https://doi.org/10.31449/inf.v48i11.5941 Informatica 48 (2024) 15 –32 15 
 
Profit Estimation Model and Financial Risk Prediction Combining 
Multi-Scale Convolutional Feature Extractor and BGRU Model 
 
Yangyan Liu
1*
, Bolin Pan
2
 
1
School of Economics and Management, Guangxi Science & Technology Normal University, Laibing 546199, China 
2
School of Economic Management, Ningxia Institute of Science and Technology, Shizuishan 753000, China 
E-mail: liuyangyan001@163.com  
*
Corresponding author 
Keywords: convolutional neural network, feature extraction, bidirectional gated recurrent unit, profit estimate, 
financial risk 
Received: March 21, 2024 
In response to the inaccuracy of financial risk prediction and profit prediction for enterprises, a 
financial risk prediction model based on the graph networks was designed. This experiment combined 
multi-scale feature extraction and sequence analysis methods. In addition, the model adopted a 
structurally concise and effective bidirectional gated recurrent unit to capture temporal relationships in 
time series data. The profit prediction model was combined multi-scale advantages and attention 
mechanisms. The latter enhanced the recognition and utilization of influential features, which improved 
predictive ability and practical value. These results confirmed that the accuracy of this model 
significantly improved to 98.03% after iterative training. The F1 score of the financial risk prediction 
model reached 0.98, demonstrating an excellent performance. The profit prediction model performed 
better than other models in both regression and classification problem indicators, with an error close to 
0 and a mean square error of 0.0232. This indicated that the model had extremely high prediction 
accuracy. Therefore, both models have strong predictive ability and have practical application 
significance. 
Povzetek: Razvit je finančni model, ki združuje multi-lestvični konvolucijski ekstraktor lastnosti in     
BGRU model za natančno napovedovanje dobička in finančnih tveganj podjetij. Model z uporabo 
pozornostnih mehanizmov in sekvenčne analize dosega visoko zmogljivost in praktično vrednost v 
finančni analizi ter obvladovanju tveganj.
1 Introduction 
In the current globalized economic environment, the 
estimation of corporate profitability and the prediction of 
financial risk becomes the important links in financial 
analysis and investment decision-making. On the one 
hand, it is crucial to accurately estimate a company's 
profit prospects for investors. Because it directly relates 
to the investment and return of capital. On the other hand, 
financial risk prediction is particularly important for 
corporate management. Because it can help enterprises 
identify potential financial problems in a timely manner, 
formulate corresponding countermeasures, and ensure the 
stable development of the enterprise [1-3]. As 
information technology develops, data analysis 
technology becomes a key force driving innovation, 
especially the application of artificial intelligence in the 
financial field. In recent years, deep learning technologies 
have emerged and developed. These technologies have 
made significant achievements in multiple fields such as 
image recognition, speech processing, and natural 
language processing. In this context, the application of 
deep learning techniques to profit estimate and financial 
risk prediction begins to receive widespread attention 
from academia and industry. Specifically, Convolutional 
Neural Network (CNN) shows great potential in 
processing time-series data due to its excellent feature 
extraction capabilities [4-6]. Gated Recurrent Unit (GRU) 
is an effective variant of recurrent neural networks, which 
has value in processing sequence information [7]. The 
financial data of enterprises are high-dimensional and 
contain rich time-series characteristics. However, 
traditional machine learning techniques are inadequate in 
dealing with such problems. In addition, there are many 
factors that need to be considered due to the complexity 
of financial data. How to effectively integrate these 
factors has become a significant challenge [8-10]. In this 
context, this study aims at combing multi-scale 
convolutional feature extractions with bio-heuristic for a 
bidirectional gated recurrent single model. A novel profit 
estimate and financial risk prediction framework is 
constructed, aiming to achieve more accurate and robust 
performance in predicting future profits and risks. The 
study first proposes the research purpose. Secondly, the 
financial risk prediction and profit estimate models are 

16   Informatica 48 (2024) 15 –32                                                            Y . Liu et al. 
designed. Then the model performance is validated. 
Finally, the research conclusion is obtained. 
In recent years, academic discussions on financial 
prediction automation continue to increase. Peng et al. 
proposed a novel predictive tool that utilized a model to 
describe the volatility of financial returns as numerous 
potential statistical distributions. Meanwhile, the 
predictive accuracy was optimized by analyzing the best 
case of these distributions. These results confirmed that 
the model performed better than most traditional 
prediction methods [11]. Cao et al. analyzed and 
predicted the financial crisis of listed companies from the 
perspective of constructing financial risk early warning 
systems for e-commerce enterprises, using methods based 
on deep learning technology. Meanwhile, they designed a 
financial stability warning system to achieve the early 
detection and handling of crises [12]. 
In profit forecasting, Sirisha et al. fused three neural 
network models to form a comprehensive model. This 
model was used to process past financial data to predict 
the future economic situation of the enterprise. These 
results confirmed that the model exhibited an accuracy 
prediction rate of over 90% [13]. Kim et al. developed a 
multivariate model under self-attention mechanism, 
aiming to plot the Bitcoin prices. These results confirmed 
that the prediction results of this model had high accuracy 
[14]. Shajalal et al. proposed an innovative deep learning 
network architecture aimed at predicting the stock risk of 
market products. They used a series of data balancing 
techniques to train data balance. Then the balanced data 
were used as the basis for training deep neural networks. 
These results confirmed the effectiveness and accuracy of 
the model in predicting product shortages [15]. 
He et al. developed two flood forecasting models in 
the application of Bidirectional Gated Recurrent Unit 
(BGRU). These two models were based on BGRU and 
Backpropagation Neural Network (BPNN), as well as 
bidirectional Long Short-Term Memory (LSTM) and 
BNN, respectively. These two models surpassed both 
unidirectional and bidirectional models in terms of 
prediction accuracy and reliability [16]. Dangut et al. 
developed a rare fault recognition model that relied on 
deep hybrid learning methods. They conducted empirical 
analysis using large-scale records from the central 
maintenance system of large aircraft. These experiments 
confirmed that this model improved accuracy by 18% and 
recall by 5%, effectively identifying rare faults [17]. Li et 
al. created a deep learning model that integrated 
end-to-end CNN and BGRU for extracting and 
classifying spatiotemporal information. These 
experiments confirmed that the model could effectively 
distinguish healthy rats from diseased rats, with a 
classification accuracy of up to 98.73%. This study had 
important reference significance for the clinical diagnosis 
of Parkinson's disease [18]. 
In summary, in recent years, researchers have 
increasingly focused on building efficient, accurate, and 
stable early warning systems in automated financial 
forecasting. The improved neural networks and deep 
learning-based methods are important. These methods 
play important roles in enterprise financial risk prediction 
and the development of financial stability warning 
systems. There are various prediction tools, such as 
self-attention mechanisms based on multivariate long 
short-term models. From the perspective of prediction, 
the proposal of these tools undoubtedly opens up new 
avenues for financial prediction. This study combines the 
method of multi-scale convolutional feature extraction, 
which has advantages over traditional models in 
information acquisition. The addition of BGRU enables 
the prediction model to better capture time series 
information, effectively improving the accuracy of profit 
estimate and financial risk predictions. The literature 
summary is shown in Table 1. 
 
