Stock and Futures Market Prediction Using Deep Learning Approach

Abstract

In recent years, numerous studies have been devoted to predict the price fluctuations of financial markets. Taiwan 50 Exchange Traded Funds (ETF) is one of the important indicators to measure the volatility of the component stocks of the Taiwan 50 Index. With the development of the financial market, the trading volume of Taiwan Stock Index Futures is also increasing. The three markets play the important roles of economic development in the Taiwan. This study predicts the trend of Taiwan 50 ETF and Taiwan index futures applying machine learning and deep learning approaches which have excellent data exploration capabilities. This study applies the support vector regression (SVR), artificial neural networks (ANN), recurrent neural network (RNN), and long short-term memory network (LSTM) to predict the trend of the Taiwan stock market. This study uses various financial and technical factors as inputs, and extract variables from the factors affecting Taiwan’s economy to build models, and compares the benefits between models to explore future market.

1. Introduction

Stock market prediction has been commonly regarded as a challenging topic. Various econometric or statistical models were widely used in the prior studies [1, 2]. Although the prediction of stock market returns is a non-linear problem, many studies have proved that the non-linear models are not necessarily superior to the linear models [3, 4, 5]. Moreover, the linear technical indices are proved to help investors make investment strategies to improve investment performance. Zhu et al. used the variable moving average (VMA) on the basis of the moving average (MA) to successfully predict the stock price trends of Shanghai Stock Exchange Composite Index (SSEC) and the Shenzhen Stock Exchange Component Index (SZSE) [6].

Recent studies used the machine learning methods to solve the non-linear problems to capture the fluctuation of stock price. Chen and Hao proposed the hybridized framework of the feature weighted support vector machine (FWSVM) and the feature weighted K-nearest neighbor (FWKNN) to effectively predict the stock market indices of the SSEC and the SZSE [7]. They found that that these models can achieve a better prediction capability to both the SSEC and the SZSE in the short, medium and long term. Paiva et al. embedded support vector machine into an investment decision to predict the São Paulo Stock Exchange Index (Ibovespa) [8]. However, machine learning would require taking a lot of time in feature extraction, regularizing the extracted features, and then interpreting the roles through the expert knowledge.

Compared with the machine learning, deep learning autonomously regularizes the information features to reduce many steps in the machine learning process, and thus greatly saving labor costs and time. Recurrent neural network (RNN) which is on approach of deep learning approach is specifically suitable for dealing with the time series issues. Chen et al. proposed a hybrid model combining the RNN and the adaptive boosting (AdaBoost) to predict the stock indices, where the experiment results showed that the model has good prediction performance [9].

Taiwan 50 Exchange Traded Funds (ETF) was founded by the Yuanta Funds in 2003, consisting of 50 largest stocks by market value of the Taiwan Stock Exchange Corporation (TSEC). The Taiwan Stock Exchange Index Futures (FITX) was first launched by the Taiwan Futures Exchange (TAIFEX) in 1998, targeting on the weighted stock index issued by the TSEC, and changing in accordance with the fluctuation of the weighted stock index. The FITX has always been the largest trading commodity in Taiwan’s future market. This study applies the support vector regression (SVR), the artificial neural networks (ANN), the RNN and the long short-term memory network (LSTM) to predict the trend of the Taiwan stock market. The trained models are not only used to predict the fluctuation trend of the stock market, but also compared to decide which one has the best performance. In this study, the Taiwan 50 ETF and the FITX are taken as the study samples. This study calculates the technical indices based on the stock data provided by the Taiwan Economic Journal (TEJ).

The remainders of this study are organized as follows. The literatures are reviewed in the Section 2. In Section 3, this study develops data collection and research models. Section 4 describes the experiment results and performance evaluation. The Section 5 concludes this work.

