The Prediction Method of Stock Index Based on Sentiment Analysis

of Investors' Comments

Jiaming Zhao

1,* a

and Ke Wen

School of Management and Economics, Beijing Institute of Technology, Liangxiang East Road, Beijing, China

School of Management and Economics, Beijing Institute of Technology, Beijing, China

Keywords: Sentiment Analysis, Natural Language Processing, Deep Learning, Stock Index Prediction.

Abstract: The research content of this paper is to obtain the sentiment of retail investors through text analysis of their

comments, and apply it into stock price prediction. In this paper, a retail investor sentiment analysis model

Fin-Bert-TextCNN is proposed to construct the sentiment index, exploring the relationship between retail

investor sentiment and the stock market, and predict the rise and fall of A shares. This paper finds that Fin-

Bert-TextCNN model has higher prediction accuracy and is more suitable for the analysis of financial

commentary texts. The deep learning technology used in this paper can well depict the nonlinear relationship

between investor sentiment and stock index, which is helpful to prevent and resolve major financial risks.

1 INTRODUCTION

In modern society, with the development of enterprise

financing, the stock exchange market has injected

extraordinary strength into social development.

However, in the gradual development of the stock

market, many problems have also been exposed.

Among them, the financial security problems caused

by the excessive sentiments of retail investors have

attracted extensive attention in the industry. In

particular, the financial crisis or the short selling of

international capital may greatly affect the confidence

of retail investors and then turn into irreversible

turbulence in the stock market. Therefore, it is very

important to study the impact of retail investor

sentiment on stock trading. In some cases, it is even

possible to prevent and defuse major financial risks

based on the sentiment of retail investors.

In principle, investor sentiment has a great impact

on stock trading. The stock price not only depends on

the value of the enterprise, but also related to the

investor sentiment to some extent. How to predict the

behavior of investors through their sentiments has

great research value for ensuring the stability of the

stock market. Investor sentiment, especially for the

secondary stock market, has a large number of retail

https://orcid.org/0000-0003-23338-4444

https://orcid.org/0000-0003-2377-4000

investors, whose sources of information are not only

news and industry research, but also subject to a large

change in short-term stock price rise and fall.

Compared with the news text, the comments of

individual investors can better represent their

sentiments at a certain time.

However, the existing sentiment analysis model is

difficult to extract the sentiment of retail investors in

stock reviews. Most of the texts used in the pre-

training process are not financial related, and it is

difficult to analyze the flexible text of comments in

the comment area. Therefore, the core of this paper is

to mine the comment sentiment of retail investors

through model innovation and then forecast the stock

index.

2 LITERATURE REVIEW

There is a strong association between the sentiment

of retail investors and stock price movements. Cowles

is an early example of stock forecasting by analyzing

news texts Cowles (Cowles 1933). He classified Peter

Hamilton's editorial articles as "bullish", "bearish"

and "uncertain", and then used these classifications to

predict the future return of Dow Jones Industrial

478

Zhao, J. and Wen, K.

The Prediction Method of Stock Index Based on Sentiment Analysis of Investors’ Comments.

DOI: 10.5220/0012035200003620

In Proceedings of the 4th International Conference on Economic Management and Model Engineering (ICEMME 2022), pages 478-483

ISBN: 978-989-758-636-1

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

Average. Lee et al. (2002) used generalized

autoregressive conditional heteroscedasticity to

verify the significant positive correlation between

investor sentiment and stock returns. Dragos and

Laura used consumer confidence index to measure

investor sentiment and found it was positively

correlated with stock returns (Oprea&Brad 2014).

Yang et al. (2020) found that news sentiment has a

negative impact on credit default swap spreads, and

proved that there was a very close relationship

between negative news and the 2008 financial crisis.

