Design and Application of Macro Economic Intelligent Prediction Decision Support System Based on Agent System

Economic forecasting is a prospective speculation on the future economic prospects based on scientific methods and means and the mastery of certain information materials, according to the laws of economic development. Macroeconomic forecasting plays an important role in economic decision-making [1-4]. At the government level, macroeconomic forecasts are an important basis for the state to formulate macroeconomic policies, prepare economic development plans, and carry out macroeconomic regulation and control. At the enterprise level, the judgment of macroeconomic trends is the primary premise for enterprises to make business decisions on production, sales and investment [5-8]. At the household level, a certain knowledge of macroeconomic forecasting helps households to better organize their consumption, savings and investment [9-10].

However, macroeconomic forecasting is a complex task that requires the consideration of many factors, including policy, industry and market trends. Since these factors are very complex and interrelated, it is difficult for traditional models to accurately predict economic trends [11-14]. Intelligent macroeconomic intelligent forecasting based on artificial intelligence technology provides more accurate macroeconomic forecasts by being able to process large amounts of data and utilizing techniques such as machine learning and natural language processing, especially the decision support system for intelligent macroeconomic forecasting based on Agent systems [15-18].

However, artificial intelligence in economic forecasting is not all about advantages, and there are still some challenges. One of the most important challenges is that the economic data itself is often dynamic and changes frequently, which may affect the accuracy of economic forecasting models [19-22]. For example, changes in government policies can have a significant impact on the market, and AI forecasting models may not be able to take such changes into account. In addition, since the predictions of AI models are based on historical data, the AI models may miss some emerging markets or chance events, which may reduce their accuracy [23-26].

Literature [27] discusses the progress in the application of AI to behavioral macroeconomic axes. It is pointed out that AI behavioral models have potential applications in areas such as forecasting accuracy, and the findings lay the foundation for future research and applications in this area. Literature [28] examines the role of AI in key areas such as macroeconomics and its impact on the global economy, and analyzes the opportunities and challenges it poses and considers its potential future development. Literature [29] proposes a new approach to incorporate sentiment from newspaper articles into macroeconomic forecasting. It also presented a thematic data filtering method for Bi-LSTM for extracting sentiment scores, emphasizing that emotions related to happiness and anger have the strongest predictive power for the predicted variables. Literature [30] describes predictive models for economic development applying machine learning techniques, especially support vector machines. And Bayesian approach and multiple kernel functions are taken to optimize the hyperparameters. And the effectiveness of the model is evaluated. Literature [31] proposed a macroeconomic forecasting model of LSTM neural network, aiming to realize intelligent and efficient macroeconomic forecasting. By constructing the evaluation index system and other steps, this forecasting method is emphasized to be able to carry out macroeconomic forecasting effectively. Literature [32] introduced a macroeconomic indicator forecasting method based on social network and semantic analysis techniques. It was found that the quantity and tone of news are important predictors of GDP, business and consumer confidence indices. Literature [33] examined the use of big data in economic forecasting and conducted a critical review of empirical studies utilizing big data sources, pointing out the limitations of the use of big data in macroeconomic forecasting. It is indicated that future work using big data should focus on improving the quality and accessibility of its use. Literature [34] utilized the LSTM algorithm to improve the efficiency of economic forecasting in order to solve the overfitting problem in macroeconomic forecasting. It reveals that the improved LSTM model can be effectively used in macroeconomic forecasting. Literature [35] proposed an economic forecasting method combining AI and big data analysis. And the factors such as political and social environment are combined to form the main body that affects the economy. It reveals that the proposed model can be used as a base model for statistics, analysis, decision-making and other functions of the economy.

It can be seen that machine learning technology, semantic analysis technology and so on have been applied in macroeconomic forecasting, which shows that intelligent macroeconomic forecasting has been developed into a more mature field, but the Agent system as a field of artificial intelligence has not been embodied in the relevant research in this field, which also shows that the forecasting system, technology and other aspects of intelligent macroeconomic forecasting are not comprehensive enough, so based on the Agent This also shows that the forecasting system, technology and other aspects of intelligent macroeconomic forecasting are not comprehensive enough, so the research on intelligent macroeconomic forecasting based on Agent system makes up for the shortcomings in this field.

The Agent-based Macroeconomic Intelligent Forecasting Decision Support System proposed in this paper is guided by the theory and methodology of dealing with complex macroeconomic problems, and consists of many different types of software Agents such as Management Agent, Interaction Agent, and Decision Agent. The mathematical model used in the system is the mixed-frequency data model, on the basis of which the mixed-frequency vector autoregressive model MF-VAR is proposed, and the mixed-frequency VAR model constructed with low-frequency quarterly variables and high-frequency monthly variables commonly used in macroeconomics is taken as an example for simple analysis, and Kalman smoothing operator is used to forecast macroeconomic indicators. The dynamic factor model and Bayesian estimation are introduced to the MF-VAR model to construct the MF-VAR model combined with the factor model and the BMF-VAR model. The former combines the state space of mixed-frequency data model with factor variables for macroeconomic forecasting, and the latter Gibbs sampling in the framework of Bayesian to realize macroeconomic forecasting decision. The model system is constructed from Agent technology, and the cooperation between model Agents is explored to finalize the construction of macroeconomic intelligent forecasting decision support system. The comprehensive validity of the system in this paper is examined in terms of reliability, responsiveness, and decision support error rate. A complete macroeconomic sample of China from January 1996 to December 2022 is used to carry out macroeconomic forecasting decision-making before and after the new crown epidemic in China.

