2024 Volume 33 Issue 11
Article Contents

Kehao Yang, Song Zheng, Tianhu Yu, Aceng Sambas, Muhamad Deni Johansyah, Hassan Saberi-Nik, Mohamad Afendee Mohamed. Dynamical analysis, control, boundedness, and prediction for a fractional-order financial risk system[J]. Chinese Physics B, 2024, 33(11): 110501. doi: 10.1088/1674-1056/ad7afb
Citation: Kehao Yang, Song Zheng, Tianhu Yu, Aceng Sambas, Muhamad Deni Johansyah, Hassan Saberi-Nik, Mohamad Afendee Mohamed. Dynamical analysis, control, boundedness, and prediction for a fractional-order financial risk system[J]. Chinese Physics B, 2024, 33(11): 110501. doi: 10.1088/1674-1056/ad7afb

Dynamical analysis, control, boundedness, and prediction for a fractional-order financial risk system

  • Received Date: 06/07/2024
    Accepted Date: 06/09/2024
    Available Online: 01/10/2024
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Figures(28)  /  Tables(2)

Article Metrics

Article views(492) PDF downloads(0) Cited by(0)

Access History

Dynamical analysis, control, boundedness, and prediction for a fractional-order financial risk system

Abstract: This paper delves into the dynamical analysis, chaos control, Mittag–Leffler boundedness (MLB), and forecasting a fractional-order financial risk (FOFR) system through an absolute function term. To this end, the FOFR system is first proposed, and the adomian decomposition method (ADM) is employed to resolve this fractional-order system. The stability of equilibrium points and the corresponding control schemes are assessed, and several classical tools such as Lyapunov exponents (LE), bifurcation diagrams, complexity analysis (CA), and 0–1 test are further extended to analyze the dynamical behaviors of FOFR. Then the global Mittag–Leffler attractive set (MLAS) and Mittag–Leffler positive invariant set (MLPIS) for the proposed financial risk (FR) system are discussed. Finally, a proficient reservoir-computing (RC) method is applied to forecast the temporal evolution of the complex dynamics for the proposed system, and some simulations are carried out to show the effectiveness and feasibility of the present scheme.

1.   Introduction
  • Since fractional-order calculus more close to reality, it has been adopted to model various biological, medical, chemical, and industrial phenomena, and so on.[1–5] In fact, fractional-order modeling provides a more dependable method for predicting, describing, and managing the behaviors of dynamic systems. Many works showed the fractional models, such as the recently developed hyperchaos complex system of fractional-order,[6] neural networks featuring hub structure, and multiple rings of fractional-order,[7] bank data model of -order integrating two unequal time delays,[8] fractional order monkeypox virus model,[9] and fractional buffering model,[10] could offer more accurate representations compared with their integer-order counterparts. Up to date, chaos in fractional-order dynamic systems, which exhibit some excellent and interesting dynamical behaviors, has been emerged as a fascinating subject for the researchers.[11–15]

    Estimating the boundaries of solutions for chaotic systems is one of recently interesting focal points to investigate the dynamic behavior in a system. It is commonly known that the ultimate bounds and major dynamical system is crucial to be determined, in particular with practical considerations in locating hidden attractors, chaos synchronization, and chaos control.[16,17] Leonov[18,19] first proposed the computation of the Lorenz system’s global bound. Later, Swinnerton–Dyer[20] applied Lyapunov functions to study the solution bounds of the Lorenz equations. Liao et al.[21] further designed a family of generalized Lyapunov functions, where they derived a positive invariant set of the Lorenz system and a new global attractive set by the application of synchronization and chaos control. Jian et al.[22] discussed the MLB and provided estimates of the global MLAS and MLPIS for a fractional-order chaotic system. Peng et al.[23] extended their exploration into the MLB and synchronization of fractional-order plasma chaotic systems, employing fractional differential inequalities and generalized Lyapunov functions. Ren et al.[24] employed the Lagrange coefficient method to establish a precise four-dimensional (4D) Rabinovich hyperchaotic system’s ultimate bound set.

