Computational Economics (2025) 65:241–270

[Link]

Design of Neuro‑Stochastic Bayesian Networks

for Nonlinear Chaotic Differential Systems in Financial
Mathematics

Farwah Ali Syed1 · Kwo‑Ting Fang1 · Adiqa Kausar Kiani2 ·

Muhammad Shoaib3 · Muhammad Asif Zahoor Raja2

Accepted: 3 March 2024 / Published online: 28 March 2024

© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature
2024

Abstract
The research community’s treatise on computational economics and financial mod-
els has promising interest for the exploration and exploitation of artificial intelli-
gence (AI)-based computing paradigm to offer enriched efficacies for business
stratagems, consumer utility, and scarce resource management for enriched society
evolution. In this study, AI-based neuro-stochastic Bayesian networks (NSBNs) are
presented for mathematical models that govern the dynamics of nonlinear chaotic
financial differential systems (NCFDSs). The descriptive expressions for NCFDS are
portrayed through multi-class differential compartments for macroeconomic agents
in terms of interest rate, investment demand, and price index. The reference data
acquisition for the execution of the multi-layer structure of NSBNs is performed
with Adams numerical procedure for sundry scenarios of NCFDSs by varying the
cost per investment, saving amount, as well as, commercial market demand elastic-
ity. The designed NSBN outcomes consistently overlap with the reference solutions
having negligible magnitude of error for each scenario of NCFDS. The efficacy of
proposed NSBNs is presented through mean square error based convergence curves,
illustrations for adaptive controlling parameters, 2D–3D visual depictions, error his-
togram studies, and regression indices for variants of nonlinear chaotic differential
systems in mathematical finance.

Keywords Nonlinear chaotic financial differential systems · Artificial neural

networks · Adams numerical procedure · Bayesian regularization networks · Error
analysis

Extended author information available on the last page of the article

Vol.:(0123456789)
242 F. A. Syed et al.

1 Introduction

Numerous affiliated fields are derived from the study of chaos and other related
fields, including chaos in mathematics, chaos in meteorology and chaotic eco-
nomics (Shi et al., 2021). Chaos in economics was first observed by analysts in
1985 (Shao et al., 2019), although the government has implemented the neces-
sary regulating measures with relatively restricted effects due to the chaotic cir-
cumstances, which makes the macroeconomic movements haphazard. However,
the link between different financial system variables may be understood by the
development of the financial-based system model analysis, thus the factors in the
financial system can be preserved within the range by alterations of their constitu-
ent variables. (Huang et al., 2018; Li et al., 2023).
Traditional management strategies have failed to portray the dynamics of stiff
nonlinear economic and financial systems representing complex and challenging
aspects, as a result, the investigation of utilizing novel methodologies recently
has acquired significant prominence (Jahanshahi et al., 2021a, 2021b; Kumar &
Singh, 2021; Lei et al., 2023; Zhou et al., 2021). Analysts and intellectuals have
utilized simultaneous ordinary differential order systems to mimic accurately
the nonlinearity, stiffness, and chaotic characteristics in these models through-
out the past few decades (Jin & Yang, 2021; Sattar & Braik, 2023; Shi et al.,
2022; Jahanshahi et al., 2021a, 2021b; Yousepour & Jahanshahi, 2019). Modern
economic paradigms argue that in order to deal with market conditions, financial
models should account not only for all current aspects, but also the market deci-
sions, the pattern of prior experiences, and agent conduct (Samitas et al., 2022;
Taleizadeh et al., 2022; Zhang et al., 2022). A crucial aspect of averting finan-
cial crises and slowing the market is understanding the dimensions and structures
of economic and financial methods (Ali et al., 2023; Wang et al., 2021; Youscri
& Mirjalili, 2020). Economic hazards and substantial financial damages are fre-
quently the outcome of unanticipated changes in financial systems. Creating a
thorough and accurate model of economic systems is the first and most important
step in effectively recognizing chaotic occurrences in these systems (Wang et al.,
2022a, 2022b; Zhou et al., 2023). Applying cutting-edge methodologies is essen-
tial for their modeling for the reason that financial and economic-based systems
are complicated (Wang et al., 2022a). Besides these, the mathematical models
for economic and financial systems portrayed with a set of differential equations
have been extensively reported in the literature, please see (Atangana et al., 2020;
Çalış et al., 2022; Milovanović & von Sydow, 2020; Wang et al., 2022b; Wen &
Yang, 2019) and reference cited therein. Besides that, the artificial intelligence
(AI) based computing paradigm using neural networks and learning systems
looks promising to be exploited to address the chaotic economic and financial
models (Polyzoz et al., 2020; Samitas et al., 2020).
Emulation, often referred to as supervised learning and surrogate modeling, is
crucial for comprehending and managing complex systems (Bazrkar et al., 2023;
Bukhari et al., 2023; Charpentier et al., 2023; Santur, 2023). Modeling with arti-
ficial neural networks (ANNs) (Abdelkareem et al., 2022; Kordkheili et al., 2023;
Design of Neuro‑Stochastic Bayesian Networks for Nonlinear… 243

Mak et al., 2018; Taloba, 2022), fractional ANNs (Hadian Rasanan et al., 2020;
Lodhi et al., 2019), wavelet ANNs (Faheem et al., 2021; Liu et al., 2022), swarm-
ing intelligence (Tang et al., 2021; Verma et al., 2023), evolutionary algorithmic
approaches (EA) (Bao et al., 2023; Singh et al., 2023), boosting trees (Douiba
et al., 2023; Oyedele et al., 2023), radial basis functions (Piazzola et al., 2023;
Wu et al., 2023), Gaussian procedures (Annaby et al., 2023; Ping et al., 2023), are
only a few of the emulation methods that have been established at the moment.
Each of these tactics has a unique mix of benefits and drawbacks. An algorith-
mic system called a network of neurons, which functions exactly like the brain of
human beings, makes an effort to identify hidden links in gathered data (Anandita
Iyer & Umadevi, 2023; Wichmann & Geirhos, 2023). The ANNs have a signifi-
cant impact on the financial industry due to their competency (Platt, 2022; Pou-
finas et al., 2023) to offer reliable time-series prediction (Polyzos et al., 2020;
Samitas et al., 2020), securities categorization, algorithmic trading, credit risk-
based modeling (Ghoddusi et al., 2019; Gogas & Papadimitriou, 2021; Sariev
& Germano, 2020), the formulation of customized indexes, and cost derivatives
(Kurani et al., 2023; Magazzino & Mele, 2022). Neural networks must thus be
supplemented with other approaches in complicated situations. HIV-infection-
based model (Nisa et al., 2023; Seboka et al., 2023), mosquito’s dispersal model
(Umar et al., 2020), Emen-flower fourth order equation (Sabir et al., 2020), hybrid
nanofluidic systems (Ilyas et al., 2021), porous fin model (Ahmad et al., 2019),
dusty plasma system (Bukhari et al., 2020), and astrophysics (Szabó et al., 2022).
On the other hand, the probabilistic computing paradigm based on AI procedures
is yet to be executed for the analyses of the nonlinear chaotic financial differen-
tial systems (NCFDS) for which the studies are currently available for analyzing
the dynamics of deterministic numerical and analytical solvers. This study pre-
sents an in-depth study of intelligent computation based on neural networks with
detailed analysis to solve financial systems. The objectives of the current research
work are:

• To study the dynamics of the mathematical model for NCFDS, portrayed the
macroeconomic agents in terms of interest rate, investment demand, and price
index.
• To construct AI-based neuro-stochastic Bayesian networks (NSBNs) with feed-
forward neural networks backpropagated by the Bayesian regularization.
• To acquire the reference data for the execution of the multi-layer structure of
NSBNs numerically for variants of NCFDSs.
• To evaluate the performance of NSBN on exhaustive experimentation for solving
NCFDS with broad range of accuracy and convergence operators.

The contributions, insights, and novelty of the presented NSBN approach are
summarized as follows:

• A novel AI-based neuro-stochastic Bayesian networks are presented for math-

ematical models that govern the dynamics of nonlinear chaotic financial differen-
tial systems
244 F. A. Syed et al.