2

Macroeconomic Intelligent Forecasting Decision Support System

Macroeconomic Intelligent Forecasting Decision Support System is an open and complex giant system with the participation of many people and involving many factors, which has some essential specialties such as giant system, openness and uncertainty. The Agent-based macroeconomic intelligent forecasting decision support system proposed in this paper is guided by the theoretical method of dealing with complex giant systems, and utilizes Agent to closely combine the theoretical method of dealing with complex systems, decision theory, information collection and knowledge discovery technology, macroeconomic forecasting and simulation to form a multi-Agent-based human-machine cooperation system to better help users to make macroeconomic decisions.

2.1

System Composition and Architecture

Agent-based macroeconomic decision support system consists of many different types of software Agents, which mainly include the following. 1)

Management Agent responsible for system management

2)

Interactive Agent with a friendly interface to interact with people

3)

Decision-making Agent composed of a decision-making method to make decisions

4)

Evaluation Agent that evaluates the decision-making scheme

5)

Combination Agent that generates new decision-making solutions by utilizing existing decision-making solutions.

6)

responsible for information collection, organization of data mining and knowledge discovery of the information Agent system includes a management Agent, a number of interactive Agents interacting with the system administrator, decision-making users and a number of experts, a number of decision-making Agents, information Agents, an evaluation Agent and a combination of Agentsp [36].

All Agents are distributed on networked computers and communicate using a network of networked calculators. Theoretically, each Agent can be distributed on any computer connected to the Internet, but the general distribution structure is to concentrate all the other Agents except the Expert Interaction Agent in the Intranet dedicated to the Macroeconomic Research Department, so that on the one hand to facilitate the management of the system, and on the other hand, for security and confidentiality considerations. The Expert Interaction Agent communicates with other Agents through the Internet.

2.2

Agent Functions and Operation

1)

Management Agent

It is the first Agent that runs up in the system, and other Agents joining the system must register with it. Registered information including Agent name, address (it is located in the network computer name or address), the type of Agent and its main functions. It also provides information about the Agent query function, so that an Agent can be queried to find the appropriate communication and cooperation objects.

2)

Interaction Agent

It has to interact with three types of users: system administrator, decision-making user and expert, which are called administrator interaction Agent, decision-making user interaction Agent and expert interaction Agent. system administrator first establishes management Agent when installing the system, and later administrator can directly interact with management Agent to help manage the system, but in order to operate more conveniently, administrators are generally However, in order to operate more conveniently, the administrators usually establish an administrator interaction Agent on the computer they often use, and communicate with the management Agent through the interaction Agent to complete the system management.

3)

Decision-making Agent

It consists of a decision-making method and the knowledge required for that decision-making method. For example, an Agent consisting of a likelihood-satisfaction multi-objective decision-making method and its corresponding decision-making knowledge can be called a P-S decision-making Agent. The Decision Agent is the main body of the computer to make decisions, it gets the decision problem from the Interaction Agent and gets some information needed to make decisions through the Information Agent.

4)

Combination Agent

The decision scheme obtained by the decision-making Agent is combined using certain methods to obtain a new decision scheme. By combining Agents, multiple decision-making methods can be combined.

5)

Evaluation Agent

Save the decision programs made by each decision-making Agent or expert, and use certain evaluation criteria to evaluate each program. And the decision-making program is simulated using the macroeconomic economic simulation system to test whether the decision-making program can get the expected results, so as to make the decision support system linked to the simulation system. Finally, each decision scheme and its evaluation results are sent to the decision-making user interaction Agent, and the user selects the final decision scheme.

6)

Information Agent

Information Agent firstly collects and organizes the data and information needed for macroeconomic decision-making, and there are two main sources of information and data: one is obtained from users through interaction with them. The other is through the Internet search to find useful information and data and fulfill the necessary procedures to obtain them.

2.3

Mathematical modeling in the system

2.3.1

Mixed Frequency Data Modeling

The research and application of mixing data models can be broadly categorized into the following three main groups. 1)

Mixed-frequency data conversion

Mixed-frequency data processing and the comprehensive use of mixed-frequency data results actually do not fully use the information contained in the high-frequency data in the mixed-frequency data, mainly including (1) high-frequency data converted to low-frequency data preprocessing methods, (2) low-frequency data converted to high-frequency data preprocessing methods.

2)

Distributed lag model

Distributed lag model can directly use mixed-frequency data for model construction and estimation, MIDAS model is evolved from the distributed lag model, MIDAS model is technically the mathematical expectation of the distribution lag of high-frequency explanatory variables under specific conditions [37].

The form of a simple linear distributional lag model can be expressed as: (1) $y_{t} = α_{0} + \sum_{j = 1}^{k} \sum_{i = 0}^{n} β_{i j} x_{j t - i / m} + ε_{t}$

where y_t denotes the low-frequency explanatory variable, x_jt denotes the jrd high-frequency explanatory variable, k denotes the number of high-frequency explanatory variables, n denotes the maximum lag order of the high-frequency explanatory variable, and m denotes the frequency multiplicity of the mixed-frequency data. Distributed lag models often add an autoregressive term to transform the model into an autoregressive distributed lag model. 3)

Mixed Frequency Data Model

Mixed-frequency data models are models that can make use of information from mixed-frequency data in a non-destructive way, without the need to pre-process the data or over-parameterize it, and can incorporate a large number of existing advanced measures. The two main ones, the MIDAS model and the MF-VAR model, which are the focus of the chapters of this thesis, are discussed below. (1)

The MIDAS model

The research and application of MIDAS model mainly focuses on the empirical research of MIDAS in finance and economic theory, and also includes the improvement and extension of the theoretical model of MIDAS. The MDAS model handles the problem of non-smooth mixing data with the help of the mixing error correction model.