With the development of financial markets, many

scholars have studied the models of stock price

forecasting. Time series model, used as a traditional

method, has been difficult to achieve satisfactory

results due to the strong model assumptions and

inflexible parameter settings (Xu & Tian 2021). On

the contrary, literature review shows that deep

learning can process higher dimensional data and has

stronger representation ability. It has made

outstanding achievements in different fields such as

speech recognition (Chiu et al., 2018), computer

vision (He et al., 2016) and natural language

processing (Deng and Yang, 2017). Among them,

LSTM model is suitable for modeling time series

data, and BERT model has great advantages in

financial text sentiment analysis, which can reflect

the dynamic changes of stock related companies in

real time. Therefore, this paper creatively applies

BERT model to the sentiment analysis task of

financial comments from investors in China, and

proposes a stock index prediction method based on

LSTM model.

3 SENTIMENT ANALYSIS

MODEL

This paper adopts the Fin-Bert-CNN model when

analyzing the sentiment of retail investors. The Fin-

Bert-CNN model is divided into two parts: Fin-Bert

and TextCNN. BERT is a word embedding model,

which is responsible for converting investor's

sentiment comment texts into 768 dimensional word

feature vectors. The CNN layer model will convert

comment texts into sentiment scores after pre-training

by extracting the features of word vectors.

3.1 Fin-Bert Word Embedding Model

3.1.1 Bert Word Embedding Model

With the development of machine learning, especially

deep learning, more choices have been made in the

construction of sentiment models. The first step of

these models for analyzing text of sentiments is to

convert the text into word vectors, which is called

word embedding. Common words embedding models

include Word2Vector (Mikolov et al., 2013) and

Glove (Pennington et al., 2014).

Bert (Bidirectional Encoder Representation from

Transformers) word embedding model is a language

representation model based on Transformer

mechanism launched by Google in 2018 (Devlin et

al., 2019). BERT, known as a pre-trained natural

language processing (NLP) model, uses a new MLM

natural language representation algorithm to pre-train

bidirectional transformers.

Figure 1: One way Transformer model.

Figure 2: Two way Transformer model.

Figure 3: Bert Structure.

The Prediction Method of Stock Index Based on Sentiment Analysis of Investors’ Comments

479

As shown in Figure 1 and Figure 2, Bert creatively

adopts a two-layer Transformer architecture, and its

unique bidirectional coding enables it to depict the

representation more accurately. Different from other

natural language representation models, Bert does not

need a large number of corpora for pre-training. It

only needs to add an additional output layer for fine-

tune, which can play an outstanding role in various

aspects of language processing.

Bert structure is shown in the Figure 3, where E1,

E2, ..., En is the text input of the model, T1, T2, ...,

Tn is the output vector of the model. The core of

bidirectional Transformer feature extractor included

in Bert model is self attention mechanism.

3.1.2 Fin-Bert Word Embedding Model

Based on Financial Text Pre-Training

The materials used in the traditional BERT model

pre-training include the contents of many industries.

Although it covers a wide range of fields, it has poor

ability to process the financial field corpus. Value

simplex re pre-trained BERT prototype by using

financial corpora, and made the Fin-Bert word

embedding model open-source. Compared with the

original Bert-base-Chinese model, the Fin-Bert

model is more suitable for financial texts. It can

reduce noise and improve accuracy to a certain extent

when processing word embedding for retail investor

comments.

3.2 Textcnn Convolution Neural

Network Model

Convolutional Neural Network (CNN) is a commonly

used model in Deep Learning. It includes the input

layer, hidden layer and full connection layer (output

layer). Multiple neurons in the input layer are

responsible for receiving nonlinear input information.

In this paper, it is the 768 dimensional word vector

output by the Fin-Bert model. The hidden layer is

composed of multiple neuron links between the input

layer and the output layer. In the hidden layer, the

convolution layer and pooling layer appear

repeatedly. The final full connection layer DNN is the

output layer neuron with the softmax activation

function, and the output result is the probability of the

possible result, which takes the highest probability as

the final output result.

3.3 Steps to Acquire Sentiment Index

We use Fin-Bert-TextCNN model to complete the

construction of retail investor sentiment index in the

following four steps. Step 1: Fin-Bert-TextCNN

model pre-training. Step 2: Fin-Bert model converts

comments into eigenvectors. Step 3: CNN model

extracts feature vector characteristics and outputs

comment sentiment scores. Step 4: Construct and

calculate the sentiment index according to daily retail

investor sentiment.

