.

3. According to the image characteristics shown in Table 2, 30 images of various types were created for each month with RStudio software.

4. After the images were created, a label for each was automatically generated with the name of the image, which was then compared with the closing stock price 5 days later. Please see the following example.

5. Finally, if the image name rate of return was greater than 0, 1 was automatically added to the end of the label; otherwise, 0 was automatically added.

This text has been added to section 4.1.

[Q8] The information provided about image characteristic is ambiguous.

[A8] The information has been clarified in Table 2 and image characteristics have been explained better in the text. In the experiment for this paper, deep learning was performed by generating various chart images according to the characteristics listed in Table 2.

[Q9] Model validation is weak. The author needs to specify the size of the increase dataset, decrease dataset, and Size of training, validation, and testing dataset explicitly.

[A9] In order to obtain a set of images for learning, a dataset was created with daily data for all companies listed on the KOSPI Index, and a stock price chart for each image characteristic was created. Total amount of data: 12 (months)*5 (years)*789 (number of companies)*30 (chart image types). Then, 90% of the training set and 10% of the test set were selected randomly so learning and inference could be performed.

This information has been added in the revised manuscript.

[Q10] Detail of CNN network training parameters is missing such as loss function.

[A10] We added the following information about parameters in the revised manuscript (p. 14).

Parameters:

Loss function : RMSE

Activation function: {ReLU, softmax}

Image size = 28 * 28

[Q11] Since the proposed CNN has very limited layers, the network overfitting might be the issue, the author needs to verify this with network generalization capability.

[A11] The stock price images created for this study were relatively simple. In our experience, the more layers we set, the worse the performance, and the more parameters we included, the greater the number of calculations. Therefore, we concluded that there was no issue of overfitting.

[Q12] The comparison of the presented result with the previous result is not comprehensive. The reference cited in Table 7 (Namely Gunduz et.al -2017, Nelson et.al-2017) are missing in the reference list.

[A12] Thank you. The reference is newly added.

[Q13] The validity of result presented in the paper need to be tested on benchmark datasets as author proposed their own CNN, instead of standard CNN such as VGG, and AlexNet, there are many CNN parameters still need to be optimized in the network such as numbers of layers, optimizers, activation functions, and normalizations.

[A13] The authors optimized the network as follows:

Parameter: {Values}

Filter size: {3 x 3}

Size of max-pooling: {(2, 2)}

Optimizer: {Adam}

Activation function: {combination of activation function}

Epochs: {2000}

Batch size: {200}

Dropout rate: {0.5,0.25}

Number of layers: 3

This information has been added to the revised version of the paper (Table 4).

[Q14] The image dataset generation process needs to be discussed in detail.

[A14] The process of creating an image is as follows.

1. Daily data for 5 years (2015-2019) was collected on 789 companies included in the KOSPI Index calculation in the Dataguide program.

2. Images were composed of data collected on a monthly basis.

3. According to the image characteristics shown in Table 2, 30 images of various types were created for each month with RStudio software.

4. After the images were created, a label for each was automatically generated with the name of the image, which was then compared with the closing price of the stock price 5 days later. Please see the following example.

5. Finally, if the image name rate of return was greater than 0, 1 was automatically added to the end of the label; otherwise, 0 was automatically added.

This text has been added to section 4.1.

[Q15] There are many grammatical and syntactic errors inside. A native speaker can fix it.

[A15] Thank you. The paper was edited by a native speaker who is familiar with technical papers.

[Reviewr 2]

[Q1] Scope of the manuscript is limited. The research work uses only image features for price prediction.

[A1] The study of predicting stock prices with images has begun. Many example studies have been cited in the paper. However, there is still no research on how the characteristics of a chart affect the performance of deep learning-based stock price predictions. This is the novelty of this paper.

The results (Table 8) suggest that using chart images is more useful for stock price prediction than numeric and text-based methods.

Also, the authors are considering a method of combining image-based prediction and technical indicator-based prediction in future research. This information has been newly added in the discussion section.

