Financial Risk Prediction of Listed Companies based on Text and
Financial Data
Xu Wei and Yonghui Chen
Southwest University of Science and Technology School of Computer Science and Technology, Mianyang, China
Keywords: Deep Learning, CNN, Financial Risk, Annual Report Text.
Abstract: This paper uses the relevant financial data of 4348 A-share listed companies from 2010 to 2019 and the
discussion and analysis of operation in the annual report as the research sample, and uses the Pytorch
framework to build a neural network model to predict whether the listed companies fall into financial crisis.
The experimental results show that when text data is combined with traditional financial index data, the
prediction accuracy of the deep learning model can reach about 85%, which can significantly improve the
accuracy of financial risk prediction compared with using only financial data.
1 INTRODUCTION
The financial crisis of listed companies is one of the
main drivers of the financial risk and gains primary
attention from creditors and investors. The financial
risks of listed companies may further propagate
recession and thus jeopardize the economy at large
(Ding 2012). Therefore, it is of great significance to
pay attention to the financial situation of listed
companies.
The research of financial risk prediction is
usually based on the company's stock market
transaction information and the company's financial
index data (Ye 2017, Song 2019, Hosaka 2019,
Zhang 2021). However, as an unstructured and
qualitative data form, the annual report text of listed
companies also plays a very important role in how to
convey information to the public. For example, the
annual report submitted by a listed company to the
regulator contains all the information about the
company in the past year (Feng 2019). Among them,
“Discussion and Analysis of Operating Conditions”
mainly analyzes the operating conditions, operating
models, and operating strategies of listed companies.
It is a summary of the overall operating conditions
of the past year and a generalization of future
operating directions. Recent research has proven that
qualitative company reports contain important
information that can predict financial risks
(Campbell 2014). Since it is difficult to obtain and
quantify text data, it is still challenging to effectively
combine financial text with financial data in a
financial model (Lang 2015). In this research, we
propose a new deep learning method that combines
financial text and financial data to predict financial
risks. We found that text data can supplement
traditional accounting and market-based variables to
predict financial risk, and the deep learning model
that combines financial text and financial data has
higher prediction accuracy than a model that uses a
single type of input.
2 MATERIALS AND METHODS
2.1 Data
2.1.1 Financial Indicator Data
In this paper, financial indicator data of A-share
listed companies from 2010 to 2019 were obtained
from the Jukuan database. Since financial indicators
are subject to change during the year, the data on
financial indicators as of the date of disclosure of the
company's annual report by the listed companies are
taken.
Based on relevant studies, a total of 30 financial
indicators are selected in this paper to analyze the
solvency, operating capacity, development capacity,
profitability, risk level, cash flow, and ratio structure
indicators of enterprises. Based on
general indicators,
as many financial factors affecting the business
conditions of enterprises are considered as possible,
240
Wei, X. and Chen, Y.
Financial Risk Prediction of Listed Companies based on Text and Financial Data.
DOI: 10.5220/0011172500003440
In Proceedings of the International Conference on Big Data Economy and Digital Management (BDEDM 2022), pages 240-244
ISBN: 978-989-758-593-7
Copyright
c
2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
and rich indicators are selected for feature learning
to avoid artificially removing features that have a
significant impact on the prediction results. The
selected financial indicators include current ratio,
quick ratio, cash ratio, gearing ratio, cash flow ratio,
accounts receivable turnover, inventory turnover,
current asset turnover, total asset turnover, operating
income growth rate, operating profit growth rate, net
profit growth rate, net flow from operating activities
growth rate, total assets growth rate, return on assets,
total net asset margin, return on net assets, net
operating margin, return on investment, financial
leverage, operating leverage, consolidated leverage,
net cash flow from operating activities per share,
cash recovery rate, operating index, cash asset ratio,
working capital ratio, fixed asset ratio, equity
concentration index and z-index(Table 1).
Table 1: Financial index system of financial risk prediction.