• The descriptive expressions for NCFDS are portrayed viably through multi-class
differential compartments for macroeconomic agents in terms of interest rate,
investment demand, and price index.
• The synthetic data acquisition for NSBNs is effectively performed with Adams
numerical procedure for sundry scenarios of NCFDSs for cost per investment,
saving amount, and commercial market demand elasticity.
• The efficacy of proposed NSBNs is presented through MSE based convergence
curves, illustrations for adaptive controlling parameters, 2D–3D visual depic-
tions, error histogram studies, and regression indices for variants of NCFDSs.

The introduction of the financial model along with the relevant literature on intel-
ligent computation techniques is given in Sect. 1. The mathematical description of
the nonlinear chaotic financial differential systems is presented in Sect. 2, Meth-
odology is provided in Sect. 3 along with the description of the proposed NSBNs
The comprehensive analyses of the neural networks back-propagation with Bayes-
ian regularization optimization approaches to study the dynamics of financial model
governed with three differential classes are portrayed through multi-class differen-
tial compartments for macroeconomic agents in terms of interest rate, investment
demand and price index are presented in Sect. 3 though variety of graphical illus-
trations and discusses the simulation results for different scenarios of the NCFDSs
using NSBNs for its solution. Finally, the conclusions of the study are provided in
Sect. 4 that concludes the current investigation with a few potential future research
directions.

2 Nonlinear Chaotic Financial Differential System

A financial model based on NCFDSs as shown in Fig. 1 is governed by the fol-

lowing mathematical expressions as reported in (Atangana et al., 2020; Chen, 2008;
Chen et al., 2020; Hajipour et al., 2018; Wang et al., 2011, 2020)

Fig. 1  Schematic of nonlinear chaotic financial differential systems

Design of Neuro‑Stochastic Bayesian Networks for Nonlinear… 245

dx
= z + (y − a)x, x(0) = x0
dt
dy
= 1 − by − x2 , y(0) = y0 (1)
dt
dz
= −x − cz, z(0) = z0
dt
where x0 , y0 , z0 , represents the initial values for the rate of interest x, investment
of demand y, and price index z. which stimuli the distribution price variance, sav-
ing amount a, per-investment cost b, and demand elasticity of commercials c are
positive constant indices, respectively. The further related information of NCFDS
regarding theoretical analysis of the local, and global stability, as well as the appro-
priate bound of the parameters can be seen in (Chen, 2008). Different research works
have been carried out on such dynamic financial models such as (Milovanović &
von Sydow, 2020; Platt, 2022; Polyzos et al., 2023; Wen & Yang, 2019). Schematic
of nonlinear chaotic financial differential systems illustrated in underlying Fig. 1.

2.1 Mathematical Representation of NCFDS

The NCFDS model from Eqs. (1) is outlined mathematically in the following sec-
tion for four different scenarios and cases as shown in Table 1. These four scenarios
are built using different values for the financial economic indicators that indicate of
saving amount, and the commercial market’s demand elasticity for different cases.
NCFDSs model in Eq. (1) that corresponds to the 1–3 scenarios with 3 proposed
cases and variation of the initial condition-based cases in the 4th scenario as tabu-
lated in Table 1 was constructed as follows to plaid the precision, reliability and
stability of the envisioned NSBNs scheme.

Table 1  Index of cases and Scenario (scr) Case (c) Parameters Initial con-
scenarios of NCFDSs ditions
a b c x0 y0 z0

1 1 1.5 0.10 1 2 3 2
2 3.0 0.10 1 2 3 2
3 5 0.10 1 2 3 2
2 1 3 0.05 1 2 3 2
2 3 0.10 1 2 3 2
3 3 0.15 1 2 3 2
3 1 3 0.10 0.5 2 3 2
2 3 0.10 1 2 3 2
3 3 0.10 1.5 2 3 2
4 1 3 0.10 1 1 2 1
2 3 0.10 1 2 3 2
3 3 0.10 1 3 4 2
246 F. A. Syed et al.

The following is a description of one of the NCFDS models (1) for a = 3, b = 0.1
and c = 1.0, x0 = 2, y0 = 3 and z0 = 2 as reported in (Chen, 2008), while few variants
of NCFDS is portrayed in Appendix Table A1.
dx
= z + (y − 3.0)x, x(0) = 2
dt
dy
= 1 − 0.1y − x2 , y(0) = 3 (2)
dt
dz
= −x − 1.0z, z(0) = 2
dt
Similarly, the rest of the equations of the NCFDS model can be formulated for
each of the four scenarios and their respective three cases in the same manner as
discussed in this section.

3 Methodology for the Analysis of Dynamic Financial Systems

The methodology for analysis of dynamic financial system (1) along with the math-
ematical representation in the last section is provided in two stages; firstly, the syn-
thesis data generation with a numerical solver, and secondly, the necessary proce-
dure of the proposed neuro-stochastic Bayesian scheme is provided. The details for
implementing and reproducing outcomes with proposed intelligent neuron-super-
vised networks via NSBNs are presented for the interested readers. The proposed
scheme is implemented as portrayed in Fig. 2 which illustrates the general process
connection diagram of the proposed NSBN for solving the variants of NCFDSs in
terms of the following steps: the problem, the governing differential systems, refer-
ence numerical solution, Adams numerical method, network diagram, network sim-
ulations, process workflow, convergence on MSE, variation of controlling param-
eters, analysis of results with the aid of error histogram and regression analysis. The
necessary elaborative details for the two stages implemented in this study are given
below:
Stage 1. Dataset acquisition: Firstly, the synthetic datasets for NSBNs are
acquired for NCFDS presented in set of Eq. (1) numerically via Adams method. In
order to put together the reference dataset for the proposed intelligent computing
with NSBN for determination of the approximate findings for the NCFDSs as indi-
cated in Eqs. (1) and the few specific scenarios offered in Table 1 for diversified var-
iants, Adams scheme was engaged. The detailed mathematical procedure along with
the settings of accuracy and convergence controlling metrics of the Adams method
is adopted as in the aforementioned study (Hanif et al., 2023; Shoaib et al., 2023).
The Adams numerical technique is executed with the automatic adjustments of input
grids ranging between (0, 200) by the aid of the ‘NDSolve’ function for numeri-
cal outcomes of ordinary differential equations (ODEs) and partial differential equa-
tions (PDEs), which is available in the software package of Mathematica for solving
NCFDSs variation.
Design of Neuro‑Stochastic Bayesian Networks for Nonlinear… 247

Fig. 2  Graphical abstract of the NCFDS study using NSBNs approach

Stage 2. Developing neuro-stochastic supervised networks: The dataset created

using the Adams numerical process served as the NSBNs default target output and
input grids for solving the variants of NCFDSs. The NSBNs are formulated on
the basis of a single neuron model that consists of input, hidden, and output lay-
ers exploiting log-sigmoidal activation or windowing function, as illustrated in
Fig. 3; i.e., subfigure 3(a), while the overall block structure is presented in Fig. 2.
The developed NSBN models were put into practice by using the software package
MATLAB for the toolbox of a neural network via ‘ntstool’ routine. After exhaustive
simulations are performed to formulate the overall structure of networks as shown
in subfigure 3b. Utilizing the backpropagation effectiveness of Bayesian regulariza-
tion techniques (Sariev & Germano, 2020), network weights were trained, tuned, or
learned with the procedure and setting of the scheme on a similar pattern as reported
in (Gouravaraju et al., 2023; Hajeb et al., 2023).
248 F. A. Syed et al.