[Q2] The manuscript uses only CNN on the image data for prediction. There are state-of-the-art methods that should also be used and compare with CNN.

[A2] We agree with your statement. In the current study, only the CNN algorithm is used, but in future work, we will include algorithms such as BI-GRU and LSTM among RNNs. For the purposes of this paper, the main focus is differences in the performance of deep learning algorithms according to image characteristics. Our results inspire us to further research in this area.

This information has been added to the section discussing the limitations of this paper.

[Q3] The main idea of this research work is to use image features for prediction. But there is not enough information, how the images and there features are generated. So in section 4.1, please provide some more information how the chart images were created.

[A3] The process of creating an image is as follows.

1. Daily data for 5 years (2015-2019) was collected on 789 companies included in the KOSPI Index calculation in the Dataguide program.

2. Images were composed of data collected on a monthly basis.

3. According to the image characteristics shown in Table 2, 30 images of various types were created for each month with RStudio software.

5. Finally, if the image name rate of return was greater than 0, 1 was automatically added to the end of the label; otherwise, 0 was automatically added.

This text has been added in section 4.1.

Attachment

Submitted filename: Revision Summary2021.01.12.docx

Click here for additional data file.^{(33.3KB, docx)}

PLoS One. doi: 10.1371/journal.pone.0253121.r003

Decision Letter 1

Thippa Reddy Gadekallu

4 May 2021

PONE-D-20-29152R1

Impact of Chart Image Characteristics on Stock Price Prediction with a Convolutional Neural Network

PLOS ONE

Dear Dr. Kwon,

==============================

ACADEMIC EDITOR:

Based on the comments received from the reviewers and my own observation, I recommend minor revisions for the manuscript.

==============================

Please submit your revised manuscript by Jun 18 2021 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.
A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.
An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Thippa Reddy Gadekallu

Academic Editor

PLOS ONE

Journal Requirements:

Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #1: All comments have been addressed

Reviewer #3: (No Response)

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

Reviewer #1: Yes

Reviewer #3: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #3: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

Reviewer #1: No

Reviewer #3: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

Reviewer #1: Yes

Reviewer #3: Yes

**********

6. Review Comments to the Author

Reviewer #1: I appreciate that author attempted to address the reviewer's comments well in the revised manuscript and now the manuscript is much improved. However, the authors still didn't provide sufficient detail on model training and ablation study. It will be worthy for publication after such minor revision.

Reviewer #3: • The Wide ranges of applications need to be addressed in Introductions

• The objective of the research should be clearly defined in the last paragraph of the introduction section.

• Add the advantages of the proposed system in one quoted line for justifying the proposed approach in the Introduction section.

• The motivation for the present research would be clearer, by providing a more direct link between the importance of choosing your own method.

The authors can cite the following references

Analysis of dimensionality reduction techniques on big data

Deep neural networks to predict diabetic retinopathy

Antlion re-sampling based deep neural network model for classification of imbalanced multimodal stroke dataset

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #3: No

PLoS One. 2021 Jun 23;16(6):e0253121. doi: 10.1371/journal.pone.0253121.r004

Author response to Decision Letter 1

14 May 2021

Reviewer 1

[Q1] The authors still didn't provide sufficient detail on model training and ablation study. It will be worthy for publication after such minor revision.

[A1] Thank you for your comments. Details on model training have been added in Table 4 (activation function, learning rate, number of layers).

Reviewer 2

[Q1]

The objective of the research should be clearly defined in the last paragraph of the introduction section. • Add the advantages of the proposed system in one quoted line for justifying the proposed approach in the Introduction section.

[A1]

The objective of the research has been added in the last paragraph of the introduction.

The advantages of the proposed system have been outlined in the last paragraph of the introduction.

[Q2]

The motivation for the present research would be clearer, by providing a more direct link between the importance of choosing your own method.

The authors can cite the following references

Analysis of dimensionality reduction techniques on big data

Deep neural networks to predict diabetic retinopathy

Antlion re-sampling based deep neural network model for classification of imbalanced multimodal stroke dataset • The motivation for our research has been described at the end of the introduction section.