Step 1 input more than 5000 manually scored

comments into the sentiment analysis model, and

conduct pre-training for the entire model. The Fin-

Bert model in step 2 converts the text of investor

comments into word vectors. Step 3 is responsible for

converting the word vector output from Bert layer

into a single sentence score in the same format as the

pre-training tag set. For example, 0, 1, 2 represent

negative, neutral, and positive respectively.

TextCNN hidden layer includes many alternating

convolution layers and pooling layers. Suppose Bert

has 𝑠 words for a single comment. And the word

vector output by Bert model has 𝑑 dimensions. Then

for this sentence, we can get the matrix A of row

𝑠 and column 𝑑.

The convolution kernel is a matrix w with a width

of d and a height of h. Then 𝑤 has ℎ * 𝑑 parameters

that need to be updated. For a sentence, matrix A can

be obtained after embedding layer. A [i: j] represents

lines 𝑖 to 𝑗 of A, so the convolution operation can be

expressed by the following formula:

The single word window matrix is expressed as:

𝑜



=𝑤⋅

𝐴

[𝑖:𝑖+ ℎ− 1],𝑖

=1,2,…,𝑠−ℎ+1

(1)

Activate with tan function after stacking offset:

𝑐



=𝑡𝑎𝑛



𝑜



+𝑏



(2)

The convolution result is:

𝑐=

[

𝑐



,𝑐



,…,𝑐



]

(3)

where 𝑏 represents the offset parameters. Through

pooling operation, max pooling is performed on each

matrix to obtain a vector 𝑧 with length 𝑚, where 𝑚 is

the number of filters in the convolution process.

𝑧=

[

𝑐



,𝑐



,…,𝑐



]

(4)

The final result is input to the full connection layer

and output through the softmax function.

Step 4: Construct the sentiment index according

to the daily sentiment of retail investors. Since there

are numbers of likes after each comment, it is

necessary to calculate the sentiment index through

weighted processing.

ICEMME 2022 - The International Conference on Economic Management and Model Engineering

480

4 STOCK INDEX PREDICTION

MODEL

4.1 LSTM Model

LSTM is an improved RNN model, which has a

unique forgetting gate mechanism to solve the

problem of gradient disappearance and divergence of

RNN in long-term information processing.

LSTM is composed of several consecutive

memory cells. The main components of a single

memory cell include input gate, output gate and

forgetting gate. 𝑋



represents the input at time t, and

ℎ



represents the state of memory cells at time 𝑡. 𝜎

represents the feedforward network layer of the

activation function sigmod, and 𝑡𝑎𝑛ℎ represents the

activation function.

In the actual operation process, the input gate

inputs data at time t, the input gate and the forgetting

gate calculate the data respectively:

𝑖



=𝛿



𝑊



∗



𝑋



,ℎ





+𝑏





(5)

𝐶





=𝑡𝑎𝑛ℎ



𝑊



∗



𝑋



,ℎ





+𝑏





(6)

𝑓



=𝛿𝑊



∗



𝑋



,ℎ





+𝑏





(7)

Where, 𝑖



represents the value of the output gate,

𝐶





represents the candidate state of the current cell, 𝑓



represents the value of the forgetting gate, 𝑊



and 𝑏



represents the weight and offset parameters of the

forgetting gate respectively. Finally, the status of

memory cells is updated and calculates the value of

the output gate:

𝐶



=𝑖



∗𝐶





𝑓



∗

𝐶



(8)

𝑂



=𝛿



𝑊



∗



𝑋



,ℎ





+𝑏





(9)

The value of the output gate is output through the

activation function tan:

ℎ



=𝑂



∗𝑡𝑎𝑛ℎ



𝐶





(10)

4.2 LSTM Stock Index Prediction

Based on Investor Sentiment

After the Fin-Bert-TextCNN model obtains the

sentiment index of retail investors, it is necessary to

predict the stock index in combination with other

variables. This paper selects six basic characteristics

of stock data, namely, the highest price, the lowest

price, the opening price, the closing price, the trading

volume and the trading price.

The steps of LSTM stock index prediction based

on investor sentiments are as follows. First, acquire

the basic characteristics of stock data and conduct

data preprocessing. Second, input a large amount of

stock fundamental data and the corresponding

sentiment index into LSTM stock index prediction

model for training. Third, predict the stock price for a

period of time.

