Prediction of Shanghai Stock Index Based on Investor Sentiment and CNN-LSTM Model

Yi SUN; Qingsong SUN; Shan ZHU

doi:10.21078/JSSI-2022-620-13

PDF(3369 KB)

Journal of Systems Science and Information ›› 2022, Vol. 10 ›› Issue (6) : 620-632. DOI: 10.21078/JSSI-2022-620-13

Prediction of Shanghai Stock Index Based on Investor Sentiment and CNN-LSTM Model

Author information +

History +

Abstract

In view of the breakthrough progress of the depth learning method in image and other fields, this paper attempts to introduce the depth learning method into stock price forecasting to provide investors with reasonable investment suggestions. This paper proposes a stock prediction hybrid model named ISI-CNN-LSTM considering investor sentiment based on the combination of long short-term memory (LSTM) and convolutional neural network (CNN). The model adopts an end-to-end network structure, using LSTM to extract the temporal features in the data and CNN to mine the deep features in the data can effectively improve the prediction ability of the model by increasing investor sentiment in the network structure. The empirical part makes a comparative experimental analysis based on Shanghai stock index in China. By comparing the experimental prediction results and evaluation indicators, it verifies the prediction effectiveness and feasibility of ISI-CNN-LSTM network model.

Key words

convolution neural network / long short-term memory / investor sentiment / stock price forecasting

Cite this article

EndNote

Ris (Procite)

Bibtex

Download Citations

Yi SUN , Qingsong SUN , Shan ZHU. Prediction of Shanghai Stock Index Based on Investor Sentiment and CNN-LSTM Model. Journal of Systems Science and Information, 2022, 10(6): 620-632 https://doi.org/10.21078/JSSI-2022-620-13

1 Introduction

With the rapid development of computer science and financial technology, predicting the changing trend of stock prices has become the main exploration and research direction of many industry scholars. The efficient market hypothesis has been regarded as one of the cornerstones of modern financial economics theory^[1], and the emergence of large number of financial anomalies means that traditional financial theories have certain limitations^[2-4]. Different from the research of traditional financial theory, behavioral finance focuses on analyzing various behavior deviations of investors in the securities market during the transaction process. It also emphasizes the important role of investors' emotional factors and their trading behaviors in the analysis and forecasting of securities returns. As Baker and Wurgler said, "The question today is not whether investor sentiment can affect stock returns, but to what extent investor sentiment affects stock returns^[5-7]!" Large number of subsequent studies have indeed proved this point. Compared with Europe and the United States' modern stock markets, China's stock market has developed in a short period of time, and is dominated by retail investors, supplemented by institutional investors. Many analysis results show that China's stock market has not reached the weakly effective market pattern of European and American stock markets. At the same time, the research results of many experts and scholars show that investor sentiment has had a greater impact on the price changes in China's stock market.

Therefore, we attempts to apply investor sentiment to the Shanghai Stock Index forecasting research, and apply the popular deep learning model and neural network prediction model in recent years, especially the recurrent neural network model to predict the Shanghai Stock Index closing price.

The rest of the paper is as follows: The second part mainly includes the construction and methods of investor sentiment and its impact on the price of stocks and other assets, and the introduction of time series forecasting related models (traditional and modern). The third part describes in detail the method ISI-CNN-LSTM proposed in this article. The fourth part is the empirical results and analysis of the model and data. The fifth part is the conclusion, summary on the train of thought and the achievement.

2 Literature Review

De Long determines the role of investor sentiment in the financial market: If ignorant noise traders make investment decisions based on investor sentiment, risk-averse arbitrageurs will encounter arbitrage restrictions^[8]. Changes of investor sentiment can lead to more noisy transactions, greater mispricing, and excessive volatility. In Baker's opinion, the problem we encountered is no longer about whether investor sentiment will affect stock prices, but how to accurately measure changes in investor sentiment and quantify its impact on the stock market^[5-7].

By combing the literature directly and indirectly, there are six ways to build investor sentiment as follows^{[9, 10]}:

1)

Directly use market transaction data to construct^{[11, 12]};

2)

Use principal component analysis to obtain common factors of indicators^[13-15];

3)

Use relevant survey and research data^[16-18];

4)

Extract keywords that contain investor sentiment in the content of media reports^[19-21];

5)

Use weather, important game results, changes in the political environment, etc. The construction of events that affect public sentiment^[22-24];

6)

Construction of theoretical models related to investor sentiment, etc.^[25-27]. Investor sentiment index (ISI) is in line with the purpose of stock market return forecasting^[28].

