Neural Networks Usability Analysis in Economic Security Indicators
Dynamics Forecasting
Anna G. Shmeleva
a
, Evgenii S. Mityakov
b
and Andrey I. Ladynin
c
Institute of Integrated Safety and Special Instrument Engineering, Informatics department, MIREA – Russian Technological
University, 78 Vernadsky Avenue, 119454, Moscow, Russia
Keywords: Economic security, Economic security indicators, Neural networks, Time series forecasting, Machine learning.
Abstract: The paper is devoted to neural network applicability analysis in economic security indicators forecasting
models. Exchange indicators usage makes it possible to introduce tools for economic security operational
monitoring, since they largely determine the economic environment dynamics development. Neural networks
training and various models comparative analysis for economic time series values predictions were carried
out using the dollar / ruble exchange rate example. The paper presents two initial data sets analysis with
information content of different amounts. Predictive indicators calculation was carried out using three neural
networks different models and gradient boosting method. The results obtained in the work make it possible
to identify the best neural network model for target indicators prediction, as well as to analyse neural networks
approaches effectiveness’ in the problems under consideration, depending on the dimension of the initial data.
Based on the simulation’s results, there is a conclusion, that neural network methods in economic security
indicators forecasting can be justified only on a significant sample size.
1 INTRODUCTION
Modern technological capabilities open up prospects
for new tools usage regarding economic processes
studies in systems of various hierarchy. It seems
intuitive and scientifically substantiated that
operational forecasting and strategic control require
the automated information systems widespread
implementation, including those based on artificial
intelligence methods and machine learning
algorithms (Tomashevskaya, 2020). Thus, in modern
realities, machine learning methods are highly
involved in solving various economic problems and
are being implemented in almost all spheres of human
life. The corresponding algorithms find their
application both for economic processes in a broad
sense analysis, and for solving specific practical
business problems, becoming one of the management
problems top-notch analysis tools (Shamin, 2019).
One of the key and widely discussed problem
classes, which analysis mechanisms allow modern
machine learning and artificial intelligence tools
a
https://orcid.org/0000-0003-2300-3522
b
https://orcid.org/0000-0001-6579-0988
c
https://orcid.org/0000-0001-7659-2581
usage is economic indicators forecasting, since, as a
rule, the analyzed values are quantitatively set by time
series.
An important task in various hierarchies systems
economic processes studies can undoubtedly be the
ensuring economic security problem. Analysis,
forecasting, monitoring and economic security
management problems regarding national economy
subjects are quite acute nowadays, due to endogenous
and exogenous environment constantly changing
situation. Appropriate scientific and methodological
basis development will allow in a fairly accurate
manner to predict relevant indicators values and, as a
result, make scientifically grounded management
decisions in order to minimize certain processes
negative impact. New challenges and threats dictate
new breakthrough technologies needs in management
tasks at all national economy levels.
Corresponding economic security indicators
operational control, analysis and dynamics
forecasting in order to ensure an appropriate
considering object safety level, along with traditional
Shmeleva, A., Mityakov, E. and Ladynin, A.
Neural Networks Usability Analysis in Economic Security Indicators Dynamics Forecasting.
DOI: 10.5220/0010695400003169
In Proceedings of the International Scientific-Practical Conference "Ensuring the Stability and Security of Socio-Economic Systems: Overcoming the Threats of the Crisis Space" (SES 2021),
pages 149-154
ISBN: 978-989-758-546-3
Copyright
c
2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
149
methods, require modern threats predicting
algorithms usage and effective strategies choice,
oriented on possible consequences prevention
(Shmeleva, 2019). Considering social and economic
processes, world’s digital transformation, computing
systems technological capabilities exponential
growth alongside with databases and information
storages significant development, machine learning
could be effectively used as an approach towards the
raised problem solution (Chio, 2020; Hastie, 2009;
Flah, 2015).