2. Literature review

2.1 Stock market and foreign exchange market

The three major institutional traders in Taiwan stock market refer to the foreign investors, the investment trusts and the dealers. Foreign investors specifically refer to the investors of the foreign stock markets except Taiwan, the investment trusts stand for the securities investment companies, and the dealers shall be the securities companies certified by the government to use their own company capital to trade on the TSEC. Wang analyzed the impact of institutional investors’ sentiment on the stock market trading, and used four volatility models to analyze the stocks on the American Stock Exchange (AMEX) [10]. They found that the institutional investors take lower investment risks when they are pessimistic; on the contrary, they are take higher risks when they are optimistic about the stock market.

The foreign exchange market is a financial market distributed among countries to trade currencies. Except for the holidays and major festivals, the foreign exchange markets continuously operate and trade various currencies throughout the day. The foreign exchange rates fluctuate complexly as the foreign exchange market would be affected by various external factors, making it not easy to predict the currency exchange rates. Yong et al. proposed the Gaussian mixture model initialized neuro fuzzy (GMMINF) model to predict the exchange rates of the Australian dollar, the euro and the British pound against the US dollar, and obtain good prediction results at the closing price [11]. Henryíquez and Kristjanpoller suggested a hybrid model of an independent component analysis in combination with the neural network to predict the exchange rates of the euro, the British pound, the Japanese dollar, the Swiss franc and the Canadian dollar against the US dollar, where the results showed that the accuracy of the proposed hybrid model is higher than other econometric models [12].

2.2 Technical indices

Grodinsky proposed fundamental analysis and technical analysis to analyze the stock market price. Investors can use the fundamental analysis to evaluate the firm value or predict the fluctuation of the financial market through the economic theories and the financial statements analysis [13]. The fundamental indicators include earnings per share (EPS), price-to-earnings ratio (PER), price-book ratio (PBR), earnings before interest and tax (EBIT) and return on equity (ROE). Bunea et al. used the linear regression model to observe the financial statements of 1253 firms in Romania, where the results showed that the asset turnover and the PER have the greatest impact on the stock return [14]. Lončarski and Vidovič used the principal component analysis (PCA) to analyze the financial statements of the companies in the S & P 500 index, where the profitability, the growth prospects and the risk can be effectively reflected through the analysis of the financial statements [15].

The technical analysis was originated from the Dow Theory of Robert [16], where the theory proposed that the stock prices will fluctuate with the stock market. Elliott subsequently suggested the Elliott Wave Theory, explaining that the end time of the trend can be predicted [17]. Afterwards, Granville mentioned the moving average, providing the basis for most of the technical indices [18]. Common technical indices are also the moving average convergence diversity (MACD), the relative strength index (RSI) and the stochastic oscillator (KD).

Lin used 14 technical indices calculated to predict the fluctuation of the stock market, and regression analysis was used to analyze their relationship [19]. He found that the technical analysis can significantly improve the predictability of the stock price from the investment portfolio. Jiang et al. used the technical indices and macroeconomic characteristics as the inputs, and used the decision tree algorithms and deep learning to predict the S & P 500, the Dow 300 and the Nasdaq 100 indices, where the results showed that these models have achieved good performance [20]. Picasso et al. used 10 technical indices from the Nasdaq 100 index and the financial news related to the Nasdaq 100 index from the Intrinio API, where the emotional features were extracted from the texts, and were further combined with the technical indices as the training variable inputs to the Random Forest, SVM, feed forward neural network models [21]. The results showed that the proposed method can effectively classify the uptrend and downtrend in the stock portfolio.

2.3 Support vector regression

Machine learning is an innovative method combining statistics and computer science, by which allows the machine to learn from the past data and experience, find out the rules and make its learning as smart as the human beings. Nowadays, it has become popular in various application fields, such as visual images, natural language processing, search engine, voice handwriting recognition, data exploration and securities market analysis and prediction. In terms of securities market analysis and prediction, Patel et al. compared four models trained by the artificial neural network, the SVM, the Random Forest and the Bayesian in combination with the technical indices, where the results showed that the Random Forest has higher accuracy than the other models [22]. Lee et al. proposed the financial network index of investment strategy, using the logical regression, the SVM and the Random Forest, which justified that the proposed scheme can effectively predict the fluctuation of the stock market, specifically with the best performance in the short-term investment [23].

