Introduction

Due to the volatile nature of the stock market, predicting stock price movements for intraday trade is a challenging task. Interpretation of stock-related critical information makes trading more profitable [1]. The efficient market hypothesis (EMH) states that the prices of trade assets such as stocks already reflect all publicly available information, and thus, if your investments are based on publicly available information, you will not systematically outperform the market over time. Investors are making money by either taking a chance or making riskier investments [13]. However, most of the stock-related studies [4, 5] stated that the stock market can be predicted. Stock analysis can be carried out in two ways. The first is fundamental analysis, and the second is technical analysis. The fundamental analysis involves an examination of financial data, management, business track record, competition, earnings, macroeconomic factors, and overall economic conditions. The technical analysis is a method of making stock measurements based on historical price. This method involves various statistical analyses of stock data based on stock prices.

The extant literature suggests that the technical indicator is an essential parameter for evaluating stock prices [4, 7]. The first contribution of this paper is the identification of a trend in data by using a combination of candlestick data and technical indicators. The existing research has not considered these combinations. The second contribution of this paper is using DNN to classify and accurately predict a stock price’s up and down movements. The existing research is based on a three-layer artificial neural network (ANN), which is unable to classify the up and down movements of stock prices.

Ranjeeta et al. [4] studied stock movement prediction on global indices, namely the Hang Seng index, Financial Times Stock Exchange, and Bombay Stock Exchange. The study considered the variational mode decomposition method for stock data, and machine learning methods were used to identify a nonlinear pattern in data. Paraskevi et al. [14] studied stock mid-price movement prediction using a limit order of stock price data. The work considered a set of features which include the prices, volumes, and ask and bid price of the stock. The performance of extracting features was evaluated based on a machine learning algorithm. Weng et al. [11] proposed multi-filter neural network method to extract stock price feature. The task considered stock-related features, namely open, close, high, low, and volume. Jing et al. [18] proposed sliding window method to cut raw the historical stock price data into the predefined length and proposed an unsupervised heuristic method for stock price prediction. Suryoday et al. [3] considered six technical indicators to predict the stock price using random forests and gradient boosted decision tree methods. Arash et al. [8] proposed a graph-based semi-supervised approach to predict up and down movements of stock prices. However, most of the existing work considered technical indicators to evaluate the stock price movements. Hsun [12] proposed candlestick-based approach for profitable trading in the stock market. However, most of the work considered technical indicators to evaluate the stock prices. Therefore, we have combined technical analysis and candlestick approach in this study.

A comparison between ANN and support vector machine (SVM) classifiers found that ANN performs better than SVM in forecasting the Istanbul Stock Exchange [7]. In another study, ten technical indicators were used to classify the up and down movements of stocks using ANN, SVM, random forest, and Naive Bayes, and it found that the random forest model outperformed the others [15]. The future performance of stock indices was studied using fuzzy time series modeling [16]. Zhang et al. [19] studied an optimization method with a backpropagation ANN to predict stock movements in prices. However, existing trading rules did not predict future periods when market conditions change dynamically. The augmented Dickey–Fuller test has been used to verify and assess the performance of trading rules [6]. Fei Wu et al. [17] discussed face attributes and considered different subspaces for learning the pattern. PCA-based feature extraction technique [9, 10] has been studied to recognize handwritten character, and SVM with different kernel parameters was considered for classification. Mallika et al. [2] proposed PCA method for facial emotion recognition and considered a random forest method for classification. Most of the work considered SVM and ANN classifier to predict the stock price up and down movements.

An overall framework for a stock prediction model described in Fig. 1. The following steps have been considered for extracting the technical indicators.

  • Step 1 Input National Stock Exchange (NSE) data

  • Step 2 Apply ten technical indicators [7]

  • Step 3 Calculate the technical indicators based on formulas [7]

  • Step 4 Output continuous technical indicator values

The following steps have been considered for extracting the candlestick pattern.

  • Step 1 Input NSE data

  • Step 2 Extract day low, high, open, and close price

  • Step 3 Apply Japanese's candlestick pattern on daily basis

  • Step 4 Output candlestick pattern

We have proposed a method to identify a trend in data using a combination of candlestick data and technical indicators, and it is described in Algorithm 1. In Algorithm 1, step 4, if SMA technical values are trading above the past 10-day average value of the stock price and candlestick pattern data are greater than the mean value of stock price, then it is considered as up movement, otherwise down movement (https://in.tradingview.com/ideas/candlestick/). In step 5, if EMA technical values are trading above the past 10-day average value of the stock price and candlestick pattern data are greater than the mean value of the stock price, then it is considered as up movement, otherwise down movement. In step 6, if MOM technical values are trading above the past 10-day average value of the stock price and candlestick pattern data are greater than the mean value of the stock price, then it is considered as up movement, otherwise down movement. In step 7, if STCK technical values are trading above STCD 14-day average value of the stock price and candlestick pattern data are greater than the mean value of the stock price, then it is considered as up movement, otherwise down movement.