Machine learning mathematical tools have been
increasingly and effectively used in social and
economic processes studies. In various fields of
scientific knowledge, supervised and unsupervised
learning, reinforcement learning methods are widely
used (Bata, 2020). Machine learning methodology in
economic problems is being explored in many
authors’ works. For example, works (Adadi, 2021;
Athey, 2018) analyze machine learning impact on
economy development as a whole, meanwhile article
(Mullainathan, 2017) describes relationships between
economics and "Big Data" technologies.
In economic systems security indicators analysis
and prediction, artificial intelligence methodology
and machine learning is presented mainly at the micro
level, mainly regarding taxation, lending and banking
fields (Andini, 2019; Andini, 2018; Chakraborty,
2017). In macro-level problems studies, named
toolkit is presented mainly in exchange rates and
indicators analysis, while regional and sectoral
economy level has a lesser extent. The paper is
devoted to neural network models set up and its
following usage in economic time series values
prediction comparative analysis based on dollar /
ruble exchange rate historical indicators. Later, this
toolkit can be used to predict other economic security
indicators that have initial data sufficient amount.
2 MATERIALS AND METHODS
Quite often, in order to obtain a forecast, it is required
to solve a regression problem, which consists of
quantitative values predictions based on known data
in the past (Drejper, 2007). Within the economic
security framework studies the regression problem
may appear when building predictive models for
threats occurrence and economic security indicators
values calculation in the foreseeable time interval.
For example it is required to create a model predicting
economic security indicators dynamics based on the
initial data. Building such model can be attributed
towards supervised machine learning task, since
initially there is a retrospective data set (previous
periods indicator values dynamics), thus model
results should consist of indicators predicted values in
a given forecasting horizon. The described problem
structure unambiguously refers to regression methods
solving.
One of the most common and relevant machine
learning methods for forecasting time series are
models based on neural networks mathematical
apparatus. During its operation, algorithms based on
neural networks usage are affected by learning, which
is a key advantage, allowing to be used in processing
and analyzing information. Technically, training is
characterized by finding the model connections
coefficients, identifying the dependencies between
input and output data. The class of mathematical
methods based on neural network modeling is quite
extensive, since neural networks can have different
characteristics, model types, hyperparameters
variable values sets. Consequently, neural network
one and only optimal type selection for predicting
economic dynamics is a complex, relevant scientific
task that goes beyond the scope of this study. In
addition, optimal model choice largely depends on
the amount and dynamics of the initial input data,
imposing additional difficulties in the analysis.
Nevertheless, in this work, three different neural
networks comparative performance analysis, as well
as the gradient boosting method, is carried out.
Solving economic forecasting problems using
neural network methods is often an alternative to
simulation and econometric analysis. At the same
time, it is known that neural network models for
economic processes are often characterized by
redundant information of a significant amount. There
are practically no systematic characteristics of the
described phenomenon essence. The neural network
approach states the fact that it is possible to transform
a set of input parameters into an output variable with
a sufficiently high accuracy. Meanwhile, the input
variables number can be quite significant (several
thousand), and training and tuning a neural network
requires many model experiments. In this regard, the
expediency question usage such tools in specific
analysis practical problems and forecasting in various
levels economic security systems remains open.
In our opinion, the neural networks apparatus
usage is advisable in combination with traditional
econometric analysis methods. Moreover, neural
network mathematical apparatus implementation is
justified only with a significant sample size initial
data, which is far from standard regarding on the
indicators adopted in official documents basis, since
they usually have a measurement frequency of one
SES 2021 - INTERNATIONAL SCIENTIFIC-PRACTICAL CONFERENCE "ENSURING THE STABILITY AND SECURITY OF
SOCIO - ECONOMIC SYSTEMS: OVERCOMING THE THREATS OF THE CRISIS SPACE"
150
month or more. Thus, it is advisable to use neural
network forecasting methods as an auxiliary tool in
stock indicators or exchange rates analysis likewise
other indicators with a fairly extensive measurements
base. Nevertheless, comparative analysis using
results, obtained with traditional econometrics and
machine learning models can significantly enrich
research conclusions, as well as enchance results
explanation and interpretation abilities with neural
network modeling capabilities.