Table 1: Literature summary 
Author Literature dimension Content Conclusion 
Li et al. 
Risk prediction 
Design and improve a neural 
network model, using data from 
listed companies from 2017 to 2020 
as the basic dataset for empirical 
analysis 
The prediction accuracy of the 
improved model exceeds 80% 
Peng et al. 
Design a risk prediction model based 
on financial volatility as a statistical 
distribution 
The model performs better than most 
traditional prediction tools 
Cao et al. 
Construct an early warning model 
for e-commerce enterprise financial 
risk based on deep learning 
The model can achieve early risk 
detection and risk handling 
Sirisha et al. 
Profit forecast 
Design a comprehensive model that 
integrates three types of neural 
networks to predict the future profit 
situation of enterprises 
The model accuracy prediction rate 
exceeds 90% 
Kim et al. 
Design a multivariable model based 
on self-attention mechanism to 
The prediction results of the model 
have high accuracy and superiority 

Profit Estimation Model and Financial Risk Prediction Combining … Informatica 48 (2024) 15 –32 17 
predict the profitability of Bitcoin 
Shajalal et al. 
Propose a deep learning network 
architecture to predict the risk of 
product shortage and product profit 
in the market 
The effectiveness and accuracy of the 
model are superior 
He et al. 
Application of BGRU 
model 
Propose a flood prediction model 
that integrates BGRU and BNN 
The model has advantages over 
traditional prediction methods in terms 
of prediction accuracy and reliability 
Dangut et al. 
Design a rare fault recognition model 
for large aircraft based on deep 
hybrid learning method 
This model improves prediction 
accuracy by 18% and recall by 5% 
Li et al. 
Design a deep learning model that 
integrates end-to-end CNN and 
BGRU for clinical diagnosis of 
Parkinson's disease 
The model can effectively distinguish 
healthy and diseased rats, with a 
classification accuracy of up to 
98.73% 
 
Current research has predicted enterprise risks and 
profits from multiple perspectives. However, all methods 
are single risk or profit predictions, which have not 
formed an integrated system for predicting risks and 
profits simultaneously. This study presents a concise and 
effective bidirectional GRU and attention mechanism. 
This method not only improves the predictive ability of 
the model, but also integrates risk prediction and profit 
prediction. Then enterprises can make comprehensive 
decisions and avoid information conflicts caused by 
different prediction systems. 
 
2 Design of financial risk and profit 
prediction models 
A novel financial risk prediction model, namely 
Multi-Scale Bidirectional Gated Recurrent Unit (M-BU), 
is designed based on BGRU. In addition, a profit estimate 
model is designed, which is the Multi-Scale Adaptive 
Bidirectional Gated Recurrent Unit (M-AU). The 
financial risk prediction model is based on the graph 
networks and adopts a strategy combining multi-scale 
feature extraction technology and sequence analysis 
methods. Convolutional layers with different sizes of 
convolutional kernels are used to extract rich features. 
Meanwhile, advanced techniques such as Batchnorm and 
Mish functions are used to prevent overfitting while 
maintaining the model's generalization ability. BGRU can 
effectively capture temporal relationships in time series 
data. The profit estimate model combines multi-scale 
advantages and attention mechanisms. The latter 
enhances the recognition and utilization of influential 
features by assigning different weights to different 
features. 
 
2.1 Design of financial risk prediction model 
The first step is the analysis and reorganization of the 
sample data in all financial risk prediction model designs. 
The essential step before data analysis is to determine 
which variables should be included in the research scope 
[19-21]. The study sets a criterion to exclude variables 
with a missing degree exceeding one-third in the dataset. 
Because these variables can provide insufficient 
information to facilitate the analysis. The missing data 
will be filled in. The dynamic data imputation method is 
adopted instead of simply replacing the mean or 
minimum values [22-24]. The study observes that the 
values of many financial variables mostly fluctuate 
around 30% of their mean after in-depth analysis of the 
data distribution of various financial variables. Therefore, 
the strategy for handling missing data is as follows: a 
random number based on the fluctuation range is used to 
replace the average value of the variables. The calculation 
method is represented by formula (1). 
Missing value
Mean value+Randomness Mean value
=

 (1) 
A distortion rate is defined higher than 20% when 
encountering samples with severe data missing in 
generating a dataset. This means that the information of 
these samples is not sufficient to support a 
comprehensive analysis. Samples containing numerous 
missing values will be excluded under these conditions. 
There may still be situations where data is scarce despite 
going through layers of screening. For example, some 
corporate samples may only cover some financial 
variables. The study also decides to delete the incomplete 
data samples. Dealing with outliers is equally essential. 
Then this study applies the rule based on triple standard 
deviation to identify and remove outliers. The study uses 
the upper and lower quartiles to define and handle 
outliers based on the distribution of data from each year 
given the differences in data from different years. The 
key to this step is to eliminate data points that may distort 
the analysis results. The final stage of data processing is 
quantitative standardization. Different variables may have 
different scales and distribution intervals. Therefore, the 
study uses the min-max method to transform all financial 
variables into a unified standard interval of 0 to 1. 
Min-max standardization refers to adjusting each variable 
to a uniform size. This can eliminate the impact of 

18   Informatica 48 (2024) 15 –32                                                            Y . Liu et al. 
different measurement units and maintain comparability 
between different variables. Standardization involves 
subtracting the minimum value of each variable from its 
value across all data points. Meanwhile, the variable is 
divided by the amplitude of the variable range to 
calculate the standardized data points [25-27]. The 
calculation method is represented by formula (2). 
 
( )
( ) ( )
min
max min
xx
x
xx
−
 =
−
 (2) 
In formula (2), ( ) min x and ( ) max x represent 
the minimum and maximum values of the financial index, 
respectively. A completely new pattern is designed. This 
pattern enhances the ability to mine important 
information in financial data by integrating two main 
methods: multi-level feature recognition and sequence 
analysis. Firstly, a feature extraction module with 
multi-scale perception capability is designed to capture 
risk conditions associated with numerous financial 
indicators. The innovative inspiration for this strategy 
comes from the concept of graph networks. This module 
can integrate convolution kernels of various sizes and 
successfully extract rich feature levels from financial data. 
Figure 1 shows the structure of convolution and feature 
recognition. 
 
f11 f12
f21
x11
x21
f22
x12
x22
x31 x32
x13
x23
x33
o11 o12
o21 o22
S*S n*n
CBM
Concat
 
Figure 1: Convolutional and feature recognition structures 
 
Small-scale convolution kernels distinguish subtle 
local patterns, while large-scale convolution kernels help 
to interpret external environmental correlations. This 
model can capture and apply the complex interactions 
between various financial indicators. The Batchnorm 
method is utilized to prevent overfitting of the model. 
Firstly, the mean is calculated using formula (3). 
 
1
1 m
Bi
i
x
m

=
=

 (3) 
Afterwards, the data variance is calculated using 
formula (4). 
 ( )
2
2
1
1 m
B i B
i
x
m

=
=−

 (4) 
Normalization is represented by formula (5). 
 
2
ˆ
iB
i
B
x
x


−
=
+
 (5) 
The final calculation result can be obtained, 
represented by formula (6). 
 
ˆ
ii
yx  =+ (6) 
In formula (6), 

 represents the scaling variable. 

 is a translation variable. The Mish function is utilized 
as the activation function, represented by formula (7). 
 ( ) ( )
tanh ln 1
x
Mish x e =  +
 (7) 
In multi-scale feature extraction, convolution 
operation is first required to obtain the feature 
representation vector, represented by formula (8). 
 