Classic models of stock market prediction include ARIMA^[29-31], GARCH^[32-34], VAR^[35] and Markov chain^[36-38]. In the past few years, with the widespread application of artificial intelligence technology, many investors apply artificial intelligence technology and machine learning methods to stock price prediction. Mancuso combines various previous time series forecasting models and forecasting methods based on machine learning technology, and make full use of the advantages of various models on this basis to improve the accuracy of forecast results as much as possible^[39]. Zhang believes that the use of machine learning to make predictions can obtain higher accuracy of prediction results than historical simulation methods in the past^[40]. At the same time, machine learning methods greatly improve the safety of uncertainty analysis results. Kim proposed that in order to improve the forecasting effect of time series models, various time series forecasting models based on statistical methods can be improved, and various forecasting methods can be compared experimentally^[41]. In order to select representative time series data to test various forecasting methods, Routledge applied the key technology of time series forecasting based on machine learning to the business intelligence model^[42]. It can learn from historical data of asset price time series and predict the value of asset price in the future, and formulate trading strategies to provide investors with trading decision recommendations. Similarly, Chen believes that machine learning methods are the research foundation of artificial intelligence technology^[43]. And also, this method has obvious advantages in carrying out complex time series analysis. Therefore, the research of machine learning has a good role in promoting time series analysis. Dai uses text mining technology and sentiment analysis methods to generate positive and negative investor sentiment time series data, and uses machine learning methods and neural network prediction models to verify stock market price changes^[44]. Like this, it is convenient for the government, listed companies and investors to make investment decisions.

Further, some scholars have begun to try to combine investor sentiment with machine learning for forecasting research. They believe that they can improve the judgment and decision-making on complex financial markets by combining traditional finance theories with behavioral finance theories. Chu believes that the application of investor sentiment index can improve the overall level of stock price prediction, especially the combination of investor sentiment index and time series forecasting model^[45]. Since the time series prediction model can describe the linear relationship of the stock price time series, further improve the prediction accuracy of the time series model, we will introduce the investor sentiment index. In addition to the basic value of stocks, investor sentiment that reflects investor expectations will also have an important impact on stock prices. Based on the time series model combined with investor sentiment index to predict the stock price, we found that investor sentiment will affect the fluctuation of stock price to a large extent^[46]. Hu comprehensively uses big data analysis, text mining technology and traditional time series models to evaluate changes in investor sentiment by constructing an effective indicator system^[47]. And considering the influence of domestic macroeconomic factors, analyze and study the influence mechanism of investor sentiment on the price fluctuation of Chinese stock market.

Other authors have done similar research, Li^[48] uses a time series model to test investor sentiment and stock returns, and analyze whether there is a two-way influence mechanism between investor sentiment and stock price changes. The results show that there is a clear correlation between investor sentiment and stock price fluctuations. Xiong^[49] believes that the use of time series models to study the short-term impact of investor sentiment on stock price volatility can supplement long-term research on the impact of investor sentiment on stock market returns to a certain extent. Sun^[50] believes that investor sentiment is an important factor in the price of securities. By combining the traditional time series model with the investor sentiment index, it can be concluded through empirical analysis that the relationship between investor sentiment and stock price fluctuations is interactive.

If the time series model does not include the investor sentiment index for forecasting, the accuracy of the stock price forecast will decrease. After adding investor sentiment index to the time series model, the model can capture abnormal stock price changes, indicating that investor sentiment will have a certain impact on stock price changes^[51]. Jin believes that the investor sentiment index based on neural network algorithm has obvious advantages in explaining the changes in the Shanghai Stock Exchange Index^[52]. Investor sentiment has an asymmetrical return distribution to the different changing trends of the stock market, and is significantly predictable to the extreme returns of the downward trend. Yi believes that the sentiments of both long and short investors have an asymmetric effect on the trading volume and yield of the trading market^[53]. Investor short-side sentiment has an obvious impact on market turnover than long-side sentiment. The combination of stock technical analysis practice and academic research on investor sentiment can provide a quantitative standard for investor sentiment that is applicable to both the stock market and individual stocks. Meanwhile by applying the investor sentiment index to the time series model, the model can be applied to stock price forecasting^[54].

In the last few years, deep learning method has become popular in various fields, the method has been widely used in predictive models. It can make up for some shortcomings in traditional machine learning and capture highly nonlinear relationships. Especially, the convolution neural network and Long Short-Term Memory (CNN-LSTM) model^{[55, 56]}, this model is very suitable for stock price prediction because of its good selectivity, memory and other characteristics. Hence, the paper has proposed to combine investor sentiment with CNN-LSTM model to conduct an empirical analysis of Shanghai Stock index prices in China.

3 Methodology

Shanghai stock index prediction is a comprehensive index based on subjective experience judgment and comprehensive information. Its prediction research involves multivariable time series. Therefore, this paper constructs a series CNN-LSTM model for prediction, taking into account irregular time-varying information, complex nonlinearity and multi-dimensional data spatial feature extraction and analysis.