[A2]

• The motivation for our research has been described at the end of the introduction section.

• The following sentences have been added in the revised version of the manuscript:

In fact, performing the CNN algorithm using image datasets or multimodal datasets has been actively attempted in various domains such as medicine (Reddy et al., 2020b). In addition, prediction performance has been improved by selecting an appropriate preprocessing method according to the domain (Gadekallu et al., 2020; Reddy et al., 2020a). However, only a few studies have mentioned a preprocessing method that selects the optimal image characteristics in advance for stock price prediction.

• Correspondingly, we also added the following references:

Reddy, G. T., Reddy, M. P. K., Lakshmanna, K., Kaluri, R., Rajput, D. S., Srivastava, G., & Baker, T. (2020a). Analysis of dimensionality reduction techniques on big data. IEEE Access, 8, 54776-54788.

Gadekallu, T. R., Khare, N., Bhattacharya, S., Singh, S., Maddikunta, P. K. R., & Srivastava, G. (2020). Deep neural networks to predict diabetic retinopathy. Journal Of Ambient Intelligence and Humanized Computing, 1-14.

Reddy, T., Bhattacharya, S., Maddikunta, P. K. R., Hakak, S., Khan, W. Z., Bashir, A. K., ... & Tariq, U. (2020b). Antlion re-sampling based deep neural network model for classification of imbalanced multimodal stroke dataset. Multimedia Tools and Applications, 1-25.

Attachment

Submitted filename: Revision Summary2021.05.14.docx

Click here for additional data file.^{(19.2KB, docx)}

PLoS One. doi: 10.1371/journal.pone.0253121.r005

Decision Letter 2

Thippa Reddy Gadekallu

17 May 2021

PONE-D-20-29152R2

Impact of Chart Image Characteristics on Stock Price Prediction with a Convolutional Neural Network

PLOS ONE

Dear Dr. Kwon,

==============================

ACADEMIC EDITOR:

I guess the authors submitted the revision in a hurry. They claimed that they have addressed all the comments of teh reviewers in the response sheet but in the manuscript they are not reflected. For instance in teh response sheet the authors claimed that:

• Correspondingly, we also added the following references: Reddy, G. T., Reddy, M. P. K., Lakshmanna, K., Kaluri, R., Rajput, D. S., Srivastava, G., & Baker, T. (2020a). Analysis of dimensionality reduction techniques on big data. IEEE Access, 8, 54776-54788. Gadekallu, T. R., Khare, N., Bhattacharya, S., Singh, S., Maddikunta, P. K. R., & Srivastava, G. (2020). Deep neural networks to predict diabetic retinopathy. Journal Of Ambient Intelligence and Humanized Computing, 1-14. Reddy, T., Bhattacharya, S., Maddikunta, P. K. R., Hakak, S., Khan, W. Z., Bashir, A. K., ... & Tariq, U. (2020b). Antlion re-sampling based deep neural network model for classification of imbalanced multimodal stroke dataset. Multimedia Tools and Applications, 1-25

But this is not reflected in the paper. I recommend the authors to check all their manuscript carefully, address all the comments and then submit the revised manuscript

==============================

Please submit your revised manuscript by Jul 01 2021 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.
A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.
An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

We look forward to receiving your revised manuscript.

Kind regards,

Thippa Reddy Gadekallu

Academic Editor

PLOS ONE

Journal Requirements:

[Note: HTML markup is below. Please do not edit.]