Speaking of time series forecasting, several
approaches were used allowing to analyze initial data
using different neural network models in order to
achieve objective results comparison. The long short-
term memory network known as the LSTM model
was chosen at first. Model’s formal definition, which
includes the so-called forgetting gates, can be noted
down as follows:
𝑓
= 𝜎
𝑊
𝑥
+𝑈
ℎ
+𝑏
,
(1.1
)
𝑖
=𝜎
𝑊
𝑥
+𝑈
ℎ
+𝑏
,
(1.2
)
𝑜
=𝜎
𝜎
𝑊
𝑥
+𝑈
ℎ
+𝑏
,
(1.3
)
𝑐
=
𝑓
∘𝑐
+
+𝑖
∘𝜎
𝜎
𝑊
𝑥
+𝑈
ℎ
+𝑏
(1.4
)
ℎ
=𝑜
∘𝜎
𝑐
,
(1.5
)
where: 𝑥
, ℎ
stand for input and output vectors
respectively, 𝑐
represents system’s states vector;
𝑊,𝑈,𝑏 are parameters matrix and vectors, 𝑓
, is a
forgetting gate vector, 𝑖
,𝑜
define input and output
gate vectors, respectively.
It should be noted that model (1.1-1.5) is widely
applicable to the time series problems analysis, which
is explained by its relatively high accuracy in relation
to the initial data under study. For comparative
analysis needs, the second chosen approach refers to
a recurrent network model, which is considered as a
basis for short-term memory networks development
(Bergmeir, 2021). Its formal record can be
represented as follows (2.1, 2.2).
ℎ
=𝜎
𝑊
ℎ
+𝑉
𝑥
+𝑏
,
(2.1
)
𝑧
=tanh
𝑊
ℎ
+𝑏
,
(2.2
)
where: ℎ
stands for hidden state, 𝑧
refers to
input and output parameters for the 𝑡 moment of time.
𝑊
и 𝑉
define weight matrices, 𝑏
– stands for bias
vector.
The third model for comparative analysis was the
RBF network, which uses radial basis functions as a
tool for neurons activation (Truc, 2018). As a rule,
RBF networks assume three layers, including an
input, nonlinear activation layer and a linear output.
The model can be represented as a real values vector,
and the output vector is determined based on the
following 𝜑 𝑅
→𝑅
mapping (3).
𝜑
𝑥
=𝑎
𝜌
|
𝑥−𝑐
|
,
(3
)
𝑁 determines hidden layer neurons number, 𝑐
is
a neuron main hidden layer and 𝑎
stands for 𝑖-th
output neuron weight.
It is advisable to make the following assumptions
(4.1-4.3) for time series analysis and forecasting:
𝜑
0
=𝑥
1
,
(4.1
)
𝑥
𝑡
≈𝜑
𝑡−1
,
(4.2
)
𝑥
𝑡+1
≈𝜑
𝑡
=𝜑
𝜑
𝑡−1
.
(4.3
)
In addition to these three models, the comparison
also included gradient boosting method (XGBoost),
which is also often used for time series forecasting,
characterizing by a relatively higher prediction
accuracy over short time intervals and data with
seasonality pronounced presence. The gradient
boosting method is associated with a corresponding
optimization function aimed to improve model
training processes efficiency (5.1, 5.2).
𝐿
=
=𝑙
𝑦
,𝑦
+
𝑓
𝑥
+Ω
𝑓
,
(5.1
)
Ω
𝑓
=𝛾𝑇+
1
2
𝜆
|
𝜔
|
,
(5.2
)
where: 𝑙 is the loss function, 𝑦
,𝑦
are 𝑖 -th
sample element values used for training and 𝑡-first
predictors sum, 𝑥
represents 𝑖 -th training sample
elemt, 𝑓
notes function trained at step 𝑡, 𝑓
𝑥
is a
prediction on 𝑖-th training sample element, Ω
𝑓
is
function 𝑓 regularization, 𝑇 formalizes vertices
number in the analyzing tree, 𝜔 represents leaves
values, 𝛾 and 𝜆 are regularization parameters.