( )
( )
s
m
s conv x
s conv x






 (8) 
In formula (8), 
x
 represents the eigenvector. Then 
the feature representation vector can be output using 
formula (9). 
 ( ) , x concat s m  =
 (9) 
The model also includes a pooling layer, aiming at 
screening and enhancing favorable features. This model 
has the ability to select the main information in the 
feature vectors and exclude non-deterministic 
interference information. The ability is achieved by 
applying filters of various sizes in the max pooling 
operation of multiple feature images [28-30]. This 
approach not only reduces the dimensionality of the data, 

Profit Estimation Model and Financial Risk Prediction Combining … Informatica 48 (2024) 15 –32 19 
but also reduces the overall complexity of the model to 
accelerate the training objectives. The key to model 
design is to adopt sequence learners after completing 
feature extraction. Figure 2 shows the pooling structure. 
13
17
5
9
25
6
0
2
31
0
17
8
86 13 39 72
9
25
88
72
72
25
9
88
 
Figure 2: Pooled structure 
 
The study utilizes the BGRU structure, which can 
capture temporal continuity in financial data and 
simultaneously learn past and future information. As a 
result, more accurate risk prediction can be achieved. The 
design of BGRU is concise and effective, reducing the 
parameter complexity of LSTM, but still retaining the 
ability to learn long-range dependencies. It is possible to 
determine when to retain or discard information by 
judging the gating mechanism. As a result, the efficiency 
of neural networks can be improved in processing time 
series data. Figure 3 shows the structure of the financial 
risk prediction model. 
S*S n*n
CBM
Concat
 
Pooling
S*S n*n
CBM
Concat
 
Pooling
GRU GRU GRU
GRU GRU GRU
GRU GRU GRU
o
1
o
2
o
3
l
1
l
2
l
n
y
1
y
2
 
Figure 3: Structure of financial risk prediction model 
 
The model reset gate is represented by formula (10). 
 ( )
1 t t xr t hr r
Z X W H W b 
−
= + +
 (10) 
In formula (10), 

 represents the activation 
function. 
1 t
H
−
 is the implicit state of the previous node. 
t
X refers to the current input. W means weight. The 
update gate is represented by formula (11). 
 ( )
1 t t xs t hs s
Z X W H W b 
−
= + +
 (11) 
The memory gate is represented by formula (12). 
( ) ( )
1
tanh ~
t t xh t t hh h
H X W R H W b
−
= + + (12) 
Finally, M-BU combines multi-level spatial features 
with the temporal features learned by BGRU. As a result, 
a comprehensive prediction of the company's financial 
risk can be achieved. The dual advantages of feature 
mining and sequence analysis provide strong data support 
for model decision-making. Firstly, a multi-level 
convolutional network performs feature extraction by 
using convolutional kernels of different sizes. Therefore, 
clear hierarchical financial data features can be extracted. 
Next, the pooling operation further optimizes the feature 

20   Informatica 48 (2024) 15 –32                                                            Y . Liu et al. 
vectors, reduces dimensionality, and removes noise. 
Finally, these features are integrated into a financial risk 
foresight model that takes into account the time 
dimension through the BU network. Table 2 shows the 
parameters of the network structure. 
 
Table 2: Structural parameters of financial risk prediction model 
Number of 
layers 
Layer name Nuclear size Output size 
Parameter 
quantity 
0 Input layer - (38, 1) 0 
1 
One-dimensional 
convolutional layer 1 
5×5 (38, 4.2) 25.2 
1 
One-dimensional 
convolutional layer 2 
3×3 (38, 4.1) 16.8 
2 
One-dimensional 
convolutional layer 3 
5×5 (38, 8.4) 168.4 
2 
One-dimensional 
convolutional layer 4 
3×3 (38, 8.2) 104.8 
3 Merge layers - (38, 16.6) 0 
4 Bidirectional GRU layer 1 64 (16, 128.4) 39600 
5 Bidirectional GRU layer 2 64 (16, 128.2) 74160 
6 Dropout - (16, 128) 0 
7 Leveling layer - (1, 2048.8) 0 
8 Fully connected layer - (1, 2) 4100 
 
2.2 Design of financial profit prediction 
model 
The initial step involves collecting economic report data 
from numerous listed entities in the stock market in the 
design of the financial profit estimate model. Incomplete 
or unrelated information is discarded in the screening 
process to maintain the accuracy of the developed model. 
Economic data contain many core indicators, such as the 
proportion of corporate capital benefits and fixed asset 
profitability. They undoubtedly have analytical 
significance, but not all indicators have equal influence 
on predicting economic benefits. The M-AU model has 
been designed as a solution in response to this situation. 
This model is committed to improving and enhancing the 
efficiency of classical methods. Traditional BGRU often 
assumes that different features have the same influence 
when processing time series data. In view of this, the 
study decides to prioritize the importance of features to 
enhance the model's discriminative ability. M-AU has 
allocated weights to different features by inheriting the 
attention mechanism and improving the model. 
Meanwhile, more computing power is applied to features 
that have a significant impact on expected results, thereby 
enhancing the model's ability to identify and evaluate 
impacts. Figure 4 shows the structure of the profit 
estimate model. 
 
S*S n*n
CBM
Concat
 
Pooling
S*S n*n
CBM
Concat
 
Pooling
GRU GRU GRU
GRU GRU GRU
o
1
o
2
o
3
l
1
l
2
l
n
y
1
y
2
Attention mechanism layer
Fully connected layer
Multi scale 
feature 
extraction
Sequence 
learning
 

Profit Estimation Model and Financial Risk Prediction Combining … Informatica 48 (2024) 15 –32 21 
Figure 4: Structure of profit forecast model 
 
This mechanism does not exist alone, but can be 
combined with other network structures. The 
identification rate of key information can be improved by 
applying it in various stages of model processing. More 
precisely, the typical design of M-AU includes using 
filters of various scales to perform deep feature mining of 
financial data and identify data characteristics of different 
scales. Subsequently, this model is able to learn the 
connections between time series data by utilizing the 
bidirectional cyclicity of the BGRU structure. The 
attention mechanism introduced on this basis assigns 
specific weights to each hidden layer state, represented by 
formula (13). 
 ( )
tt
a score h =
 (13) 
In formula (13), 
t
h represents the output of the 
multiscale extractor. These weights are normalized 
through the softmax function, representing the weighted 
impact of each feature on the final estimation result. This 
model integrates weighted features in subsequent stages 
and obtains the final prediction result through the fully 
connected layer and softmax function. The normalized 
weights are represented by formula (14). 
 ( )
1
exp /
n
t t t
t
p a a
=
=

 (14) 
Finally, resource allocation is based on the attention 
mechanism, represented by formula (15). 
 
1
n
t t t
i
y p h
=
=

 (15) 
Overall, M-AU is committed to improving the 
accuracy of financial data analysis through exquisite 
structural selection and algorithm integration. This 
provides a more reliable scientific basis for financial 
decision-making. Figure 5 shows the model training 
process. 
Allocate 
attention 
resources
Entering 
GRU for 
learning
Hidden layer 
calculation 
completed?
Calculate 
hidden 
layer values
Calculate 
loss 
function
Calculate 
loss 
function
Output 
layer bias 
value
Calculate 
loss 
function
Partial 
derivative of 
loss function
Update 
weights
End
Y
N
Start
Different 
scale 
convolution 
kernel 
convolution 
operations
Input 
training 
data
Data 
reconciliation
Obtain 
multi-scale 
feature 
vectors
Calculate 
vector 
weights
Vector 
weight 
normalization
Calculate 
hidden 
layer values
Hidden layer 
deviation 
value
 
Figure 5: Model training process 
 
The M-AU architecture innovatively integrates 
hierarchical financial risk tools and focusing techniques 
in the latest financial data analysis and expectation 
framework. This model has new insights in financial 
information assessment and trend prediction. This mode 
focuses on decoding continuous data and can identify the 
main characteristics closely related to prediction 
performance, thereby improving prediction accuracy and 
reliability. M-AU processes financial information with 
meticulous techniques under this approach and can tailor 
precise enterprise value evaluations for fund investors. 
M-AU adopts a dual scale filter and analyzes the input 
financial information layer by layer in the structural 
design. Therefore, the richness and complexity of the data 

22   Informatica 48 (2024) 15 –32                                                            Y . Liu et al. 
can be ensured while mining the information. This stage 
is crucial in capturing patterns that may be overlooked, 
such as using filters of different scales to simultaneously 
reveal short-term and long-term financial trends, adding a 
three-dimensional perspective to the decision-making 
process. BGRU is used to analyze historical and potential 
trends after the financial risk analysis. BGRU provides a 
comprehensive analytical perspective for understanding 
and predicting complex financial time series and 
successfully captures the temporal dependencies of data. 
The introduction of a focus layer can improve the pattern 
prediction function. Hierarchy, as a filter in the network, 
emphasizes the recognition of important features by 
weighting the feature vectors of convolutional layers. 
This indicates that in decision-making, patterns can 
automatically identify the most relevant data without 
relying on artificially set feature importance. Therefore, 
the mode's ability to adjust and formulate strategies can 
be significantly enhanced in dealing with complex 
backgrounds. The network architecture, as an ultimate 
output, presents the superimposed characteristics as 
predicted values through the combination of compact 
layers and softmax incentive function. Then the values 
are further transformed into net profit prediction for the 
next cycle. The output results provides potential 
economic performance predictions for fund operators and 
decision-makers. Meanwhile, potential risks can be 
quantitatively measured and returned on funds. Table 3 
shows the structural parameters. 
 