3.1 CNN Model

The neural network based on convolution layer is the CNN. With its strong feature extraction and recognition ability, CNN first achieved excellent performance in the fields of computer vision and natural language processing. Generally, one-dimensional convolution is mainly used for temporal data processing^[56-58], two-dimensional convolution is used for image spatial convolution, and three-dimensional convolution is mainly used for three-dimensional spatial convolution. In this paper, the convolution kernel in CNN is a one-dimensional structure. Multivariate time series are input through different channels to retain information to the greatest extent. CNN is mainly composed of three modules: CNN is a feedforward neural network. Its basic structure is composed of input layer, convolutional layer, pooling layer, fully connected layer (FC) and output layer, as shown in Figure 1.

Figure 1 Structure of one-dimensional convolutional neural network

Full size|PPT slide

Input of CNN is composed of

N

eigenvectors in m consecutive time steps. Let

x_{t}

represent the eigenvector of the

t

-th time step, then the sequence form constructed in this section is:

X_{t} = [x_{(t - M + 1)}, x_{(t - M + 2)}, \dots, x_{t}]^{T}

That is

X_{1} = [x_{1}, x_{2}, \dots, x_{M}]^{T}

X_{2} = [x_{2}, \dots, x_{(M + 1)}]^{T}

, other sequences can be deduced in turn.

In the convolution layer, the convolution kernel can be regarded as a sliding window on the time series, and the short-time features between the sequences are extracted.

Assuming that

y_{j}^{l}

is the

j

-th Feature Map of the convolution layer, the operation formula is:

\begin{array}{rcl} y_{j}^{l} = f (w_{i j}^{(l)} \otimes \sum_{i = 1}^{k} y_{i}^{l - 1} + b^{l}), \end{array}

(1)

where

w_{i j}^{(l)}

represents the weight matrix connecting layer

l

-1 and layer

l

convolution layer,

b^{l}

is the offset value corresponding to layer

l

, and the weight is shared in all areas of the same layer.

f (\cdot)

is the activation function, and the operator

\otimes

represents convolution.

y_{i}^{l - 1}

represents the

i

th feature map of layer

l - 1

After each convolution layer is connected to the pooling layer, the dimension of the feature surface output by the convolution layer is further reduced without increasing the training parameters, so as to reduce the network parameters and improve the robustness of the model. In this paper, the commonly used maximum pooling is selected for operation, that is, the maximum values of elements in the pooling domain are extracted to form a condensed feature map.

After repeated convolution and pooling, the full connection layer concatenates all compressed feature surfaces output by the last pooling layer into feature vectors and inputs them into the full connection layer. Each neuron in the full connection layer is fully connected with the neurons in the previous layer, which can integrate the features extracted from the convolution layer and pooling layer, so as to obtain more distinguishing features.

3.2 LSTM Model

LSTM is a special recurrent neural network (RNN), which improves the memory module in the traditional RNN model, that is, a single tanh layer or sigmoid layer, and avoids the problem that effective historical information cannot be saved for a long time due to the influence of continuous data input^{[59, 60]}. Through the design of gate structure and memory unit state, LSTM can effectively update and transmit the key information in time series.

The basic unit of LSTM model is memory module, which includes memory unit and three gate structures controlling the state of memory unit, namely forgetting gate, input gate and output gate. The forgetting gate determines to forget useless historical information from the memory unit state, the input gate determines the influence of the current input data on the memory unit state, and the output gate determines the output information. The structure of LSTM module is shown in Figure 2.

Figure 2 Structure diagram of LSTM module

Full size|PPT slide

Assuming that

x_{t}

represents the input vector at time

t

h_{t - 1}

represents the output at time

t

-1, and

W_{f}, W_{i}, W_{c}, W_{o}, U_{F}, U_{I}, U_{C}, U_{O}

and

b_{f}, b_{i}, b_{c}

and

b_{o}

represent the weight matrix and offset vector respectively, the process of state update and information output by the memory module is as follows.

First, The forgetting gate forget the useless historical information:

f_{t} = σ (W_{f} ∙ x_{t} + U_{f} ∙ h_{t - 1} + b_{f}) .

(2)

Then, the status of the input gate is updated according to the input data and historical information:

{\begin{cases} i_{t} = σ (W_{i} ∙ x_{i} + U_{i} ∙ h_{t - 1} + b_{i}), \\ {\tilde{C}}_{t} = \tanh (W_{c} ∙ x_{i} + U_{c} ∙ h_{t - 1} + b_{c}), \\ C_{t} = f_{t} C_{t - 1} + i_{t} ∙ {\tilde{C}}_{t} . \end{cases}

(3)

Finally, the output gate outputs the information of the current time:

{\begin{cases} O_{t} = σ (W_{o} ∙ x_{i} + U_{o} ∙ h_{t - 1} + b_{o}), \\ h_{t} = O_{t} ∙ \tanh (C_{t}), \end{cases}

(4)

where

σ

is the logistic sigmoid function

f_{t}

i_{t}

and

o_{t}

respectively represent the output state of the forgetting gate, input gate and output gate at time

t

, and

c_{t}

epresents the memory unit state at time

t

3.3 CNN-LSTM Model

The overall structure of CNN-LSTM model for predicting Shanghai stock index in this paper is shown in Figure 3.