    Chaotic economics, a recent theory within the realm of physical economics, can identify organized patterns and systematic behavior obscured within seemingly random economic phenomena. Many studies have explored the application of economic and FR models[25–28] by chaotic systems theory. Zhang et al.[25] constructed a risk management process of financial system, studied some interesting dynamical behaviors, and obtained the conditions for stability of the proposed system, respectively. Tacha et al.[26] investigated dynamical analysis, control, and circuit simulation of a financial system with an absolute function term. Furthermore, Tacha et al.[27] also discussed dynamical analysis of a fractional-order finance system with negative parameters. Jahanshahi et al.[28] examined the dynamical behavior of a 4D FR hyperchaotic system with variable-order fractional derivatives, and designed the predictive control scheme for suppressing the chaotic behavior while stabilizing the close-loop system. These investigations suggest that economics and financial systems adhere to a highly complex nonlinear system characteristics and cannot be simply modulated by integer-order calculus.

    Recently, the time series forecasting has been employed in several fields including agriculture, commerce, biology, medicine, and so on.[29–31] There has been a proliferation of time series prediction methods in the field of artificial neural networks, such as recurrent neural networks (RNN), feed-forward neural networks (FNN), and so on. Time series prediction has made extensive use of long short-term memory neural networks (LSTM), a unique kind of recurrent neural networks.[32,33] The improvement of LSTM network is the use of extra gates and better control over the processing information in the hidden units. Therefore, LSTM networks outperform traditional recurrent neural networks in their ability to learn long-term associations in time series and to predict sequence data. However, certain intricate nonlinear behaviors are challenging to handle by using LSTM networks and it has high computational costs. Due to the shortcomings of traditional recurrent neural networks, RC algorithm known as echo state network[34–36] has been developed. It is extensively applied in time series prediction because of its simpler structures, less computational costs, and higher accuracy. For example, Song et al.[37] implemented the RC algorithm for real-time power load forecasting, demonstrating superior predictive performance compared to FFNN. Na et al.[38] proposed a hierarchical echo state network with sparse learning and demonstrated the effectiveness based on experiments.

    The motivation of this work can be summarized as follows: The financial risk system is characterized by variability and complexity. Traditional integer-order calculus has serious limitations in predicting and controlling the risks of the financial system and cannot accurately predict the time of occurrence and the degree of risk. Therefore, the application of fractional-order calculus is essential in the study of financial risk system. We delve deeper into the dynamical behavior and discuss LE, bifurcation diagrams, the C0 algorithm,[39] SE algorithm,[40] and 0–1 test.[41] Then the global MLAS and MLPIS are studied. We also prove the attractiveness of the ultimate bound set of the proposed FOFR system. By modifying system parameters and other conditions, a multitude of attractive sets can be obtained. On the other hand, the prediction of financial risks is very favorable for the prevention and control of financial crises by an accurate and simpler algorithm like RC scheme. To our knowledge, the fractional dynamical behavior, MLB, and forecasting of the FOFR system with one absolute function term have not been studied thus far. Therefore, further in-depth study of the FOFR system with one absolute function term is warranted. The contributions of this paper are of four-fold.

    (i) Proposing a fractional-order financial risk (FOFR) system featuring an absolute function term.

    (ii) Investigating the dynamical analysis and complexity analysis using Lyapunov exponents, bifurcation analysis, 0–1 test, C0, and SE.

    (iii) The attractiveness of the ultimate bound set of the FOFR system is obtained.

    (iv) An efficient RC approach is employed to predict the temporal evolution of the complicated dynamics.

    The remainder of this paper is structured as follows. Section 2 presents some basic definitions, lemmas, and modelling. In Section 3, according to ADM, we investigate the stability analysis of equilibrium points and the dynamical behaviors in the proposed system, respectively. In Section 4, the linear controller will be designed to stabilize system at equilibrium points. Section 5 focuses on estimation of the global MLAS and MLPIS for the proposed system. Section 6 introduces the mathematical model for the RC technique along with numerical results. Furthermore, comparisons with the LSTM technique are conducted. The study is then summarized, and concluding remarks are provided in the last section.