The support vector machine (SVM) was derived by Vapnik and Chervonenkis [24]. It is a supervised learning algorithm combined with classification and regression. In this study, the support vector regression (SVR), an extension of the SVM, is used. SVR has the capable of processing continuous data, and is suitable for deal with the time series data of stocks. Lahmiri used the SVR to analyze the stock data of a listed personal computer company, and the results showed that the accuracy of the model trained by the SVR is higher than other fundamental analysis models [25]. Mishra and Padhy selected 92 stocks in the Indian stock market, and used SVR to predict the stock market prices, where the results showed that the predicted prices are close to the actual prices, and the investment transactions of the selected stocks have high returns [26].

2.4 Artificial neural network and deep learning

Artificial neural network (ANN) is one kind of the machine learning, and simulates the human brain to make judgment. Rezaee et al. integrated the Fuzzy C-Means, the data envelopment analysis and the ANN to analyze the listed company data of the Tehran Stock Exchange (TSE) in Iran, where the results showed this method can help investors make decisions fast [27]. Wang et al. combined price volatility network with ANN to predict the copper price in the New York Mercantile Exchange (NYMEX), where the results showed that the proposed hybrid technology is more accurate than the traditional neural network in terms of the price prediction [10].

Deep learning extends from ANN, the continuously increases its hidden layers in addition to the input layer, the hidden layer and the output layer constituting the traditional ANN. LeCun et al. [28] proposed convolutional neural network (CNN) to process pictures, while Williams and Zipser [29] proposed RNN to deal with texts. Gunduz et al. [30] used the CNN to predict the stock price of the Borsa Istanbul Stock Exchange (BIST) in Turkey, where the results showed that it can successfully capture the hourly fluctuation of the stocks. Xu et al. used the RNN to classify the features of China Securities Index 300 (CSI 300) [31]. Chen et.al proposed a hybrid model combined with the RNN and the adaptive boosting to predict the CSI 300 Index, where the results showed that the hybrid model is superior to ANN and SVR [9].

The long short-term memory (LSTM) is an artificial recurrent neural network (RNN) architecture. The first RNN was proposed by Williams and Zipser [29]. When the first sequence of the RNN is performing the output operation, an additional set of vector information is also being output concurrently as the input to the subsequent sequence point. Therefore, the second sequence point have the previous sequence point and its own set of vector information. These two sets of vector information are subject to certain calculation before being output, which means that in addition to the latest information, the RNN model would simultaneously receive the previous information to achieve the effect of memory. Nevertheless, the RNN has one drawback—although this model has the effect of memory, it cannot keep long-term memory but only short-term memory. To resolve the aforesaid problem, Hochreiter and Schmiduber proposed an artificial neural network named LSTM [32]. The biggest difference between the LSTM and the RNN lies that the LSTM has many cell states, which can receive the long-term memories of the past. Fischer and Krauss used the LSTM, an ANN for sequence learning, to predict the S & P 500 index, where the results showed that the accuracy of the model was higher than that of the machine learning [33]. Cao et al. proposed a hybrid model combined with the empirical mode decomposition (EMD) and the LSTM to predict the stock prices using four stock market indices in the United States, where the results showed that the proposed model performs better than the fundamental analysis model [34].

3. Research methodology

According to prior literature, this study chooses the technical indices, the institutional investor and the foreign exchange market of Taiwan 50 ETF and the Taiwan Stock Exchange Index Futures (FITX), and then reduce these data dimensionalities to create the input variables. After training by the SVR, ANN, RNN and LSTM respectively, this study obtains the best model to predict the stock price fluctuation, and further verify the accuracy of each model to judge which model has high accuracy for the stock prediction.

3.1 Research framework

This study classifies each of Taiwan 50 ETF and the FITX into its own control group and experimental group. The control group is the fundamental data, while the experimental group includes the foreign exchange, the major institutional traders and the technical index. This study investigates the correlation of the research object against its own individual inputs, select the best model by analyzing the training results, use the best model to predict the price fluctuation of the research objects, and then evaluate the performance of the model. The details for the experimental design are illustrated in the following Figures 1 and 2.