PLoS One. 2021 Jun 23;16(6):e0253121. doi: 10.1371/journal.pone.0253121.r006

Author response to Decision Letter 2

28 May 2021

[Q]

They claimed that they have addressed all the comments of the reviewers in the response sheet but in the manuscript they are not reflected. For instance in teh response sheet the authors claimed that:

Correspondingly, we also added the following references: Reddy, G. T., Reddy, M. P. K., Lakshmanna, K., Kaluri, R., Rajput, D. S., Srivastava, G., & Baker, T. (2020a). Analysis of dimensionality reduction techniques on big data. IEEE Access, 8, 54776-54788. Gadekallu, T. R., Khare, N., Bhattacharya, S., Singh, S., Maddikunta, P. K. R., & Srivastava, G. (2020). Deep neural networks to predict diabetic retinopathy. Journal Of Ambient Intelligence and Humanized Computing, 1-14. Reddy, T., Bhattacharya, S., Maddikunta, P. K. R., Hakak, S., Khan, W. Z., Bashir, A. K., ... & Tariq, U. (2020b). Antlion re-sampling based deep neural network model for classification of imbalanced multimodal stroke dataset. Multimedia Tools and Applications, 1-25

But this is not reflected in the paper.

[A]

Thank you for your careful review. We added missing references and modified the bibliography form to fit PLOS-ONE. All other comments are already reflected from the R2 version of the paper.

Attachment

Submitted filename: Revision Summary2021.05.29.docx

Click here for additional data file.^{(17.5KB, docx)}

PLoS One. doi: 10.1371/journal.pone.0253121.r007

Decision Letter 3

Thippa Reddy Gadekallu

1 Jun 2021

Impact of Chart Image Characteristics on Stock Price Prediction with a Convolutional Neural Network

PONE-D-20-29152R3

Dear Dr. Kwon,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at http://www.editorialmanager.com/pone/, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to-date. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Thippa Reddy Gadekallu

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

Reviewers' comments:

PLoS One. doi: 10.1371/journal.pone.0253121.r008

Acceptance letter

Thippa Reddy Gadekallu

3 Jun 2021

PONE-D-20-29152R3

Impact of Chart Image Characteristics on Stock Price Prediction with a Convolutional Neural Network

Dear Dr. Kwon:

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department.

If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact onepress@plos.org.

If we can help with anything else, please email us at plosone@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Dr. Thippa Reddy Gadekallu

Academic Editor

PLOS ONE

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Attachment

Submitted filename: Revision Summary2021.01.12.docx

Click here for additional data file.^{(33.3KB, docx)}

Attachment

Submitted filename: Revision Summary2021.05.14.docx

Click here for additional data file.^{(19.2KB, docx)}

Attachment

Submitted filename: Revision Summary2021.05.29.docx

Click here for additional data file.^{(17.5KB, docx)}

Data Availability Statement

[pone.0253121.ref001] 1.Carpenter GA, Grossberg S, Markuzon N, Reynolds JH, Rosen DB, et al. Fuzzy ARTMAP: A neural network architecture for incremental supervised learning of analog multidimensional maps. IEEE Transactions on Neural Networks 1992;3(5):698–713. doi: 10.1109/72.159059 [DOI] [PubMed] [Google Scholar]

[pone.0253121.ref002] 2.Sim HS, Kim HI, Ahn JJ. Is Deep Learning For Image Recognition Applicable to Stock Market Prediction?. Complexity 2019:1–10. 10.1155/2019/4324878. [DOI] [Google Scholar]

[pone.0253121.ref003] 3.De Gooijer JG, Hyndman RJ. 25 Years of Time Series Forecasting. International Journal of Forecasting 2006; 22(3):443–473. 10.1016/j.ijforecast.2006.01.001. [DOI] [Google Scholar]

[pone.0253121.ref004] 4.Menon VK, Vasireddy NC, Jami SA, Pedamallu VTN, Sureshkumar V, Soman. Bulk Price Forecasting Using Spark Over Use Data Set. In International Conference on Data Mining and Big Data 2016; 9714:137–146. 10.1007/978-3-319-40973-3_13. [DOI] [Google Scholar]

[pone.0253121.ref005] 5.Box GEP, Jenkins GM, Reinsel GC, Ljung GM, et al. Time Series Analysis: Forecasting and Control. John Wiley and Sons Inc 2015; 37(5):709–712. 10.1111/jtsa.12194. [DOI] [Google Scholar]