Figure 3 CNN-LSTM model structure of Shanghai Stock Index Prediction

Full size|PPT slide

The model is mainly composed of CNN and LSTM in series. Among them, the data spatial features of multivariable time series are extracted through CNN convolution and pooling layer, and transferred to LSTM layer to remove noise at the same time. The LSTM layer models the irregular time information transmitted by the CNN layer. Finally, CNN-LSTM generates prediction data after fully connecting the layer, and then the prediction model evaluates and analyzes it through error measurement.

Specifically, CNN-LSTM model can extract complex features from confidence index and its multiple influence variables and conduct predictive analysis^[61]. Firstly, the starting layer of the model is composed of CNN, which accepts a variety of variable inputs affecting the confidence index, such as relevant time factors and economic factors. After the variables enter the input layer, CNN extracts data features through the hidden layer and outputs them. In this process, there may be multiple hidden layers. The more layers, the deeper the network depth, and the final output is used as the input of LSTM. The hidden layer consists of convolution layer, relu layer (activation function) and pooling layer. Realization of multivariate time series variables by convolution layer.

Convolution operation and output the results to the next layer. Convolution operation simulates the response of a single neuron to visual stimulation, that is, when processing time series, each neuron processes the received relevant sequence data, and extracts and expresses the data features. Convolution operation can reduce the number of parameters and make CNN-LSTM network deeper.

3.4 Model Prediction and Evaluation Method

In this paper, the prediction and evaluation methods of CNN-LSTM model mainly use the commonly used error mean square MSE, average absolute percentage error MAPE and goodness of fit

R^{2}

to evaluate and analyze the prediction.

Suppose

Y_{i}

is the observed value of the closing price of Shanghai stock index,

\tilde{Y_{i}}

is the predicted value obtained by the model, and the mean value of

\tilde{Y_{i}}

\bar{Y}

, then:

M S E = \frac{1}{n} \sum_{i = 1}^{n} {(Y_{i} - \tilde{Y_{i}})}^{2},

(5)

\begin{array}{rcl} MAPE = \frac{1}{n} \sum_{i = 1}^{n} | \frac{Y_{i} - {\tilde{Y}}_{i}}{Y_{i}} | \times 100 %, \end{array}

(6)

\begin{array}{rcl} R^{2} = \sum_{i = 1}^{n} {({\tilde{Y}}_{i} - \bar{Y})}^{2} / \sum_{i = 1}^{n} {(Y_{i} - \bar{Y})}^{2} . \end{array}

(7)

The smaller the of MSE and MAPE, the smaller the prediction error. The higher the value of

R^{2}

, the better the fitting degree.

4 Empirical Results

4.1 Experimental Process

The experiment is mainly composed of data downloading, data processing, model training, and fine-tuning parameters. The specific flow chart is shown in Figure 4.

Figure 4 Experimental flowchart

Full size|PPT slide

4.2 Data

The experimental data in this article is the 302 trading days of the Shanghai Stock Exchange Index from January 12, 2019 to March 31, 2020, downloaded using the Tushare finance interface package. Tushare is a free and open source python financial data interface package. The data set includes: Transaction date, the highest price of the day, the lowest price of the day, the opening price of the day, the closing price of the day, the fluctuation limit, the fluctuation range, the trading volume, the trading limit and other time series data. We use the Chinese Investor Sentiment Index (CISI) compiled by the National Development Research Institute of Peking University as the data source for the investor sentiment index in this article (The URL is https://www.nsd.pku.edu.cn/zsfb/zgtzzqxzs/250262.htm).

4.3 Data Preprocessing

First, sort the downloaded dataset by date to obtain a complete dataset without disorder. Secondly, these data are standardized by

z

-score to eliminate the impact of different magnitudes of data. Then, take the closing price of the second day as the tag value, the first 200 days in the dataset as the training set data, and the last 52 days as the test set data.

4.4 Parameter Setting

In the training of the CNN-LSTM stock prediction model, we set the number of iterations to 50 times, and set the time window span to 1–5 days, so as to select time windows of different spans for comparison experiments. The experimental comparison results are shown in Figures 5

\sim

In this paper, the average absolute error and the goodness of fit between the predicted data and the real data (

R^{2}

) are selected as the evaluation indicators of the model performance. By using different time windows for comparative testing, we compared the performance of the CNN-LSTM stock prediction model under different parameter settings. The results are shown in Table 1.