Fig. 3  NSBN architecture development for solving differential equations

Performance evaluation: The Bayesian regularization-based optimizers set the

weights of networks by minimizing mean square error (MSE) between actual and
estimated responses. The MSE, as well as the absolute error (AE) measure, are used
to assess the performance of the proposed NSBN for solving variants of NCFDS.
The necessary definition of these operators is given as follows:
M M M
1 ∑ 1 ∑ 1 ∑
MSEx = (xm − x̂ m )2 , MSEy = (ym − ŷ m )2 , MSEz = (z − ẑ m )2
M m=1 M m=1 M m=1 m
(3)
| | | | | |
AEx = |xm − x̂ m |, AEy = |ym − ŷ m |, AEz = |zm − ẑ m |, (4)

where MSEx, MSEy and MSEz are the respective MSE’s for interest rate x, cost per
investment y and price index z, xm is the mth index for instance of the original target,
while approximate output is represented with x̂ m on the similar pattern AEx, AEy and
AEz are AE’s of respective variables in NCFDS model.
Design of Neuro‑Stochastic Bayesian Networks for Nonlinear… 249

4 Simulation Results with Analysis

The dataset is generated with Adams scheme for case (c) and scenario (scr) of
NCFDSs as provided in Table 1 by taking t ϵ (0, 200) with a step size of 0.25,
i.e., 801 grid points. The generated data for three cases and four scenarios is seg-
mented as 80% of 801 input grid points, i.e., 641, are utilized randomly for train-
ing, 10% of total 801 instances, i.e., 80, for testing and validation, for solving
NCFDSs in Eq. (1) by execution of the proposed NSBN procedure via ‘ntstool’
with 50 hidden-neurons, log-sigmoid activation function, and input, output, and
hidden layers as shown in Fig. 3b, while overall schematic process flow of a neu-
ral network is illustrated as block diagram in sub-figure 3(a) and necessary details
are mentioned in the last section. Outcomes of the designed NSBNs are shown in
Figs. 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14 and 15 for different variations of NCFDSs
problems here, and to avoid the redundancy in the presentation of the results the
outcomes are also illustrated in Appendix section in Figures (A1–A12). Here it is
essential to annotate that there are four scenarios with each having three cases in
the dynamic financial model (DFS). However, only scr-1:c-1, scr-2:c-1, scr-1:c-2,

Fig. 4  Performance curves on MSE to solve the variants of dynamic financial model using NSBN
approach. Note This figure shows the iterative convergence with step increment in epoch number for the
proposed NSBNs for solving each scenario of the NCFDS. The iterative training and testing lines con-
verge consistently for each case of NCFDSs and the best compromise on the basis of over or under-fitting
is achieved at 1000 epoch number with a value of MSE around ­10–10–10–12
250 F. A. Syed et al.

Fig. 5  NSBN adaptive control parameters for solving NCFDS scenarios. Note Various adaptive control-
ling metrics have been observed in terms of, step size, i.e., mu, gradient, i.e., first-order optimality, num
parameter, i.e., a number of tuning parameters, and the sum of squared param, i.e., a sum of square errors
parameters for regularization, are monitored iteratively for step increase in the epoch index, i.e., an inde-
pendent execution for networks buildup, for smooth and error-free implementation of NSBNs algorithm
portrayed in this figure

scr-2:c-3, respectively are chosen to represent graphically to discuss the results

of the proposed study as shown in subfigure of Figs. 4, 5, 6, 7 and 8, while the
illustrations for scr-1:c-3, scr-3:c-1 scr-3:c-3, scr-4:c-1 are provided in Figures
(A1–A5) of Appendix). Additionally, due to similarity in patterns of the out-
comes is seen for the rest of variation, therefore, here are not displayed in the
pictorial form. Results based on the mean squared error (MSE), i.e., the mean
absolute deviation of estimated targets and original targets available through
standard numerical results, adaptive controlling parameters, solution plots with
error analysis, histograms, and regression are inserted along with the dynamics of
NCFDS in terms of 2D and 3D graphical illustrations with necessary elaborative
interpretations in this section.
The content portrayed in Fig. 4 and its subfigures 4a, 4b, 4c, and 4d for four
scenarios represented by scr-1:c-1, scr-2:c-1, scr-1:c-2, scr-2:c-3, respectively
while plots for scr-1:c-3, scr-3:c-1 scr-3:c-3, scr-4:c-1 are given respectively in
subfigure 4e, 4f, 4g, and 4h of Figure A1 of Appendix, that shows the iterative
convergence with step increment in epoch number for the proposed NSBNs for
solving a set of equations represented in (1). It is the visual representation of the
Design of Neuro‑Stochastic Bayesian Networks for Nonlinear… 251

Fig. 6  Comparative solutions plots of NSBN from numerical results for sundry scenarios of NCFDSs. Note The
training state response graph in this figure is epitomized with labels including training targets, training outputs,
test targets, test outputs, errors, and fit on the vertical scale that measures the accuracy of the proposed NSBN
for the prediction of numerical outcomes of the NCFDS. Whereas the time steps or data points at which the
error value provides the measure of the accuracy of the network in quantitative terms at each step of time can be
visualized along a horizontal axis For the recognition of error patterns zoom plots are also provided to closely
analyze the fitness process of the algorithm. It can be observed that even the maximum error is as low as pos-
sible and ranges around ­10–3–10–5 with the employment of the NSBN scheme for the solution NCFDS scenarios
252 F. A. Syed et al.

Fig. 6  (continued)
Design of Neuro‑Stochastic Bayesian Networks for Nonlinear… 253

Fig. 7  Error distribution pictorial representation for comparative analysis of NSBN for the solution of
NCFDSs. Note Error histogram visualizes the distribution of errors for 2403 instances, calculated as a
difference between the model’s estimated and exact results for all three explanatory variables i.e., the rate
of interest ‘x’, investment of demand ‘y’, and price index ‘z. in all cases of the given four scenarios in this
figure, the majority of the error values fall under the range where the error value is close to zero i.e., it
reflects the ideal circumstance in which the model’s predicted outcomes exactly match the actual values

performance, during the training process of the neural network i.e., the perfor-
mance metric that plots the objective function, in our case MSE on vertical-axis
and on horizontal-axis shows the number of training iterations, i.e., epochs. The
iterative training and testing lines converge consistently for each case of NCFDSs
and the best compromise on the basis of over or under-fitting is achieved at
1000 epoch number with a value of MSE around 4.9277 × ­10–12, 1.8991 × ­10–11,
3.2369 × ­10–10, for c-1, c-2 and c-3, respectively, of scr-1, 5.1045 × ­ 10–12,
–11 –12
1.8991 × ­10 , and 1.5134 × ­10 , for cases c-1, c-2 and c-3, respectively, of scr-
2, 3.3859 × ­10–12, 1.8991 × ­10–11, 3.2369 × ­10–12 for cases c-1, c-2 and c-3, respec-
tively, of scr-3 and 4.4458 × ­10–12, 1.8991 × ­10–11, 1.1756 × ­10–11 for cases c-1,
c-2 and c-3, respectively, of scr-4 of NCFDS Moreover, in all of the four sce-
narios the two curves i.e., testing and training, are near to each other such that
the gap between them is not large that confirms NBLNS performance is accurate
for both testing and training data. However, the performance on training data is
relatively better than testing which is understandable because the training is per-
formed on prior known targets while testing is carried out on unbiased data. One
can easily visualize the robust and stable execution of proposed NSBNs with no
254 F. A. Syed et al.

Fig. 8  Regression illustrations for comparative study of NCFDS economic systems. Note The value of
the regression coefficient R, which measures how well the suggested NSBNs scheme performs consist-
ently across sample points in training and testing, is in close vicinity of the ideal unity value. As a result,
the value of R equal to 1 can be inferred as the NSBN algorithm fits the data perfectly and explains the
NCFDS differential model i.e., each of three cases of the four proposed scenarios

evidence of large fluctuations in the performance metric i.e., the network does not
struggle to learn and no further adjustment is needed to train the model.
Various adaptive controlling metrics have been observed in terms of, step size,
i.e., mu, gradient, i.e., first-order optimality, num parameter, i.e., a number of tun-
ing parameters, and sum of squared param, i.e., sum of square errors parameters
for regularization, are monitored iteratively for step increase in the epoch index,
i.e., an independent execution for networks buildup, for smooth and error free
implementation of NSBNs. The pictorial representation of adaptive controlling is
shown in the subplots of Fig. 5 i.e., 5(a) for scr-1:c-1, 5(b) for scr-2:c-1, 5(c) for
scr-1:c-2, 5(d) for scr-2:c-3, while A2(e) for scr-1:c-3, A2(f) for scr-3:c-1, A2(g)
for scr-3:c-3, and A2(h) for scr-4:c-1 in Figure A2 of Appendix, for the presented
financial differential system. As shown by the low magnitude of gradient for the
case 1 of scenario 1 i.e., 5.0806 × ­10–7 at the epoch value equals to 1000, the
parameters have been regularized in a consistent manner that the performance of
the proposed NSBN algorithm is efficient and stable. Similar pattern is observed
for rest of cases with gradient value ranges between 1­ 0–6 and ­10–7 in the sub-
plots of the training state plots that further gives the evidence that the param-
eters are adjusted effectively allowing the network with lower possible error to
be moved towards the configuration of the NCFDS model numerical solutions.
Moreover, the adaption parameter step size termed as “mu” corresponds the value
of 50,000 at 1000 epoch in all cases of the NCFDS scenarios. as the step size
taken during a successive iteration of the optimization course is being affected
by the damping factor “mu” and is able to control the stability and convergence
of the model. The number of tuned metrics ranges from 373.5042 to 637.1007 at
Design of Neuro‑Stochastic Bayesian Networks for Nonlinear… 255

Fig. 9  Comparison of solutions dynamics of NCFDS for scenario 1. Note This figure portrayed the
effects of saving amount ‘a’ = 1.5, 3, and 5 on the rate of interest ‘x’, investment of demand ‘y’, and price
index ‘z’ in the spam of [0, 200]. While keeping the values fixed for parameters i.e., per-investment cost
‘b’ = 0.10, demand elasticity of commercials ‘c’ = 1

epoch number equals 1000 for the entire set of subplots representing the train-
ing state. The tuned parameters serve as a scaling factor, balancing the effect of
the regularization term in relation to the model’s size and complexity. The sum
squared parameter aids in preventing overfitting to the trained data by limiting the
complexity and sensitivity of the model. It can be depicted from the subfigures
256 F. A. Syed et al.