Figure 1.

The experimental design for Taiwan 50 ETF.

Figure 2.

The experimental design for FITX.

3.2 Datasets collection

This study use the data of Taiwan 50 ETF, the FITX, the foreign exchange rate, the technical indices and the three major institutional traders. The research period is from 2015 to 2018.

All data related to the Taiwan 50 ETF and the FITX come from TEJ database. According to Cao et al. [34], Xu et al. [31] as well as Mishra and Padhy [26], we use the opening price, the highest price, the lowest price and the closing price for the Taiwan 50 ETF and the FITX.

For the exchange rates of the foreign exchange market, this study uses the six foreign currencies are converted to the TWD in terms of the cash exchange rate and the spot exchange rate respectively, thereby a total of 12 variables. According to Yong et al. [11] and Baffour et al. [35], the exchange rates of the euro, the British pound and the US dollar were selected as the variables. The renminbi (CNY), the Japanese yuan (JPY) and the Hong Kong dollar (HKD) circulating in the regions near Taiwan are also included.

For the technical index, this study also applies eight parameters from KD, MACD and RSI. Jiang et al [20] and Picasso et al. [21] suggested that KD, MACD and RSI are important technical indexes. The eight parameters include RSV, K, D, EMA12, EMA26, DIF, MACD and RSI.

For the three major institutional traders, the data come from the TAIFEX, and include six variables. The three major institutional traders are defined as the dealers, the investment trusts and the foreign investors. Only the long and short trading volumes of the three institutional traders are selected as the input variables of the model. This study only selects trading volume because that the trading volume positively correlates with the contract value.

3.3 Dimensionality reduction and feature selection

All data are subject to the dimensionality reduction suggested by Noryani et al. [36] and normalization according to Cao et al. [34] before using as the inputs of the models.

To avoid decrease the accuracy of model, we adopt the stepwise regressions that for dimensionality reduction. The selection criterion is the group with the largest value of adjusted R square. In the models, we use current dependent variables, but use lagged independent variables. Finally, we select 53 variables, including 13 variables for the opening prices of the Taiwan 50 ETF, 14 variables for the closing prices of the Taiwan 50 ETF, 13 variables for the opening prices of the FITX, and 13 variables for the closing prices of the FITX.

Regarding the Taiwan 50 ETF, the variables of opening prices refer to five variables related to foreign exchanges (HKD, GBP, EUR, JPY, and CNY), two variables related to three major institutional traders (Long trading volumes of the foreign investors and short trading volumes of the dealers), and six variables related to technical indices (EMA26, MACD, EMA12, RSI, RSV, and D). This study adds USD into the variables of closing prices of the Taiwan 50 ETF, others are similar to the opening prices of the Taiwan 50 ETF.

Regarding the FITX, the variables of opening prices refer to fix variables related to foreign exchanges (HKD, GBP, EUR, JPY, CNY and USD), two variables related to three major institutional traders (Long trading volumes of the foreign investors and short trading volumes of the dealers), and five variables related to technical indices (MACD, EMA12, RSI, RSV, and D). The variables for closing prices of the FITX are the same as for opening prices of the FITX.

4. Experimental results and performance evaluation

4.1 Experimental environment

The experimental environment of this study was developed under the Windows 10 system (i7 CPU, 16G RAM); the application was programmed with the Python, while the stepwise regression analysis in the SPSS (Statistical Product and Service Solutions) was selected to perform the dimensionality reduction of the data; the free value-added open source version of Anaconda software was used to create the artificial intelligence model, under which the Tensorflow environment was established; the Keras kit was used to create the models with the ANN, the RNN and the LSTM, while the machine learning model was built with the Scikit-Learn kit to create the SVR model.