(2)

Mixed-frequency state-space modeling

The research and application of frequency-mixing state-space model is mainly to regard the low-frequency data in the frequency-mixing data as unobservable (with default values) high-frequency data, and then apply the state-space model and Kalman filtering to obtain the missing high-frequency data in the low-frequency data.

2.3.2

Mixed-frequency vector autoregressive modeling

The MF-VAR model represents the mixed-frequency data as a state-space expression and treats the low-frequency data as unobservable state variables, and then estimates the model using either Kalman filtering or Bayesian estimation [38]. The following section describes the mixed-frequency VAR model estimated using Kalman filter and the mixed-frequency VAR model estimated using Bayesian methods, and introduces the mixed-frequency VAR model combined with the factor model. 1)

MF-VAR model

For analytical simplicity, the mixed-frequency VAR models constructed from low-frequency quarterly variables and high-frequency monthly variables commonly used in macroeconomics are presented below. (1)

Monthly and quarterly mixed-frequency VAR models

The mixed-frequency VAR model views the low-frequency data as a high-frequency data model with default values and obtains the monthly GDP metrics from the quarterly GDP through time decomposition and monthly metrics. The value of monthly GDP growth rate $y_{t}^{(m)}$ can be obtained from quarterly GDP growth rate y_t through the following summing relationship: (2) $y_{t} = \frac{1}{3} y_{t}^{(m)} + \frac{2}{3} y_{t - 1 / 3}^{(m)} + y_{t - 2 / 3}^{(m)} + \frac{2}{3} y_{t - 1}^{(m)} + \frac{1}{3} y_{t - 4 / 3}^{(m)}$

Then the potential monthly GDP growth rate $y_{t}^{(m)}$ and the monthly indicator $x_{t}^{(m)}$ can form a bivariate VAR system, viz: (3) $Φ (L^{(m)}) (\begin{matrix} y_{t}^{(m)} - μ_{x}^{(m)} \\ x_{t}^{(m)} - μ_{x} \end{matrix}) = u_{t}^{(m)}$

where $Φ (L^{(m)}) = \sum_{i - 1}^{p} Φ_{i} L^{(i / m)}$ , and $u_{t}^{(m)} ~ N (0, Σ)$ .

In order to obtain the state space expression of MF-VAR, the following two state vectors are defined, i.e: (4) $S_{t_{m}} = (\begin{matrix} Z_{t_{m}} \\ ⋮ \\ Z_{t_{m} - 4} \end{matrix}) and Z_{t_{m}} = (\begin{matrix} y_{t_{m}}^{*} - μ_{y}^{*} \\ x_{t_{m}} - μ_{x} \end{matrix})$

From the above equation, the following state-space representation of the MF-VAR model can be obtained, i.e: (5) $s_{t_{m} + 1} = A s_{t_{m}} + B v_{t_{m}}$ (6) $(\begin{matrix} y_{t_{m}} - μ_{y} \\ x_{t_{m}} - μ_{x} \end{matrix}) = C s_{t_{m}} (\begin{matrix} y_{t_{m}} - μ_{y} \\ x_{t_{m}} - μ_{x} \end{matrix}) = C s_{t_{m}}$

where $v_{t_{m}} ~ N (0, I_{2})$ , and $μ_{y} = 3 {\dot{μ}}_{y}$ always holds, and experience shows that estimating the mean parameters μ_y and μ_x in the framework of the state space is very difficult, so we estimate the centered parameters, at which point the system matrix becomes: (7) $\begin{array}{l} A = [\begin{matrix} A_{1} \\ A_{2} \end{matrix}] \\ A_{1} = [\begin{matrix} Φ_{1} & \dots & Φ_{p} & 0_{2 \times 2 (5 - p)} \end{matrix}] \\ A_{2} = [I_{8} 0_{3 \times 2}] \\ B = [\begin{matrix} Σ^{1 / 2} \\ 0_{3 \times 2} \end{matrix}] \\ C = [H_{0} \dots H_{4}] \end{array}$

where C is the matrix containing the $H (L_{m}) = \sum_{i} - 0^{4} H_{i} L_{m}^{i}$ -lag polynomials, i.e: (8) $H (L_{m}) = [\begin{matrix} 1 / 3 & 0 \\ 0 & 1 \end{matrix}] + [\begin{matrix} 2 / 3 & 0 \\ 0 & 0 \end{matrix}] L_{m} + [\begin{matrix} 1 & 0 \\ 0 & 0 \end{matrix}] L_{m}^{2} + [\begin{matrix} 2 / 3 & 0 \\ 0 & 1 \end{matrix}] L_{m}^{2} + [\begin{matrix} 1 / 3 & 0 \\ 0 & 0 \end{matrix}] L_{m}^{4}$

According to the constraints of the summation equations, we consider only the matrices A and B for p ≤ 4. Similarly, it is possible to obtain the expression at p > 4 by modifying the state and system matrices.