Summarizing, these tools make it possible to carry
out predictive information qualitative characteristics
comparative analysis obtained on the neural network
basis. Obtained results were interpreted in the
graphical form, ready for further analysis.
3 RESULTS AND DISCUSSION
Neural networks training and forecast data obtaining
were carried out according to the dollar / ruble
exchange rate historical values. This indicator
belongs to the exchange-traded, as well as oil price or
RTS index. Exchange indicators usage makes it
Neural Networks Usability Analysis in Economic Security Indicators Dynamics Forecasting
151
possible to introduce tools for economic security
operational monitoring, since they highly determine
corresponding conjuncture progress dynamics.
The first dataset contains dollar value in rubles
starting in January 1996 up to December 2020, with
observations made every month, so the total
observations number in the first dataset is 300. The
second dataset included dollar value in rubles from
March 21, 2000 to March 21, 2020, with daily
observations, 4968 observations in total. Despite the
higher observation frequency second set has many
missing days, such as weekends and holidays, as well
as occasional observations absences. Thus, second set
represents significantly large data amount, but it is
still incomplete and cannot be considered ideal. Both
data sets were divided into training and test sets in
accordance with the model requirements (Figure 1).
Values up to November 2019 were assigned to the
training set and observations starting from November
1, 2019 inclusive - to the test set.
Figure 1: Training and test samples dataset distribution
When the models were trained using the first set
consisting of 300 observations, as expected, the data
turned out to be insufficient for acceptable accuracy
predictions. The forecast using the XGBoost method
returned closest to the real data result, although it did
not repeat the trends. The radial basis function
network presented the worst result. Forecast
comparison results with real data on a segment of the
test sample are shown in Figure 2.
Figure 2: First dataset forecast results comparison
Based on presented forecasting models, it can be
concluded that gradient boosting method is highly
effective in studies based on priori information small
amount. However, it is obvious that such forecasts
have no practical sense with such a low accuracy.
Figure 3 compares all methods root mean square and
mean absolute errors.
Figure 3: First dataset forecast errors comparison
Same models training on the second dataset with
a much higher observations frequency (and,
accordingly, a large number of them) provided a
qualitatively different result. All three neural network
models follow original data trends, in contrast to the
gradient boosting method, which shows average
value that is weakly correlated with real data over the
entire interval. The most accurate prediction was
recorded using the LSTM model with mean square
and mean absolute errors of 0.254327 and 0.4625752,
respectively. Apart from the XGBoost library, the
least accurate forecast was obtained using a network
of radial basis functions, however, it also came close
to real data near peak values in the segment’s end, as
shown in Figure 4.
Figure 4: Second dataset forecast results comparison
Obviously, in this case, neural network
predictions are much more accurate and more likely
to have practical use. It is expected that with large
training set representation, neural networks face
overfitting problem. However, in accordance with
SES 2021 - INTERNATIONAL SCIENTIFIC-PRACTICAL CONFERENCE "ENSURING THE STABILITY AND SECURITY OF
SOCIO - ECONOMIC SYSTEMS: OVERCOMING THE THREATS OF THE CRISIS SPACE"
152
this work’s purposes, aiming to show fundamental
differences in neural network models efficiency, the
comparison was made exclusively between the basic
models, without parameters precise manual
adjustment. Predicted indicators errors comparison
after second dataset training is shown in Figure 5.
Figure 5: Second dataset forecast errors comparison
4 CONCLUSIONS
Based on the presented analysis, the following
assumptions can be made: a dataset with a dimension
of 300 observations is clearly not enough to train
neural networks in order to predict such volatile time
series. Neural networks with LSTM and sRNN
models showed the best results in both cases,
respectively. The radial basis function network,
which performed the worst on the first set, gave a
much more accurate result due to training on a larger
sample. In both cases, gradient boosting method
returned a fairly average result, weakly correlating
with real data trends.