5 RESULTS

5.1 Data Acquisition and

Pre-Processing

In this paper, the stock index prediction model is

based on the market trading index and investor

sentiment index.

Market trading indexes are the six basic

characteristics of stock trading data, which are the

highest price, lowest price, opening price, closing

price, trading volume and trading price. These data

can be downloaded from CSI 300 and used after

normalization. The method of normalization is

selected to use the Min-Max Normalization (Patro et

al., 2015), which is also called deviation

standardization.

The specific sample data is as follows:

Table 1. Transaction data sample.

Date Open High Low Close Volume

2022/3/7 3438.56 3438.56 3360.74 3372.86 3.94

2022/3/8 3372.55 3383.63 3287.34 3293.53 4.16

2022/3/9 3303.71 3321.48 3147.68 3256.39 4.66

2022/3/10 3312.18 3326.58 3291.24 3296.09 3.78

2022/3/11 3259.32 3315.66 3217.42 3309.75 3.84

2022/3/14 3271.89 3297.8 3223.53 3223.53 3.38

2022/3/15 3192.36 3196.92 3063.97 3063.97 4.65

For data representing investor sentiment, this

paper uses crawler code to crawl the comment area

from websites such as Eastmoney and Ping An

Fortune. This paper has obtained more than 90000

The Prediction Method of Stock Index Based on Sentiment Analysis of Investors’ Comments

481

comments, including the date, likes, user ID, and

publishing time of each comment. The time span is

44 months.

Due to the obvious noise of comments proposed

by retail investors, this paper used the cleaning data

algorithm to clear the comments that are blank, or

with no actual meaning. Moreover, the disordered

symbols in the comments were also removed.

Finally, 92239 qualified comment data were

obtained. 4975 of them were selected to be scored

manually as the labeled training set of the sentiment

analysis model, and 995 of them were used as the test

set and verification set.

5.2 Model Pre-Training

The pre-training of the model mainly refers to the

investor sentiment model, including fine-tune of Fin-

Bert model and CNN model tagged set data. In this

part, this paper selected 5000 consecutive data from

Eastmoney stock comments for double blind scoring.

The scoring standard refers to the sentiment keywords

in the industry research papers published by

Eastmoney, Wind and other institutions, with the

combination of the opinions of experts with years of

stock trading experience. Hence, investor comments

are divided into three different sentiments with 0, 1, 2

labels representing negative, neutral and positive

respectively.

5.3 Results and Verification

In essence, the analysis of investor sentiment is a

classification problem. This paper selects the

accuracy, precision, recall, and F1 values as

evaluation indicators, all of which are commonly

used in classification problem. The index calculation

method is as follows:

Accuracy

𝑇𝑃+ 𝑇𝑁

𝑇𝑃

𝐹𝑃

𝑇𝑁

𝐹𝑁

(11)

Precision

𝑇𝑃

𝐹𝑃

(12)

Recall

𝑇𝑃

𝐹𝑁

(13)

𝐹



2 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛× 𝑅𝑒𝑐𝑎𝑙𝑙

Precision

Recall

(14)

Where, 𝑇𝑃 represents the numbers of positive and

correct samples predicted, 𝑇𝑁 represents the

numbers of negative and correct samples predicted,

𝐹𝑃 represents the number of positive and wrong

samples predicted and 𝐹𝑁 represents the number of

negative and wrong samples predicted. The results

are shown in Table 2:

Table 2. Results of sentiment model test set.

Accurac

Precision Recall F1 value

73.17 55.75 55.33 59.97

As it can be seen from Table 2 that the accuracy

rate, precision rate, recall rate and F1 value of this

model are greatly improved compared with other

models, and the accuracy rate of Bert model is more

than 70%. At the same time, the results also reversely

verifies the theoretical basis of this model, that is,

sentiment will have a greater impact on investors'

performance and stock index.