Fig. 10  Comparison of solutions dynamics of NCFDS for scenario 2. Note This figure portrayed the
effects of per-investment cost ‘b’ = 0.05, 0.10, and 0.15 on the rate of interest ‘x’, investment of demand
‘y’, and price index ‘z’ in the spam of [0, 200]. While keeping the values fixed for parameters i.e., saving
amount ‘a’ = 3.0, demand elasticity of commercials ‘c’ = 1

of the training state plots; the sum squared param value ranges between 39.0539
and 79.449 at epoch value 1000. Furthermore, the validation checks do not fail
even for a single time during the entire course of training i.e., the value of the
validation error is consistently equal to zero, for each sample plot of the proposed
schemes of NCFDSs.
Design of Neuro‑Stochastic Bayesian Networks for Nonlinear… 257

Fig. 11  Comparison of solutions dynamics of NCFDS for scenario 3. Note This figure portrayed the
effects of demand elasticity of commercials ‘c’ = 0.5, 1, and 1.5 on the rate of interest ‘x’, investment of
demand ‘y’, and price index ‘z’ in the spam of [0, 200]. While keeping the values fixed for parameters
i.e., saving amount ‘a’ = 3.0, per-investment cost ‘b’ = 0.10

The dynamics of each case of all four NCFDS variations are also analyzed, and
the NCFDS results are presented in the sub-figures of Fig. 6 exclusively with the ref-
erence numerical outputs. Figure 6 and its associated sub-figures 6(a) for scr-1:c-1,
6(b) for scr-2:c-1, 6(c) for scr-1:c-2, 6(d) for scr-2:c-3 (A3(e) for scr-1:c-3, A3(f)
for scr-3:c-1, A3(g) for scr-3:c-3, and A3(h) for scr-4:c-1 are presented in Appendix
258 F. A. Syed et al.

Fig. 12  Comparison of solutions dynamics of NCFDS for scenario 4. Note This figure portrayed the
effects of the variations of the values of the initial conditions for x­ 0 = 1, 2, 3 for ­y0 = 2,3,2 and ­z0 = 1,2,2
on the rate of interest ‘x’, investment of demand ‘y’, and price index ‘z’ in the spam of [0, 200]. While
keeping the values fixed for parameters i.e., saving amount ‘a’ = 3.0, per-investment cost ‘b’ = 0.10,
demand elasticity of commercials ‘c’ = 1

section Figure A3) depicts the plots of fitness with regards to output as well as error
from standard results for the four scenarios of NCFDS differential model. One
can gain insights by visualizing the plot that the network learning is accurate and
it converges towards the optimal solution of the desired NCFDS outcomes during
Design of Neuro‑Stochastic Bayesian Networks for Nonlinear… 259

Fig. 13  Graphical illustration of neural network training of NSBN for NCFDSs model. Note Demonstra-
tions of how training state, performance, and other metrics of presented NSBNs algorithm are evolving
with each specified epoch step are illustrated in this figure for each of the proposed four scenarios and its
three cases the interpretation of trained neural network outcomes for NCFDS models is revealed in rela-
tion to the MSE-based fitness values, the number of computing epochs, the computation duration, and
the NSBNs parameters such as gradient and learning rate

the process of training. response, behavior or even performance of NSBN network

during different phases of training can be visualized through the fitness graph. The
fitness function is epitomized with labels including training targets, training out-
puts, test targets, test outputs, errors, and fit on the vertical scale that measures the
accuracy of the proposed algorithm for the prediction of numerical outcomes of
NCFDS differential system. Whereas the time steps or data points at which the error
value provides the measure of the accuracy of the network in quantitative terms at
each step of time can be visualized along horizontal axis. Moreover, for the pur-
pose of detailed insights or fine details for the recognition of error patterns zoom
plots are also provided to closely analyze the fitness process of the algorithm. It can
be observed that even the maximum error is as low as possible and ranges around
­10–3–10–5 with the employment of NSBN scheme for the solution of each case of
four proposed NCFDS scenarios.
Histogram illustration plots and regression index for each NCFDS variation were
used to conduct a comprehensive investigation of the NSBNs performance. Fig-
ures 7 and 8 demonstrate the outcomes of error histogram plots and regression illus-
trations, respectively. Error histogram visualizes the distribution of errors for 2403
instances, calculated as a difference between the model’s estimated and exact results
for all three explanatory variables. The vertical axis of each sub-figure of Fig. 7 dis-
plays the frequency with which each error value occurs, while the horizontal-axis,
260 F. A. Syed et al.

Fig. 14  2-Dimensional parametric plot representation for NCFDSs scheme with the employment
of NSBNs. Note: In a nonlinear NCFDS system, unpredictable and irregular patterns of motion can
develop. This refers to the chaotic behavior of the state variables i.e., rate of interest ‘x’, investment of
demand ‘y’, and price index ‘z’ as analyzed commencing 2-dimensional illustrations in this figure
Design of Neuro‑Stochastic Bayesian Networks for Nonlinear… 261

Fig. 15  3-Dimensional parametric plot representation for NCFDSs scheme with the employment of
NSBNs. Note For certain values of the system parameters like saving amount, cost per investment, and
demand elasticity of commercial market, 3D illustrations in this figure spectacles the chaotic behavior of
the state variables rate of interest ‘x’, investment of demand ‘y’, and price index ‘z’, the chaotic behavior
is characterized by erratic and seemingly random patterns of motion that can be challenging to anticipate
or manage

it displays the error values. Here, the overall error range is divided into 20 smaller
precise bins. The amount of errors falling within a given error range is shown by the
height of each bar and the error axis’s (horizontal-axis) zero error value is repre-
sented by a zero error line. The outcomes are displayed in the sub-figures 7a, 7b, 7c,
and 7d (for subfigures A4 (e), A4 (f), A4 (g), and A4 (h) in Figure A4 of Appendix)
for different cases of the 4 scenarios of dynamic financial scheme. The effectiveness
of the NSBN algorithm on results for the specified NCFDS differential models has
been assessed using error histograms. For example, in the first case of scenario 1
of the NCFDS, the training dataset is below but close to 1800, and the test dataset
262 F. A. Syed et al.