(2)

Estimation and prediction of default values

After obtaining the estimation results from the above model, the Kalman smoothing operator can be used to forecast the GDP growth rate, and the use of Kalman smoothing operator can take the latest published data of the monthly indicators into account in the forecast, so that it is possible to realize the use of monthly indicators and Kalman smoothing operator to forecast the GDP growth rate in the case that the data of the season is not available.

2)

MF-VAR model combined with factor modeling (1)

Dynamic factor model [39].

Consider the following n smooth high-frequency vectors $x_{t}^{(m)} = {(x_{1, t}^{(m)}, ..., x_{n, t}^{(m)})}^{'}$ that have been standardized with zero mean and variance of 1. The dynamic factor model can then be expressed as: (9) $x_{t}^{(m)} = Λ f_{t}^{(m)} + ξ_{t}^{(m)}, ξ_{t}^{(m)} ~ N (0, Σ_{ξ})$ (10) $ζ_{t}^{(m)} = B η_{t}, η_{t} ~ N (0, I_{q})$ (11) $f_{t}^{(m)} = \sum_{i = 1}^{p} A f_{t - i}^{(m)} + ζ_{t}^{(m)}$

where equation (9) is the factor equation and equation (10) defines the law of motion of the latent factor, which can be seen to be driven by a q-dimensional sequence of standard white noise according to the definitions of (10) and (11).

For the purpose of predictive analysis, defining $y_{t}^{(m)}$ as the high-frequency latent variable of the observable low-frequency variable y_t and assuming that the stochastic high-frequency factor $f_{t}^{(m)}$ is smooth, the relationship between the high-frequency latent variable and the high-frequency cofactor can be formulated as follows: (12) $y_{t}^{(m)} = β^{'} f_{t}^{(m)}$

And the relationship between the observable low-frequency variable $y_{t}^{(m)}$ and the high-frequency latent variable y_t can be expressed as follows: (13) $y_{t} = \frac{1}{m} \sum_{i = 1}^{m} y_{t}^{(m)}$

(2)

The state-space representation of the mixed-frequency data model combining factor variables expresses equations (14) and (15) as a state-space model as follows: (14) $[\begin{matrix} x_{t}^{(m)} \\ y_{t} \end{matrix}] = [\begin{matrix} Λ & 0 & 1 \\ 0 & 0 & 1 \end{matrix}] [\begin{matrix} f_{t}^{(m)} \\ y_{t}^{(m)} \\ y_{t} \end{matrix}] + [\begin{matrix} ξ_{t}^{(m)} \\ ε_{t} \end{matrix}]$ (15) $[\begin{matrix} I_{r} & 0 & 0 \\ - β^{'} & 1 & 0 \\ 0 & - \frac{1}{m} & 1 \end{matrix}] [\begin{matrix} f_{t + 1}^{(m)} \\ y_{t + 1}^{(m)} \\ y_{t + 1} \end{matrix}] = [\begin{matrix} A_{1} & 0 & 0 \\ 0 & 0 & 0 \\ 0 & 0 & Ξ_{t + 1}^{(m)} \end{matrix}] [\begin{matrix} f_{t}^{(m)} \\ y_{t}^{(m)} \\ y_{t} \end{matrix}] + [\begin{matrix} ζ_{t + 1}^{(m)} \\ 0 \\ 0 \end{matrix}]$

3)

BMF-VAR model

The BMF-VAR model is a study of vector autoregressive models containing different frequencies in a Bayesian framework. Unlike the recursive expressions of Kalman filtering, the BMF-VAR model articulates solutions to the likelihood and posterior problems faced under Gibbs sampling and the resulting classical inference based on the likelihood function. In addition, the Markov characteristic makes block Gibbs sampling possible, which makes the sampling of balanced low-frequency data somewhat faster to compute [40].

Assuming that there exists a k-dimensional latent variable ${Y_{t}^{*}}_{t - 1}^{T}$ obeying a simplified VAR(p) model, the BMF-VAR model can be expressed as: (16) $Y_{t}^{*} = c + \sum_{i = 1}^{p} ϕ_{i} Y_{t - i}^{*} + ε_{t}$

where ε_t ~ N(0, Ω), the time units shown in equation (16) t are based on the highest frequency data in the VAR system. The column vector $Y_{t}^{*}$ is unobservable, where some variables may be able to be observed at low frequencies and may change sampling frequency at any time.

It is therefore necessary to define a convention for observable data. Suppose ${Y_{t}^{*}}_{t - 1}^{T}$ is the first sequence of variables in the VAR system of ${Y_{t}^{*}}_{t - 1}^{T}$ . Suppose that the high-frequency latent variables $Y_{a}^{*}, Y_{a + 1}^{*}, \dots, Y_{b - 1}^{*}, Y_{b}^{*}$ are summed into a single observable variable ${\bar{Y}}_{a, b} = \sum_{j - 0}^{b - a} Y_{a + j}^{*}$ in the interval [a, b], i.e., 1 ≤ a ≤ b ≤ T, a, b ∈ N. In this paper, we will refer to this interval as the “summing period”. We can then construct a data series ${Y_{t}}_{t - 1}^{T}$ , where Y_a = Y_a+1 = ⋯ = Y_b−1 = N.A., and $Y_{b} = {\bar{Y}}_{a, b}$ . As a special case, a = b implies that the high-frequency data are directly observable. The data series ${Y_{t}}_{t - 1}^{T}$ contains both observable data and a summing structure, since the computation of a series of NA. Elements reflects a summing cycle. There are times when high-frequency data are averaged rather than summed to obtain low-frequency data, e.g., ${\bar{Y}}_{a, b} = \frac{1}{b - a + 1} \sum_{j - 0}^{b - a} Y_{a + j}^{*}$ . In that case, it is simply notated as $Y_{b} = (b - a + 1) {\bar{Y}}_{a, b}$ .