[pone.0253121.ref006] 6.Selvin S, Vinayakumar R, Gopalakrishnan EA, Menon VK, Soman KP. Stock Price Prediction Using LSTM, RNN and CNN-sliding Window Model. In International Conference on Advances in Computing, Communications and Informatics 2017; 23(1):1643–1647. 10.1109/ICACCI.2017.8126078. [DOI]

[pone.0253121.ref007] 7.Hoseinzade E, Saman H. CNNpred: CNN-based Stock Market Prediction Using A Diverse Set of Variables. Expert Systems with Applications 2019; 129:273–285. 10.1016/j.eswa.2019.03.029. [DOI] [Google Scholar]

[pone.0253121.ref008] 8.Cao J, Wang J. Stock Price Forecasting Model Based on Modified Convolution Neural Network and Financial Time Series Analysis. International Journal of Communication Systems 2019; 32(12):e3987. 10.1002/dac.3987. [DOI] [Google Scholar]

[pone.0253121.ref009] 9.Jain S, Gupta R, Moghe A. Stock Price Prediction on Daily Stock Data Using Deep Neural Networks. In International Conference on Advanced Computation and Telecommunication 2018; 10:1–13. 10.1109/ICACAT.2018.8933791. [DOI]

[pone.0253121.ref010] 10.Di Persio L, Honchar O. Artificial Neural Networrs Architectures for Stock Price Prediction: Comparisons and Applications. International journal of circuits, systems and signal processing 2016; 10:403–413. Available from:https://www.semanticscholar.org/paper/41487776364a6dddee91f1e2d2b78f46d0b93499. [Google Scholar]

[pone.0253121.ref011] 11.Guo S, Yan Z, Zhang K, Zuo W, Zhang L. Toward Convolutional Blind Denoising of Real Photographs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2019; 49(1):1712–1722. 10.1109/CVPR.2019.00181. [DOI]

[pone.0253121.ref012] 12.Sezer OB, Ozbayoglu AM. Financial Trading Model with Stock Bar Chart Image Time Series with Deep Convolutional Neural Networks. In Intelligent Automation and Soft Computing 2019; 26(2):2007–2012. 10.31209/2018.100000065. [DOI] [Google Scholar]

[pone.0253121.ref013] 13.Lan R, Zou H, Pang C, Zhong Y, Liu Z, Luo X. Image Denoising Via Deep Residual Convolutional Neural Networks. Signal, Image and Video Processing 2019;15(10):1–8. 10.1007/s11760-019-01537-x. [DOI] [Google Scholar]

[pone.0253121.ref014] 14.Lakhe S, Mariwalla R, Reddy C. Regression Analysis Based Linear Model for Predicting Stock Prices. Industrial Engineering Journal 2017; 10:154–157. 10.26488/IEJ.10.1.9. [DOI] [Google Scholar]

[pone.0253121.ref015] 15.Reddy T, Bhattacharya S, Maddikunta PKR, Hakak S, Khan WZ, Bashir AK, et al. Antlion re-sampling based deep neural network model for classification of imbalanced multimodal stroke dataset. Multimedia Tools and Applications 2020:1–25. 10.1007/s11042-020-09988-y. [DOI] [Google Scholar]

[pone.0253121.ref016] 16.Ballings M, Van den Poel D, Hespeels N, Gryp R. Evaluating Multiple Classifiers for Stock Price Direction Prediction. Expert systems with Applications 2015; 42(20):7046–7056. [Google Scholar]

[pone.0253121.ref017] 17.Bessembinder H, Kaufman HM. A Comparison of Trade Execution Costs for NYSE and NASDAQ-listed Stocks. Journal of Financial and Quantitative Analysis 1997; 287–310. [Google Scholar]

[pone.0253121.ref018] 18.Preis T, Kenett DY, Stanley HE, Helbing D, Ben-Jacob E. Quantifying the behavior of stock correlations under market stress. Scientific reports 2012; 2(1):1–5. doi: 10.1038/srep00752 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0253121.ref019] 19.Hagenau M, Liebmann M, Neumann D. Automated News Reading: Stock Price Prediction Based On Financial News Using Context-capturing Features. Decision Support Systems 2013; 55(3):685–697. 10.1016/j.dss.2013.02.006. [DOI] [Google Scholar]