4.2 Parameters setting and performance evaluation index

This study chooses the linear kernel function with the C (cost) adjusted to 10 and the learning rate of kernel algorithm adjusted to 0.001. Except for the kernel function, the other parameters were tested by trial and error method. The ANN is divided into four layers, one of which was the input layer, two the hidden layers, and one the output layer. The number of iterations is 200. This study uses the adaptive moment estimation (ADAM) as the optimizer and the MAE in the Loss equation with the learning rate 0.001 and the activation function Linear. All the parameter settings are also tested by trial and error method.

In the RNN model, there are five layers, one of which was the input layer, two the RNN layers, one the fully connected layer, and one the output layer. The number of layers is set according to Jiang et al. [20]. The number of iterations is 200. We select the ADAM as the optimizer and the MAE in the Loss equation with the learning rate 0.001 and the time step 3. The activation function of the RNN layer is Tanh, and the activation function of the output layer is Linear. All the parameter settings are tested by trial and error method.

In the LSTM model, there are also five layers, one of which was the input layer, two the LSTM layers, one the fully connected layer, and one the output layer. The number of layers is set according to Cao et al. [34]. The number of iterations is 200. Again, we select the ADAM as the optimizer and the MAE in the Loss equation with the learning rate 0.001 and the time step 3. The activation functions of the LSTM layer are Sigmoid and Tanh, and output layer activation functions is Linear. All the parameter settings are tested by trial and error method.

This study selects the mean absolute error (MAE), the mean square error (MSE) and the mean absolute percent error (MAPE) as the evaluation indices, where the formulas are listed as follows:

$$\mathit{MAE}=\frac{1}{n}\sum _{i=1}^n\mid y_i-\widehat{y}\mid$$E1

$$\mathit{MSE}=\frac{1}{n}\sum _{i=1}^n{\left(y_i-\widehat{y}\right)}^2$$E2

$$\mathit{\text{MAPE}}=\frac{1}{n}\sum _{i=1}^n\mid \frac{y_i-\widehat{y}}{y_i}\mid$$E3

Where ${\mathit{y}}_{\mathit{i}}$ is the actual value, $\widehat{\mathit{y}}$ the predicted value, n the total number of data, i a certain amount of data. The smaller the values of MAE, MSE and MAPE are, the better the accuracy of the predictive model.

4.3 Experimental results

In the experimental results of this study, the mean absolute error (MAE), the mean square error (MSE) and the mean absolute percent error (MAPE) were selected as the evaluation indices of the model, where the smaller the values of MAE, MSE and MAPE are, the better the accuracy of the predictive model.

Tables 1–4 show the results of the Taiwan 50 ETF. This study gets the different results among fundamental data, foreign exchange, the technical indices and the major institutional traders. For the fundamental data, the LSTM gives the best predictive power, followed by the RNN, and finally the ANN. Figures 3 and 4 illustrates the trend prediction for the closing price and opening price with different machine learning and deep learning approaches, respectively. For the foreign exchange and the technical indices, the ANN exhibits the best predictive power, followed by the LSTM, and finally the RNN. Figures 5 and 6 illustrates the foreign exchange trend prediction for the closing price and opening price with different machine learning and deep learning approaches, respectively. Figures 7 and 8 are represented for the closing price and opening price with technical indices. For the major institutional traders, the LSTM presents the best results, followed by the ANN, and finally the RNN. The details are also illustrated in Figures 9 and 10.

Models	Prediction	MAE	MSE	MAPE
SVR	Closing price	1.984	4.587	2.406
ANN	Closing price	0.628	0.788	0.772
RNN	Closing price	0.627	0.772	0.769
LSTM	Closing price	0.620	0.763	0.765
SVR	Opening price	4.696	23.298	5.699
ANN	Opening price	0.413	0.379	0.497
RNN	Opening price	0.409	0.367	0.493
LSTM	Opening price	0.403	0.378	0.488

Table 1.

The results of the fundamental data of the Taiwan 50 ETF.