Repeating the above steps for each of the k variables in the VAR system, the union of the above variables results in a k × T data matrix Y.

The main tasks of the BMF-VAR model are to estimate parameter Θ ≡ {c, ϕ₁, …ϕ_p, Ω} in the model and to recover potential Y^* from the observable dataset Y. Gibbs sampling in a Bayesian framework is capable of performing these two tasks.

2.4

Models and Agents

A model is a simplified representation and embodiment of something objective, an abstract description of the objective laws of nature. The mathematical model in the model is to use mathematical tools to understand the objective world. The role of the mathematical model is to get the unknown results based on the known data. Considering the whole process of the mathematical model from the establishment to the end of the operation, it can be found that the mathematical model has the basic elements such as the applicable environment, the model function, the constraints, the model parameters, the input data, the solution method and the output results. Thus the mathematical model can be defined in the form of a collection of octets:

Model=<En, Fu, Re, Pa, In, Me, Ou, Ev>

Where.

En - Applicable environment, refers to the conditions of applicability of the model, each identified model has different conditions of applicability.

Ev - advantages and disadvantages of the model, a qualitative description of the model.

Fu - model function, the function that constitutes the model, is a mathematical expression.

Re - constraints, constraints on the model function, can be a mathematical expression.

Pa - Parameters required by the model, various parameters required in the model function or constraints.

In - the data required as input for the model.

Me - the method of solving the model, the solution algorithm suitable for the model.

Ou - the output of the model, the result of the model obtained after solving.

Agent weak concept understands Agent from a broad perspective that it should be autonomous, social, reactive and behavioral; while strong concept adds certain human characteristics such as knowledge, beliefs, intentions, etc. to the above definition. By analyzing these characteristics of an Agent, they can be grouped into three basic characteristics: knowledge, goals and capabilities. An Agent with the basic characteristics can be represented as follows using a ternary:

Agent=<Kn, Go, Ca>

Where.

Kn - Agent knowledge.

Go - Agent goal.

Ca - Agent capability.

2.5

Agent-based modeling system

Past research has generally not carried out an in-depth study of the specific implementation process of the Agent of the model, the use of Agent technology to build the model system, representation of the model is subject to further research, this section gives such a model. 1)

Model to Agent mapping

According to the mapping from model to model Agent to construct model Agent, this method stores the applicable conditions, advantages and disadvantages of each model, as well as the exogenous variables it needs as the knowledge of the model Agent in the knowledge base of the model Agent, so that the model system changes from the traditional static way to a model system that can respond to and adapt to the environment, thus making the problem solving way change from the traditional model system in which the system calls the model system to the model system to the model system. In the traditional modeling system, the system calls the model passive solution to the model Agent to use its own knowledge to automatically match the active solution of the problem, in order to make the decision support system to solve the problem with flexibility.

2)

Composition of the modeling system

After the prediction models needed for system construction are represented by model Agents, the model system becomes a multi-agent system, all model Agents are managed by the central control Agent, the model Agent and the central control Agent are subordinate and superior, and the model Agents are equal to each other in the same level.

This model system is an organic part of MEIFDSS, which includes three types of Agents in the model system part, the center control Agent, the model construction Agent and the model Agent. The management of the model system parts based on multi-model Agent is simpler than that of the traditional model library system, and does not need a complex model management system. The center control Agent is responsible for the management of each model Agent, including storing the basic information of each model Agent, registering the new model Agent, and the model construction is done by the model construction Agent.

2.6

Cooperation between Model Agents

There are more or less cooperation ways between multiple Agents in multi-agent systems, and the existing cooperation ways are realized through the communication between Agents.

In solving complex problems, more or less cooperation is needed between multi-model Agents, and such complex problems are very common in macroeconomic forecasting, so the way of cooperation between model Agents in the model system directly affects the effective operation of the system, and it is necessary to study the cooperation between model Agents.

In multi-agent systems, we can categorize the relationship between Agents into cooperation and competition in general. In the existing research collaboration, consultation, negotiation and other relationships between Agents can be attributed to the cooperative relationship, this cooperative relationship can occur for one and the same task, or for their own tasks, there are many systems based on such Agents, in this relationship, the Agents can work together to survive and profit. Competitive relationship is a relationship in which Agents exclude each other for their own purposes, and the strategy of survival of the fittest is practiced among Agents in a competitive environment. In MEIFDSS, the tasks of all model Agents are run to accomplish the forecasting of macroeconomics and at the same time to provide decision support to decision makers, and the relationship between them is cooperative.

2.7

Implementation of the MEIFDSS modeling system

In MEIFDSS, the multi-objective dynamic input-output optimization model and artificial neural network forecasting model are the core, and there are also system dynamics and other models. Among them, the system dynamics model predicts the population in macroeconomy, and the multi-objective dynamic input-output optimization model and artificial neural network prediction model are used to run together on the main indicators of macroeconomy and cross-referenced to each other, and finally the results are integrated comprehensively.