2.   Preliminaries and modeling
  • Numerous definitions available out there for fractional order calculus,[1,2] such as Caputo, Riemann–Liouville, and Grunwald–Letnikov. In this work, we apply the Caputo definition. The following definitions apply to the Caputo integral and derivative in Refs. [1,2].

    Definition 1[1,2] Given f(t), its Caputo fractional integral can be defined as

    where Γ (⋅) is a gamma function, and t ≥ t0, q > 0.

    Definition 2[1,2] The Caputo fractional derivative of order q is given as follows:

    where m–1 < q < m, t0 and t are integers that indicate the operator’s bounds. t0CDtq. Also, when 0 < q < 1, the fractional-order derivative of f(t) can be concluded as

    Definition 3[1,2] The following equations defines two parameters Mittag–Leffler function:

    for some complex number m. Simple operation gives

    Definition 4[22] Consider the Lyapunov function Vλ(t) = Vλ (x(t)) with λ > 0, if there exist constants Lλ > 0, rλ > 0, and x0 ∈ Rm such that

    for t ≤ t0, Vλ (t0) > Lλ, Vλ (t) > Lλ, then the set Ψλ = {x(t) ∈ Rm| Vλ (x(t)) ≤ Lλ} is said to be a global MLAS of the system. If for any t > t0, x(t,t0, x0) ∈ Ψλ, t > t0, or x0 ∈ Ψλ, then Ψλ is said to be an MLPIS, where x(t) = (x1,x2,…,xm)T,x0 = x(t0).

    Lemma 1[3] Consider a nonlinear dynamical system of fractional-order t0CDtqx(t)=f(x(t)), with 0 < q < 1, x(t) ∈ Rm, and f(x(t)) as a continuous differential nonlinear function. Moreover, let a respective Jacobian matrix J = ∂f/∂x with eigenvalues λi at the equilibrium point, if all eigenvalues satisfy |arg (λi)| > qπ/2, where i = 1,2,…,m, then the system is local asymptotic stability.

    Lemma 2[22] Let f(t) ∈ R be a smooth function, then t0CDtq(f2(t))≤2f(t)t0CDtq(f(t)), t ≥ t0, 0 < q < 1.

    Lemma 3[22] Let q ∈ (0,1) and a constant K ∈ R. If a continuous function f(t) satisfies t0CDtq(f(t))≤Kf(t), then f(t) ≤ f(0)Eq(Ktq), for the case of t ≥ 0.

  • Muhamad Deni Johansyah et al.[42] proposed a new FR system can be defined as

    where {x1,x2, and x3} = {occurrence value risk, analysis value risk, and control value risk} for the current market. Inspired by the integer-order FR system in Ref. [42]. This study applies the Caputo fractional derivative operator to the system (1). Consequently we can derive the FR system in the form of fractional-order differential equations as follows:

    where q ∈ (0,1) represents the order of derivative. In case of q = 1, system (2) degenerates into an integer-order differential equation. An approximate solution by using the ADM[43] to the FOFR system (2) can be simplified into the following equation:

    where j = 1, 2, 3, while the detailed derivation of bj1,bj2,…,bj6 is available in Appendix A.

    Based on this approximate solution, when h = 0.01, q = 0.98, a = 10, b = 28, c = 0.4, d = 8/3 with the initial values (10,10,10), the phase diagram of the FOFR system (2) can be obtained as shown in Fig. 1, which exhibits chaotic characteristics.

  • To solve the equilibrium points for system (2), let

    we obtain a better understanding of the system’s dynamics by solving Eq. (4).

    When x2 > 0, if bd–d–cd < 0, the system has a unique equilibrium point E1 = (0,0,0);bd – d – cd ≥ 0, the system has two equilibrium points E1 = (0,0,0) and E2=(bd−d−cd,bd−d−cd,b−c−1).