Models	Prediction	MAE	MSE	MAPE
SVR	Closing price	3.306	16.054	4.032
ANN	Closing price	0.603	0.732	0.743
RNN	Closing price	0.618	0.756	0.753
LSTM	Closing price	0.610	0.743	0.751
SVR	Opening price	5.137	32.284	6.126
ANN	Opening price	0.386	0.359	0.478
RNN	Opening price	0.398	0.384	0.488
LSTM	Opening price	0.391	0.373	0.481

Table 2.

The results of the foreign exchange of the Taiwan 50 ETF.

Models	Prediction	MAE	MSE	MAPE
SVR	Closing price	3.164	10.928	3.842
ANN	Closing price	0.638	0.759	0.783
RNN	Closing price	0.674	0.829	0.832
LSTM	Closing price	0.633	0.748	0.767
SVR	Opening price	4.937	25.623	5.897
ANN	Opening price	0.421	0.380	0.512
RNN	Opening price	0.488	0.469	0.604
LSTM	Opening price	0.419	0.381	0.510

Table 3.

The results of the major institutional traders of the Taiwan 50 ETF.

Models	Prediction	MAE	MSE	MAPE
SVR	Closing price	3.934	16.596	4.796
ANN	Closing price	0.608	0.721	0.751
RNN	Closing price	0.621	0.746	0.761
LSTM	Closing price	0.611	0.829	0.776
SVR	Opening price	5.303	29.290	6.447
ANN	Opening price	0.383	0.346	0.474
RNN	Opening price	0.393	0.368	0.491
LSTM	Opening price	0.388	0.356	0.484

Table 4.

The results of the technical indices of the Taiwan 50 ETF.

Figure 3.

The prediction of the Taiwan 50 ETF closing price of the fundamental data. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

Figure 4.

The prediction of the Taiwan 50 ETF opening price of the fundamental data. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

Figure 5.

The prediction of the closing price of the foreign exchange rate. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

Figure 6.

The prediction of the opening price of the foreign exchange rate. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

Figure 7.

The prediction of the closing price of the technical indices. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

Figure 8.

The prediction of the opening price of the technical indices. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

Figure 9.

The prediction of the closing price of the major institutional traders. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

Figure 10.

The prediction of the opening price of the major institutional trader group. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

Tables 5–8 present show the results of the FITX. For the fundamental data, the LSTM gives the best predictive power, followed by the RNN, and finally the ANN. Besides, fundamental data, foreign exchange, the technical indices and the major institutional traders have the same priority of predictive power. This study finds that the LSTM exhibits the best results, followed by the ANN and finally the RNN. The details are also illustrated in Figures 11–18.

Models	Prediction	MAE	MSE	MAPE
SVR	Closing price	205.683	51803.42	1.925
ANN	Closing price	82.635	13881.782	0.794
RNN	Closing price	81.351	13511.3	0.786
LSTM	Closing price	80.893	13315.75	0.770
SVR	Opening price	406.140	180288.4	3.822
ANN	Opening price	55.214	6979.852	0.531
RNN	Opening price	54.520	6807.414	0.523
LSTM	Opening price	53.960	6701.942	0.518

Table 5.

The results of the fundamental data of the FITX.

Model	Prediction	MAE	MSE	MAPE
SVR	Closing price	385.580	168158.3	3.065
ANN	Closing price	79.325	12857.62	0.758
RNN	Closing price	80.150	13731.32	0.761
LSTM	Closing price	78.087	11314.08	0.743
SVR	Opening price	180.467	46233.91	1.604
ANN	Opening price	51.591	6326.661	0.483
RNN	Opening price	53.664	6724.012	0.520
LSTM	Opening price	52.117	6865.473	0.495

Table 6.

The results of the foreign exchange of the FITX.

Model	Prediction	MAE	MSE	MAPE
SVR	Closing price	199.591	49900.01	1.873
ANN	Closing price	84.654	14017.21	0.828
RNN	Closing price	85.681	14851.94	0.823
LSTM	Closing price	82.742	13731.18	0.803
SVR	Opening price	413.321	184692.3	3.903
ANN	Opening price	57.731	6659.928	0.552
RNN	Opening price	59.927	7734.202	0.621
LSTM	Opening price	55.975	6390.768	0.536

Table 7.