Thus, we can conclude that neural network
methods usage in economic security indicators
forecasting is justified only on a significant sample
size. Therefore, machine learning methods in
economic security analysis and forecasting problems
have not yet become widespread, since typical
research mainly involves indicators with a month or
more discrete interval. At the same time, machine
learning methodology in multi-level objects
economic security analysis will be more and more in
demand. This is mainly due to the information arrays
exponential growth and initial data structuring needs.
On one hand, machine learning algorithms provide
fast and accurate results, on the other hand, additional
resources are required.
Mathematical forecasting methods usage in
economic security indicators studies based on the
machine learning apparatus in combination with more
traditional analysis methods allow to reach adequate
comprehensive assessment results. This work results
make it possible to identify the most relevant neural
network model for economic systems indicators
prediction, as well as to compare various forecasting
methods effectiveness depending on the initial data
dimension. Thus, we advise to keep caution aimed to
secure management decisions effectiveness and
validity based on machine learning methods
evidence-based verification results. Compliance with
this requirement will create new perspectives in
analysis and forecasting, as well as reduce decision-
making associated risks.
REFERENCES
Tomashevskaya, V.S., Yakovlev, D.A., 2020. Additivity
condition for reference kiosk based on page load time.
In Russian Technological Journal.
Shamin, R.V., 2019. Machine learning in economics.
“Green print”. Moscow. p. 140.
Shmeleva, A.G., Ladynin, A.I. 2019. Industrial
Management Decision Support System: from Design to
Software. In Proceedings of the 2019 IEEE Conference
of Russian Young Researchers in Electrical and
Electronic Engineering (EIConRus).
Chio, K., Frimjen, D., 2020. Machine learning and security.
DMK Press. Moscow.
Hastie, T., Tibshirani, R., Friedman, J., 2009. The Elements
of Statistical Learning. Springer-Verlag.
Flah, P., 2015. Machine Learning. The science and art of
building algorithms that extract knowledge from data.
"DMK Press". Moscow.
Bata, M., Carriveau, R. & Ting, D.SK., 2020. Short-term
water demand forecasting using hybrid supervised and
unsupervised machine learning model. In Smart Water.
5. 2.
Adadi, A., 2021 A survey on data‐efficient algorithms in
big data era. In J Big Data. 8. 24.
Athey, S., 2018. The impact of machine learning on
economics. In The Economics of Artificial Intelligence:
An Agenda. University of Chicago Press.
Mullainathan, S., Spiess, J., 2017. Machine learning: an
applied econometric approach. In Journal of Economic
Perspectives.
Andini, M., 2019. Machine learning in the service of policy
targeting: the case of public credit guarantees. Bank of
Italy, Economic Research and International Relations
Area.
Andini, M., 2018. Targeting with machine learning: An
application to a tax rebate program in Italy. In Journal
of Economic Behavior & Organization.
Chakraborty, C., 2017. Machine learning at central banks.
Bank of England. Working Paper.
Drejper, N., Smit, G., 2007. Applied regression analysis.
«Vil'jams», Moscow.
Neural Networks Usability Analysis in Economic Security Indicators Dynamics Forecasting
153
Hansika Hewamalage, 2021. Recurrent Neural Networks
for Time Series Forecasting: Current status and future
directions. In International Journal of Forecasting. 37.
1.
Truc, N. V., Anh, D. T., 2018. Chaotic Time Series
Prediction Using Radial Basis Function Networks. In
4th International Conference on Green Technology and
Sustainable Development (GTSD).
SES 2021 - INTERNATIONAL SCIENTIFIC-PRACTICAL CONFERENCE "ENSURING THE STABILITY AND SECURITY OF
SOCIO - ECONOMIC SYSTEMS: OVERCOMING THE THREATS OF THE CRISIS SPACE"
154