Table 3: Structural parameters of profit prediction model 
Number of 
layers 
Layer name Nuclear size Output size 
Parameter 
quantity 
0 Input layer - (38, 1) 0 
1 
One-dimensional convolutional 
layer 1 
5×5 (38, 4) 24.5 
1 
One-dimensional convolutional 
layer 2 
3×3 (38, 4) 16.8 
2 
One-dimensional convolutional 
layer 3 
5×5 (38, 8) 169.2 
2 
One-dimensional convolutional 
layer 4 
3×3 (38, 8) 105.6 
3 Merge layers - (38, 16) 0 
4 Bidirectional GRU layer 1 64 (16, 128) 39608 
5 Bidirectional GRU layer 2 64 (16, 128) 74256 
6 Fully connected layer 1 - (128, 16) 275.5 
7 Attention mechanism layer - (16, 128) 0 
8 Flatten layer - (1, 2048) 0 
9 Fully connected layer 2 - (1, 1) 2051.3 
 
The exquisite design of M-AU provides efficient 
analytical tools for financial analysis professionals at the 
practical application level. Therefore, a thorough 
understanding of the economic situation and revenue 
potential of enterprises can be enhanced. Overall, the 
economic benefit estimation framework based on M-AU 
continuously analyzes and practices the latest data 
analysis technologies to enhance the certainty of 
economic analysis. A framework is formed by combining 
multi-level convolutional networks, bidirectional control 
loops, and focusing mechanisms. This framework can 
efficiently process time series data, identify complex 
economic patterns, and accurately predict the company's 
future economic performance. The M-AU prediction 
mechanism is expected to transform into a disruptive tool 
in the financial industry with the increase of data and the 
continuous refinement of patterns. This provides 
theoretical support for business strategy formulation and 
opens up new perspectives for financial investors in risk 
and return assessment. 
First, the data are divided into training and test sets 
according to 70% and 30% arm strength to ensure that the 
model has sufficient generalization ability for unseen data. 
In addition, the study introduces L2 regularization 
technique to reduce the risk of overfitting. Regularization 
works by adding a penalty term to the loss function, 
which can limit the weight of the model and avoid 
excessive model complexity. The early stop method is 
used to control overfitting. This means that if the loss of 
the validation set does not improve within 10 consecutive 
training cycles, the training process will be stopped. 
Indicators closely related to the recent financial 
health and liquidity ability of the enterprise are selected 
in the risk prediction, such as asset liability, current, and 
quick ratios. Because these indicators can fully reflect the 
solvency and cash flow status of the enterprise. These 
indicators are closely related to enterprise risks, which 
can reflect the enterprise's risk situation in real-time. 
Indicators directly related to the company's profitability 
are selected in the profit forecasting, such as revenue 
growth rate, gross profit margin, net profit margin, etc. 
These indicators can directly reflect the company's 
profitability and potential, which can predict the 
company's future profitability performance. 

Profit Estimation Model and Financial Risk Prediction Combining … Informatica 48 (2024) 15 –32 23 
 
 
 
 
3 Effectiveness of financial risk 
prediction and profit prediction 
models 
The performance analysis of financial risk prediction and 
profit estimate model included three parts. Firstly, the 
performance of financial risk prediction was analyzed. 
Secondly, the performance of the profit estimate model 
was analyzed. Finally, a comprehensive prediction effect 
analysis was conducted using actual enterprise cases. 
3.1 Financial risk prediction effect 
The aim of this experiment is to analyze and predict data 
from Special Treatment (ST) and non-ST companies 
using machine learning models. The study divided the 
dataset into training and testing sets based on standard 
proportions to effectively evaluate the performance of the 
model. Table 4 shows the experimental setup. 
 
Table 4: Structural parameters of profit prediction model 
Dataset settings 
Dataset categories Number of ST companies Number of non-ST companies Total 
Training set 252 11500 11752 
Test set 85 4760 4845 
Total 335 16260 16595 
Parameter settings 
Set project Detail settings 
Dataset partitioning ratio Training set: Test set=7:3 
Learning rate 0.001 
Optimizer Adam 
Training rounds 50 
Dropout application layer Fully connected layer 
 
In Table 4, the training set includes 252 data points 
from ST company and 11500 data points from non-ST 
company, totaling 11752 samples. The test set includes 
85 data points from ST company and 4760 data points 
from non-ST company, totaling 4845 samples. The entire 
dataset consists of 335 ST company data points and 
16260 non-ST company data points, totaling 16595 
samples. In addition, the Dropout technique was 
introduced in the fully connected layer of the model to 
improve the generalization ability of the model and 
prevent overfitting. This means that a certain proportion 
of neurons will be randomly discarded during training. 
This can increase the sparsity of the model structure, 
forcing the model to learn effectively through a small 
number of features, thereby improving its performance on 
unseen data. The models used for comparison are as 
follows: CNN, GRU, Multi-Scale Gated Recurrent Unit 
(M-GU), Support Vector Machine (SVM), FastText, 
Deep Long Short-Term Memory (D-LSTM), 
Character-Level Convolutional Neural Network 
(Char-CNN), Character-Level Convolutional Recurrent 
Neural Network(Char-CRNN), Word-Level 
Convolutional Neural Network with 
Attention(Word-CNN-Att), Deep Recurrent Neural 
Network (DRNN), and Deep Bidirectional Gated 
Recurrent Unit (D-BU).  
First, the study collected data from multiple 
financial databases and publicly available company 
information for ST and non-ST companies in the 
experimental process. The focus was on financial 
indicators during the data collection process, such as debt 
ratio and current ratio, and data cleaning, preprocessing. 
Meanwhile, feature selection steps were carried out. Then 
the experiment began by using the Adam optimizer and 
setting appropriate learning rates and training rounds. The 
experiment was divided into three stages. Firstly, the 
learning mode and generalization ability of the model 
were trained and tested, followed by financial risk 
prediction. Secondly, financial indicators were accurately 
predicted and profit forecasts were made through 
parameter adjustments and test data validation. Finally, J 
company was chosen as the empirical research object to 
evaluate the actual effectiveness of predictions. 
Figure 6 shows the parameter variation effect. 

24   Informatica 48 (2024) 15 –32                                                            Y . Liu et al. 
2 0 4 6
Epoch
0.092
0.094
0.096
0.098
0.100
0.102
0.104
0.106
8 10
(b) Change in loss function
2 0 4 6
Epoch
0.97800
0.97825
0.97850
0.97875
0.97900
0.97925
0.97950
0.97975
Accuracy
8 10
(c) Change in accuracy
0.2 0 0.4 0.6
Dropout
90
91
92
93
94
96
97
98
0.8 1.0
(a) Dropout value impact
Accuracy
Trough 
point
Accuracy
 
Figure 6: Parameter variation effect 
 
In Figure 6, the loss function values in the database 
showed a significant decreasing trend as the iteration 
increased. This indicated that the model's predictive 
ability gradually improved in consistency with the actual 
situation. Meanwhile, the accuracy of the model also 
significantly improved as the loop tests increased. The 
proportion of values close to 1 continued to increase, 
further confirming the improvement in the performance 
of the research model. In addition, the classification 
accuracy was the highest when the dropout value was 0.4, 
reaching 98.03%. Table 5 shows a comparison of vertical 
improvements. 
 