is between 1800 and 2500. In this situation, the zero error point falls under the bin
with the center value of 9.4 × ­10–7. Accordingly, one can infer from subfigures 7
that the majority of instances are populated in the error bar of values − 2.9 × ­10–6
(scenario 2, case 1), − 1.9 × ­10–6 (scenario 1, case 2), − 7.3 × ­10–7 (scenario 2, case
3), 6.62 × ­10–5 (scenario 1, case 3), − 0.0001, − 4.01 × ­10–7 (scenario 3, case 1),
4.01 × ­10–7 (scenario 3, case 3) and 1.49 × ­10–5 (scenario 4, case 1), having the ref-
erence zero-error line for NCFDS in set (1), respectively and similar patterns are
achieved for the rest of the remaining scenarios and its cases, which shows the rea-
sonable precision or accuracy of NSBN algorithm. Moreover, histogram statistics
indicate that the majority of dataset samples have an error close to zero that falls
within the following range. Sub-figures of Fig. 7 illustrate that in all cases of the
given 4 scenarios, the majority of the error values fall under the range where the
error value is close to zero i.e., it reflects the ideal circumstance in which the mod-
el’s predicted outcomes exactly match the actual values.
Likewise, regression values for all four NCFDS variations solved by the NSBNs
was equal to unity for each training, testing and all dataset samples, as presented
through Fig. 8 and its subfigures (a) for scr-1:c-1, (b) for scr-2:c-1, (c) for scr-1:c-2,
8(d) for scr-2:c-3, (referred Appendix section for subfigures A5(e) for scr-1:c-3,
A5(f) for scr-3:c-1, A5(g) for scr-3:c-3, and A5(h) for scr-4:c-1 of Figure A5).
The value of the regression coefficient R, which measures how well the suggested
NSBNs scheme performs consistently across sample points in training and testing,
is in close vicinity of the ideal unity value. As a result, the value of R equal to 1 can
be inferred as NSBN model fits the data perfectly and explains NCFDS differential
model i.e., each three cases of the four proposed scenarios. However, if the ratios
of the sampling data are drastically altered, for example, to 50% both for training
and testing, the reliable precision for training samples would still be attained, i.e., R
approach unity, while the regression matrix for all data samples would still be accu-
rate. The value of the proposed NSBN was further supported by regression plots of
the comprehensive numerical exploration with R near its desired value that depicts
the accuracy of its performance through a minimized sum of squared errors between
the actual and predicted outcomes. Likewise, we conducted a comprehensive eval-
uation of recently published studies on supervised neural networks (Ahmad et al.,
2022; Nonaka et al., 2023; Sarı et al., 2023) and discovered a similar regression plot
trend.
The solution of underlying dynamics of each of the four scenarios, as depicted
in Figs. 9, 10, 11 and 12, spectacles is consistent and analogous as well as overlap-
ping of the proposed NSBN and mathematical revelations along with the absolute
errors graph as depicted in Figures A6–A9 of Appendix for four scenarios and its
associated three cases, respectively. From Fig. 9 part a, b, c for scr-1:c-1, scr-1:c-2,
scr-1:c-3, Fig. 10 part a for scr-2:c-1, part b for scr-2:c-2, and part c for scr-2:c-3,
Similarly scr-3:c-1, scr-3:c-2, scr-3:c-3 are depicted in the Fig. 11 while scr-4:c-1,
scr-4:c-2 and scr-4:c-3 of Fig. 12, here similar patterns can be observed in each
case of four variants of NCFDS. As a result, by visualizing the trend of the pre-
dicted outcomes of the targeted outputs and comparison of it with the actual values
provides reasonable evidence that the NSBN performance for the numerical solu-
tion of the financial differential model is consistently accurate. Each of subplots of
Design of Neuro‑Stochastic Bayesian Networks for Nonlinear… 263

Figures A6–A9 in Appendix depicts the absolute error magnitude for each NCFDS
scenario in order to determine the level of accuracy. Absolute error graphically rep-
resents the difference between the target value and the “original” value for each step
of time i.e., the degree of measurement inaccuracy. By analyzing the absolute error
graph, one can understand the model’s strengths and weaknesses, compare different
models, identify problematic time steps or data patterns, and monitor the model’s
performance over time. Absolute error values for NCFDS in Eqs. (1) for the three
variables of interest symbolized as x, y, and z, ranged from ­10–9 to 1­ 0–5, ­10–8 ­to10−2,
­10–9 to ­10–5, ­10–10 to ­10–5, ­10–8 to ­10–2, ­10–8 to ­10–5, ­10–9 to ­10–5, and ­10–9 to ­10–4, as
shown in Figures A6–A9 of Appendix i.e., scr-1:c-1, scr-1:c-2, scr-1:c-3, scr-2:c-1,
scr-2:c-2, scr-2:c-3, scr-3:c-1,scr-3:c-2, scr-3:c-3, scr-4:c-1, scr-4:c-2 and scr-4:c-3,
respectively. It can be seen that the minimum range corresponds to 1­ 0–10 (i.e., lower
limit of absolute error values) from the reported absolute error values of different
DFS scenarios. Whereas the error range for maximum absolute error is ­10–2, also
corresponds to the upper limit of the error values for all NCFDS variants in Eq. (1).
It is inferred that the accuracy of the NSBN scheme evaluated by the absolute meas-
ure for the given financial differential models of the present study. The NSBN accu-
racy is consistent and no significant deteriorations are examined with regard to the
model’s performance. In an ideal situation the value of the error should be minimal,
at this point, by monitoring the graph of each scenario of the DFS model, the value
of absolute error appears to be as low as possible i.e., NSBN predicting the time
series accurately. It can be identified that the NSBN does not need to be adjusted or
retrained i.e., the optimal accuracy is already maintained.
Demonstrations of how training state, performance, and other metrics of pre-
sented NSBNs algorithm are evolving with each specified epoch step are illustrated
in the sub-Figure 13(a) to 13(f) of Fig. 13 for each of the proposed four scenar-
ios and its three cases (i.e., scr-1:c-1, scr-1:c-2, scr-1:c-3, scr-2:c-1, scr-2:c-2, scr-
2:c-3, and in Figure A10 of Appendix sub-figures A10(g), A(h), A(i), A(j), A(k),
and A10(l) for scr-3:c-1, scr-3:c-2, scr-3:c-3, scr-4:c-1, scr-4:c-2 and scr-4:c-3,
respectively.) In Fig. 13, the interpretation of trained neural network outcomes for
NCFDS models is revealed in relation to the MSE-based fitness values, the num-
ber of computing epochs, the computation duration, and the NSBNs parameters
such as gradient and learning rate. It is clearly visible through the naked eye that
the NSBN outcomes are accurate with minimum possible error and achieves per-
formances of 4.9277 × ­10–12, 1.8991 × ­10–11, 3.2369 × ­10–10, in times of 0:09:08,
0:08:55, and 0:11:06 based on 1000 epochs for cases 1, 2 and 3, respectively, of
scr-1, 5.1045 × ­10–12, 1.8991 × ­10–11, and 1.5134 × ­10–12, in times of 0:08:04,
0:08:55, and 0:09:04 based on 1000 epochs for cases 1, 2 and 3, respectively, of
scr-2, 3.3859 × ­10–12, 1.8991 × ­10–11, 3.2369 × ­10–12 in times of 0:09:23, 0:08:55,
and 0:14:04 based on 1000 epochs for cases 1, 2 and 3, respectively, of scr-3 and
4.4458 × ­10–12, 1.8991 × ­10–11, 1.1756 × ­10–11 in times of 0:11:19, 0:08:55, and
0:10:41 based on 1000 epochs for cases 1, 2 and 3, respectively, of scr-4 of NCFDS,
respectively.
In a nonlinear DFS system, unpredictable and irregular patterns of motion can
develop. This is referred to the chaotic behavior of the state variables as analyzed
commencing 2D(2-dimensional) and 3D(3-dimensional) parametric evaluation
264 F. A. Syed et al.

illustrated in Figs. 14 and 15. The parametric behavior between two variables
including x, y and z can be visualized through 2D plots of Fig. 14 i.e., 14(a) for
scr-1:c-1, 14(b) for scr-1:c-2, 14(c) for scr-1:c-3, 14(d) for scr-2:c-1, 14(e) for scr-
2:c-2, and 14(f) for scr-2:c-3, (in Figure A11 of Appendix for A11(g) for scr-3:c-1,
A11(h) for scr-3:c-2, A11(i) for scr-3:c-3, A11(j) for scr-4:c-1, A11(k) for scr-4:c-2,
and A11(l) for scr-4:c-3, respectively). While x, y and z parametric behavior of all
three variables is visualized through 3D graphs of Fig. 15 i.e., 15(a) for scr-1:c-1,
15(b) for scr-1:c-2, 15(c) for scr-1:c-3, 15(d) for scr-2:c-1, 15(e) for scr-2:c-2, 15(f)
for scr-2:c-3, while the subfigures A12(g) for scr-3:c-1, A12(h) for scr-3:c-2, A12(i)
for scr-3:c-3, A12(j) for scr-4:c-1, A12(k) for scr-4:c-2, and A12(l) for scr-4:c-3,
respectively, of Figure A12 in Appendix. Extreme sensitivity to the opening circum-
stances is a characteristic of chaotic behavior, which means that modest changes in
the system’s initial conditions can result in radically divergent long-term paths. For
certain values of the system parameters like saving amount, cost per investment, and
demand elasticity of commercial market, the 2D and 3D illustrations spectacles the
chaotic behavior of the state variables x, y, and z. In the 2D and 3D demonstrations,
the chaotic behavior is characterized by erratic and seemingly random patterns of
motion that can be challenging to anticipate or manage.