The construction of the model Agent is realized by the object-oriented method in VC++, and the specific model Agent is generated by the inheritance mechanism during the running process according to the need. Object-oriented inheritance can be divided into four forms from the content: substitution inheritance, where the subclass is copied completely from the parent class; containment inheritance, where the subclass contains all the characteristics of the parent class; restricted inheritance, where the subclass has some of the characteristics of the parent class; and specialized inheritance, where the subclass is a special part of the parent class; the subclass is a special part of the parent class. Inheritance, the child class is a special individual of the parent class.

In the specific realization process, the above input-output model Agent class is constructed first, and its goal is the objective function of the above input-output model and the result obtained when reaching the objective function, and the model Agent class has multiple goals. The knowledge of the model Agent class includes the above constraints, the known importance of each objective, the various parameters needed, the raw data required, and the test methods for the solution results.

3

Macroeconomic Intelligent Forecasting Decision-Making Simulation Experiment

In order to verify the comprehensive effectiveness of the multi-agent-based macroeconomic intelligent forecasting decision support system designed in this paper, simulation experimental tests are needed, and the comparison systems chosen are the C/S structure-based macroeconomic intelligent forecasting decision support system (referred to as “C/S structure-based system”), the CDSS-based Macroeconomic Intelligent Forecasting Decision Support System (“CDSS-based system”). The specific experimental environment is a MacBook Pro laptop with Mac OS 10.14.1, I72.5GHz processor and 6GB RAM.

3.1

Reliability analysis

The next experimental tests firstly set the reliability as the test index, and the specific experimental comparison results are shown in Fig. 1. As can be seen from the figure, the reliability of the system designed in this paper is significantly higher than that of the system based on C/S structure and the system based on CDSS. When the number of test samples is 100, 200, 300, 400 and 500, the reliability of the system in this paper reaches 98.27%, 99.53%, 99.4%, 99.27%, 99.31%, respectively, and always maintains higher than 98%, whereas the reliability of the system based on the C/S structure and the system based on the CDSS both decreases significantly with the increase of the number of samples.

3.2

Response time analysis

The following experimental tests will be macroeconomic intelligent forecasting decision support response time as a comparison object, the specific experimental comparison results are shown in Table 1. Analyzing the experimental data in the table, it can be seen that the macroeconomic intelligent forecasting decision support response time of each system will increase with the increase of time. However, compared with the other two systems, the system in this paper predicts the market demand in time, provides a solid theoretical basis for the enterprise marketing decision support, effectively reduces the response time of the system, and is the lowest among the three systems. When the number of experiments is 100 times, the response time of this paper’s system is only 28ms, while the response time of the system based on C/S structure and the system based on CDSS reaches 46ms and 53ms.

Table 1.

Response time

Number of experiments	Response time of macroeconomic intelligent forecasting decision support(ms)
Number of experiments	System of this article	System based on C / S structure	System based on CDSS
10	13	15	17
20	14	18	20
30	16	21	23
40	18	25	27
50	21	29	31
60	23	32	35
70	25	36	38
80	26	39	42
90	27	43	48
100	28	46	53

3.3

Decision support error rate analysis

The internal structure of each system is completely different, which leads to the fact that when it carries out the macroeconomic intelligent forecasting decision support, there may be the situation of decision support error. The following experiments set the macroeconomic intelligent prediction decision support error rate as the evaluation index, and the experimental comparison results are specifically shown in Figure 2. As can be seen from the figure, due to the system in this paper timely forecast for the actual macroeconomic situation, and combined with the prediction results of real-time adjustment of the decision support program, effectively reduce the decision support error rate, the decision support error rate is always less than 1.2%. In contrast, the decision support error rate of the system based on C/S structure and the system based on CDSS reaches 1.69% and 2.01% respectively, which is a high decision support error rate.

4

Macroeconomic Intelligent Forecasting Decision-making System (MIFDDS) application practice

This chapter will apply the Macroeconomic Intelligent Forecasting Decision Support System constructed in this paper to perform real-time forecasting and short-term forecasting of the macroeconomy. The application practice will use the complete macroeconomic sample of China from January 1996 to December 2022 to make real-time forecasts and short-term forecasts of the macroeconomy before and during the New Crown epidemic in China. Using the MF-VAR model as the base model, the MF-VAR model combined with the factor model and the BMF-VAR model will be applied to the real-time forecasting and short-term forecasting before and after the New Crown epidemic, respectively.

4.1

Analysis of real-time forecast results

In this paper, we use the sample interval of January 1996-December 2019 before the Xin Guan epidemic, of which the forecast period is January 2015-December 2019, a period in which China’s macroeconomic environment was not affected by major events. This paper uses the MF-VAR model as the base model for the empirical evidence in this chapter, and uses this as the benchmark model to do a comparison of China’s macroeconomic forecasting accuracy with the MF-VAR model combined with the factor model and the BMF-VAR model. In this paper, the root mean square forecast error (RMSFE) is used to measure the forecasting performance between the two models. When the RMSFE is less than 1, it means that the MSFE of the MF-VAR model combined with the factor model and the BMF-VAR model is smaller than that of the baseline MF-VAR model, i.e., the forecasting effect of the former two models is better than that of the latter. And the RMSFE is greater than 1, i.e., the prediction effect of MF-VAR model and BMF-VAR model combined with factor model is worse than the benchmark MF-VAR model. The prediction results of different macroeconomic indicator points are specifically shown in Table 2. 1)

For the point prediction results of the three variables of fixed asset investment completion, freight transportation, currency and quasi-currency, the MF-VAR model combined with the factor model, and the BMF-VAR model can lead to a small increase in prediction accuracy.