Table 1 Parameter selection

Time window	Mean absolute error	Mean square error	R²
2	0.13766	0.02986	0.40335
3	0.13243	0.03268	0.4155
4	0.13956	0.03691	0.31617
5	0.14207	0.03392	0.34479

The experimental results show that when the time window of the model is set to 2, the prediction accuracy of the model is greatly improved. This is because when the value of the time window increases, the predictive performance of the model will also increase as the information learned by the model increases. When the time window is 4 or 5, the performance of the model is degraded. This is because when the time window is too large, it will lead to a decline in the value of information and make the information mediocre, thus ignoring the small-scale impact of short-term information on the stock market.

We integrated the average absolute error and the changes in the goodness of fit

R^{2}

, and finally set the time window to 3. The data of the 8 attributes of the previous 3 days are used as the input layer of the neural unit. The closing prices on the 4th and the second day are used as the training label values.

4.5 Experimental Results

Taking Chinese Investor Sentiment Index (CISI) as a new input variable, using the CNN-LSTM model under optimal parameters to predict the upper positive real closing price, and comparing it with SVM and LSTM, the results are shown in Figures 9

\sim

12.

Figure 9 SVM prediction results

Full size|PPT slide

Figure 10 LSTM prediction results

Full size|PPT slide

Figure 11 CNN-LSTM prediction results

Full size|PPT slide

Figure 12 ISI-CNN-LSTM prediction results

Full size|PPT slide

The model performances of different models on the test set are shown in Table 2. The average absolute error of the ISI-CNN-LSTM model is 0.1317, the mean square error is 0.0289, and the goodness of fit between the predicted value and the true value is

R^{2}

of 0.4932. It shows that the CNN-LSTM model that takes into account investor sentiment is the most ideal.

Table 2 Model comparison

Model name	Mean absolute error	Mean square error	R²
SVM	22.88	2912	-6.3184
LSTM	0.1392	0.0309	-0.1005
CNN-LSTM	0.1324	0.0327	0.4155
ISI-CNN-LSTM	0.1317	0.0289	0.4932

5 Conclusion

Aiming at the prediction of Shanghai stock index, this paper makes full use of the outstanding advantages of CNN depth network in multivariate data feature extraction and LSTM network in sequence data analysis, constructs the series CNN-LSTM depth network learning model formed by the direct integration of CNN and LSTM, and uses the model to predict and analyze Shanghai Stock Index. Make full use of multivariate data information and realize data continuous feature input according to time sliding window, give full play to the feature extraction advantages of CNN model, and integrate the feature vector extracted directly into LSTM model as input, which better realizes the requirements of feature extraction, timing and complex nonlinear prediction and analysis of confidence index data.

The python language and testorflow framework are used to implement the model, and the Shanghai stock index data are used to predict the closing price. Through the comparative experiment with SVM, LSTM and LSTM-CNN, it can be seen that the prediction accuracy of CNN-LSTM model considering investor sentiment has been greatly improved. It shows that the prediction of stock information by this model is feasible and effective, and brings more valuable reference to investors' quantitative investment decision-making.

References

Publishing order | Descend order by publishing year | Descend order by cited within

1	Fama E F. Efficient capital markets: A review of theory and empirical work. The Journal of Finance, 1970, 25 (2): 383- 417. Cited in this article [1]

2	Bansal R, Yaron A. Risks for the long run: A potential resolution of asset pricing puzzles. The Journal of Finance, 2004, 59 (4): 1481- 1509. https://doi.org/10.1111/j.1540-6261.2004.00670.x Cited in this article [1]

3	Campbell J Y, Cochrane J H. By force of habit: A consumption-based explanation of aggregate stock market behavior. Journal of Political Economy, 1999, 107 (2): 205- 251. https://doi.org/10.1086/250059

4	Peng L, Xiong W. Investor attention, overconfidence and category learning. Journal of Financial Economics, 2006, 80 (3): 563- 602. https://doi.org/10.1016/j.jfineco.2005.05.003 Cited in this article [1]

5	Baker M, Wurgler J. Investor sentiment and the crosssection of stock returns. Journal of Financial Economics, 2006, 80 (3): 563- 602. https://doi.org/10.1016/j.jfineco.2005.05.003 Cited in this article [2]

6	Baker M, Wurgler J. Investor sentiment in the stock market. Journal of Economic Perspectives, 2007, 21 (2): 129- 151. https://doi.org/10.1257/jep.21.2.129

7	Baker M, Wurgler J, Yuan Y. Global, local, and contagious investor sentiment. Journal of Financial Economics, 2012, 104 (2): 272- 287. https://doi.org/10.1016/j.jfineco.2011.11.002 Cited in this article [2]