5 Conclusion of the Study

In conclusion, this study focuses on the utilization of AI in the realm of mathemati-

cal financial and economic models. The study proposes the implementation of AI-
based neuro-stochastic Bayesian networks to govern the dynamics of nonlinear
chaotic financial differential systems. The descriptive expressions for these systems
are presented through multi-class differential compartments, which represent mac-
roeconomic agents such as interest rates, investment demand, and price indices. To
validate the effectiveness of the NSBNs, reference data is acquired using the Adams
numerical procedure. Various scenarios of NCFDSs are explored by adjusting the
quantities of interest in terms of saving amount, cost per investment, and elasticity
of demand. The outcomes generated by the designed NSBNs consistently align with
the reference solutions, exhibiting minimal error, i.e., in the range of 1­ 0–10–10–12
across all NCFDS scenarios. The efficacy of the proposed NSBNs is demonstrated
through several analytical techniques. Mean square error (MSE)-based convergence
curves, adaptive controlling parameters, and 2D–3D visual depictions provide a
comprehensive understanding of the NSBNs’ performance. Error histogram studies
and regression indices further reinforce the accuracy and reliability of the NSBNs
for different variants of nonlinear chaotic differential systems in mathematical
finance. Overall, this research underscores the promising potential of AI-based com-
puting paradigms, particularly NSBNs, in enhancing efficiencies in business strate-
gies, consumer utility, and scarce resource management. By leveraging the power of
AI, we can foster the evolution of a healthier society by providing enriched tools for
decision-making and optimization in financial and economic domains. On the other
hand, ANNs-based models have complicated systems that have a negative impact
on the performance as well. One frequent flaw shared by the majority of neural
Design of Neuro‑Stochastic Bayesian Networks for Nonlinear… 265

networks is concerned with overfitting, i.e., over-training of the data (Bejani & Gha-
tee, 2021; Bilbao & Bilbao, 2017). Additionally, neural networks may not function
effectively when receiving erroneous and noisy input data (Ribli et al., 2019). From
a future perspective, one should investigate the improvement of ANN not only to
give remedies to the pointed concerns (Poufinas et al., 2023) but also to make ANN
capable of solving stiff nonlinear financial models of paramount interest (Bas et al.,
2023; Behera et al., 2023; Khan et al., 2024). Moreover, the proposed AI-based
computing paradigm via neuro-stochastic Bayesian networks is equally applicable to
predict the time series modeling of financial differential systems on the availability
of real or survey cross-sectional, pooled, and panel data.
Supplementary Information The online version contains supplementary material available at [Link]
org/​10.​1007/​s10614-​024-​10587-4.

Data availability No data is associated with this manuscript.

Declarations
Conflict of interest No any conflict of interest in this study.

References
Abdelkareem, M. A., Soudan, B., Mahmoud, M. S., Sayed, E. T., AlMallahi, M. N., Inayat, A., & Olabi,
A. G. (2022). Progress of artificial neural networks applications in hydrogen production. Chemical
Engineering Research and Design, 182, 66–86.
Ahmad, I., Ilyas, H., Raja, M. A. Z., Cheema, T. N., Sajid, H., Nisar, K. S., & Abbas, M. (2022). Intel-
ligent computing based supervised learning for solving nonlinear system of malaria endemic model.
AIMS Mathematics, 7(11), 20341–20369.
Ahmad, I., Zahid, H., Ahmad, F., Raja, M. A. Z., & Baleanu, D. (2019). Design of computational intel-
ligent procedure for thermal analysis of porous fin model. Chinese Journal of Physics, 59, 641–655.
Ali, K., Hongbing, H., Liew, C. Y., & Jianguo, D. (2023). Governance perspective and the effect of eco-
nomic policy uncertainty on financial stability: Evidence from developed and developing econo-
mies. Economic Change and Restructuring, 56(3), 1971–2002.
Anandita Iyer, A., & Umadevi, K. S. (2023). Role of AI and its impact on the development of cyber
security applications. In: Sarveshwaran, V., Chen, [Link]., Pelusi, D. (eds.) Artificial Intelligence and
Cyber Security in Industry 4.0 (pp. 23–46). Springer. [Link]
Annaby, M. H., & Al-Abdi, I. A. (2023). A Gaussian regularization for derivative sampling interpolation
of signals in the linear canonical transform representations. Signal, Image and Video Processing, 17,
2157–2165.
Atangana, A., Bonyah, E., & Elsadany, A. A. (2020). A fractional order optimal 4D chaotic financial
model with Mittag–Leffler law. Chinese Journal of Physics, 65, 38–53.
Athalye, V., & Haven, E. (2023). Causal viewpoint and ensemble interpretation: From physics to the
social sciences. Philosophical Transactions of the Royal Society A, 381(2252), 20220279.
Bao, C., Gao, D., Gu, W., Xu, L., & Goodman, E. D. (2023). A new adaptive decomposition-based evo-
lutionary algorithm for multi-and many-objective optimization. Expert Systems with Applications,
213, 119080.
Bas, E., Egrioglu, E., & Tunc, T. (2023). Multivariate picture fuzzy time series: New definitions and
a new forecasting method based on Pi-sigma artificial neural network. Computational Economics,
61(1), 139–164.
Bazrkar, M. J., & Hosseini, S. (2023). Predict stock prices using supervised learning algorithms and par-
ticle swarm optimization algorithm. Computational Economics, 62(1), 165–186.
266 F. A. Syed et al.