    When x2 < 0, if bd–d+cd < 0, the system has a unique equilibrium point E1 = (0,0,0); bd – d + cd ≥ 0, the system has two equilibrium points E1 = (0,0,0), E2=(−bd−d+cd,−bd−d+cd,b+c−1).

    To summarize, there are three equilibrium points when the conditions are satisfied:

    By considering a = 10, b = 28, c = 0.4, and d = 8/3 such that bd–d–cd ≥ 0, bd – d + cd ≥ 0, system (2) exhibits the three equilibrium points to be

    The Jacobian matrix of the system (2) at the equilibrium point of E* is

    By substituting E1 = (0,0,0) into the Jacobian matrix (6), the eigenvalues are λ1 = 11.828, λ2 = 2.667, λ3 = −22.828. By Lemma 1, we have |arg(λ1)| = 0, |arg(λ2)| = 0, and |arg(λ3)| = π, therefore E1 is said to be unstable.

    By substituting E2 = (8.4221,8.4221,26.6) into the Jacobian matrix (6), the eigenvalues are λ1 = −14.044,λ2,3 = −0.0113132 ± 10.0506i. By Lemma 1, we obtain |arg(λ1)| = π and |arg(λ2)| = |arg (λ3)| = 1.5719, hence E2 is locally asymptotically stable.

    By substituting E3 = (−8.5479,−8.5479,27.4) into the Jacobian matrix (6), the eigenvalues are λ1 = −13.6725,λ2,3 = 0.202887 ± 10.3364i, so we have |arg(λ1)| = π, |arg(λ2)| = 1.5506, |arg(λ3)| = 1.5506, by Lemma 1, E3 is local asymptotic stability when q < 2|arg (λ2,3)|/π = 0.9871.

3.   Dynamical analysis of the FOFR system
  • This section analyzes the dynamical properties of the proposed model in relation to its parameters. We examine the equilibrium point characteristics by utilizing various techniques based on bifurcation diagram and LE. The FOFR system (2) exhibits a range of dynamic behaviors that includes stable, periodic, and chaotic motions.

  • Figures 2 and 3 respectively depict the bifurcation diagram and spectrum of LE when d takes the values between [0,3.5]. When d ∈ (0.15,0.5), the system exhibits periodic behavior and has two negative LE and one zero LE. When d ∈ (0,0.15) ∪ (0.5,3.2), the system exhibits one positive LE which concludes the chaoticity of the system. When d ∈ (3.2,3.5), there are three negative LEs which means the system is stable.

  • Following a similar approach, we conduct further analysis by varying parameter q. The bifurcation diagram is displayed in Fig. 4 when the parameter q ∈ (0.75,1) is considered. It can be observed that the system exhibits a stable state when q ∈ (0.75,0.93), while it demonstrates chaotic behavior when q ∈ (0.93,1). Additionally, the LE spectrum shown in Fig. 5 confirms the chaotic nature of the system, aligning with the findings from the bifurcation diagram. Specifically, when q ∈ (0.75,0.93), all three LEs are negative, indicating a stable state. However, when q ∈ (0.93,1), the system displays chaos with one positive LE.

  • By varying d, the phase portraits and the p–s dynamics of the 0–1 test are plotted as in Fig. 6. Apparently, the system exhibits periodicity when d = 0.15. When d = 2 and d = 3.3, the system is chaotic and stable. It shows that the system goes through periodic behavior and becomes gradually stable as d improves. Additionally, results from numerical imaging corroborate with the C and SE algorithms, found in Fig. 7.

  • Figure 8 provides phase portraits and p–s dynamics using the 0–1 test for specific values of q. It can be observed that the system exhibits chaos when q = 0.94 and q = 0.98, while it shows stable behavior when q = 0.84. Furthermore, figure 9 presents the C0 and SE complexity values belonging to system (2). Given smaller values of q ∈ (0.75,0.93), the complexity values are low, whereas oscillations are observable at higher values of q ∈ (0.93,1). This suggests that as the parameter q improves, the system gradually becomes more chaotic. It indicates that as the fractional order q gradually increases, the financial system is out of control and more damaging to the economy. Figure 10 shows the corresponding time series diagram of the system. As shown in figure, the smaller the order q, the faster the system reaches the stable state and the occurrence value risk fluctuate in the smaller range. In practice, this means that financial risks that spread on a small scale can be effectively controlled in a relatively short period of time. Thus, it indicates that the fraction-order system more accurately reflects the real-world financial risk system compared to integer-order one.