Table 5: Vertical improvement comparison 
Model Data set Accuracy F1 score 
Number of 
parameters 
Training time 
Memory 
footprint 
SVM A 0.9498 0.945 15000 100.50s 500MB 
SVM B 0.9376 0.938 15000 95.75s 500MB 
CNN A 0.9672 0.967 250000 25.09s 1.2GB 
CNN B 0.9549 0.955 250000 30.17s 1.2GB 
GRU A 0.9532 0.953 80000 197.78s 750MB 
GRU B 0.9485 0.948 80000 205.89s 750MB 
BGRU A 0.9702 0.97 120000 349.05s 1GB 
BGRU B 0.9667 0.966 120000 360.11s 1GB 
M-GU A 0.9742 0.974 100000 237.45s 900MB 
M-GU B 0.9689 0.968 100000 245.96s 900MB 
M-BU A 0.9803 0.98 150000 377.93s 1.1GB 
M-BU B 0.9757 0.975 150000 390.07s 1.1GB 
 
In Table 5, M-BU achieved an accuracy of 98.03% 
on dataset A, which was superior to other models such as 
BGRU's 97.02%, GRU's 95.32%, and the classical CNN 
model's 96.72%. M-BU also performed well on dataset B, 
achieving an accuracy of 97.57%. M-BU still maintained 
a leading position compared to BGRU's 96.67%, GRU's 
94.85%, and CNN's 95.49%. This high accuracy meant 
that M-BU more accurately identified potential risk 
points when predicting financial risk. In addition, M-BU 
also performed excellently in F1 scores, reaching 0.980 
and 0.975 on datasets A and B, respectively, significantly 
higher than other models. This indicated that M-BU had 

Profit Estimation Model and Financial Risk Prediction Combining … Informatica 48 (2024) 15 –32 25 
higher reliability in predicting positive and negative 
samples. The training time of M-BU was longer 
compared to other models, with 377.93 seconds and 
390.07 seconds on datasets A and B, respectively. This 
reflected the additional time investment required for 
model training, partly due to the predictable complexity 
of the M-BU structure compared to general GRU. 
However, this additional time investment was reasonable 
and beneficial considering that training was a one-time 
cost and the model was used multiple times after 
deployment. M-BU had 150000 parameters in terms of 
parameter quantity, slightly more than other competitive 
models such as GRU's 80000. The larger number of 
parameters corresponded to the learning ability of the 
model, increasing the risk of overfitting and the demand 
for computing resources. M-BU was approximately 
1.1GB in terms of memory usage. This indicated that 
M-BU could perform effective training with certain 
computing resources. Figure 7 shows a horizontal 
comparison. 
68
70
72
74
76
78
Types
Accuracy (%)
(e) Yah. A.
82
84
86
88
90
92
94
96
98
FastText
D-LSTM
Char-CNN
Char-CRNN
Word-CNN-Att
DRNN
D-BGRU
M-BU
Accuracy (%)
Types
(a) AG
94
95
96
97
98
99
100
Accuracy (%)
Types
(b) DBP
FastText
D-LSTM
Char-CNN
Char-CRNN
Word-CNN-Att
DRNN
D-BGRU
M-BU
90
91
92
93
94
95
96
97
98
99
Accuracy (%)
(c) Yelp P.
Types
FastText
D-LSTM
Char-CNN
Char-CRNN
Word-CNN-Att
DRNN
D-BGRU
M-BU
56
58
60
62
64
66
68
70
72
Accuracy (%)
Types
(d) Yelp R.
FastText
D-LSTM
Char-CNN
Char-CRNN
Word-CNN-Att
DRNN
D-BGRU
M-BU
FastText
D-LSTM
Char-CNN
Char-CRNN
Word-CNN-Att
DRNN
D-BGRU
M-BU
 
Figure 7: Horizontal comparison 
 
In Figure 7, the performance of M-BU was excellent 
on different datasets in a series of recent horizontal 
comparison analysis of models. M-BU not only 
demonstrated excellent processing capabilities compared 
to other cutting-edge technologies in comparative studies, 
but also reflected unique insights into prediction accuracy. 

26   Informatica 48 (2024) 15 –32                                                            Y . Liu et al. 
M-BU achieved a classification accuracy of 95.7% in the 
AG dataset, surpassing the 95.2% of D-BU in the same 
series. M-BU had shown enormous potential in the DBP 
dataset, leading with an accuracy of 99.6%, which was 
more accurate than D-BU's 99.5%. M-BU achieved an 
accuracy of 98.0% in the Yelp P. dataset, which was the 
highest among all comparison models and significantly 
better than the DRNN model's 97.3%. The improvement 
of this data indicated that M-BU had better recognition 
ability when dealing with complex text content. M-BU 
also demonstrated strong predictive performance in the 
Yelp R. dataset, with a classification accuracy of 70.5%, 
surpassing D-BU's 69.3%. This indicated that M-BU 
could more accurately capture semantic details and user 
emotional tendencies in handling user comments. M-BU 
still maintained a leading position in the Yah. A. dataset, 
with an accuracy of 77.6% being the best performance 
among all comparison models. This ratio was 
significantly higher than D-BU's 76.7%, demonstrating 
its high applicability and stability in a wide range of 
fields. 
 
3.2 Profit forecast effect 
The study first analyzed the effect of parameter changes 
on M-AU in the profit estimate effect analysis. Figure 8 
shows the variation pattern of the loss function value with 
increasing iteration. 
 
2 0 4 6
Epoch
-0.10
0.00
0.10
0.20
0.30
0.40
0.50
0.60
Loss
8 10
(b) MS-ABGRU
2 0 4 6
Epoch
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
Loss
8 10
(c) MS-BGRUA
0.2 0 0.4 0.6
Epoch
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.8 1.0
(a) Dropout value impact
Mse
Trough point
Smooth point
Smooth point
 
Figure 8: The parameter variation effect of M-AU 
 
Figure 8 shows the variation pattern of the loss 
function value with increasing iteration. These data 
confirmed that as iterations increased, the loss function 
value significantly decreased, approaching a small value 
close to 0. This change indicated that this model was 
increasingly approaching the optimal fit to the training 
data. A Mean Square Error (MSE) of 0.0232 was 
obtained by evaluating the performance of the model on 
the test set. This provided an important basis for studying 
and evaluating the prediction accuracy of the model. On 
the other hand, experiments with different dropout values 
showed that they had a significant impact on the 
prediction error of the model. The model reached the 
lowest MSE on the test set especially when the dropout 
value was set to 0.6, which was 0.0349. This result 
suggested that the model could better generalize to 
unseen data and avoid overfitting under this setting. 
Therefore, it was recommended to set dropout to 0.6 to 
optimize model performance under the current 
experimental conditions. Figure 9 shows the horizontal 
comparison of M-AU. 

Profit Estimation Model and Financial Risk Prediction Combining … Informatica 48 (2024) 15 –32 27 
0.6
0.7
0.8
0.9
0.5
100 200 300 400
D-LSTM
F1
1.0
500 600
(a) AG
Word-CNN-Att
Char-CRNN
Epochs
Our model
0.6
0.7
0.8
0.9
0.5
100
20
0
300 400
F1
1.0
500 600
(b) DBP
Epochs
0.6
0.7
0.8
0.9
0.5
100 200 300 400
F1
1.0
500 600
(c) Yelp P.
Epochs
0.6
0.7
0.8
0.9
0.5
100 200 300 400
F1
1.0
500 600
(d) Yelp R.
Epochs
D-LSTM
Word-CNN-Att
Char-CRNN
Our model
D-LSTM
Word-CNN-Att
Char-CRNN
Our model
D-LSTM
Word-CNN-Att
Char-CRNN
Our model
 
Figure 9: Horizontal comparison of M-AU 
 
In Figure 9, the design model consistently 
maintained an advantage in F1 values among these four 
subgraphs representing different datasets, with the curve 
located above the F1 value curve of the other models. 
Therefore, the design model performed better on 
commonly used datasets. Table 6 shows a comparison of 
the vertical improvements of M-AU. 
 