Selected indicators Abbreviation Calculation description
Solvency
X1 Current ratio RA Current assets / current liabilities
X2 Quick ratio RAT Quick assets / current liabilities
X3 Cash ratio CR Monetary capital / current liabilities
X4 Asset liability ratio LEV Total liabilities / total assets
X5 Cash flow ratio CASHCL Net operating cash flow / current liabilities
Operating
capacity
X6 Accounts receivable turnover ARTURNOV
Average balance of operating income / accounts
receivable
X7 Inventory turnove
r
INTURNOV Operating income / inventory
X8 Turnover rate of current assets VOL
Net income from main business / average balance of
current assets
X9 Total asset turnove
r
TATO Operating income / total assets
Development
capacity
X10 Growth rate of operating revenue RG
Increase in operating income / operating income of
the previous period
X11 Operating profit growth rate PGR
Increase in operating profit / operating profit of the
previous perio
d
X12 Net profit growth rate EG
Increase in net profit / net profit of the previous
period
X13 Growth rate of net flow from
operating activities
GNFOA
Increase in net operating cash flow / net operating
cash flow of the previous perio
d
X14 Growth rate of total assets TAGR
Increase in total assets / total assets at the beginning
of the period
Profitability
X15 Return on assets ROA
Profit without financial expenses / average total
assets
X16 Net interest rate of total assets ROT Net profit / total assets
X17 Return on net assets ROE After tax profit / net assets
X18 Net operating interest rate NPM Net profit / operating income
X19 Return on investment ROI Average total profit / total investment
Risk level
X20 financial leverage DFL
Change rate of earnings per share of common stock /
change rate of EBIT
X21 Operating leverage DOL
Change rate of EBIT / change rate of production and
sales volume
X22 Integrated lever DTL
Change rate of net profit / change rate of main
b
usiness income
Cash flow
analysis
X23 Net cash flow from operating
activities per share
NCFOPS
Net operating cash flow / number of common shares
outstanding
X24 Cash recovery rate CRA
Net operating cash flow / total assets at the end of the
period
X25 Operating index OI Net operating cash flow / gross operating cash flow
Ratio
structure
X26 Cash asset ratio CAR Cash assets / total assets
X27 Working capital ratio WCR Total working capital / assets
X28 Fixed assets ratio LTCR Fixed assets / total assets
Internal
governance
X29 Equity concentration index HERF
Number of shares of the largest shareholder / total
number of shares of the company
X30 Z index Z
Number of shares of the first largest shareholder /
number of shares of the second largest shareholde
r
Financial Risk Prediction of Listed Companies based on Text and Financial Data
241
In the actual data, the numerical ranges of
different features are different, which may appear in
the feature space. Individual features with large
values have a dominant impact on the sample. To
make all features on the same scale, it is necessary
to map them to the same scale, to improve the
accuracy of the model and speed up the fitting
speed, so the samples should be normalized in this
experiment. You can use the MinMaxScaler
provided in the preprocessing class in Sklearn, the
specific formula is (1). Where min represents the
minimum value of each feature in the data, and max
is the maximum value of each feature.
𝑋
scale
(1)
2.1.2 Text Data
We obtained the annual reports of all listed
companies from 2010 to 2019 in PDF format from
Eastmoney.
Python provides many class libraries for parsing
PDF files, among which PDFMiner and
PDFPlumber are some of the most common parsing
methods. This paper finally uses PDFMiner to
analyze the annual report in PDF format. After
obtaining the text, clean it first, and use the regular
formula to extract the chapter of “Discussion and
Analysis of Operating Conditions” in the annual
report. It is worth mentioning that, according to the
requirements of the CSRC, “Discussion and
Analysis of Operating Conditions” was not used as a
separate chapter before 2016, but inserted into other
chapters. Therefore, different methods should be
taken to extract the annual reports in different
periods. In this study, we exclude the blank samples
in the section of “Discussion and Analysis of
Operating Conditions” and finally obtain 28549 text
data.
2.2 Methods
2.2.1 Text Representation
Before combining text with financial indicator data
into the model, we need to consider how to represent
the text. The usual approach is to convert text into
vectors. That is, the text is mapped into a new space
and represented by multi-dimensional continuous
real number vectors.
We use the word2vec model based on Skip-gram
(Deoras 2013). The goal of the Skip-gram model is
to make a word predict the words around it. For a
sequence of words W1, W2, W3, ..., Wn, we
maximize log p.
max
∑
∑
,
log 𝑝𝑤
∣𝑤
(2)
In formula (2), c represents the number of words
before and after the word is considered.
In this paper,
according to the size of the text data training set, the
dimension of the word vector is set at 100 and the
fixed-length is set at 5. Negative sampling is used to
calculate log p, and the sub-sampling of words is
proportional to their inverse frequency. In
word2vec, words with similar semantics have high
cosine similarity and allow vector calculation of
words.
This paper uses Jieba word segmentation to
segment our financial text data set to obtain the
financial text corpus. We use the general vocabulary
training set to train a general word vector model,
and then use it for the second training. The final
word vector contains financial background
information, which is more suitable for our task. We
calculate the average of all word vectors in
each text,
i.e., each financial text is represented by a
100-dimensional vector.
2.2.2 Model Building
We use Pytorch to combine financial index data
with text data to build our model. The model
structure is shown in Figure 1:
Figure 1 Schematic diagram of financial indexes and financial text combination model.
BDEDM 2022 - The International Conference on Big Data Economy and Digital Management
242
The input of the model is a 130-dimensional
vector, which consists of two parts:
Financial indicator data: including 30
financial indicators, each of which is
normalized;
Financial text: the section of
“business
discussion and analysis” in the annual report.