2)

For imports, exports and industrial value added, the forecasts seem to be more difficult to improve by adding other factors, and the introduction of the dynamic factor model, Bayesian method of estimation did not lead to a significant improvement in the forecasting performance of the MF-VAR model.

3)

For power generation and the 7-day interbank lending weighted average interest rate, the forecasting performance of the BMF-VAR model is significantly better than that of the MF-VAR model, and the BMF-VAR model does not show any obvious disadvantage compared with the benchmark MF-VAR model, although from the point of view of the real-time forecasting results, neither of the two paths shows a clear advantage, but in the short-term forecasting of the indicator shows obvious advantages.

4)

For total retail sales of consumer goods, GDP and value added of the tertiary industry, the forecasting performance of the MF-VAR model combined with the factor model is significantly better than that of the MF-VAR model, and the addition of the dynamic factor model to the MF-VAR model shows an absolute advantage in the short-term forecasting in all periods. Meanwhile, the BMF-VAR model does not show a significant advantage in forecasting, but it does not lead to a loss in forecasting accuracy.

Table 2.

Real-time forecast results

Model	Prediction range
Model	0	2	4	6	8
Completed investment in fixed assets
MF-VAR model combined with factor model	0.94	0.98	0.9	0.95	0.95
BMF-VAR model	0.77	0.82	0.87	0.84	0.8
Exports
MF-VAR model combined with factor model	1	1	1	1	1
BMF-VAR model	1.03	1.02	1.02	1.03	1
Power generation
MF-VAR model combined with factor model	1.02	1	1	1	1
BMF-VAR model	1.12	0.95	1	0.98	1
Imports
MF-VAR model combined with factor model	1	1	1	1	0.98
BMF-VAR model	1.05	1.06	1.04	1.04	1.02
7-day interbank weighted average interest rate
MF-VAR model combined with factor model	1	1.01	1.05	1.03	1.05
BMF-VAR model	1.05	0.98	0.97	1	0.94
Currency and quasi-currency
MF-VAR model combined with factor model	0.9	0.9	0.82	0.84	0.77
BMF-VAR model	0.97	0.89	0.8	0.79	0.8
Freight volume
MF-VAR model combined with factor model	0.97	1	1	1	1
BMF-VAR model	0.98	1	1	1	0.97
Total retail sales of social consumer goods
MF-VAR model combined with factor model	0.98	0.98	0.97	0.96	0.96
BMF-VAR model	0.84	0.94	1.05	1	0.95
Gross domestic product
MF-VAR model combined with factor model	0.98	1	0.97	0.92	0.92
BMF-VAR model	1.32	1.1	1.25	1.08	0.96
Industrial added value
MF-VAR model combined with factor model	1.02	1	1	1.02	1
BMF-VAR model	1.38	1.14	1.17	1.05	1
The added value of tertiary industry
MF-VAR model combined with factor model	1.02	0.98	0.93	0.85	0.85
BMF-VAR model	0.75	1.16	1	0.93	0.85

Overall, the dynamic factor model in the MF-VAR model will improve the model’s prediction accuracy. Therefore, it can be concluded that when China’s economic environment is not affected by major events, the introduction of the dynamic factor model into the MF-VAR model can improve the accuracy of forecasting to a certain extent.

4.2

Comparative analysis of forecast results

In this section, the sample interval used is extended to January 1996-December 2022 after the New Crown outbreak, where the forecasting period is January 2015-December 2022, and in order to make the pre-New Crown and post-New Crown forecasting results comparable, the mean of the densities is still used in this section as the Root Mean Squared Forecasting Error (RMSFE) of the point forecasts to assess the true realized values, and the results are shown in Table 3. 1)

First, the forecasting performance of the BMF-VAR model is significantly better than that of the MF-VAR model for the completed fixed asset investment, power generation, freight transportation, and total retail sales of consumer goods. However, unlike the prediction results before the Xin Guan epidemic, the advantage of the prediction performance of the MF-VAR model combined with the factor model has declined, and the prediction accuracy is not as significant as that before the epidemic, but it is still better than the MF-VAR model on the whole.

2)

For currency and quasi-currency and GDP, the forecasting performance of the MF-VAR model combined with the factor model and the BMF-VAR model are both significantly better than that of the MF-VAR model, especially for currency and quasi-currency, the introduction of the dynamic factor model, Bayesian method of estimation greatly improves the forecasting accuracy of the MF-VAR model for all periods of the short-term forecast.

3)

For the three variables of exports, imports and industrial value added, the introduction of the dynamic factor model, Bayesian approach estimation instead leads to worse accuracy of real-time forecasting for the three variables, but the disadvantage is not significant. For these three variables, the advantage of forecasting accuracy was also not revealed when extreme data were not added before the epidemic, but relatively speaking, this disadvantage became less significant when the economy was hit by a major shock.

Table 3.