 
Table 6: Comparison of M-AU vertical improvements 
Dataset 
name 
Model MSE MAE MAD RMSE 
R² 
score 
Accuracy Recall 
F1 
score 
Financial 
dataset A 
CNN 0.1978 0.5684 0.5044 0.4447 0.81 0.75 0.72 0.735 
Financial 
dataset A 
GRU 0.1898 0.5447 0.484 0.4356 0.83 0.78 0.74 0.76 
Financial 
dataset A 
BGRU 0.1547 0.4439 0.3944 0.3933 0.86 0.82 0.79 0.805 
Financial 
dataset A 
M-BU 0.1089 0.3125 0.2777 0.33 0.9 0.85 0.83 0.84 
Financial 
dataset A 
M-AU 0.0232 0.1191 0.1058 0.1523 0.96 0.92 0.91 0.915 
Financial 
dataset B 
CNN 0.2056 0.5792 0.5164 0.4533 0.78 0.73 0.7 0.715 
Financial 
dataset B 
GRU 0.1945 0.5512 0.4998 0.441 0.81 0.76 0.73 0.745 
Financial 
dataset B 
BGRU 0.1602 0.4589 0.4129 0.4002 0.85 0.81 0.78 0.795 
Financial 
dataset B 
M-BU 0.1137 0.3291 0.2905 0.3372 0.88 0.83 0.81 0.82 
Financial 
dataset B 
M-AU 0.0263 0.1256 0.1128 0.1621 0.94 0.9 0.88 0.89 
 
  

28   Informatica 48 (2024) 15 –32                                                            Y . Liu et al. 
In Table 6, the comparison indicators are MSE, 
Mean Absolute Error (MAE), Mean Absolute Deviation 
(MAD), and Root Mean Squared Error (RMSE). M-AU 
was only 0.0232 on MSE on financial dataset A, which 
was about 88% lower than CNN's 0.1978. This 
demonstrated its high accuracy in estimating true profit 
values. The MAE also significantly decreased from 
CNN's 0.5684 to 0.1191, further confirming that M-AU 
had higher prediction accuracy at the average level. 
Similarly, M-AU led the MAD index at 0.1058, while 
CNN was 0.5044. This trend was equally evident in 
RMSE and R2 scores. The RMSE of M-AU reached 
0.1523, which was much lower than CNN's 0.4447, 
reflecting that M-AU had smaller errors in predicting 
various data points. Meanwhile, the R2 score of M-AU 
reached 0.96, close to perfect predictive ability, far 
exceeding CNN's 0.81. This demonstrated that M-AU 
was significantly better than other models in overall data 
fitting. In terms of financial dataset B, the performance of 
all models has decreased compared to dataset A. This 
might be attributed to the complexity of dataset B, but 
M-AU continued to maintain its superiority. The MSE 
and MAE of M-AU were 0.0263 and 0.1256, respectively, 
slightly increasing compared to dataset A, but still 
leading other models. M-AU demonstrated strong 
performance in classification performance indicators such 
as accuracy, recall, and F1 score on both datasets. The 
accuracy of M-AU reached 0.92 and 0.9, recall rates were 
0.91 and 0.88, and F1 scores were 0.915 and 0.89, 
respectively. These were all higher than other models. 
For example, the accuracy of CNN on dataset A was 0.75, 
the recall was 0.72, and the F1 score was 0.735, which 
highlighted the significant improvement of M-AU in 
classification accuracy. 
 
3.3 Comprehensive case analysis 
The J company selected for the study is listed electronic 
technology Co., Ltd., which focuses on the manufacturing 
of high-end consumer electronics products and has 
approximately 200 employees. This company's main 
products include smartphones and wearable devices to 
meet the market's demand for high-quality electronic 
devices. Table 7 shows the case results. 
 
 
Table 7: Case results 
Financial risk prediction 
Risk indicators Description Estimation Prediction interval Remarks 
Financial risk level 
Model-based risk 
classification 
Low risk - 
Risk is effectively 
managed 
Asset liability ratio 
The percentage of 
liabilities to assets 
0.45 43%-47% Control at a safe level 
Current ratio 
Display short-term 
solvency 
2.1 2.0-2.2 Good short-term liquidity 
Quick ratio 
Exclude the solvency 
of inventory 
1.5 1.4-1.6 
Quick assets in good 
condition 
Interest coverage ratio 
Operate profit before 
interest expenses 
6 5.5-6.5 
Interest payment without 
pressure 
Cash flow ratio 
Cash flow to current 
debt ratio 
1.2 1.1-1.3 Good cash flow situation 
Profit forecast 
Profit indicators Description 
Predictive 
value 
Prediction interval Remarks 
Revenue growth rate 
Expected annual 
revenue growth rate 
0.05 4%-6% Stable business growth 
Return on equity 
Return on equity to 
shareholders 
0.1 9%-11% Good investment return 
Total asset turnover 
rate 
The efficiency of 
generating revenue 
from total assets 
0.8 0.7-0.9 
Reasonable asset 
utilization efficiency 
Net profit attributable 
to shareholders of the 
parent company 
Actual net profit 
attributable to 
shareholders 
32 million 
yuan 
31 million to 33 
million yuan 
Expected good 
profitability 
Operating cost ratio 
The proportion of 
operating costs to 
operating revenue 
0.85 80%-90% 
Reasonable gross profit 
margin 
Earnings per share Earnings per share 1.5yuan 1.4-1.6 yuan Good earnings per share 
 

Profit Estimation Model and Financial Risk Prediction Combining … Informatica 48 (2024) 15 –32 29 
In Table 7, the financial risk forecast showed a debt 
level of 45%, which was at the center of the forecast 
range of 43%-47%. This indicated that companies were 
cautious in the application of financial leverage, with 
lower risks, which were effectively managed. The current 
ratio and quick ratio were 2.1 and 1.5, respectively, both 
within their predicted ranges of 2.0-2.2 and 1.4-1.6. This 
indicated that both short-term liquidity and quick assets 
were in good condition, indicating that the company had 
sufficient ability to cope with short-term debt. In addition, 
the interest coverage ratio was 6 times, safely within the 
predicted range of 5.5-6.5. This meant that the enterprise 
could easily pay interest expenses with operating profits. 
The cash flow ratio was 1.2, which fell within the 
predicted range of 1.1-1.3. This indicated that the 
enterprise had healthy cash flow and was conducive to 
ensuring smooth daily operations and investment 
activities. The predicted revenue growth rate was 5% in 
the profit estimate, which was consistent with the 
predicted range of 4%-6%. This demonstrated stable 
business growth for the enterprise. The return on equity 
was 10%, accurately falling within the expected range of 
9%-11%, proving that investors could expect a good 
return on investment. The total asset turnover rate was 0.8, 
which was within the predicted range of 0.7-0.9. This 
indicated that the enterprise had a reasonable asset 
utilization efficiency. The predicted value of 32 million 
reflected the expected good profitability of the company 
for the net profit attributable to shareholders of the parent 
company. This prediction was close to the edge of its 
prediction range of 31 million to 33 million. The 
operating cost rate was controlled at 85%, with a 
predicted range of 80%-90%, maintaining a reasonable 
gross profit margin. Finally, the earnings per share 
forecast was 1.5 yuan, within the forecast range of 1.4 to 
1.6 yuan, indicating a good earnings per share. In 
summary, the financial forecasting model applied 
provided excellent and reliable predictions of financial 
risk and profitability. These predicted values matched the 
actual indicators and were mostly within a safe prediction 
range. This indicated that the model had high accuracy 
and reliability in understanding the financial health status 
of enterprises. Enterprises could adjust their strategies in 
a timely manner, manage potential risks, and ensure 
sustained positive growth through these data-driven 
insights. 
 