Behera, S., Nayak, S. C., & Kumar, A. P. (2023). A comprehensive survey on higher order neural
networks and evolutionary optimization learning algorithms in financial time series forecasting.
Archives of Computational Methods in Engineering, 30, 4401–4448.
Bejani, M. M., & Ghatee, M. (2021). A systematic review on overfitting control in shallow and deep
neural networks. Artificial Intelligence Review, 54, 6391–6438.
Bilbao, I., & Bilbao, J. (2017). Overfitting problem and the over-training in the era of data: Particu-
larly for Artificial Neural Networks. In 2017 Eighth International Conference on Intelligent
Computing and Information Systems (ICICIS) (pp. 173–177). IEEE.
Bukhari, A. H., Shoaib, M., Kiani, A. K., Chaudhary, N. I., Raja, M. A. Z., & Shu, C. M. (2023).
Dynamical analysis of nonlinear fractional order Lorenz system with a novel design of intelligent
solution predictive radial base networks. Mathematics and Computers in Simulation. [Link]
org/​10.​1016/j.​matcom.​2023.​06.​005
Bukhari, A. H., Sulaiman, M., Raja, M. A. Z., Islam, S., Shoaib, M., & Kumam, P. (2020). Design of
a hybrid NAR-RBFs neural network for nonlinear dusty plasma system. Alexandria Engineering
Journal, 59(5), 3325–3345.
Çalış, Y., Demirci, A., & Özemir, C. (2022). Hopf bifurcation of a financial dynamical system with
delay. Mathematics and Computers in Simulation, 201, 343–361.
Charpentier, A., Elie, R., & Remlinger, C. (2023). Reinforcement learning in economics and
finance. Computational Economics, 62, 425–462. [Link]
Chen, S. B., Jahanshahi, H., Abba, O. A., Solís-Pérez, J. E., Bekiros, S., Gómez-Aguilar, J. F., &
Chu, Y. M. (2020). The effect of market confidence on a financial system from the perspective of
fractional calculus: Numerical investigation and circuit realization. Chaos, Solitons & Fractals,
140, 110223.
Chen, W. C. (2008). Nonlinear dynamics and chaos in a fractional-order financial system. Chaos, Soli-
tons & Fractals, 36(5), 1305–1314.
Douiba, M., Benkirane, S., Guezzaz, A., & Azrour, M. (2023). An improved anomaly detection model
for IoT security using decision tree and gradient boosting. The Journal of Supercomputing,
79(3), 3392–3411.
Faheem, M., Raza, A., & Khan, A. (2021). Collocation methods based on Gegenbauer and Bernoulli
wavelets for solving neutral delay differential equations. Mathematics and Computers in Simula-
tion, 180, 72–92.
Ghoddusi, H., Creamer, G. G., & Rafizadeh, N. (2019). Machine learning in energy economics and
finance: A review. Energy Economics, 81, 709–727.
Gogas, P., & Papadimitriou, T. (2021). Machine learning in economics and finance. Computational
Economics, 57, 1–4.
Gouravaraju, S., Narayan, J., Sauer, R. A., & Gautam, S. S. (2023). A Bayesian regularization-back-
propagation neural network model for peeling computations. The Journal of Adhesion, 99(1),
92–115.
Hadian Rasanan, A. H., Bajalan, N., Parand, K., & Rad, J. A. (2020). Simulation of nonlinear fractional
dynamics arising in the modeling of cognitive decision making using a new fractional neural net-
work. Mathematical Methods in the Applied Sciences, 43(3), 1437–1466.
Hajeb, M., Hamzeh, S., Alavipanah, S. K., Neissi, L., & Verrelst, J. (2023). Simultaneous retrieval of
sugarcane variables from Sentinel-2 data using Bayesian regularized neural network. International
Journal of Applied Earth Observation and Geoinformation, 116, 103168.
Hajipour, A., Hajipour, M., & Baleanu, D. (2018). On the adaptive sliding mode controller for a hyper-
chaotic fractional-order financial system. Physica a: Statistical Mechanics and Its Applications,
497, 139–153.
Hanif, A., Kashif Butt, A. I., & Ahmad, W. (2023). Numerical approach to solve Caputo-Fabrizio-frac-
tional model of corona pandemic with optimal control design and analysis. Mathematical Methods
in the Applied Sciences, 46, 9751–9782.
Huang, C., Cai, L., & Cao, J. (2018). Linear control for synchronization of a fractional-order time-
delayed chaotic financial system. Chaos, Solitons & Fractals, 113, 326–332.
Ilyas, H., Ahmad, I., Raja, M. A. Z., & Shoaib, M. (2021). A novel design of Gaussian WaveNets for
rotational hybrid nanofluidic flow over a stretching sheet involving thermal radiation. International
Communications in Heat and Mass Transfer, 123, 105196.
Jahanshahi, H., Orozco-López, O., Munoz-Pacheco, J. M., Alotaibi, N. D., Volos, C., Wang, Z., & Chu,
Y. M. (2021a). Simulation and experimental validation of a non-equilibrium chaotic system. Chaos,
Solitons & Fractals, 143, 110539.
Design of Neuro‑Stochastic Bayesian Networks for Nonlinear… 267

Jahanshahi, H., Sajjadi, S. S., Bekiros, S., & Aly, A. A. (2021b). On the development of variable-order
fractional hyperchaotic economic system with a nonlinear model predictive controller. Chaos, Soli-
tons & Fractals, 144, 110698.
Jin, T., & Yang, X. (2021). Monotonicity theorem for the uncertain fractional differential equation and
application to uncertain financial market. Mathematics and Computers in Simulation, 190, 203–221.
Khan, H. A., Ghorbani, S., Shabani, E., & Band, S. S. (2024). Enhancement of neural networks model’s
predictions of currencies exchange rates by phase space reconstruction and Harris Hawks’ optimiza-
tion. Computational Economics, 63, 835–860. [Link]
Kordkheili, M. S., & Rahimpour, F. (2023). Artificial neural network and semi-empirical modeling of
industrial-scale Gasoil hydrodesulfurization reactor temperature profile. Mathematics and Comput-
ers in Simulation, 206, 198–215.
Kumar, Y., & Singh, V. K. (2021). Computational approach based on wavelets for financial mathematical
model governed by distributed order fractional differential equation. Mathematics and Computers in
Simulation, 190, 531–569.
Kurani, A., Doshi, P., Vakharia, A., & Shah, M. (2023). A comprehensive comparative study of artificial
neural network (ANN) and support vector machines (SVM) on stock forecasting. Annals of Data
Science, 10(1), 183–208.
Lei, Y., Qiaoming, H., & Tong, Z. (2023). Research on supply chain financial risk prevention based on
machine learning. Computational Intelligence and Neuroscience, 2023, 6531154. [Link]
1155/​2023/​65311​54
Li, X., Wang, J., & Yang, C. (2023). Risk prediction in financial management of listed companies based
on optimized BP neural network under digital economy. Neural Computing and Applications, 35(3),
2045–2058.
Liu, Y., Wang, Z., & Huang, X. (2022). Multistability analysis of state-dependent switched Hopfield neu-
ral networks with the Gaussian-wavelet-type activation function. Mathematics and Computers in
Simulation, 196, 232–250.
Lodhi, S., Manzar, M. A., & Raja, M. A. Z. (2019). Fractional neural network models for nonlinear Ric-
cati systems. Neural Computing and Applications, 31, 359–378.
Magazzino, C., & Mele, M. (2022). Can a change in FDI accelerate GDP growth? Time-series and ANNs
evidence on Malta. The Journal of Economic Asymmetries, 25, e00243.
Mak, S., Sung, C. L., Wang, X., Yeh, S. T., Chang, Y. H., Joseph, V. R., & Wu, C. J. (2018). An effi-
cient surrogate model for emulation and physics extraction of large eddy simulations. Journal of the
American Statistical Association, 113(524), 1443–1456.
Milovanović, S., & von Sydow, L. (2020). A high order method for pricing of financial derivatives using
radial basis function generated finite differences. Mathematics and Computers in Simulation, 174,
205–217.
Nisa, S. U., Mahmood, A., Ujager, F. S., & Malik, M. (2023). HIV/AIDS predictive model using random
forest based on socio-demographical, biological and behavioral data. Egyptian Informatics Journal,
24(1), 107–115.
Nonaka, M., Agüero, M., & Kovalsky, M. (2023). Machine learning algorithms predict experimental out-
put of chaotic lasers. Optics Letters, 48(4), 1060–1063.
Oyedele, A. A., Ajayi, A. O., Oyedele, L. O., Bello, S. A., & Jimoh, K. O. (2023). Performance evalua-
tion of deep learning and boosted trees for cryptocurrency closing price prediction. Expert Systems
with Applications, 213, 119233.
Piazzola, C., Tamellini, L., Pellegrini, R., Broglia, R., Serani, A., & Diez, M. (2023). Comparing multi-
index stochastic collocation and multi-fidelity stochastic radial basis functions for forward uncer-
tainty quantification of ship resistance. Engineering with Computers, 39(3), 2209–2237.
Ping, M., Jia, X., Papadimitriou, C., Han, X., Jiang, C., & Yan, W. (2023). A hierarchical Bayesian frame-
work embedded with an improved orthogonal series expansion for Gaussian processes and fields
identification. Mechanical Systems and Signal Processing, 187, 109933.
Platt, D. (2022). Bayesian estimation of economic simulation models using neural networks. Computa-
tional Economics, 59(2), 599–650.
Polyzos, E., Samitas, A., & Rubbaniy, G. (2023). The perfect bail‐in: Financing without banks using
peer‐to‐peer lending. International Journal of Finance & Economics. [Link]
2838.
Polyzos, S., Samitas, A., & Katsaiti, M. S. (2020). Who is unhappy for Brexit? A machine-learning,
agent-based study on financial instability. International Review of Financial Analysis, 72, 101590.
268 F. A. Syed et al.