Fig. 1
figure 1

Overall framework of proposed model

Full size image

In step 8, if RSI technical values are trading above 30 and candlestick pattern data are greater than the mean value of the stock price, then it is considered as up movement, otherwise down movement. In step 9, if MACD technical values are trading above MACD 26-day average value of the stock price and candlestick pattern data are greater than the mean value of the stock price, then it is considered as up movement, otherwise down movement. In step 10, if R technical values are trading below \(-80\) and candlestick pattern data are greater than the mean value of the stock price, then it is considered as up movement, otherwise down movement. In step 11 if A/D technical values are trading below \(-100\) and candlestick pattern data are greater than the mean value of the stock price, then it is considered as up movement, otherwise down movement. In step 12 if CCI technical values are trading above 100 and candlestick pattern data are greater than the mean value of the stock price, then it is considered as up movement, otherwise down movement. In step 13 if EMA technical values are trading above past 20-day average value of the stock price and candlestick pattern data are greater than the mean value of the stock price, then it is considered as up movement, otherwise down movement.

The output of this algorithm is the modified technical indicators, which are inputs to the DNN prediction model.

figure a

DNN is one of the popular methods for learning lots of different types of pattern. The proposed model composed of more than one neural network layer. The neural network is designed based on human brains. Human brains consist of billions of small cells known as neurons. Kara et al. [7] considered a three-layer artificial neural network model for stock movements’ classification, and it is described in Fig. 2. The proposed work, DNN, is considered to classify stock price’s up and down movements, as shown in Fig. 3. It is five layers of architecture with three hidden layers, and each layer transfers the data between the neurons. It takes the input of ten technical indicator values, and it is denoted by \(x_i\). Each of these technical indicators' values is multiplied by weight \(W_i\), and it denoted as \(W_i x_i\). Initially, some random values are assigned. Each of these weights is assigned to input value and weighted sum performed, and it is defined in Eq. (1). The bias is typical for each neuron. There is one bias per neuron. An activation function takes the weighted sum of input and bias. Kara et al. [7] study considered the sigmoid activation function. The sigmoid function is used for models where we have to predict the probability as an output. It exists between 0 and 1 and is represented by \(f(x) = \frac{e^{x}}{e^{x}+1}\). The ReLU activation function is considered in the experiment. ReLU is the most popular function used to deal with nonlinear data. It gives an output of x if x is positive and 0 otherwise, and it is represented as \(f(x) = \max (0,x)\). There is no upper limit in ReLU. After training the DNN, it uses backpropagation method to improve the performance of the network. Cost function helps to reduce the error rate. The cost value is the difference between DNN predicted output and the actual output. The least cost value is obtained by making adjustments to the weights and biases iteratively throughout the training process. Stochastic gradient descent (SGD) mechanism identifies how to change the weight and bias to minimize the cost functions. The DNN training process is nothing but finding the best value of weight and biases for each neuron in the network to maximize accuracy.

$$\begin{aligned} \alpha = \bigg (\sum _{i=1}^{n} W_i x_i + B \bigg ) \end{aligned}$$
(1)
Fig. 2
figure 2

Three-layer ANN prediction model by Kara et al. [7]

Full size image
Fig. 3
figure 3

Framework of proposed five-layer deep neural network

Full size image

In this work, stock price data are collected from NSE, India, and the website URL is (http://www.nseindia.com/stock). The stocks considered herein are Infosys, Reliance, HDFC, and HDFC Bank. The datasets range from the year 2008 to 2018. We have studied two commonly used metrics referenced in the literature, namely accuracy and F measures, and they are described in Eqs. (2) and (3). These metrics are considered to evaluate the performance of DNN model. We use tenfold cross-validation in the experiment.

$$\begin{aligned} {{\mathrm{Accuracy}}}= & {} \frac{{{\mathrm{TP}}} +{{{\mathrm{TN}}}}}{{{\mathrm{TP}}} +{{\mathrm{TN}}} + {{\mathrm{FP}}} + {{\mathrm{FN}}}} \end{aligned}$$
(2)
$$\begin{aligned} {{{ F}}} {{\mathrm{Measure}}}= & {} 2\times \frac{ {{\mathrm{Precision}}} \times {{\mathrm{Recall}}}}{{{\mathrm{Precision}}} + {{\mathrm{Recall}}}} \end{aligned}$$
(3)

Table 1 shows that the highest accuracies for Reliance stock, Infosys, HDFC Bank, and HDFC are 0.8361, 0.8178, 0.8211, and 0.8349, respectively. The performance of the proposed DNN model is better than the existing work and described in Table 1. The reason for the better performance is that the data preparation uses a combination of candlestick data and technical indicators. In addition, we have built a model on a five-layer deep neural network; the existing model was built on a three-layer artificial neural network.

Table 1 Results' comparison with the existing work
Full size table

In this study, we make predictions of the stock price’s up and down movements in the NSE. The first contribution of this paper is the identification of a trend in data by using a combination of candlestick data and technical indicators. The existing research has not considered these combinations. The second contribution of this paper is using DNN to classify and accurately predict a stock price's up and down movements. The proposed deep neural network prediction model outperforms the ANN model. This paper focuses on intraday predictions. In the future work, we examine short-term and long-term prediction along with fundamental data analysis, such as a stock’s price–earnings ratio, book value, and a company’s quarterly profit.