The test indicators of stock index prediction are

RMSE, SAE and Accuracy. The calculation methods

of RMSE and MAE are as follows:

𝑅𝑀𝑆𝐸=



𝑁







𝑋

redictio

,

−𝑋

rea

,





(15)

𝑀𝐴𝐸=

𝑁









𝑋

rediction,



−𝑋

real,





(16)

This paper selects three representative market

indexes for verification, and the results are shown in

Table 3:

Table 3. Prediction results.

Name

Window

length

Forecasting

time

Accurac

Shenzhen

Stock

Index

1017 60 58.90

Shanghai

Stock

Index

1015 60 60.11

NASDAQ

Index

1075 60 78.81

6 CONCLUSIONS

The rise of deep learning technology provides an

effective solution for studying the stock financial

market and preventing financial security risks. This

paper creatively extracts sentiment from the

comments of retail investors by constructing Fin-

Bert-CNN model and forecasts the stock index

through LSTM model.

In the process of exploring the sentiment of retail

investors, this paper not only uses the latest pre-

trained Fin-Bert model, but also uses the tag set

scored under the specialized standard in training

process. The results show that this model can extract

investor sentiment very well, with an accuracy rate of

about 70%. Compared with other natural language

ICEMME 2022 - The International Conference on Economic Management and Model Engineering

482

processing algorithms and models, this model is more

suitable for the financial industry, especially for retail

investors.

Moreover, this paper also plays a positive role in

studying the relationship between investor sentiment

and stock index. The relationship between investor

sentiment and stock index is not a simple linear

relationship, but a very complex nonlinear

relationship. The deep learning model can well depict

this complex relationship. This study also proves that

we can predict the rise and fall of stock index with

high accuracy after considering the sentiment of

investors, especially retail investors. In the future, the

study can be improved in the following aspects.

Firstly, exclude the survivor bias in investor

comments, since profitable investors tend to comment

while others not. Secondly, improve the size of data

sample, especially the data in tag set, which can make

the results more accurate and close to reality.

REFERENCES

Chiu C.C., Sainath T.N., Wu Y., et al. (2018). Stateof-the-

art Speech Recognition with Sequence-to-sequence

Models, 2018 IEEE International Conference on

Acoustics, Speech and Signal Processing.

Cowles A. (1933). Can Stock Market Forecasters Forecast?

Econometrica, 1(3).

Deng, L., Yang, L. (2018). Deep Learning in Natural

Language Processing, Singapore: Springer.

Devlin, J. , Chang, M.W., Lee, K., Toutanova, K. (2019).

BERT: Pre-training of Deep Bidirec-tional Trans

formers for Language Understanding, Proceedings of

NAACL-HLT.

He, K., Zhang, X., Ren, S., Sun, J. (2016). Deep Residual

Learning for Image Recognition, 2016 IEEE

Conference on Computer Vision and Pattern

Recognition.

Lee Y.W., Jiang X.C., Indro C.D. (2002). Stock Market

Volatility, Excess Return, and the Role of Investor

Sentiment. Journal of Banking & Finance, 26(12),

2277 -2299.

Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S.,Dean

J. (2013). Distributed Re presenltationsof Words and

Phrases and 'Their Com positionality. Proceedings of

Advances in Neural Information Pro-cessing Systems.

Oprea, D.S., Brad, L. (2014). Investor Sentiment and Stock

Returns: Evidence from Romania. International

Journal of Academic Research in Accounting Finance

and Management Sciences, 42, 19-25.

Pennington, J., Socher, R., Manning, C.D., (2014). Glove:

Global Vectors for Word Representation, Proceedings

of the 2014 Conference on Empirical Methods in

Natural Language Processing.

S. Gopal Krishna Patro, Kishore Kumar Sahu. (2015).

Normalization: A Preprocessing Stage. CoRR, abs,

1503.06462.

Xu, X.C., Tian, K. (2021). A Novel Financial Text

Sentiment Analysis-Based Approach for Stock Index

Prediction. The Journal of Quantitative & Technical

Economics, 38(12), 124-145.

Y, S.X., L, Z.C., Wang, X.J. (2020). News sentiment credit

spreads and information asymmetry. North American

Journal of Economics and Finance, 52.

The Prediction Method of Stock Index Based on Sentiment Analysis of Investors’ Comments

483