8	De Long J B, Shleifer A, Summers L H, et al. Noise trader risk in financial markets. Journal of Political Economy, 1990, 98 (4): 703- 738. https://doi.org/10.1086/261703 Cited in this article [1]

9	Gao Y, Xiang X, Feng X, et al. Investor sentiment and financial market. Journal of Tianjin University, 2018, 20 (6): 481- 494. Cited in this article [1]

10	Simon D P, Wiggins R A. S & P futures returns and contrary sentiment indicators. Journal of Futures Markets, 2001, 21 (5): 447- 462. https://doi.org/10.1002/fut.4 Cited in this article [1]

11	Cornelli F, Goldreich D, Ljungqvist A. Investor sentiment and preIPO markets. The Journal of Finance, 2006, 61 (3): 1187- 1216. https://doi.org/10.1111/j.1540-6261.2006.00870.x Cited in this article [1]

12	Neupane S, Paudyal K, Thapa C. Firm quality or market sentiment: What matters more for IPO investors?. Journal of Banking & Finance, 2014, 44, 207- 218. Cited in this article [1]

13	Sibley S E, Wang Y, Xing Y, et al. The information content of the sentiment index. Journal of Banking & Finance, 2016, 62, 164- 179. Cited in this article [1]

14	Yang L Z, Goh J, Chiyachantana C. Valuation uncertainty, market sentiment and the informativeness of institutional trades. Journal of Banking & Finance, 2016, 72, 81- 98.

15	Chung S L, Hung C H, Yeh C Y. When does investor sentiment predict stock returns?. Journal of Empirical Finance, 2012, 19 (2): 217- 240. https://doi.org/10.1016/j.jempfin.2012.01.002 Cited in this article [1]

16	Schmeling M. Investor sentiment and stock returns: Some international evidence. Journal of Empirical Finance, 2009, 16 (3): 394- 408. https://doi.org/10.1016/j.jempfin.2009.01.002 Cited in this article [1]

17	Chiu C J, Harris R D F, Stoja E, et al. Financial market volatility, macroeconomic fundamentals and investor sentiment. Journal of Banking & Finance, 2018, 92, 130- 145.

18	Lee W Y, Jiang C X, Indro D C. Stock market volatility, excess returns, and the role of investor sentiment. Journal of Banking & Finance, 2002, 26 (12): 2277- 2299. Cited in this article [1]

19	Smales L A. News sentiment and bank credit risk. Journal of Empirical Finance, 2016, 38, 37- 61. https://doi.org/10.1016/j.jempfin.2016.05.002 Cited in this article [1]

20	Siganos A, Vagenas-Nanos E, Verwijmeren P. Divergence of sentiment and stock market trading. Journal of Banking & Finance, 2017, 78, 130- 141.

21	Renault T. Intraday online investor sentiment and return patterns in the US stock market. Journal of Banking & Finance, 2017, 84, 25- 40. Cited in this article [1]

22	Chang S C, Chen S S, Chou R K, et al. Local sports sentiment and returns of locally headquartered stocks: A firm-level analysis. Journal of Empirical Finance, 2012, 19 (3): 309- 318. https://doi.org/10.1016/j.jempfin.2011.12.005 Cited in this article [1]

23	Addoum J M, Kumar A. Political sentiment and predictable returns. The Review of Financial Studies, 2016, 29 (12): 3471- 3518. https://doi.org/10.1093/rfs/hhw066

24	Hwang B H. Country-specific sentiment and security prices. Journal of Financial Economics, 2011, 100 (2): 382- 401. https://doi.org/10.1016/j.jfineco.2010.10.020 Cited in this article [1]

25	Benhabib J, Liu X, Wang P. Sentiments, financial markets, and macroeconomic fluctuations. Journal of Financial Economics, 2016, 120 (2): 420- 443. https://doi.org/10.1016/j.jfineco.2016.01.008 Cited in this article [1]

26	Bethke S, Gehde-Trapp M, Kempf A. Investor sentiment, flight-to-quality, and corporate bond comovement. Journal of Banking & Finance, 2017, 82, 112- 132.

27	Zheng Y, Osmer E, Zhang R. Sentiment hedging: How hedge funds adjust their exposure to market sentiment. Journal of Banking & Finance, 2018, 88, 147- 160. Cited in this article [1]

28	Ruan Q, Wang Z, Zhou Y, et al. A new investor sentiment indicator (ISI) based on artificial intelligence: A powerful return predictor in China. Economic Modelling, 2020, 88, 47- 58. https://doi.org/10.1016/j.econmod.2019.09.009 Cited in this article [1]

29	Li Z, Liu S. Prediction comparison based on ARIMA model, grey model and regression model. Statistics & Decision, 2019, 35 (23): 38- 41. Cited in this article [1]

30	Pektaş A O, Cigizoglu H K. ANN hybrid model versus ARIMA and ARIMAX models of runoff coefficient. Journal of Hydrology, 2013, 500 (1): 21- 36.