Poufinas, T., Gogas, P., Papadimitriou, T., & Zaganidis, E. (2023). Machine learning in forecasting motor
insurance claims. Risks, 11(9), 164.
Ribli, D., Pataki, B. Á., Zorrilla Matilla, J. M., Hsu, D., Haiman, Z., & Csabai, I. (2019). Weak lensing
cosmology with convolutional neural networks on noisy data. Monthly Notices of the Royal Astro-
nomical Society, 490(2), 1843–1860.
Sabir, Z., Saoud, S., Raja, M. A. Z., Wahab, H. A., & Arbi, A. (2020). Heuristic computing technique for
numerical solutions of nonlinear fourth order Emden-Fowler equation. Mathematics and Computers
in Simulation, 178, 534–548.
Samitas, A., Kampouris, E., & Kenourgios, D. (2020). Machine learning as an early warning system to
predict financial crisis. International Review of Financial Analysis, 71, 101507.
Samitas, A., Kampouris, E., & Polyzos, S. (2022). Covid-19 pandemic and spillover effects in stock mar-
kets: A financial network approach. International Review of Financial Analysis, 80, 102005.
Santur, Y. (2023). A novel financial forecasting approach using deep learning framework. Computational
Economics, 62, 1341–1392.
Sarı, B., Türkeş, S., Güney, H., & Keskinkan, O. (2023). The utilization and modeling of photo-fenton
process as a single unit in textile wastewater treatment. CLEAN—Soil, Air, Water, 51(1), 2100328.
Sariev, E., & Germano, G. (2020). Bayesian regularized artificial neural networks for the estimation of
the probability of default. Quantitative Finance, 20(2), 311–328.
Sattar, D., & Braik, M. S. (2023). Metaheuristic methods to identify parameters and orders of fractional-
order chaotic systems. Expert Systems with Applications, 228, 120426.
Seboka, B. T., Yehualashet, D. E., & Tesfa, G. A. (2023). Artificial intelligence and machine learning
based prediction of viral load and CD4 status of people living with HIV (PLWH) on anti-retroviral
treatment in Gedeo Zone public hospitals. International Journal of General Medicine, 16, 435–451.
[Link]
Shao, K., Zhou, L., Guo, H., Xu, Z., & Chen, R. (2019, July). Finite-time synchronization and param-
eter identification of fractional-order Lorenz chaotic system. In 2019 Chinese Control Conference
(CCC) (pp. 1120–1124). IEEE.
Shi, J., He, K., & Fang, H. (2022). Chaos, Hopf bifurcation and control of a fractional-order delay finan-
cial system. Mathematics and Computers in Simulation, 194, 348–364.
Shi, W., Chen, Q., Li, N., & Ca, G. (2021). Control and synchronization of hyperchaotic financial system
based on computer simulation. In 2021 Second International Conference on Electronics and Sus-
tainable Communication Systems (ICESC) (pp. 385–388). IEEE.
Shoaib, M., Anwar, N., Ahmad, I., Naz, S., Kiani, A. K., & Raja, M. A. Z. (2023). Neuro-computational
intelligence for numerical treatment of multiple delays SEIR model of worms propagation in wire-
less sensor networks. Biomedical Signal Processing and Control, 84, 104797.
Singh, R. K., Singh, A. R., & Yadav, R. K. (2023). A balanced-quantum inspired evolutionary algorithm
for solving disassembly line balancing problem. Applied Soft Computing, 132, 109840.
Szabó, R., Szklenár, T., & Bódi, A. (2022). Machine learning in present day astrophysics. Europhysics
News, 53(2), 22–25.
Taleizadeh, A. A., Safaei, A. Z., Bhattacharya, A., & Amjadian, A. (2022). Online peer-to-peer lending
platform and supply chain finance decisions and strategies. Annals of Operations Research, 315(1),
397–427.
Taloba, A. I. (2022). An artificial neural network mechanism for optimizing the water treatment process
and desalination process. Alexandria Engineering Journal, 61(12), 9287–9295.
Tang, J., Liu, G., & Pan, Q. (2021). A review on representative swarm intelligence algorithms for solving
optimization problems: Applications and trends. IEEE/CAA Journal of Automatica Sinica, 8(10),
1627–1643.
Umar, M., Raja, M. A. Z., Sabir, Z., Alwabli, A. S., & Shoaib, M. (2020). A stochastic computational
intelligent solver for numerical treatment of mosquito dispersal model in a heterogeneous environ-
ment. The European Physical Journal plus, 135(7), 1–23.
Verma, S., Sahu, S. P., & Sahu, T. P. (2023). Two-stage hybrid feature selection approach using Levy’s
flight based chicken swarm optimization for stock market forecasting. Computational Economics.
[Link]
Wang, B., Jahanshahi, H., Arıcıoğlu, B., Boru, B., Kacar, S., & Alotaibi, N. D. (2022a). A variable-order
fractional neural network: Dynamical properties, Data security application, and synchronization
using a novel control algorithm with a finite-time estimator. Journal of the Franklin Institute, 360,
13648–13670.
Design of Neuro‑Stochastic Bayesian Networks for Nonlinear… 269

Wang, B., Liu, J., Alassafi, M. O., Alsaadi, F. E., Jahanshahi, H., & Bekiros, S. (2022b). Intelligent
parameter identification and prediction of variable time fractional derivative and application in a
symmetric chaotic financial system. Chaos, Solitons & Fractals, 154, 111590.
Wang, S., He, S., Yousefpour, A., Jahanshahi, H., Repnik, R., & Perc, M. (2020). Chaos and complexity
in a fractional-order financial system with time delays. Chaos, Solitons & Fractals, 131, 109521.
Wang, Y. L., Jahanshahi, H., Bekiros, S., Bezzina, F., Chu, Y. M., & Aly, A. A. (2021). Deep recurrent
neural networks with finite-time terminal sliding mode control for a chaotic fractional-order finan-
cial system with market confidence. Chaos, Solitons & Fractals, 146, 110881.
Wang, Y. S., Jiang, X., & Liu, Z. J. (2016). Bank failure prediction models for the developing and devel-
oped countries: Identifying the economic value added for predicting failure. Asian Economic and
Financial Review, 6(9), 522–533.
Wang, Z., Huang, X., & Shi, G. (2011). Analysis of nonlinear dynamics and chaos in a fractional order
financial system with time delay. Computers & Mathematics with Applications, 62(3), 1531–1539.
Wen, C., & Yang, J. (2019). Complexity evolution of chaotic financial systems based on fractional calcu-
lus. Chaos, Solitons & Fractals, 128, 242–251.
Wichmann, F. A., & Geirhos, R. (2023). Are deep neural networks adequate behavioral models of human
visual perception? Annual Review of Vision Science, 9, 501–524.
Wu, J., Fang, L., Dong, G., & Lin, M. (2023). State of health estimation of lithium-ion battery with
improved radial basis function neural network. Energy, 262, 125380.
Youscri, D., & Mirjalili, S. (2020). Fractional-order cuckoo search algorithm for parameter identification
of the fractional-order chaotic, chaotic with noise and hyper-chaotic financial systems. Engineering
Applications of Artificial Intelligence, 92, 103662.
Yousefpour, A., & Jahanshahi, H. (2019). Fast disturbance-observer-based robust integral terminal slid-
ing mode control of a hyperchaotic memristor oscillator. The European Physical Journal Special
Topics, 228, 2247–2268.
Zhang, H., Zang, Z., Zhu, H., Uddin, M. I., & Amin, M. A. (2022). Big data-assisted social media analyt-
ics for business model for business decision making system competitive analysis. Information Pro-
cessing & Management, 59(1), 102762.
Zhou, J., Chen, S. L. P., Shi, W. W., & Kanrak, M. (2023). Cruise supply chain risk mitigation strategies:
An empirical study in Shanghai, China. Marine Policy, 153, 105600.
Zhou, S. S., Jahanshahi, H., Din, Q., Bekiros, S., Alcaraz, R., Alassafi, M. O., & Chu, Y. M. (2021).
Discrete-time macroeconomic system: Bifurcation analysis and synchronization using fuzzy-based
activation feedback control. Chaos, Solitons & Fractals, 142, 110378.

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under
a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of such publishing agreement and
applicable law.

Authors and Affiliations

Farwah Ali Syed1 · Kwo‑Ting Fang1 · Adiqa Kausar Kiani2 ·

Muhammad Shoaib3 · Muhammad Asif Zahoor Raja2

* Adiqa Kausar Kiani

adiqa@[Link]
Farwah Ali Syed
d11123010@[Link]
Kwo‑Ting Fang
fangkt@[Link]
270 F. A. Syed et al.

Muhammad Shoaib
[Link]@[Link]
Muhammad Asif Zahoor Raja
rajamaz@[Link]
1
Department of Information Management, National Yunlin University of Science
and Technology, 123 University Road, Section 3, Douliou 64002, Yunlin, Taiwan, ROC
2
Future Technology Research Center, National Yunlin University of Science and Technology,
123 University Road, Section 3, Douliou 64002, Yunlin, Taiwan, ROC
3
AI Center, Yuan Ze University, Taoyuan 320, Taiwan, ROC