[pone.0253121.ref020] 20.Qiu M, Song Y. Predicting the Direction of Stock Market Index Movement Using an Optimized Artificial Neural Network Model. PloS One 2016;11(5):e0155133. doi: 10.1371/journal.pone.0155133 . [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0253121.ref021] 21.Fischer T, Krauss C. Deep Learning with Long Short-term Memory Networks for Financial Market Predictions. European Journal of Operational Research 2018; 270(2):654–669. https://ojs.aaai.org/index.php/AAAI/article/view/9354. [Google Scholar]

[pone.0253121.ref022] 22.Deng S, Zhang N, Zhang W, Chen J, Pan JZ, Chen H. Knowledge-driven stock trend prediction and explanation via temporal convolutional network. InCompanion Proceedings of The 2019 World Wide Web Conference 2019 May 13 (pp. 678–685).

[pone.0253121.ref023] 23.Jin X, Chen Z, Yang X. Economic Policy Uncertainty And Stock Price Crash Risk. Accounting Finance 2019;58(5):1291–1318. 10.1111/acfi.12455. [DOI] [Google Scholar]

[pone.0253121.ref024] 24.Wang J, Wang J. Forecasting Stock Market Indexes using Principle Component Analysis and Stochastic Time Effective Neural Networks. Neurocomputing 2015;156:68–78. 10.1016/j.neucom.2014.12.084. [DOI] [Google Scholar]

[pone.0253121.ref025] 25.Kara Y, Boyacioglu MA, Baykan ÖK. Predicting Direction of Stock Price Index Movement Using Artificial Neural Networks and Support Vector Machines: The Sample of the Istanbul Stock Exchange. Expert Systems with Applications 2011; 38(5):5311–5319. 10.1016/j.eswa.2010.10.027. [DOI] [Google Scholar]

[pone.0253121.ref026] 26.Guresen E, Kayakutlu G, Daim TU. Using Artificial Neural Network Models in Stock Market Index Prediction. Expert Systems with Applications 2011; 38(8):10389–10397. 10.1016/j.eswa.2011.02.068. [DOI] [Google Scholar]

[pone.0253121.ref027] 27.Shastri M, Roy S, Mittal M. Stock Price Prediction Using Artificial Neural Model: An Application of Big Data. EAI Endorsed Transactions on Scalable Information Systems 2019; 6:156085. 10.4108/eai.19-12-2018.156085. [DOI] [Google Scholar]

[pone.0253121.ref028] 28.Zhong X, Enke D. Forecasting Daily Stock Market Return using Dimensionality Reduction. Expert Systems with Applications 2017; 67:126–139. 10.1016/j.eswa.2016.09.027. [DOI] [Google Scholar]

[pone.0253121.ref029] 29.Patel J, Shah S, Thakkar P, Kotecha K. Predicting Stock Market Index Using Fusion of Machine Learning Techniques. Expert Systems with Applications 2015; 42(4):2162–2172. 10.1016/j.eswa.2014.10.031. [DOI] [Google Scholar]

[pone.0253121.ref030] 30.Naik N, Mohan B. Intraday Stock Prediction Based on Deep Neural Network. National Academy Science Letters 2020; 43:241–246. 10.1007/s40009-019-00859-1. [DOI] [Google Scholar]

[pone.0253121.ref031] 31.Chung H, Shin KS. Genetic Algorithm-optimized Multi-channel Convolutional Neural Network for Stock Market Prediction. Neural Computing & Applications 2020; 32(12): 7897–914. 10.1007/s00521-019-04236-3. [DOI] [Google Scholar]

[pone.0253121.ref032] 32.Nelson DM, Pereira AC, de Oliveira RA. Stock Market’s Price Movement Prediction with LSTM Neural Networks. In Neural Networks (IJCNN) IEEE 2017:1419–1426. doi: 10.1109/ICORR.2017.8009447 [DOI] [PubMed] [Google Scholar]