The results of the major institutional traders of the FITX.

Model	Prediction	MAE	MSE	MAPE
SVR	Closing price	273.919	86916.05	2.586
ANN	Closing price	79.087	12014.09	0.746
RNN	Closing price	80.301	13448.05	0.765
LSTM	Closing price	79.153	12442.942	0.748
SVR	Opening price	500.666	265107.2	4.734
ANN	Opening price	53.546	6435.225	0.514
RNN	Opening price	53.153	6411.562	0.511
LSTM	Opening price	52.344	6433.198	0.499

Table 8.

The results of the technical indices of the FITX.

Figure 11.

The prediction of the FITX closing price of the fundamental data. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

Figure 12.

The prediction of the FITX opening price of the fundamental data. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

Figure 13.

The prediction of the FITX closing price of the foreign exchange rate. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

Figure 14.

The prediction of the FITX opening price of the foreign exchange rate. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

Figure 15.

The prediction of the FITX closing price of the major institutional traders. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

Figure 16.

The prediction of the FITX opening price of the major institutional traders. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

Figure 17.

The prediction of the FITX closing price of the technical indices. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

Figure 18.

The prediction of the FITX opening price of the technical indices. (a) SVR, (b) ANN, (c) RNN and (d) LSTM.

4.4 Prediction of the price fluctuation

Finally, this study picks out the model with the best training performance according to the above results. This study then further predicts the price fluctuation of 2018 for all trained models with the same parameter settings. The trading results are converted into a confusion matrix, whose equations are listed as follows:

$$\mathit{\text{Precision}}=\mathit{TP}/\left(\mathit{TP}+\mathit{FP}\right)$$E4

$$\mathit{\text{Recall}}=\mathit{TP}/\left(\mathit{TP}+\mathit{FN}\right)$$E5

$$F1-\mathit{\text{score}}=2\times \mathit{TP}/\left(2\ast \mathit{TP}+\mathit{FP}+\mathit{FN}\right)$$E6

$$\mathit{\text{Accuracy}}=\left(\mathit{TP}+\mathit{TN}\right)/\left(\mathit{TP}+\mathit{FP}+\mathit{TN}+\mathit{FN}\right)$$E7

The TP herein is defined as True Positive, FP as False Positive, TN as True Negative, and FN as False Negative. Where TP represent that the actual value is correct and the predicted value is correct, FP the actual value wrong and the predicted value correct, TN the actual value wrong and the predicted value wrong, as well as FN the actual value correct and the predicted value wrong. The precision, the recall, the F1-score and the accuracy are the evaluation indices of the confusion matrix, where the larger the value of the evaluation index, the better the accuracy of the model.

Tables 9 and 10 show the model with the highest confusion matrix values of Taiwan 50 ETF and FITX respectively, and are further used to predict the price fluctuation of the next day. This study takes the Taiwan 50 ETF and the FITX as the targets. TP represents that the price rises actually and is predicted to rise, FP represents that the price falls actually but is predicted to rise, TN represents that the price falls actually and is predicted to fall, and FN represents that the price rises actually but is predicted to fall. In Table 9, the experimental results find that the prediction accuracies of fluctuation for Taiwan 50 ETF are about 70%, which proves that this prediction of the price fluctuation is successful. In Table 10, the experimental results also find that the prediction accuracies of price fluctuation for the FITX are about 70%, which again proves that this prediction of the price fluctuation is successful and is of certain commercial value.

Group	Model	Benchmark	Precision	Recall	F1-score	Accuracy
Fundamental	LSTM	Opening price	0.733	0.687	0.709	0.707
Foreign exchange	ANN	Opening price	0.758	0.694	0.725	0.731
Major institutional trader	LSTM	Opening price	0.686	0.648	0.666	0.668
Technical index	ANN	Opening price	0.775	0.720	0.746	0.743

Table 9.

The confusion matrix of Taiwan 50 ETF.