31	Zhao C B, Mao C M. Forecast of Chinese carbon emission intensity based on the combined model of ARIMA and BP neural network. Resources and Environment in the Yangtze Basin, 2012, 21 (6): 665- 671. Cited in this article [1]

32	Zhou A M, Han F. Research on risk spillover effect between stock market and foreign exchange market: Based on GARCH-time-varying Copula-CoVaR model. International Finance Research, 2017, 11, 54- 64. Cited in this article [1]

33	Zheng Z L, Yang W. The dynamic correlation of financial asset returns: An empirical study based on the DCC multivariate GARCH model. Contemporary Finance, 2012, 7, 41- 49.

34	Wang S S, Wang J B, Xu T T, et al. Research and prediction of intraday volatility of Shanghai and Shenzhen 300 stock index futures based on ARMA-GARCH-SN model. Operations Research and Management, 2018, 27 (4): 153- 161. Cited in this article [1]

35	Jiang T. Measurement and analysis of security market risk based on GARCH and semiparametric VAR model: Empirical evidence from Chinese Shanghai stock market. Financial Research, 2010, 6, 103- 111. Cited in this article [1]

36	Zou Y, Yang D L, Pan Y C. An improved grey Markov chain model with ANN error correction and its application in gross domestic product forecasting. Journal of Intelligent & Fuzzy Systems, 2021, 40 (6): 1- 11. Cited in this article [1]

37	Amanatidis G, Kleer P. Rapid mixing of the switch Markov chain for strongly stable degree sequences. Random Structures & Algorithms, 2020, 57 (3): 1- 21.

38	Zhao Z J, Chu L Y, Tao D C, et al. Classification with label noise: A Markov chain sampling framework. Data Mining and Knowledge Discovery, 2019, 33 (5): 1- 37. Cited in this article [1]

39	Mancuso P, Piccialli V, Sudoso A M. A machine learning approach for forecasting hierarchical time series. Expert Systems with Applications, 2021, 182, 1- 17. Cited in this article [1]

40	Zhang R, Jia H. Production prediction method of water drive reservoir based on multivariate time series and vector autoregressive machine learning model. Petroleum Exploration and Development, 2021, 48 (1): 175- 184. Cited in this article [1]

41	Kim J, Yang M, Yoon S. The long-term agricultural weather forcast methods using machine learning and GloSea5:On the cultivation zone of Chinese cabbage. Journal of Digital Convergence, 2020, 18 (4): 1- 8. Cited in this article [1]

42	Routledge B R. Machine learning and asset allocation. Financial Management, 2019, 48 (4): 1- 26. Cited in this article [1]

43	Chen C, Li R, Shu L, et al. Predicting future dynamics from short-term time series using an anticipated learning machine. National Science Review, 2020, 7 (6): 1079- 1091. https://doi.org/10.1093/nsr/nwaa025 Cited in this article [1]

44	Dai D B, Lan Y S, Fan T J, et al. Research on stock index prediction and decision-making based on text mining and machine learning. China Soft Science, 2019, 4 (4): 166- 175. Cited in this article [1]

45	Chu X, Wu C, Qiu J. A nonlinear Granger causality test between stock returns and investor sentiment for Chinese stock market: A wavelet-based approach. Applied Economics, 2016, 48 (21): 1- 10. Cited in this article [1]

46	Jing N, Wu Z, Wang H F. A hybrid model integrating deep learning with investor sentiment analysis for stock price prediction. Expert Systems with Applications, 2021, 178, 1- 12. Cited in this article [1]

47	Hu J S, Sui Y Y, Ma F. The measurement method of investor sentiment and its relationship with stock market. Computational Intelligence and Neuroscience, 2021, 3, 1- 11. Cited in this article [1]

48	Li Y, Li W P. Firm-specific investor sentiment for the Chinese stock market. Economic Modelling, 2021, 97, 1- 16. https://doi.org/10.1016/j.econmod.2021.01.012 Cited in this article [1]

49	Xiong X, Liu M H, Chen S N. Dynamic characterization of the impact of investor sentiment and market interest rates on stock market index returns. Operations Research and Management, 2018, 27 (9): 156- 169. Cited in this article [1]

50	Sun L Y, Zhang J L. Research on the correlation between investor sentiment and stock market returns -Based on the multifractal analysis method. Financial Review, 2017, 9 (5): 92- 102. Cited in this article [1]

51	Meng W D, Zhang M Y, Lu J. Research on conditional asset pricing of AH shares based on investor sentiment. Journal of Chongqing University, 2016, 22 (2): 80- 89. Cited in this article [1]