4.   Chaos control
  • This section begins with the employment of linear feedback control method for the purpose of stabilizing system (2) towards its equilibrium points. In so doing, consider the following system:

    where w1, w2, w3 are control parameters and E* is the equilibrium point of the system (2).

    The Jacobian matrix of the system (7) at the equilibrium point of E* is

    For this Jacobian matrix J(E*), its characteristics equation is given by

    When a = 10, b = 28, c = 0.4, d = 8/3, we first select w1 = 12, w2 = 14, w3 = 10 to stabilize system (2) at equilibrium point E1 (0,0,0). The characteristic polynomial is given as

    then, the eigenvalues of the characteristic polynomial (10) are λ1 = −35.5953, λ2 = −1.4047, λ3 = −12.6667. Since |arg(λ1)| = |arg (λ2)| = |arg (λ3)| = π, the equilibrium point E1 (0,0,0) is locally asymptotically stable for all 0 < q < 1. Furthermore, in order to get a simpler control scheme, we can consider the case where one control parameter is 0 or two control parameters are 0. Therefore, we can choose w1 = 18, w2 = 20, w3 = 0 or w1 = 0, w2 = 62, w3 = 0. The eigenvalues of the Jacobian matrix (8) are λ1 = −41.5953, λ2 = −7.4047, λ3 = −2.6667 and λ1 = −5.1591, λ2 = −67.8409, λ3 = −2.667, respectively. Since |arg(λ1)| = |arg (λ2)| = |arg (λ3)| = π, the equilibrium point E1 (0,0,0) is local asymptotic stability for all 0 < q < 1 in both cases.

    The effectiveness of the designed control method is demonstrated in Figs. 11–13. These figures indicate that the designed controllers are capable of effectively controlling chaos in the FOFR system (2). Moreover, as the fractional order decreases, the system appears to move much more rapidly towards the equilibrium point E1 (0,0,0).

5.   Boundedness of the FOFR system
  • In this section, we will estimate the global MLAS and MLPIS for the proposed system (2).

    Theorem 1 When x2 > 0, let a > 0, c > −1, d > 0, λ1 > 0, μ1 > λ1b, the global MLAS and MLPIS of system (2) is as follows:

    where

    When x2 < 0, let a > 0, c < 1, d > 0, λ2 > 0, μ2 > λ2b, the global MLAS and MLPIS of system (2) is as follows:

    where

    Proof When x2 > 0, specify the subsequent generalized Lyapunov function, which is both positively definite and radically unbounded

    where λ1 > 0, μ1 > 0. By applying the fractional derivative to Vλ1,μ1 and using Lemma 2, we can obtain

    where

    Evidentially,

    we have

    According to Lemma 3, we have

    Based on Definition 4, we can conclude that Ψλ1,μ11 with a > 0, c > −1, d > 0, λ1 > 0, μ1>λ1b is an MLAS and MLPIS for system (2) when x2 > 0.

    When x2 < 0, choose the following Lyapunov function:

    where λ2 > 0,μ2 > 0. By applying the fractional derivative to Vλ2,μ2 and using Lemma 2, we can obtain

    where

    Apparently,

    Thus,

    According to Lemma 3, we have

    Based on Definition 4, we can conclude that Ψλ2,μ22 with a > 0, c < 1, d > 0, λ2 > 0, μ2 > λ2b is an MLAS and MLPIS for system (2) when x2 < 0. This completes the proof.

    Remark 1 If we take q = 0.98, a = 10, b = 28, c = 0.4, d = 8/3, λ1 = 1, μ1 = 29, λ2 = 1, μ2 = 32, then

    Figure 14 shows phase portraits and MLAS of system (2).