Group	Model	Benchmark	Precision	Recall	F1-score	Accuracy
Fundamental	LSTM	Opening price	0.611	0.738	0.669	0.670
Foreign exchange	ANN	Opening price	0.664	0.773	0.714	0.711
Major institutional trader	LSTM	Opening price	0.597	0.720	0.653	0.654
Technical index	LSTM	Opening price	0.656	0.8	0.721	0.723

Table 10.

The confusion matrix of FITX.

5. Conclusion

In this study, there are three findings. After using the stepwise regression method for data dimensionality to reduce by select the variables, the first experimental results showed that the foreign exchange and the technical indices are superior to the fundamental data, and the major institutional traders is inferior to the fundamental data. Secondly, this study compares the results of SVR, the ANN, the RNN and the LSTM models, and then use them to predict the stock market dynamics. The experimental results find that the SVR performs the worst. Specifically, the LSTM has the highest model accuracy in the control group, and the highest accuracy is given mostly by the LSTM but partly by the ANN in the experimental group. Thirdly, this study predicts the price fluctuation and obtains the accuracy about 70%.

5.1 Dimensionality reduction and feature selection

In this study, an experimental group and a control group are distinguished. In the experiments, stepwise regression analysis is used to select important variables from each data to achieve data dimensionality reduction. In the experimental group, basic stock market information is combined with many additional variables in an attempt to improve the accuracy of the model. The control group used the basic information of the stock market to compare the results with the experimental group. The experimental results showed that the model accuracy of the foreign exchange group and the major institutional trader group was higher than that of the basic group, while the major institutional trader group was lower than that of the basic group, which means that this experiment is successful. The additional variables of the foreign exchange group and the major institutional trader group help to increase the accuracy of the model, but adding the major institutional trader group actually reduces it. According to the study Abdioglu et al. [1], it shows that the three major legal institutional trader exist among the investors. Information asymmetries may cause the legal team to be unable to learn effectively during training.

5.2 Models comparison

This study uses SVR, ANN, RNN, and LSTM models to predict the stock market dynamics. The experimental results show that SVR is the worst among all testing results. It is speculated that SVR may not be able to adapt to this data. In the control group, the LSTM in predicting the opening price and closing price, both models have the highest accuracy. The experimental results show that consistent with Fischer and Krauss [33], LSTM has the best learning effect. In the experimental group, most of the LSTM models have the highest accuracy, but partly the ANN model is the highest. Overall, except for SVR, the ANN, RNN, and LSTM all perform well in predicting prices.

5.3 Forecast of uptrend and downtrend

This study predicts the price fluctuation. The price fluctuation prediction for the fundamental data, the foreign exchange, the technical indices and the major institutional traders are all have the accuracy about 70%, which is of a certain commercial value.

According to the final experimental results of this study, it can be found that the trend of uptrend and downtrend can be grasped in 2018. In the future, this study will continue to update the dataset with different time period to see whether the accuracy of the future model is as good as now. In addition, according to the parameters setting of this study, the method of predicting the opening and closing prices of the stock market was adopted. The method can also be redesigned to directly predict the future rise and fall of the stock market to explore Taiwan’s financial market. It can also be used from another deep learning methodology—CNN model in the form of pictures to predict the Taiwan’s future financial market trends.

There is still the following limitation in the experiments of this study. The simulated trading set in this study is sort of simplified without considering the actual trading situations, such as the transaction fee, the change of position and the emergency conditions, leading to the final trading results which might be slightly changed. This study provides a method to predict the fluctuation of the financial market. In the future, other scholars may take into account the conditions which will be encounter during the actual investment behavior and develop a real-time trading system.

Acknowledgments

This work was supported by the Ministry of Science and Technology, Taiwan, under Grant MOST 109-2410-H-006-116-MY2, Grant MOST110-2511-H-006-013-MY3, and Grant MOST 109-2410-H-025-016-MY2; and in part by the Higher Education Sprout Project, Ministry of Education to the Headquarters of University Advancement at National Cheng Kung University (NCKU).

Conflict of interest

The authors declare no conflict of interest.