The model splices the financial index data and
text data through the intersection of pd.merge
function of pandas library, and send them to the
convolutional neural network (CNN). The model
parameters of convolution neural network include
the number of convolution cores, the size of
convolution cores, the size of the pooling layer, and
so on. To select the best parameters to fit the model
in this paper, we reset the value range of parameters.
For example, the CNN convolution kernel size
d∈{2,3,4,5}, the number of CNN convolution cores
h∈{64,100,128,256}, the pool layer size c∈{5,6,7,8},
and the learning rate λ∈{0.01, 0.001, 0.0001},
epoch∈{5, 10, 15}, the weight value of cross-entropy
loss function f∈{0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8,
0.9, 1, 2, 3, 4}.
3 RESULTS & DISCUSSION
We compare the model proposed in this paper with
other models. These models are:
S-CNN: Feature vectors are constructed based on
financial data, and then the CNN model is used to
extract features and realize classification.
S-SVM: The model based on financial data uses
SVM to classify.
S-XGB: The model based on financial data uses
XGBoost to classify.
The evaluation results of each model are shown
in Table 2:
Table 2: Experiment summary table.
Accuracy
True Positive
Rate
True Negative
Rate
F1 Value
S-CNN 78.00% 89.02% 54.57% 0.676623
T&S-CNN 85.00% 93.38% 77.67% 0.848035
S-SVM 70.83% 75.60% 63.38% 0.689527
S-XGB 77.12% 89.02% 54.57% 0.676623
It can be seen from the table that the prediction
effect of the CNN deep learning model based on
financial data is not significantly better than the
traditional machine learning model based on
financial data. After the combination of financial
data and financial text, the CNN model is higher
than other models’ inaccuracy, true positive rate,
true negative rate, and F1 value. There may be two
main reasons:
The convolutional neural network model pays
more attention to information, which leads to
insufficient attention to important
information. After adding the financial text
features, although there is still a lot of
information, with the help of the financial text
features, important features are highlighted.
From the perspective of the financial text, the
more information combined with the data, the
better. In this way, after the combination of
important information and data, after the
screening of multi-layer neural networks, the
more important information can be selected.
4 CONCLUSIONS
As more and more financial documents appear in the
stock market, investors, regulators, and researchers
need more deep learning models to process and
analyze the information disclosures of listed
companies. Taking all A-share listed companies in
the recent ten years as samples, this paper builds a
financial risk prediction model based on financial
text and financial data. The experimental results
show that compared with using only financial data,
the F1 value of the financial risk prediction model
based on the combination of text and financial data
is significantly improved, indicating that the latest
Financial Risk Prediction of Listed Companies based on Text and Financial Data
243
progress of neural network can extract useful
information from financial text, and the financial
risk prediction combined with traditional financial
data can further improve the prediction effect.
REFERENCES
Campbell, J.L., Chen, H., Dhaliwal, D.S., et al. The
Information Content of Mandatory Risk Factor
Disclosures in Corporate Filings [J]. Social Science
Electronic Publishing.
Zhang, Chunmei, Zhao, Mingqing, Guan, Junqi. Financial
risk portfolio early warning model of manufacturing
listed companies based on lasso + SVM [J] Practice
and understanding of mathematics, 2021, 51 (5): 12.
Deoras A, Tomá? Mikolov, Kombrink S, et al.
Approximate inference: A sampling based modeling
technique to capture complex dependencies in a
language model[J]. Speech Communication, 2013,
55(1):162-177.
Ding, A.A., Tian, S., Yu, Y., et al. A Class of Discrete
Transformation Survival Models with Application to
Default Probability Prediction.
Hosaka, T. Bankruptcy prediction using imaged financial
ratios and convolutional neural networks[J]. Expert
Systems with Application, 2019, 117(MAR.):287-299.
Lang, M., Stice-Lawrence, L., Textual analysis and
international financial reporting: Large sample
evidence[J]. Journal of Accounting and Economics,
2015.
Ye, Lanzhou, Zhou J.Y., Management, S.O., Application
of Support Vector Machine in Risk Management of
Listed Companies[J]. Science Technology and
Industry, 2017.
Mai, F., Tian, S., Lee, C., et al. Deep learning models for
bankruptcy prediction using textual disclosures[J].
EUROPEAN JOURNAL OF OPERATIONAL
RESEARCH, 2019.
Song, Y., Peng, Y., A MCDM-Based Evaluation
Approach for Imbalanced Classification Methods in
Financial Risk Prediction[J]. IEEE Access, 2019, PP
(99): 1-1.
BDEDM 2022 - The International Conference on Big Data Economy and Digital Management
244