 
Table 8: Cost calculation and scalability analysis 
Characteristic Dataset size Lightweight transformer Model 
Training time 
Small (~10000 samples) 2 hours 1 hour 
Medium (~100000 samples) 12 hours 6 hours 
Large (>1000000 samples) >24 hours 12 hours 
Resource 
requirements 
CPU core 4-core 2-core 
RAM 8GB 4GB 
Recommended GPU NVIDIA GTX 1080 NVIDIA GTX 1060 
 
From Table 8, the designed model had significant 
advantages in training time compared to the lightweight 
transformer. Meanwhile, this model achieved efficient 
operation on lower computational configurations, which 
was more advantageous. 
4 Discussion 
Li X et al. used an optimized BPNN to predict the 
financial risks of listed companies. The optimized BPNN 
improved the accuracy of the prediction, with a prediction 
accuracy rate of over 80% [31]. This study is also based 
on neural networks. Moreover, more complex data can be 
handled using bidirectional gated recursive units to 
capture temporal relationships in time series data. Uddin 
MS et al. used an optimized random forest method to 
predict the credit risk of micro enterprises, which was 
more targeted but less common [32]. Meanwhile, this 
study combined multi-scale feature extraction methods to 
effectively extract and predict various types of enterprise 
information, which was more universal. Barra S et al. 
focused on the temporal nature of information similar to 
the approach of this study. Time-series encoding methods 
combined with CNN were used to transform time-series 
data into Gram angle field images for analysis [33]. The 
designed model had a simpler execution and stronger 
robustness. Uthayakumar J et al. focused on predicting 
financial crises by combining improved K-means 
clustering and chaotic genetic ant colony algorithm [34]. 
On the other hand, Bianchi D focused on bond risk by 
combining extreme tree and neural network models to 
analyze the likelihood of bond returns [35]. Both studies 
focused on one aspect of corporate risk, such as financial 
risk. On the side of bond risk, this study takes the 
perspective of enterprises as the main body to conduct 
integrated prediction analysis of enterprise risk and profit. 
The designed model performed better in classification 
and efficiency indicators than other models compared 
with the model of Li X et al. and Uddin MS et al. This 
demonstrated higher prediction accuracy, as well as more 
comprehensive and stable information processing. 
 
5 Conclusion 
A novel financial risk prediction model based on graph 
network concept was designed, which combined 
multi-scale feature extraction technology with sequence 
analysis methods. A model based on multi-scale 

30   Informatica 48 (2024) 15 –32                                                            Y . Liu et al. 
advantage and attention mechanism was also proposed 
for profit estimate. These results confirmed that the 
prediction accuracy of these two models was significantly 
improved compared to traditional financial prediction 
models. The model's accuracy had significantly improved 
after 50 rounds of training, with the highest classification 
accuracy reaching 98.03%. The financial risk prediction 
model achieved an accuracy of 98.03% and an F1 score 
of 0.980 on dataset A. The error was close to 0 on the 
profit estimate model, and the MSE was 0.0232. This 
indicated that the model could more accurately identify 
potential risks and profit points when predicting risks and 
profits. In the case study, the debt level was 45% for the 
financial risk forecast, the current ratio and quick ratio 
were 2.1 and 1.5, respectively. The actual interest 
coverage ratio and cash flow ratio were within the 
forecast range. The net profit attributable to the 
shareholders of the parent company was expected to be 
32 million in terms of profit estimate, which almost 
matched the actual indicators. Therefore, the designed 
comprehensive model is an efficient and feasible method. 
The designed model integrates risk prediction and 
profit prediction has strong predictive ability. A 
prediction system is formed for multi-feature data and 
their time characteristics. However, there are many 
differences in economic structure, market mechanisms, 
policy environment, etc. between different regions in 
practical applications. Therefore, the specific 
environment in which the model is applied may also vary, 
leading to prediction bias in universal models under 
specific conditions. Therefore, it is necessary to 
modularize the different functions of the prediction model. 
Meanwhile, targeted parameter setting and application 
solutions can be developed for different environments in 
the future. 
References 
[1] A. Lin, N. Manral, P. McElhinney, K. Aditya, H. 
Matsumoto, and J. Kwiecinski, "Deep 
learning-enabled coronary CT angiography for 
plaque and stenosis quantification and cardiac risk 
prediction: an international multicentre study," The 
Lancet Digital Health, vol. 4, no. 4, pp. e256-e265, 
2022. 
https://doi.org/10.1016/S2589-7500(22)00022-X. 
[2] W. Gleißner, T. Günther, and C. Walkshäusl, 
"Financial sustainability: measurement and 
empirical evidence," Journal of Business Economics, 
vol. 92, no. 3, pp. 467-516, 2022. 
https://doi.org/10.1007/s11573-022-01081-0. 
[3] R. Patel, S. Tanwani, and C. Patidar, "Relation 
extraction between medical entities using deep 
learning approach," Informatica, vol. 45, no. 3, pp. 
359-366, 
2021.https://doi.org/10.31449/inf.v45i3.3056. 
[4] A. Kurani, P. Doshi, A. Vakharia, and M. Shah, "A 
comprehensive comparative study of artificial neural 
network (ANN) and support vector machines (SVM) 
on stock forecasting," Annals of Data Science, vol. 
10, no. 1, pp. 183-208, 2023. 
https://doi.org/10.1007/s40745-021-00344-x. 
[5] J. Kong, S. Oh, B. O. Kang, and J. Jung. 
"Development of an incentive model for renewable 
energy resources using forecasting accuracy in 
South Korea," Energy Science & Engineering, vol. 
10, no. 9, pp. 3250-3266, 2022. 
https://doi.org/10.1002/ese3.1020. 
[6] M. Salb, M. Zivkovic, N. Bacanin, A. Chhabra, and 
M. Suresh, "Support vector machine performance 
improvements for cryptocurrency value forecasting 
by enhanced sine cosine algorithm," Computer 
Vision and Robotics, pp. 527-536, 2022. 
https://doi.org/10.1007/978-981-16-8225-4_40. 
[7] F. Liu, Y. Wu, N. Almarri, M. Habibollahi, H. T. 
Lancashire, B. Bryson, L. Greensmith, D. Jiang, and 
A. Demosthenous, "A fully implantable opto-electro 
closed-loop neural interface for motor neuron 
disease studies," IEEE Transactions on Biomedical 
Circuits and Systems, vol. 16, no. 5, pp. 752-765, 
2022. 
https://doi.org/10.1109/TBCAS.2022.3202026. 
[8] V. K. Tewari and A. Verma, "Design and application 
of double-gate MOSFET in two loops controlled 
multi-input DC-DC converter," Silicon, vol. 14, no. 
12, pp. 7321-7333, 2022. 
https://doi.org/10.1007/s12633-021-01469-7. 
[9] P. Preethi and H. R. Mamatha, "Region-based 
convolutional neural network for segmenting text in 
epigraphical images," Artificial Intelligence and 
Applications, vol. 1, no. 2, pp. 119-127, 2023. 
https://doi.org/10.47852/bonviewAIA2202293. 
[10] X. Li, J. Wang, and C. Yang, "Risk prediction in 
financial management of listed companies based on 
optimized BP neural network under digital 
economy," Neural Computing and Applications, vol. 
35, no. 3, pp. 2045-2058, 2023. 
https://doi.org/10.1007/s00521-022-07377-0. 
[11] S. Peng, S. Yang, and J. Yao, "Improving 
value-at-risk prediction under model uncertainty," 
Journal of Financial Econometrics, vol. 21, no. 1, pp. 
228-259, 2023. 
https://doi.org/10.1093/jjfinec/nbaa022. 
[12] Y. Cao, Y. Shao, and H. Zhang, "Study on early 
warning of E-commerce enterprise financial risk 
based on deep learning algorithm," Electronic 
Commerce Research, vol. 22, no. 1, pp. 21-36, 2022. 
https://doi.org/10.1007/s10660-020-09454-9. 
[13] U. M. Sirisha, M. C. Belavagi, and G. Attigeri, 
"Profit prediction using Arima, Sarima and LSTM 
models in time series forecasting: A Comparison," 
IEEE Access, vol. 10, pp. 124715-124727, 2022. 
https://doi.org/10.1109/ACCESS.2022.3224938. 
[14] G. Kim, D. H. Shin, J. G. Choi, and S. Lim, "A deep 
learning-based cryptocurrency price prediction 
model that uses on-chain data," IEEE Access, vol. 