52	Jin X, Jiang S W, Yuan Y. Research on the predictability of investor sentiment and extreme returns based on stock market information. Management Review, 2018, 30 (7): 16- 25. Cited in this article [1]

53	Yi H B, Lai J J, Dong D Y. The impact of different investor sentiments in online forums on the trading market: An empirical analysis based on the VAR model. Finance and Economics, 2015, 4 (1): 46- 54. Cited in this article [1]

54	Xiang C, Lu J. Research on the effectiveness of investor sentiment index based on technical analysis index. Management Science, 2018, 31 (1): 129- 148. Cited in this article [1]

55	Huang T T, Yu L. Application of SDAE-LSTM model in financial time series forecasting. Computer Engineering and Application, 2019, 55 (1): 142- 148. Cited in this article [1]

56	Yao H G, Mu N G. The prediction of financial time series by EMD-LSTM model. Computer Engineering and Applications, 2021, 57 (5): 239- 244. Cited in this article [2]

57	Zhao B, Lu H, Chen S, et al. Convolutional neural networks for time series classification. Journal of Systems Engineering and Electronics, 2017, 28 (1): 162- 169.

58	Jing N, Shi Z J, Shu Y M. Forecasting high frequency price of Shanghai copper futures based on attention mechanism and CNN-LSTM. Chinese Journal of Management Science, 2020, 1- 13. Cited in this article [1]

59	Cornia M, Baraldi L, Serra G, et al. Predicting human eye fixations via an LSTM-based saliency attentive model. IEEE Transactions on Image Processing, 2018, 27 (10): 5142- 5154. Cited in this article [1]

60	Geng J J, Liu Y M, Li Y, et al. Stock index prediction model based on CNN-LSTM. Statistics & Decision, 2021, (5): 134- 138. Cited in this article [1]

61	Xia M L, Jiang L L. Chinese consumer confidence index forecast based on deep network CNN-LSTM model. Statistics & Decision, 2021, (7): 21- 26. Cited in this article [1]

PDF(3369 KB)

722

Accesses

Citation

Detail

Sections

Recommended

Received	Accepted	Published
2022-02-10	2022-12-09	2022-12-25
Issue Date
2022-12-30

Please choose a citation manager

Content to export

Abstract

Key words

Cite this article

1 Introduction

2 Literature Review

3 Methodology

3.1 CNN Model

Figure 1 Structure of one-dimensional convolutional neural network

3.2 LSTM Model

Figure 2 Structure diagram of LSTM module

3.3 CNN-LSTM Model

Figure 3 CNN-LSTM model structure of Shanghai Stock Index Prediction

3.4 Model Prediction and Evaluation Method

4 Empirical Results

4.1 Experimental Process

Figure 4 Experimental flowchart

4.2 Data

4.3 Data Preprocessing

4.4 Parameter Setting

Figure 5 Time window $t = 2$

Figure 6 Time window $t = 3$

Figure 7 Time window $t = 4$

Figure 8 Time window $t = 5$

Table 1 Parameter selection

4.5 Experimental Results

Figure 9 SVM prediction results

Figure 10 LSTM prediction results

Figure 11 CNN-LSTM prediction results

Figure 12 ISI-CNN-LSTM prediction results

Table 2 Model comparison

5 Conclusion

{{custom_sec.title}}

{{custom_sec.title}}

References

{{custom_fnGroup.title_en}}

Footnotes

Share

模态框（Modal）标题

Please choose a citation manager

Content to export

Abstract

Key words

Cite this article

1 Introduction

2 Literature Review

3 Methodology

3.1 CNN Model

Figure 1 Structure of one-dimensional convolutional neural network

3.2 LSTM Model

Figure 2 Structure diagram of LSTM module

3.3 CNN-LSTM Model

Figure 3 CNN-LSTM model structure of Shanghai Stock Index Prediction

3.4 Model Prediction and Evaluation Method

4 Empirical Results

4.1 Experimental Process

Figure 4 Experimental flowchart

4.2 Data

4.3 Data Preprocessing

4.4 Parameter Setting

Figure 5 Time window t=2

Figure 6 Time window t=3

Figure 7 Time window t=4

Figure 8 Time window t=5

Table 1 Parameter selection

4.5 Experimental Results

Figure 9 SVM prediction results

Figure 10 LSTM prediction results

Figure 11 CNN-LSTM prediction results

Figure 12 ISI-CNN-LSTM prediction results

Table 2 Model comparison

5 Conclusion

{{custom_sec.title}}

{{custom_sec.title}}

References

{{custom_fnGroup.title_en}}

Footnotes

Figure 5 Time window $t = 2$

Figure 6 Time window $t = 3$

Figure 7 Time window $t = 4$

Figure 8 Time window $t = 5$