    Remark 2 When a = 10, b = 28, c = 1.4, d = 8/3, λ1 = λ2 = 1, μ1 = μ2 = 30, it is obvious that dμ12/λ1η1>dμ22/λ2η2, so the ellipsoid

    is the MLAS and MLPIS for system (2) with different fractional order q. Figure 15 shows phase portraits and MLASs of system (2) with different fractional order q.

6.   Forecasting of the FOFR system
  • The RC approach[34–36] comprises an input layer, a recurrent layer referred to as the reservoir (comprising many sparsely connected neurons) and an output layer. The connection weights of the input layer and the reservoir layer remain fixed after initialization, while the output weights are trainable and determined by solving a linear regression problem.

    In the given context, u(t) ∈ RM×1 represents the input value at time t of the time series, win ∈ RN×M (with dimension M) signifies the connection weights between the input layer and the hidden layer, wres ∈ RN×N (with N representing the number of neurons) denotes the input-to-hidden layer connection weights, and wout ∈ RL×N (with dimension L) represents the output weights from the hidden layer to the output layer. The state transition equation is expressed as follows, represented by equation

    In this equation, r(t) and y(t) represent the state of the reservoir and output, respectively. γ denotes the leaky rate, incorporating the previous state into the current one. The tanh (⋅) signifies the activation function within the reservoir.

    We study the time series solution of the proposed model in Eq. (2) by employing the RC forecasting approach. The generated time series consists of 104 data points representing a state variable. The first 7000 items is dedicated to training dataset while another 3000 is reserved for testing. To improve the reliability of the prediction, we remove the first 100 elements from both training data and testing data. Moreover, the number of RC nodes is limited to 52, while accounting only three cases of fractional order q. Figure 16 shows the phase diagrams of predicted data and test data. Figures 17–19 exhibit both results, the training and prediction processes at q = 0.97 given numerous values of state variables for the system (2). Figures 20–22 show the results of the training and prediction at q = 1, for the state variables of the system (2). Additionally, figures 23–25 illustrate the results from training and prediction processes at q = 0.93 for the state variables of the system (2). We use LSTM technique for the purpose of forecasting process of the financial time series. In so doing, the dataset is separated into the training dataset chosen from the first 7000, while a validation dataset for the other 3000. Examples of the obtained results are illustrated in Figs. 26–28. It is evident that the prediction process based on RC produces superior accuracy over extended durations and demonstrates notably reduced forecasting errors when contrasted with the LSTM method.

    Tables 1 and 2 look closer into the performance of the two approaches, specifically in terms of RMSE and MAE across various cases of q and different state variables.

7.   Conclusion
  • In this paper, we proposed an FOFR system with one absolute function term. Based on the ADM algorithm, we investigated the complexity and chaotic behavior of the system using various techniques such as phase portraits, bifurcation diagrams, Lyapunov exponent spectrum, SE, C0, and 0–1 test. Moreover, we investigated the stability of equilibrium points and explored techniques for controlling chaos. Furthermore, we provided estimations of the global MLAS and MLPIS for the FOFR model. Additionally, we presented a proficient machine learning method capable of forecasting the temporal evolution of the observed nonlinear behaviors within the proposed system, and compared with the RC-based model and the LSTM technique. The analytical results are substantiated by numerical simulations.

    Since the two important factors of time delay and fractional incommensurate-order in financial systems, the study of financial systems with time delay and fractional incommensurate-order is of great significance for understanding and managing financial markets. As such, our future work will focus on this issue.

  • Using ADM, the FOFR system (2) is decomposable into the following form:

    The results of the decomposition of the nonlinear terms are shown below, where the subscripts indicate the nonlinear terms being decomposed:

    The first set of coefficients bj0 (j = 1,2,3):

    The other sets of coefficient respectively are

    Combining the above derivations, the solution of the system can be obtained as

    where h denotes the step size.

Figure (28)  Table (2) Reference (43)

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return