Profit Estimation Model and Financial Risk Prediction Combining … Informatica 48 (2024) 15 –32 31 
10, pp. 56232-56248, 2022. 
https://doi.org/10.1109/ACCESS.2022.3177888. 
[15] M. Shajalal, P. Hajek, and M. Z. Abedin, "Product 
backorder prediction using deep neural network on 
imbalanced data," International Journal of 
Production Research, vol. 61, no. 1, pp. 302-319, 
2023. 
https://doi.org/10.1080/00207543.2021.1901153. 
[16] S. He, X. Sang, J. Yin, Y. Zheng, and H. Chen, 
"Short-term runoff prediction optimization method 
based on bgru-bp and blstm-bp neural networks," 
Water Resources Management, vol. 37, no. 2, pp. 
747-768, 2023. 
https://doi.org/10.1007/s11269-022-03401-z. 
[17] M. D. Dangut, I. K. Jennions, S. King, and Z. Skaf, 
"A rare failure detection model for aircraft 
predictive maintenance using a deep hybrid learning 
approach," Neural Computing and Applications, vol. 
35, no. 4, pp. 2991-3009, 2023. 
https://doi.org/10.1007/s00521-022-07167-8. 
[18] H. Li, Q. He, and L. Wu, "Detection of brain 
abnormalities in parkinson ’s rats by combining deep 
learning and motion tracking," IEEE Transactions 
on Neural Systems and Rehabilitation Engineering, 
vol. 31, no. 1, pp. 1001-1007, 2023. 
https://doi.org/10.1109/TNSRE.2023.3237916. 
[19] P. Weston and M. Nnadi, "Evaluation of strategic 
and financial variables of corporate sustainability 
and ESG policies on corporate finance 
performance," Journal of Sustainable Finance & 
Investment, vol. 13, no. 2, pp. 1058-1074, 2023. 
https://doi.org/10.1080/20430795.2021.1883984. 
[20] S. Wahab, M. Imran, A. Safi, and D. Kirikkaleli, 
"Role of financial stability, technological innovation, 
and renewable energy in achieving sustainable 
development goals in BRICS countries," 
Environmental Science and Pollution Research, vol. 
29, no. 32, pp. 48827-48838, 2022. 
https://doi.org/10.1007/s11356-022-18810-1. 
[21] H. N. H. Al-Hashimy, T. T. Y. Alabdullah, E. Ries, 
M. A. Ahmed, M. I. Nor, and K. A. M. Jamal, "The 
impact of financial management elements and 
behavioral intention on the financial performance," 
International Journal of Scientific and Management 
Research, vol. 5, no. 12, pp. 117-149, 2022. 
https://doi.org/10.37502/JSMR.2022.51210. 
[22] Y. Si, S. Heeringa, D. Johnson, R. J. A. Little, W. 
Liu, F. Pfeffer, and T. Raghunathan, "Multiple 
imputation with massive data: An application to the 
panel study of income dynamics," Journal of Survey 
Statistics and Methodology, vol. 11, no. 1, pp. 
260-283, 2023. 
https://doi.org/10.1093/jssam/smab038. 
[23] Z. Zhao, L. Tang, M. Fang, X. Yang, C. Li, and Q. 
Li, "Toward urban traffic scenarios and more: a 
spatio-temporal analysis empowered low-rank 
tensor completion method for data imputation," 
International Journal of Geographical Information 
Science, vol. 37, no. 9, pp. 1936-1969, 2023. 
https://doi.org/10.1080/13658816.2023.2234434. 
[24] F. Conceição, C. H. Antunes, M. Gomes, V. Silva, 
and R. Dinis, "Max-min fairness optimization in 
uplink cell-free massive MIMO using 
meta-heuristics," IEEE Transactions on 
Communications, vol. 70, no. 3, pp. 1792-1807, 
2022. 
https://doi.org/10.1109/TCOMM.2022.3144989. 
[25] Y. K. Saheed, A. I. Abiodun, S. Misra, A. S. Albahri, 
A. H. Alamoodi, B. B. Zaidan, S. Qahtan, H. A. 
Alsatar, M. S. Al-Samarraay, and A. N. Jasim, "A 
machine learning-based intrusion detection for 
detecting internet of things network attacks," 
Alexandria Engineering Journal, vol. 61, no. 12, pp. 
9395-9409, 2022. 
https://doi.org/10.1109/JBHI.2022.3167256. 
[26] V. Gómez-Escalonilla, P. Martínez-Santos, and M. 
Martín-Loeches, "Preprocessing approaches in 
machine-learning-based groundwater potential map: 
an application to the Koulikoro and Bamako regions, 
Mali," Hydrology and Earth System Sciences, vol. 
26, no. 2, pp. 221-243, 2022. 
https://doi.org/10.5194/hess-26-221-2022. 
[27] Y. Shi, X. Nie, X. Liu, L. Yang, and Y. Yin, 
"Zero-shot hashing via asymmetric ratio similarity 
matrix," IEEE Transactions on Knowledge and Data 
Engineering, vol. 35, no. 5, pp. 5426-5437, 2022. 
https://doi.org/10.1109/TKDE.2022.3150790. 
[28] R. Nirthika, S. Manivannan, A. Ramanan, and R. 
Wang, "Pooling in convolutional neural networks 
for medical image analysis: a survey and an 
empirical study," Neural Computing and 
Applications, vol. 34, no. 7, pp. 5321-5347, 2022. 
https://doi.org/10.1007/s00521-022-06953-8. 
[29] J. J. Liu, Q. Hou, Z. A. Liu, and M. M. Cheng, 
"Poolnet+: Exploring the potential of pooling for 
salient object detection," IEEE Transactions on 
Pattern Analysis and Machine Intelligence, vol. 45, 
no. 1, pp. 887-904, 2022. 
https://doi.org/10.1109/TPAMI.2021.3140168. 
[30] U. Nandi, A. Ghorai, M. M. Singh, C. Changdar, S. 
Bhakta, and R. K. Pal, "Indian sign language 
alphabet recognition system using CNN with 
diffGrad optimizer and stochastic pooling," 
Multimedia Tools and Applications, vol. 82, no. 7, 
pp. 9627-9648, 2023. 
https://doi.org/10.1007/s11042-021-11595-4. 
[31] X. Li, J. Wang, and C. Yang, "Risk prediction 
financial management of listed companies based on 
optimized BP neural network under digital 
economy," Neural Computing and Applications, vol. 
35, no. 3, pp. 2045-2058, 2023. 
https://doi.org/10.1007/s00521-022-07377-0. 
[32] M. S. Uddin, G. Chi, M. A. M. Al Janabi, and T. 
Habib, "Leveraging random forest in 
micro-enterprises credit risk modelling for accuracy 
and interpretability," International Journal of 

32   Informatica 48 (2024) 15 –32                                                            Y . Liu et al. 
Finance & Economics, vol. 27, no. 3, pp. 3713-3729, 
2020. https://doi.org/10.1002/ijfe.2346. 
[33] S. Barra, S. M. Carta, A. Corriga, A. S. Podda, and 
D. R. Recupero, "Deep learning and time 
series-to-image encoding for financial forecasting," 
IEEE/CAA Journal of Automatica Sinica, vol. 7, no. 
3, pp. 683-692, 2020. 
https://doi.org/10.1109/JAS.2020.1003132. 
[34] J. Uthayakumar, N. Metawa, K. Shankar, and S. K. 
Lakshmanaprabu, "Intelligent hybrid model for 
financial crisis prediction using machine learning 
techniques," Information Systems and e-Business 
Management, vol. 18, pp. 617-645, 2018. 
https://doi.org/10.1007/s10257-018-0388-9. 
[35] D. Bianchi, M. Büchner, and A. Tamoni, "Bond risk 
premiums with machine learning," The Review of 
Financial Studies, vol. 34, no. 2, pp. 1046-1089, 
2020. https://doi.org/10.1093/rfs/hhaa062.