[pone.0253121.ref033] 33.Pang X, Zhou Y, Wang P, Lin W, Chang V. An Innovative Neural Network Approach for Stock Market Prediction. The Journal of Supercomputing 2020; 76:2098–2118. 10.1007/s11227-017-2228-y. [DOI] [Google Scholar]

[pone.0253121.ref034] 34.Gunduz H, Yaslan Y, Cataltepe Z. Intraday prediction of borsa istanbul using convolutional neural networks and feature correlations. Knowledge-Based Systems 2017; 137:138–148. 10.1016/j.knosys.2017.09.023. [DOI] [Google Scholar]

[pone.0253121.ref035] 35.Schumaker RP, Chen H. Textual Analysis of Stock Market Prediction Using Breaking Financial News. ACM Transactions on Information Systems 2009; 27(2):1–19. 10.1145/1462198.1462204. [DOI] [Google Scholar]

[pone.0253121.ref036] 36.Saha PK, Borgefors G, di Baja GS. A Survey on Skeletonization Algorithms and Their Applications. Pattern Recognition Letters 2016; 76:3–12. 10.1016/j.patrec.2015.04.006. [DOI] [Google Scholar]

[pone.0253121.ref037] 37.Jerripothula KR, Cai J, Lu J, Yuan J. Object Co-skeletonization With Co-segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition 2017; 33(2):3881–3889. 10.1109/CVPR.2017.413. [DOI]

[pone.0253121.ref038] 38.Paracchini M, Marcon M, Villa F, Tubaro S. Deep Skin Detection on Low Resolution Grayscale Images. Pattern Recognition Letters 2020; 131:322–328. [Google Scholar]

[pone.0253121.ref039] 39.Chen Q, Chen Q, Liao Q, Jiang ZL, Fang J, Yiu S, Xi G, Li R, Yi Z, Wang X, Hui LC, Liu D. File Fragment Classification Using Grayscale Image Conversion and Deep Learning in Digital Forensics. IEEE Security and Privacy Workshops 2018: 140–147. 10.1109/SPW.2018.00029. [DOI]

[pone.0253121.ref040] 40.Keras Team, Keras 2.0 release notes 2020 [Online]. [Accessed 29 April. 2020]. Available from: https://github.com/keras-team/keras/wiki/Keras-2.0-release-notes.

[pone.0253121.ref041] 41.Eapen J, Bein D, Verma V. Novel Deep Learning Model with CNN and Bi-Directional LSTM for Improved Stock Market Index Prediction. IEEE 9th Annual Computing and Communication Workshop and Conference 2019; 9:0264–0270. 10.1109/CCWC.2019.8666592. [DOI]

[pone.0253121.ref042] 42.TensorFlow: An Open Source Machine Learning Framework for Everyone [Online]. [Accessed 29 April. 2020]. Available from: https://www.tensorflow.org/.

[pone.0253121.ref043] 43.Scikit-Learn: Machine Learning in Python [Online]. [Accessed 29 April. 2020]. Available from: https://scikit-learn.org/stable/.

[pone.0253121.ref044] 44.The Matplotlib Development Team (2018, Nov. 28). Matplotlib Version 3.0.2 [Online]. [Accessed 29 April. 2020]. Available from: https://matplotlib.org/.

[pone.0253121.ref045] 45.Mester, T. Pandas Tutorial 1: Pandas Basics [Internet]. [Cited 2020 April 29]. Available from: https://data36.com/pandas-tutorial-1-basics-reading-datafiles-dataframes-data-selection/.

[pone.0253121.ref046] 46.NumPy 2020 [Online]. [Accessed 29 April. 2020]. Available from: http://www.numpy.org/. [Google Scholar]

[pone.0253121.ref047] 47.Kingma DP, Ba J. Adam: A Method For Stochastic Optimization, Machine Learning; 2013 [Cited 2020 April 20]. Available from: https://arxiv.org/abs/1412.6980.