Predictive result

Model	Prediction range
Model	0	2	4	6	8
Completed investment in fixed assets
MF-VAR model combined with factor model	1.02	1	1	0.98	1
BMF-VAR model	0.91	0.99	0.98	1.02	1
Exports
MF-VAR model combined with factor model	0.98	1	1	1	1
BMF-VAR model	1.05	0.98	1.03	1	1
Power generation
MF-VAR model combined with factor model	0.98	0.98	1.02	1.02	1
BMF-VAR model	0.87	0.85	0.98	0.97	1
Imports
MF-VAR model combined with factor model	1	0.99	1	1	1
BMF-VAR model	1.05	1.08	1.02	1.05	0.98
7-day interbank weighted average interest rate
MF-VAR model combined with factor model	1	0.96	1.01	1.02	1.04
BMF-VAR model	0.95	1.02	1.12	1.12	1.16
Currency and quasi-currency
MF-VAR model combined with factor model	0.92	0.95	0.86	0.85	0.86
BMF-VAR model	1	1	0.95	0.84	0.85
Freight volume
MF-VAR model combined with factor model	0.98	0.99	1	1	1
BMF-VAR model	0.65	0.63	0.88	0.98	1.03
Total retail sales of social consumer goods
MF-VAR model combined with factor model	0.98	1.02	0.98	1	1.03
BMF-VAR model	0.91	0.92	0.97	0.98	1.03
Gross domestic product
MF-VAR model combined with factor model	1.24	0.95	0.94	1	0.97
BMF-VAR model	1.27	1.17	0.94	1.02	0.97
Industrial added value
MF-VAR model combined with factor model	0.95	1.01	1.01	1.02	1.02
BMF-VAR model	1.15	1.26	1.04	1	0.98
The added value of tertiary industry
MF-VAR model combined with factor model	0.77	1.06	1.15	0.96	0.95
BMF-VAR model	0.88	1.14	1.03	0.96	1

5

Conclusion

This paper builds a macroeconomic intelligent forecasting decision support system based on Agent system, which provides reliable methods and means for real-time forecasting and short-term prediction of macroeconomy.

In order to test the comprehensive effectiveness of the system in this paper, we carry out simulation experiments of macroeconomic intelligent forecasting decision-making, and select the macroeconomic intelligent forecasting decision support system based on C/S structure and CDSS as the comparative objects in the experiments. Comparing the reliability of the system based on C/S structure and the system based on CDSS, which decreases significantly with the increase of the number of samples, the reliability of the system in this paper is always maintained at a level higher than 98%. The response time is also always faster than that of the C/S-based system and the CDSS-based system, and when the number of experiments reaches 100 times, the response practice of this paper’s system is only 28 ms. At the same time, this paper’s system also has a lower decision support error rate than that of the C/S-based system and the CDSS-based system, and the decision support error rate is always lower than 1.2%. Overall, the system in this paper shows excellent performance.

The complete macroeconomic sample of China from January 1996-December 2022 is used as the experimental data to carry out the practice of macroeconomic real-time forecasting and short-term forecasting before and after China’s New Crown epidemic. During the pre-epidemic period of January 2015-December 2019, China’s macroeconomic environment was not subjected to any major shocks. The forecasting accuracy of the MF-VAR model combined with the factor model improves and outperforms that of the benchmark MF-VAR model for variables such as fixed-asset investment completion, freight transportation, currency, quasi-currency, total retail sales of consumer goods, GDP, and value added of the tertiary industry. The BMF-VAR model significantly outperforms the benchmark MF-VAR model in terms of forecasting accuracy for variables such as power generation, and the weighted average interest rate of the interbank 7-day interbank lending. The BMF-VAR model is significantly better than the benchmark MF-VAR model in terms of forecasting accuracy in the variables of power generation, interbank 7-day interbank lending weighted average interest rate, and fixed asset investment completion, freight transportation, and monetary and quasi-money variables. In the face of the forecasts of import volume, export volume and industrial value added, neither the BMF-VAR model nor the MF-VAR model combined with the factor model showed a significant advantage. Compared with the BMF-VAR model, the MF-VAR model combined with the factor model has a more obvious advantage in forecasting accuracy. In the macroeconomic real-time forecasting and short-term prediction after the Xin Guan epidemic outbreak, for the two variables of monetary and quasi-money and GDP, the forecasting performance of both the BMF-VAR model and the MF-VAR model combined with the factor model is significantly better than that of the MF-VAR model. In the variables of fixed asset investment completion, power generation, freight transportation, and total retail sales of consumer goods, the predictive performance of the BMF-VAR model is significantly better than that of the MF-VAR model, while the predictive advantage of the MF-VAR model combined with the factor model decreases, but is still better than that of the MF-VAR model as a whole. As for the three variables of export value, import value and industrial value added, both the BMF-VAR model and the MF-VAR model combined with the factor model showed worse real-time forecasting accuracy. However, in the macroeconomic context of a major economic shock during the New Crown epidemic, the disadvantages are less pronounced.

Lingua:: Inglese

Frequenza di pubblicazione:: 1 volte all'anno
Argomenti della rivista:: Scienze biologiche, Scienze della vita, altro, Matematica, Matematica applicata, Matematica generale, Fisica, Fisica, altro

Feed RSS della rivista

Design and Application of Macro Economic Intelligent Prediction Decision Support System Based on Agent System

Lijun Ma

Jiayi Li

Pubblicato online: 05 giu 2025

Ricevuto: 16 gen 2025

Accettato: 11 mag 2025

DOI: https://doi.org/10.2478/amns-2025-1040

Parole chiaveAgent system, MF-VAR model, Factor model, Kalman smoothing operator, Macroeconomy

© 2025 Lijun Ma and Jiayi Li, published by Sciendo.

This work is licensed under the Creative Commons Attribution 4.0 International License.

Parole chiave
Agent system, MF-VAR model, Factor model, Kalman smoothing operator, Macroeconomy