[pone.0253121.ref048] 48.Feurer M, Springenberg JT, Hutter F. Initializing Bayesian Optimization via Meta-Learning. In Twenty-Ninth AAAI Conference on Artificial Intelligence 2015; 14(2):135–147. https://ojs.aaai.org/index.php/AAAI/article/view/9354.

[pone.0253121.ref049] 49.Feurer M, Hutter F. Hyperparameter Optimization. In Automated Machine Learning 2019; 3–33. Available from: 10.1007/978-3-030-05318-5_1. [DOI] [Google Scholar]

[pone.0253121.ref050] 50.Wang Y, Liu M, Yang J. Data-driven Deep Learning for Automatic Modulation Recognition in Cognitive radios. IEEE Transactions on Vehicular Technology 2019; 68(4): 4074–4077. 10.1109/TVT.2019.2900460. [DOI] [Google Scholar]

[pone.0253121.ref051] 51.Loni M, Sinaei S, Zoljodi A, Daneshtalab M, Sjödin M. DeepMaker: A Multi-objective Optimization Framework For Deep Neural Networks In Embedded Systems. Microprocessors and Microsystems 2020; 73:89–93. 10.1016/j.micpro.2020.102989. [DOI] [Google Scholar]

[pone.0253121.ref052] 52.Gadekallu TR, Khare N, Bhattacharya S, et al. Deep neural networks to predict diabetic retinopathy. J Ambient Intell Human Comput 2020: 1–14. Available from: 10.1007/s12652-020-01963-7. [DOI] [Google Scholar]

[pone.0253121.ref053] 53.Reddy GT, Reddy MPK, Lakshmanna K, Kaluri R, Rajput DS, Srivastava G, Baker T, et al. Analysis of dimensionality reduction techniques on big data. IEEE Access 2020; 8:54776–54788. 10.1109/ACCESS.2020.2980942. [DOI] [Google Scholar]

[pone.0253121.ref054] 54.Song Y, Lee J. Performance Evaluation of Deep Learning Stock Price by Chart Type for Buying Policy Verification. In Fuzzy Systems and Data Mining IV 2018:646–652. 10.3233/978-1-61499-927-0-646. [Google Scholar]

PERMALINK

Impact of chart image characteristics on stock price prediction with a convolutional neural network

Guangxun Jin

Ohbyung Kwon

Roles

Abstract

I. Introduction

II. Background

2.1 Stock price prediction

Table 1. Literature on stock price prediction.

2.2 Stock price prediction based on image-based deep learning

III. Methods

Fig 1. Research model.

3.1 Image characteristics

Table 2. Image characteristics.

3.2 CNN characteristics

Fig 2. CNN network.

Table 3. CNN characteristics.

IV. Experiment

4.1 Data

Fig 3. Sample charts.

Fig 4. Sample chart image.

4.2 CNN model

Fig 5. Structure of the CNN used in the analysis.

4.3 Evaluation methodology

4.4 Experimental environment

Table 4. Experimental parameter settings.

V. Results

5.1 Differences in prediction accuracy according to image characteristics

Table 5. Results of ANOVA analysis of the difference in accuracy between stock price predictions using image characteristics.

5.2 Mediating role of filter

Table 6. Results of Chow verification for filter.

5.3 Mediating effect of dropout

Table 7. Results of Chow verification for dropout.

5.4 Performance comparison

Table 8. Comparative performance between previous and present results.

VI. Discussion

6.1 Implications

6.2 Limitations

6.3 Conclusion

Data Availability

Funding Statement

References

Decision Letter 0

Ruxandra Stoean

Roles

Author response to Decision Letter 0

Decision Letter 1

Thippa Reddy Gadekallu

Roles

Author response to Decision Letter 1

Decision Letter 2

Thippa Reddy Gadekallu

Roles

Author response to Decision Letter 2

Decision Letter 3

Thippa Reddy Gadekallu

Roles

Acceptance letter

Thippa Reddy Gadekallu

Roles

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases