Investment process intelligent method and system based on multi-agent cooperation

Abstract

The present application belongs to the technical field of decentralized finance, and particularly relates to an investment process intelligentization method and system based on multi-agent cooperation, which comprises the following steps: acquiring time series data and assigning an observation perspective; constructing a hybrid hidden market state vector containing discrete state encoding and continuous intensity factors and generating a local prior probability distribution; generating a communication topology based on marginal probability, updating beliefs by using a double-channel mechanism of transmitting local priors and Kullback-Leibler divergence, obtaining a global posterior probability distribution and generating a return distribution expression; calculating structure and parameter uncertainty indicators, and respectively mapping them into a concentration penalty term and a risk tolerance upper limit to solve optimal investment weights. The present application can effectively limit extreme risk exposure, enhance risk resistance, and achieve a good balance between returns and risks.

Description

Technical Field

[0001] This invention relates to the field of decentralized finance technology. More specifically, this invention relates to an intelligent method and system for investment processes based on multi-agent collaboration. Background Technology

[0002] In the field of intelligent investment, single-agent or single-perspective analysis models often struggle to fully identify hidden market states and their evolutionary patterns. Introducing multi-agent collaborative mechanisms into the investment decision-making process allows the system to process market information in parallel from multiple dimensions by assigning differentiated observation perspectives to different agents. However, in actual collaborative processes, the local information acquired by each agent often exhibits biases, and existing agent communication networks are mostly static structures or rely on manual experience for configuration, making it difficult to adaptively adjust to the actual market conditions. Financial markets themselves possess high uncertainty, and their hidden market states exhibit both discrete periodicity and continuous intensity characteristics.

[0003] Existing asset return forecasting methods largely focus on predicting certain points, lacking the ability to uncover and characterize the uncertainties behind the forecast results. This leads to portfolios being easily exposed to higher risks under extreme market conditions. Even when some methods attempt to use probability distribution forecasts, it is difficult to decouple the uncertainties of market states across different dimensions, let alone translate them into actionable risk control measures. During portfolio optimization, if the structural and parametric uncertainties of market states cannot be effectively characterized, the investment system struggles to reasonably constrain risk indicators such as asset concentration and conditional value at risk. Therefore, how to characterize the multiple uncertainties hidden under mixed market states and combine uncertainty indicators with return distribution expressions to drive the portfolio optimization framework, achieving robust asset weight allocation while pursuing investment returns, has become a critical technical problem that modern intelligent collaborative investment technology urgently needs to solve. Summary of the Invention

[0004] To address the technical problem that existing methods are unable to characterize multiple market uncertainties, leading to higher risks in portfolio asset allocation, this invention provides solutions in the following aspects.

[0005] In a first aspect, the present invention provides an intelligent investment process method based on multi-agent collaboration, comprising: S1, acquiring financial time-series data of the assets to be invested and assigning observation perspectives to each agent; S2, constructing a hybrid hidden market state vector composed of discrete state codes and continuous intensity factors; S3, each agent using a sequence transformation model to generate a local prior probability distribution for the hybrid hidden market state vector; S4, generating a communication topology based on the marginal probabilities of the discrete state codes in the global posterior probability distribution of the previous time step; S5, each agent performing dual-channel belief propagation, and performing calculations on the local prior probability distribution and the Kullback-Leibler divergence between its own beliefs and those of its neighbors. The primary and secondary channel messages are transmitted, and the Kullback-Leibler divergence is used to adjust the neighbor influence weights and update beliefs to obtain the global posterior probability distribution at the current time step; S6, the prediction network is driven by the global posterior probability distribution to generate prediction quantile points for the assets to be invested, forming a return distribution expression; S7, the structural uncertainty index of the discrete state code in the global posterior probability distribution and the parameter uncertainty index of the continuous intensity factor are calculated; S8, the return distribution expression is input into the optimization framework, the structural uncertainty index is mapped to the concentration penalty term, the parameter uncertainty index is mapped to the risk tolerance upper limit of the conditional value of risk, and the investment weights are solved by combining the concentration penalty term and the risk tolerance upper limit.

[0006] This invention accurately identifies market characteristics by constructing a hybrid hidden market state vector containing discrete state encoding and continuous strength factors. It achieves reliable posterior probability updates by using dual-channel communication and divergence index to adaptively adjust neighbor influence weights. It realizes non-parametric probabilistic feature measurement by using posterior probability to predict return distribution. The obtained structural uncertainty index and parameter uncertainty index are mapped to concentration penalty term and risk tolerance upper limit, respectively. In the underlying asset planning, it fully represents and combines the fuzziness of market structure and parameter estimation error, effectively limiting extreme risk exposure in high uncertainty environment, enhancing the risk resistance and disaster recovery performance of the portfolio, and achieving a good balance between return and risk.

[0007] Preferably, the step of acquiring financial time-series data of the assets to be invested and assigning observation perspectives to each agent includes: receiving time-series information containing specific investment targets from a financial database through an application programming interface; arranging the extracted data sequences in memory in ascending order of time as independent time-series data tensor record matrices to form a baseline three-dimensional tensor, and using the independently arranged time-series data tensor record matrix as a baseline observation source to segment and call to each agent; and performing slicing and mapping operations on the baseline three-dimensional tensor in terms of spatial and feature dimensions based on the pre-set heterogeneous feature preference configuration of each agent to achieve independent assignment of observation perspectives.

[0008] This invention enables concurrent invocation of heterogeneous intelligent agents on high-dimensional financial data through feature space slicing and mapping operations, avoiding redundant data transmission and repeated memory loading, and effectively improving the scheduling efficiency and data access rate of the underlying heterogeneous perspective allocation mechanism.

[0009] Preferably, the construction of the hybrid hidden market state vector composed of discrete state encoding and continuous intensity factor includes: generating a one-dimensional array in the storage space, presetting the element at index 0 of the array as an integer format variable limited to between 1 and 3, the values ​​of the integer format variables representing three macro-structure patterns respectively, to obtain the discrete state encoding; extracting the absolute difference moving average between the closing price and the opening price within a set historical time window, and the percentage deviation of the closing price relative to its historical moving average as two floating-point variables, storing them sequentially in index 1 and index 2 spaces, and the continuous intensity factor composed of the two floating-point variables.

[0010] This invention achieves seamless integration of macroscopic patterns and short-term fluctuation characteristics in the same tensor through a specific memory-point push and a high-precision one-dimensional array storage design. This not only provides a foundation for the joint deduction of multidimensional uncertainties, but also reconstructs the underlying representation of structure and continuous state parameters, providing unbiased data assurance for complex distributed operations.

[0011] Preferably, each agent employs a gated recurrent neural network model with a double forget gate as the sequence transformation model. The generation of the local prior probability distribution includes: feeding the received data tensor into the gated recurrent neural network model with a double forget gate; performing a normalized exponential transformation on the first mapping branch of the gated hidden output state matrix to output three edge prediction probability values ​​corresponding to different discrete state codes; simultaneously guiding the second mapping branch of the gated hidden output state matrix to a conditional mean square evaluation predictor with a linear fully connected mapping layer to generate the expected value of the predicted target and the distribution variance result; and encapsulating the edge prediction probability value, the expected value of the predicted target, and the distribution variance result as a local prior probability distribution.

[0012] This invention introduces a dual forget gate and a dual-branch decoupled network structure into the sequence transformation model, which accurately separates discrete classification features from continuous parameter evaluation tasks, reduces dimensionality interference in the prediction process of different hidden features, and significantly improves the convergence accuracy of local prior probability generation and the stability of temporal feature mining.

[0013] Preferably, the step of generating the communication topology based on the marginal probability of discrete state encoding in the global posterior probability distribution of the previous time step includes: extracting the marginal probability distribution matrix of discrete state encoding retained by all multi-agents in the global communication domain from the previous operation cycle; selecting any two agents to calculate the absolute difference in their marginal probability distributions and establishing connection distance values ​​based on Chebyshev distance; judging each connection distance value against a set distance limit parameter, and setting the communication permission status flag to 1 if it is lower than or equal to the distance limit parameter, and setting the communication permission status flag to 0 if it exceeds the distance limit parameter; filling in the adjacency matrix of the undirected graph according to all the communication permission status flags to complete the generation of the communication topology.

[0014] This invention assesses the cognitive differences of multiple agents at the marginal probability level through Chebyshev distance and implements dynamic routing connectivity determination by combining distance constraint parameters. This enables the underlying communication link to adaptively disconnect and reconnect in response to real macroscopic state changes, avoiding invalid information propagation and noise interference in a fixed topology.

[0015] Preferably, the step of transmitting the Kullback-Leibler divergence between one's own beliefs and those of neighbors as a secondary channel message includes: obtaining the discrete classification belief probability maintained by oneself and the discrete classification belief probability transmitted by each permitted neighboring agent; adding a preset smoothing term to the discrete classification belief probability maintained by oneself and the discrete classification belief probability transmitted by neighboring agents and re-normalizing them; calculating the Kullback-Leibler divergence based on the normalized self-belief probability and the belief probability of neighboring agents to obtain the real value of the deviation base point result; using the real value of the deviation base point result as an inconsistency feedback index to measure the mismatch distance of local neighboring nodes, and using the inconsistency feedback index to constitute the secondary channel message.

[0016] Preferably, the calculation of the structural uncertainty index and the parameter uncertainty index includes: calculating the product of the marginal probabilities of each class and the logarithm of each class in the global posterior probability distribution, summing all the product results and inverting them to obtain the Shannon entropy expression value as the structural uncertainty index; extracting the conditional variance or conditional standard deviation corresponding to the continuous intensity factor, performing dimensionless processing according to the historical scale parameter of the continuous intensity factor to obtain the dimensionless uncertainty component of each continuous intensity factor; averaging the absolute value or square value of each dimensionless uncertainty component within the estimation time window, and performing a weighted summation operation to obtain the non-negative parameter uncertainty index.

[0017] Preferably, the step of solving for investment weights by combining a concentration penalty term and a risk tolerance cap includes: extracting the allocation weight percentage of the assets to be invested in the portfolio; calculating and summing the squares of the allocation weight percentages under the condition that the sum of the allocation weight percentages is 1 to obtain the Herfindahl-Hirschman index; multiplying the structural uncertainty index by the Herfindahl-Hirschman index to construct a concentration penalty term instruction for overall objective constraints; normalizing the parameter uncertainty index to obtain a parameter uncertainty normalized value; generating a risk tolerance scaling factor based on the reciprocal of the sum of the parameter uncertainty normalized value and a preset positive bias term; multiplying the risk tolerance scaling factor by the initial risk tolerance cap to obtain an adjusted tail risk tolerance cap, so that the optimization framework constrains the value at risk with the adjusted tail risk tolerance cap.

[0018] Preferably, the prediction network driven by the global posterior probability distribution generates a return distribution expression for the investable asset by generating predicted quantile points, which includes: extracting discrete state marginal probabilities, continuous intensity factor conditional mean vectors, and conditional variance vectors from the global posterior probability distribution as context variables and inputting them into the prediction network decoder; connecting a quantile regression layer at the end of the prediction network decoder to predict the key quantile points of the investable asset in parallel at a future time step; using the key quantile points as anchor points for the return distribution expression, and constructing a piecewise monotonic quantile function by combining historical return residual samples of the same state, generating a monotonically arranged discrete sample of the return distribution; and constructing an approximate expression of the future return of the investable asset as a non-parametric empirical cumulative distribution function from the piecewise monotonic quantile function or the set of discrete sample of the return distribution.

[0019] Secondly, the present invention provides an intelligent investment process system based on multi-agent collaboration, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned intelligent investment process method based on multi-agent collaboration is implemented.

[0020] By adopting the above technical solution, the intelligent investment process based on multi-agent collaboration is generated into a computer program and stored in a memory for loading and execution by a processor. This allows for the creation of a terminal device based on the memory and processor, facilitating its use.

[0021] The beneficial effects of this invention are as follows:

[0022] This invention constructs a hybrid hidden market state vector containing discrete state encoding and continuous strength factors to represent the financial market environment. Each agent utilizes a dual-channel belief propagation mechanism to transmit probability distributions and Kullback-Leibler divergences, adjusting neighbor influence weights based on this divergence. A quantile-level return distribution is generated using the global posterior probability distribution, representing the probabilistic characteristics of future asset returns. The calculated structural uncertainty and parameter uncertainty indices are mapped to concentration penalty terms and risk tolerance caps in the optimization framework, respectively, effectively combining the fuzziness of market structure and parameter estimation errors at the decision-making level. This mechanism limits extreme risk exposure under high uncertainty environments, enhances the portfolio's resilience in the market, and achieves a good balance between return and risk. Attached Figure Description

[0023] Figure 1 This is a flowchart of the intelligent investment process method based on multi-agent collaboration in this invention; Figure 2 This is a line graph of the cumulative revenue process in this invention; Figure 3 This is a bar chart comparing the maximum concentration in this invention. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0025] This invention discloses an intelligent investment process method based on multi-agent collaboration, referring to... Figure 1 This includes steps S1-S8: S1. Obtain financial time series data and assign observation perspectives.

[0026] In an optional embodiment, financial time-series data of the assets to be invested in are acquired and corresponding observation perspectives are assigned to each agent.

[0027] Read the historical data of daily opening price, closing price, highest price, lowest price and trading volume of each investable asset in the stock pool, and calculate various financial technical indicators.

[0028] The agent set is divided into trend-following agents, momentum reversal agents, and volatility analysis agents, and each agent is assigned a moving average, relative strength index (RSI), and Bollinger Band width as its exclusive observation perspective feature. All features are normalized to obtain standardized time series tensors. It should be noted that trend-following agents, momentum reversal agents, and volatility analysis agents are the three fixed agent types used in this embodiment. Subsequent examples will focus on these three agent types, and naming conventions such as momentum tracking, price-volume game theory, and micro-volatility, which have unclear correspondences with the three agent types, will not be used.

[0029] Furthermore, the system acquires financial time-series data of the assets to be invested in and assigns corresponding observation perspectives to each agent. Specifically, it receives time-series information from a financial database via an application programming interface (API), including daily opening prices, closing prices, highest prices, lowest prices, and total daily trading volume for specific investment targets. The extracted data sequences are arranged in memory as independent time-series tensor record matrices in ascending order of natural calendar time, and these independently arranged time-series tensor record matrices are used as benchmark observation sources and distributed to each agent.

[0030] The system uses a RESTful application programming interface to periodically retrieve daily-frequency, five-dimensional basic trading data of N candidate investment portfolios within a specified time window from a financial database in JSON or CSV format. This basic trading data includes opening price variables. Closing price variable Highest price variable Lowest price variable and total daily trading volume .

[0031] The financial database includes either the Wind Financial Database or the Bloomberg Financial Database. The specified time window is preferably set to the past 252 trading days, and the candidate portfolio is set to N=50 target constituent stocks.

[0032] Upon receiving the data stream, the underlying data processing module performs preprocessing such as forward padding on missing values ​​within the data frame, and allocates the cleaned data for each target to a pre-configured system memory heap space. Based on the system memory heap space, the engine constructs a three-dimensional time-series data tensor record matrix with physical dimensions of N×T×5 according to the positive time increment attribute and caches it, setting the variable T=252 as the time step.

[0033] Based on the pre-configured heterogeneous feature preferences of each agent, the baseline three-dimensional tensor is sliced ​​and mapped in both spatial and feature dimensions to achieve independent assignment of observation perspectives. For example, the system deploys three heterogeneous agents: Agent A is a trend-following agent, which uses the closing price sequence and calculates the short-term moving average, the long-term moving average, and the difference between the short-term and long-term moving averages as trend-following perspective features; Agent B is a momentum reversal agent, which uses the closing price, highest price, and lowest price sequence and calculates the relative strength index (RSI) and overbought / oversold deviation as momentum reversal perspective features; Agent C is a volatility analysis agent, which uses the opening price, closing price, highest price, and lowest price sequence and calculates the Bollinger Band width and logarithmic return volatility as volatility analysis perspective features.

[0034] This segmentation mechanism grants specific data access permissions to corresponding multi-agent computing nodes through underlying memory address pointer mapping, enabling each multi-agent computing node to perform market hidden state inference tasks based on the assigned high-dimensional time-series data subset.

[0035] S2. Construct a hybrid hidden market state vector.

[0036] In an optional embodiment, a hybrid hidden market state vector is constructed, consisting of discrete state encoding and continuous intensity factors.

[0037] The hybrid hidden market state vector comprises two levels: state value representation and distribution parameter representation. The state value representation stores the discrete state encoding and the current value of the continuous intensity factor in the storage space. The distribution parameter representation stores the marginal probabilities of the discrete states and the conditional mean and conditional variance of the continuous intensity factor in subsequent local prior and global posterior probability distributions.

[0038] Discrete state encoding uses category numbers for storing state values ​​in the hybrid hidden market state vector. These category numbers are limited to 1 to 3, representing three macroeconomic structural patterns: bull market, bear market, and sideways market. A bull market refers to a market where prices rise overall; a bear market refers to a market where prices fall overall; and a sideways market refers to a market where prices fluctuate repeatedly within a certain range without a clear unilateral trend. During model training or probability inference, differentiable category representations can be obtained by sampling the three marginal probability vectors using Gumbel-Softmax, and the sampling results are mapped to category numbers. Thus, the category numbers are used for state indexing, and the three marginal probability vectors are used for subsequent marginal probability calculations, communication topology generation, and structural uncertainty calculations. The category numbers and the three marginal probability vectors do not constitute a data form contradiction.

[0039] The continuous intensity factor is stored using deterministic real-valued features in the state value representation of the hybrid hidden market state vector. Specifically, these features include the moving average of the absolute difference between the closing and opening prices within a historical time window, and the percentage deviation of the closing price relative to its historical moving average. In the local prior probability distribution and the global posterior probability distribution, the continuous intensity factor is further described by conditional Gaussian distribution parameters, which include the conditional mean vector. Conditional variance vector If random sampling is required during model training, a reparameterized approach is used to calculate the continuous intensity factor.

[0040] in, This is the sampling vector of the continuous intensity factor; It is the conditional mean vector; It is the square root of the conditional variance vector; The symbol for the Hadama product; This is standard normally distributed noise.

[0041] Thus, by constructing the sampling vector calculation formula for the continuous intensity factor based on the reparameterization technique in the variational autoencoder, the randomness is transferred to the standard normal distribution noise, enabling the model to achieve gradient backpropagation while maintaining the randomness of the sampling process.

[0042] Furthermore, a hybrid hidden market state vector is constructed. Specifically, a one-dimensional array is generated in the storage space, and the element at index 0 of the array is preset as an integer format variable limited to 1 to 3. The values ​​of the integer format variables represent the three macroeconomic structural patterns of bull market, bear market, and sideways market, respectively, to obtain discrete state codes. The moving average of the absolute difference between the closing price and the opening price within a set historical time window, and the percentage deviation of the closing price relative to its historical moving average are extracted as two floating-point variables and stored in index 1 and index 2 spaces respectively. The two floating-point variables together form a continuous strength factor.

[0043] The feature construction engine instantiates a high-precision floating-point one-dimensional array of physical length 3 for each asset to be analyzed at the current time step in the memory pool. This high-precision floating-point one-dimensional array serves as the core data structure for the hybrid hidden market state vector. The expression for the high-precision floating-point one-dimensional array is:

[0044] in, It is a high-precision floating-point one-dimensional array; The index 0 represents the element representing the discrete state code; A floating-point variable representing short-term potential energy intensity; A floating-point variable representing the strength of overbought or oversold prices.

[0045] The state estimation algorithm, based on the Markov chain state clustering results, selects the index 0-bit element representing the discrete state encoding. Hardcoded as an integer scalar, the value of which is locked within the integer set {1, 2, 3}. Here, value 1 maps to a bull market pattern with continuously diverging asset returns, value 2 maps to a bear market pattern with continuously declining mean, and value 3 maps to a sideways market pattern. This integer scalar serves only as a category index field; the three marginal probability vectors corresponding to the integer scalar are stored separately in a local prior probability distribution or a global posterior probability distribution for subsequent probability calculations. The three marginal probability vectors are constructed based on the discrete multinomial distribution principle in probability theory, and the specific calculation formula is as follows:

[0046] in, This is a three-term marginal probability vector; The index 0 represents the element representing the discrete state code; This represents the marginal probability of being in a bull market. This represents the marginal probability of being in a bear market. This represents the marginal probability of being in a volatile market.

[0047] The calculation module uses an extraction function to capture basic trading data within a pre-defined historical sliding time window and performs dimensionality reduction extraction of the continuous strength factor. The historical sliding time window variable W is preferably set to the past 20 trading days. This parameter value is based on the industry consensus that a month in the financial market typically contains 20 trading days, accurately reflecting short-term market characteristics at the monthly level. The microprocessor calculates the daily closing price variable sequentially according to the time series. With opening price variable The absolute difference is constructed based on the mathematical calculation of the distance between two values. The specific calculation method is as follows:

[0048] in, The absolute difference; The daily closing price is the variable; The opening price is a variable.

[0049] The microprocessor uses an arithmetic sliding window function to calculate the 20-day average of the absolute difference sequence, producing a floating-point variable representing the short-term potential strength. For example, the moving average of the difference in a certain asset is calculated to be 0.15 yuan per share.

[0050] The calculation system retrieves the percentage deviation of the closing price from its 20-day simple moving average and generates a floating-point variable representing the strength of overbought or oversold conditions. floating-point variables The calculation method is as follows:

[0051] in, A floating-point variable representing the strength of overbought or oversold prices; The closing price for the day; This is the 20-day simple moving average value; This is the deviation rate parameter.

[0052] It should be noted that this embodiment constructs a floating-point variable representing the strength of overbought or oversold conditions based on the deviation rate calculation method in the field of financial technical analysis, and converts the ratio into a percentage form to conform to the data display standards in the financial field. For example, the calculated deviation is 2.35%.

[0053] The control unit will push the two generated floating-point feature values ​​into the index 1 of the array and store them. The location and index are stored in 2 bits. The position of the vector was reconstructed through index arrangement to obtain a hybrid vector representation that fully covers both structural and continuous features. Meanwhile, the system uses a distributed parameter buffer... and The corresponding conditional mean and conditional variance fields are retained respectively, so as to facilitate the subsequent formation of the conditional Gaussian distribution parameters of the continuous intensity factor.

[0054] S3. Generate the local prior probability distribution of the hidden state.

[0055] In an optional embodiment, each agent utilizes a sequence transformation model to generate a corresponding local prior probability distribution for the hybrid hidden market state vector at each time step.

[0056] Each agent employs a gated recurrent neural network model with a dual forgetting gate as its sequence transformation model, processing standardized time-series tensors from its own observational perspective. Each agent extracts time-dependent features through long-term forgetting, short-term forgetting, and update gates, and feeds the output gated hidden state matrix into a linear classifier head network and a conditional Gaussian mean square (GMS) estimator, respectively. The linear classifier head network outputs three marginal prediction probabilities for discrete state encoding, while the GMS estimator outputs the conditional mean vector and conditional variance vector for continuous intensity factors. These three marginal prediction probabilities, the conditional mean vector, and the conditional variance vector together constitute the local prior probability distribution for the current time step.

[0057] Furthermore, each agent utilizes a sequence transformation model to generate corresponding local prior probability distributions for the hybrid hidden market state vector at each time step. Specifically, each agent feeds the received data tensor into a gated recurrent neural model with a double forget gate, performs a normalized exponential transformation on the first mapping branch of the gated hidden output state matrix, and outputs three edge prediction probability values ​​corresponding to different discrete state codes; each agent simultaneously guides the second mapping branch of the gated hidden output state matrix to a conditional mean square evaluation predictor with a linear fully connected mapping layer, generating the expected value of the target and the distribution variance; the edge prediction probability values, the expected value of the target, and the distribution variance are encapsulated, and the encapsulated result is used as the local prior probability distribution.

[0058] Each edge agent node pre-loads a gated recurrent neural network model with a double forget gate and a predetermined number of hidden layer nodes within its internal processing core for deep temporal feature mining. In deep learning, the number of hidden layer nodes in a recurrent neural network is typically set to an integer power of 2 to fully utilize the parallel computing advantages of the graphics processing unit (GPU). Therefore, the preferred configuration for the number of hidden layer nodes is 64 to 128 units. The network structure of the gated recurrent neural network model with a double forget gate consists of an input layer, a double forget hidden layer, and an output layer connected in series. The double forget hidden layer contains parallel-configured long-term and short-term forget gates, an update gate, and candidate state computation nodes. The network input of the gated recurrent neural network model is a set of data sheets of size, for example, 252×5, held according to the forward time series.

[0059] At each computation time step t, the agent iteratively feeds the network input into the input layer of the gated recurrent neural model with dual forgetting gates. The model's internal special long-term and short-term forgetting mechanisms filter redundant information in two ways, which is then extracted by the network output layer and used to generate a low-rank tensor representing the current local state as the network model's output, i.e., the gated hidden output state matrix. And hide the output state matrix of the gate. The dimension is set to 1×128. The tensor computation graph separates the output of the network model into two sets of decoupled downstream multilayer perceptron network mapping branches and performs parallel propagation.

[0060] The downstream multilayer perceptron network comprises a linear classification head network as the first mapping branch and a conditional Gaussian mean square evaluation predictor as the second mapping branch. In the parallel processing of the first mapping branch, the hidden matrix... The input to the first branch is fed forward to the linear classification head network. The linear classification head network consists of a single-layer linear fully connected network of size 128×3 and a Softmax normalized exponential activation layer cascaded after the single-layer linear fully connected network.

[0061] Through forward network computation and transformation, the linear classification head network generates the corresponding network output, which is the joint one-dimensional vector edge prediction probability corresponding to the three macroscopic state modes. The calculation method for the joint one-dimensional vector edge prediction probability is as follows:

[0062] in, For the joint one-dimensional vector edge prediction probability; For the predicted probability components corresponding to the first macroscopic state mode; For the predicted probability components corresponding to the second macroscopic state mode; This corresponds to the predicted probability component of the third macroscopic state mode.

[0063] It should be noted that this embodiment constructs a formula for calculating the joint one-dimensional vector edge prediction probability based on the probability mapping characteristics of the Softmax activation function. The characteristics of the Softmax activation function forcefully guarantee that the sum of the prediction probability components is always 1. For example, the output value of the prediction probability is... .

[0064] Meanwhile, in the second mapping branch, the hidden matrix will be... The second branch input is directed to the conditional Gaussian mean squared evaluation predictor. The network structure of the conditional Gaussian mean squared evaluation predictor includes a mean evaluation layer and a variance evaluation layer set in parallel. Both the mean evaluation layer and the variance evaluation layer are composed of a single-layer linear fully connected network, and a Softplus monotonic positive activation function layer is connected in series after the variance evaluation layer.

[0065] The mean evaluation layer calculates and linearly outputs the expected value vector of the predictions for the continuous intensity variable. Expected value vector The two-dimensional estimate corresponding to the moving average and the deviation is, for example, the value of the estimate. The variance evaluation layer uses activation function layers to constrain the nonnegativity boundary, and computes and generates the network output distribution variance vector for variance parameterization representation. The variance output value is, for example, In this invention, the network distribution parameters of the continuous intensity factor are uniformly represented by the conditional mean vector. Conditional variance vector This indicates that if the training program internally uses a log-variance vector... Before writing the local prior probability distribution, the logarithmic variance vector is converted into a conditional variance vector using an exponential function or the Softplus function. If the standard deviation is needed subsequently, the standard deviation vector is calculated using the square root formula; if the covariance matrix is ​​needed, then... Construct a diagonal covariance matrix. This invention constructs a standard deviation vector based on the statistical conversion relationship between variance and standard deviation:

[0066] in, The standard deviation vector; Let be the conditional variance vector.

[0067] The network output extracted from the first mapping branch and the network output generated from the second mapping branch are aggregated and bound. Using a hash dictionary or a heterogeneous tuple structure protocol, the aggregated and bound network outputs are encapsulated into a formatted composite message packet, generating a highly cohesive local prior probability distribution carrier that is transmitted on network nodes. The local prior probability distribution carrier includes at least a discrete-state marginal probability vector. Continuous intensity factor conditional mean vector and continuous intensity factor conditional variance vector .

[0068] S4. Generate undirected communication topology based on marginal probability.

[0069] In an optional embodiment, the communication topology is generated based on the marginal probability of the discrete state encoding in the global posterior probability distribution of the previous time step.

[0070] The global posterior probability distribution comprises two levels: the node-level posterior probability distribution matrix and the global posterior probability distribution. The node-level posterior probability distribution matrix preserves the discrete state marginal probabilities of each agent after convergence at the previous time step. These discrete state marginal probabilities are used to generate the communication topology in the next time step. The global posterior probability distribution drives payoff prediction and combinatorial optimization. Therefore, even if belief propagation has converged, differences below the convergence threshold can still be maintained between nodes, preventing the communication topology from degenerating into a constant fully connected structure.

[0071] Extract the node-level posterior probability distribution matrix corresponding to each agent in the previous time step, and marginalize the continuous intensity factor of the node-level posterior probability distribution of each agent to obtain the discrete state encoding marginal probability vector corresponding to each agent.

[0072] Calculate the Chebyshev distance between the marginal probability vectors of any two agents to construct the distance values ​​for candidate communication edges. Compare each distance value with a set distance constraint parameter. When the distance value is less than or equal to the distance constraint parameter, set the corresponding position in the communication topology matrix to 1; when the distance value exceeds the distance constraint parameter, set the corresponding position in the communication topology matrix to 0, thereby generating a sparse undirected communication topology graph for the current time step.

[0073] Furthermore, a communication topology is generated based on the marginal probabilities of the discrete state codes in the global posterior probability distribution of the previous time step. Specifically, the marginal probability distribution matrix of the discrete state codes retained from the previous operation cycle of all multi-agents in the global communication domain is extracted; the absolute difference between the marginal probability distributions of any two agents is calculated, and connection distance values ​​based on Chebyshev distance are established; each connection distance value is compared with a set distance limit parameter. If it is lower than or equal to the distance limit parameter, the communication permission status flag is set to 1; if it exceeds the distance limit parameter, the communication permission status flag is set to 0. The adjacency matrix of the undirected graph is filled according to all the communication permission status flags to complete the generation of the communication topology.

[0074] In a distributed collaborative system with M heterogeneous collaborative agents and a central scheduling controller, during the pre-preparation phase before entering a new inference time step t, the central scheduling controller retrieves and copies from the system's distributed memory shared pool the node-level discrete state marginal probability distribution matrix with a physical dimension of M×3, which was generated by the converged calculation in the previous time period (time step t-1). The node-level discrete state marginal probability distribution matrix stores the degree of identification of each node with the three state types in the previous round. The node-level discrete state marginal probability distribution matrix differs from the global posterior probability distribution; it is used to preserve the differences between nodes, while the global posterior probability distribution represents the overall system judgment.

[0075] For any two candidate connected agent nodes i and j in the global communication domain, the algorithm module reads the normalized discrete probability distribution vector of agent node i from the matrix. and the normalized discrete probability distribution vector of agent node j Normalized discrete probability distribution vector and the normalized discrete probability distribution vector The expression is as follows: ;

[0076] in, Let be the normalized discrete probability distribution vector of agent node i; Let be the marginal probability of agent node i for the first state type; Let be the marginal probability of agent node i for the second state type; Let be the marginal probability of agent node i for the third state type; Let be the normalized discrete probability distribution vector of agent node j; Let be the marginal probability of agent node j for the first state type; Let be the marginal probability of agent node j for the second state type; Let be the marginal probability of agent node j for the third state type.

[0077] Then, positional absolute difference calculations are performed on each of these three vectors, and the mathematical construction of the Chebyshev distance in the metric space is completed by extracting the largest difference component. The formula for calculating the Chebyshev distance is as follows:

[0078] in, This represents the distance between the lines. The index number for the status type; For agent node i, for the first... Marginal probability of each state type; For agent node j, for the th Marginal probability of each state type; This is the function for finding the maximum value.

[0079] It should be noted that this embodiment constructs the formula for calculating the distance between lines based on the Chebyshev distance definition in metric space mathematics. For example, assume the normalized discrete probability distribution vector of agent node i... And the normalized discrete probability distribution vector of agent node j The three absolute differences obtained from the two methods are 0.1, 0.2, and 0.1 respectively. The maximum value is taken to obtain the distance of the connecting line. It is 0.2.

[0080] After the distance metrics are constructed, the controller scheduler will connect the distance values. Perform a logical Boolean comparison with the communication distance limit parameter pre-configured in the hyperparameter table. It is recommended that the control threshold range for the communication distance limit parameter be set to a scalar floating-point number. For example, selecting a scalar floating-point number The value is set to 0.25. This threshold range is set based on the empirical distribution boundary of the financial sequence state in the probability simplex space. If the threshold is too small, the network will become too sparse and isolated, while if the threshold is too large, too much noise interference will be introduced. A value between 0.15 and 0.3 can better balance the effectiveness and independence of information transmission.

[0081] If the verification finds the connection distance value This indicates that the macroscopic logical cognitive differences between the two intelligent agent nodes have not triggered the isolation mechanism. The controller then injects a value of 1, representing the activation of the point-to-point virtual link, into the corresponding register in the link mapping table. Conversely, when the connection distance is verified... When this occurs, the communication permission status flag in the routing table for this adjacency pairing status is reset to the value 0.

[0082] After determining the loop conditions for all interaction pairs within the system domain, the scheduling center will map this set of 0 and 1 flag variables to the corresponding row and column index units, constructing and refreshing a symmetric M×M Boolean adjacency matrix graph. The diagonal loops are always set to 1, thereby achieving the task of generating the iterative network communication topology. The total number of interaction pairs is... .

[0083] S5. Perform dual-channel propagation to obtain the posterior probability distribution.

[0084] In an optional embodiment, each agent performs a two-channel belief propagation, transmitting the local prior probability distribution as the primary channel message and the Kullback-Leibler divergence between its own belief and its neighbors' beliefs as the secondary channel message. The Kullback-Leibler divergence is used to adjust the neighbor influence weights and update the beliefs until convergence yields the global posterior probability distribution for the current time step.

[0085] A message passing mechanism is initiated on the generated communication topology. Each agent packages its currently computed local prior probability distribution into a main channel message and sends the main channel message to its neighboring nodes in the topology. The main channel message includes a discrete state marginal probability vector, a continuous intensity factor conditional mean vector, and a continuous intensity factor conditional variance vector.

[0086] Simultaneously, it calculates the Kullback-Leibler divergence between its current probability belief and the received probability beliefs of its neighbors, and feeds back the Kullback-Leibler divergence as a secondary channel message to its neighbors. The Kullback-Leibler divergence measures the degree of difference between two probability distributions; a larger value indicates that the two distributions are more inconsistent, and a smaller value indicates that the two distributions are more similar.

[0087] The calculated Kullback-Leibler divergence is negatively correlated and normalized to obtain the aggregation weights for each neighbor node. Specifically, the system first uses the node's own belief as the default baseline belief for fusion, and then converts the Kullback-Leibler divergence of each neighbor node into neighbor influence weights. A larger Kullback-Leibler divergence indicates a greater difference in beliefs between the current node and the neighbor node, resulting in a lower neighbor influence weight. Conversely, a smaller Kullback-Leibler divergence indicates a smaller difference in beliefs between the current node and the neighbor node, resulting in a higher neighbor influence weight. The system normalizes the node's own influence weight and the influence weights of all neighbor nodes, ensuring that the sum of all weights participating in the fusion is 1.

[0088] Based on the normalized aggregation weights, the node's own beliefs and the main channel messages of all neighboring nodes are weighted and fused to obtain the updated local posterior beliefs. Specifically, a weighted average operation is performed on the discrete state marginal probability vector, followed by re-normalization; a weighted average operation is performed on the conditional mean vector of the continuous intensity factor; and a weighted update operation is performed on the conditional variance vector of the continuous intensity factor, combining the conditional variance of each node and the deviation of each node's conditional mean from the fused conditional mean, to obtain the updated continuous intensity factor conditional variance vector.

[0089] The system iteratively executes primary and secondary dual-channel transmit / receive operations and weight update operations, monitoring the maximum absolute error of the belief vector between two consecutive iterations. The loop terminates when the maximum absolute error is less than a set convergence threshold. The posterior belief sets retained by each agent at the convergence time are then combined to form a node-level posterior probability distribution matrix. This node-level posterior probability distribution matrix is ​​further weighted and aggregated to obtain the global posterior probability distribution. This node-level posterior probability distribution matrix and the global posterior probability distribution together constitute the global posterior probability distribution for the current time step.

[0090] Furthermore, the Kullback-Leibler divergence between the user's own beliefs and those of its neighbors is transmitted as a secondary channel message. Specifically, the user obtains the discrete classification belief probabilities maintained by itself and those transmitted by each permitted neighboring agent; a preset smoothing term is added to both the user's own and neighboring agent's discrete classification belief probabilities, and they are then re-normalized. Based on the normalized user's and neighboring agent's belief probabilities, the Kullback-Leibler divergence is calculated to obtain the real value of the deviation baseline result; the real value of the deviation baseline result is used as an inconsistency feedback index to measure the mismatch distance between local neighboring nodes, and the inconsistency feedback index constitutes the secondary channel message.

[0091] For the active node i specified in the topology and the active node i in the current round of the undirected graph with the connection flag value of 1, the network communication and buffer management thread of the active node i will load two sets of vector state information at the same time.

[0092] The first set of vector state information is a local discrete classification belief distribution array retrieved from its own computation stack memory. The second set of vector state information is a belief probability distribution array decoded and received from neighboring agent node j via the main message exchange channel. (Local discrete classification belief distribution array) And belief probability distribution array The expression is as follows: ;

[0093] in, Let i be the local discrete classification belief distribution array of the active node i; Let i be the belief probability of the active node i for the first state type; Let i be the belief probability of the active node i for the second state type; Let i be the belief probability of the active node i for the third state type; Let be the array of belief probability distributions of neighboring agent node j; Let be the belief probability of neighboring agent node j for the first state type; Let be the belief probability of neighboring agent node j for the second state type; Let be the belief probability of neighboring agent node j for the third state type.

[0094] The computational operator adds small smoothing compensation terms to both its own belief distribution vector and its neighbors' belief distribution vectors, and then performs a normalization operation after adding the smoothing compensation terms. A preferred constant offset is used. This is to avoid zero-probability terms causing logarithmic aberrations, while maintaining the non-negativity of the Kullback-Leibler divergence. The value of the constant offset is based on the empirical rule of balancing machine precision limits and computational stability in double-precision floating-point arithmetic.

[0095] When the numerical alignment process begins, the system reads the self-belief distribution array of active node i. The probability distribution array of beliefs passed to neighboring agent node j To avoid logarithmic aberrations caused by zero-probability terms, the system first adds preset smoothing compensation terms to both sets of belief probabilities, and then re-normalizes the results after adding these terms, ensuring that the normalized results still satisfy the probability distribution constraint that the sum of the probabilities of each category is 1. Subsequently, the system calls the Kullback-Leibler divergence operator, using the smoothed and normalized self-belief distribution as a benchmark, to calculate the distribution difference between the smoothed and normalized self-belief distribution and the smoothed and normalized neighbor belief distribution, obtaining a non-negative real-valued divergence. This non-negative real-valued divergence characterizes the degree of inconsistency between the two agents in their discrete state judgments, and is used as a secondary channel message in the negative correlation adjustment of the neighbor influence weights.

[0096] If the distribution vector of one's beliefs in a real-world scenario is... And the distribution of exogenous input neighbors is After smoothing and normalization, and by calling the Kullback-Leibler divergence calculation operator, the real value of the divergence is 0.0253.

[0097] The node controller extracts the reference real number, packages and encapsulates it, and directly invests it into the reserved secondary channel in the next synchronization cycle, sending the reference real number to the topologically connected subdomain. The reference real number serves as a reference benchmark for the fault-tolerant negative constraint mechanism used to adjust and attenuate the strength of interactions between nodes. When calculating neighbor weights using secondary channel messages, a negative correlation normalization method is used for attenuation processing, instead of directly using the Kullback-Leibler divergence itself as a monotonically increasing weight.

[0098] For the propagation and update of continuous intensity factors, node i synchronously receives the conditional mean vector and conditional variance vector of continuous intensity factors from each neighboring node, and includes its own conditional mean vector and conditional variance vector in the current fusion process. The system performs a weighted fusion operation on the conditional mean vectors of the node itself and each neighboring node according to normalized aggregation weights, obtaining the updated conditional mean vector of continuous intensity factors. Simultaneously, considering the magnitude of the conditional variances of the node itself and each neighboring node, as well as the deviation of each conditional mean from the fused conditional mean, a weighted update operation is performed on the conditional variance vector, obtaining the updated conditional variance vector of continuous intensity factors. Through this weighted fusion method, the distribution parameters of continuous intensity factors can propagate and converge synchronously with discrete classification beliefs, thus providing a data source for subsequent calculations of parameter uncertainty.

[0099] S6 drives the prediction network to generate a revenue distribution representation.

[0100] In an optional embodiment, the prediction network is driven by the global posterior probability distribution of the current time step to generate corresponding prediction quantile points for the assets to be invested, forming a return distribution representation.

[0101] The discrete state marginal probability, continuous intensity factor conditional mean vector, and conditional variance vector are extracted from the global posterior probability distribution at the current time step. These vectors are then used as context variables and input into the decoder of the prediction network, which is composed of a long short-term memory network. During the training phase, which requires random sampling, a hybrid hidden market state vector can be obtained by sampling based on the global posterior probability distribution.

[0102] A quantile regression layer is connected at the end of the predictive network decoder. The QuantReg algorithm from the Statsmodels library is used to set an asymmetric absolute value loss function for a specific quantile, i.e., the Pinball loss function.

[0103] The system predicts the 5%, 50%, and 95% quantiles of the return of the investable asset at a future time step in parallel. To avoid quantile crossover, the quantile regression layer applies a non-crossover constraint to each predicted quantile point, or uses a cumulative positive increment method to output the quantile points, ensuring that the 5% quantile is no higher than the 50% quantile and the 50% quantile is no higher than the 95% quantile.

[0104] To avoid directly equating a complete continuous distribution with only three discrete quantile points, this embodiment uses three key quantile points as anchor points for expressing the return distribution, and constructs a piecewise monotonic quantile function by combining historical return residual samples of the same state. Specifically, piecewise linear interpolation is used in the 5% to 50% quantile interval and the 50% to 95% quantile interval. Based on the state category with the highest marginal probability of the discrete state at the current time step, or by weighting and matching historical residual samples according to the marginal probabilities of each discrete state, extrapolation is performed using the tail slope of the corresponding historical residual samples or empirical tail samples in the tail interval below 5% and above 95% to generate K monotonically arranged discrete samples of the return distribution.

[0105] The approximate expression of the future returns of the investable asset is constructed by a non-parametric empirical cumulative distribution function based on piecewise monotonic quantile functions or discrete sample sets, rather than uniquely determining the continuous distribution function based solely on three quantile points.

[0106] S7. Calculate the uncertainty index of the structure and parameters.

[0107] In an optional embodiment, the structural uncertainty index of the discrete state code in the global posterior probability distribution of the current time step, and the parameter uncertainty index of the continuous intensity factor are calculated.

[0108] Calculate the Shannon entropy of the discrete state multinomial distribution in the global posterior probability distribution, and use the Shannon entropy value as an indicator of structural uncertainty in the multilateral transition of market states.

[0109] Simultaneously, the conditional variance vector of the conditional Gaussian distribution followed by the continuous intensity factors in the global posterior probability distribution is extracted, or the diagonal covariance matrix is ​​constructed from the conditional variance vector. The dimensionless variance components or standard deviation components of each continuous intensity factor are used as the basis for calculating parameter uncertainty.

[0110] Furthermore, the structural uncertainty index of the discrete state code in the global posterior probability distribution at the current time step, and the parameter uncertainty index of the continuous intensity factor are calculated. Specifically, the marginal probabilities of each category of the discrete state code in the global posterior probability distribution are extracted, the product of each category marginal probability and its logarithm is calculated, and all products are summed and inverted to obtain the Shannon entropy expression value, which is used as the structural uncertainty index. The conditional variance or conditional standard deviation corresponding to the continuous intensity factor in the global posterior probability distribution is extracted, and each conditional variance or conditional standard deviation is dimensionlessly processed according to the historical scale parameter of the corresponding continuous intensity factor to obtain the dimensionless uncertainty component of each continuous intensity factor. The absolute value or square value of each dimensionless uncertainty component within the estimation time window is averaged, and the averaged dimensionless uncertainty components are weighted and summed to obtain the non-negative parameter uncertainty index.

[0111] After the belief transmission operation in the graph network within the communication layer has gone through several message passing iterations and triggered a convergence interruption, the data merging center will summarize and output the global posterior probability distribution at time t, and immediately trigger the uncertainty extraction task.

[0112] Regarding the process of extracting parameters representing structural uncertainty, a normalized categorized marginal probability list representing the macroeconomic bullish / bearish logic of the entire market is extracted from the global posterior probability distribution. This list, representing a scalar set, is a normalized categorized marginal probability list of a fusion state of the macroeconomic bullish / bearish logic, such as the first type representing a bull market, the second type representing a bear market, and the third type representing a sideways market. Scalar sets satisfy the condition And the sum of all components is 1.

[0113] The hardware arithmetic cluster extracts the current component according to iterative operators and applies a natural logarithm function with built-in anti-overflow correction and an appended smoothing constant to the current component using the underlying mathematical library. The result of the natural logarithm function is then multiplied by the original scalar to obtain the result set. All real numbers in the result set are arithmetically summed and added in the internal accumulator register. The negative of the accumulated result from the internal accumulator register is then extracted to form a complete discrete information theory entropy index formula system, i.e., to calculate the Shannon entropy value.

[0114] in, This represents the Shannon entropy value. For category index conditions; For components in a scalar set; It is the natural logarithm function; To add a smoothing constant.

[0115] It should be noted that this embodiment constructs the formula for calculating the Shannon entropy value based on the definition of Shannon entropy in information theory, and introduces a small smoothing constant to avoid abnormal errors such as the true number being zero in logarithmic operations.

[0116] By constructing a complete discrete information theory entropy index formula system by summing the negative numbers of the results, it is assumed that the edge fusion ratio distribution of the three macroscopic modes obtained at the current time t is 0.7, 0.2, and 0.1, the processor output entropy value is equal to 0.8018, and the scalar index is recorded and saved as a single comprehensive parameter for evaluating the purity of the distribution of macroscopic judgment in each local node of the system.

[0117] Meanwhile, for the measurement of intrinsic parameter fluctuations of continuous factors, the system directly extracts the conditional variance vector of the continuous intensity factor from the global posterior probability distribution, and then... As a statistical representation of the continuous intensity factor distribution term. If the system internally stores a standard deviation vector, it first squares it to obtain the conditional variance vector; if the system internally stores a diagonal covariance matrix, it extracts the diagonal elements of the diagonal covariance matrix to obtain the conditional variance vector; if the system internally stores a logarithmic variance vector, it first converts the logarithmic variance vector into a conditional variance vector.

[0118] The data query agent then performs a scan operation on the stored historical evaluation window, collecting the posterior conditional variance groups evaluated each day. The parameters are set, for example, using the historical estimated range limit variable W, set to 10 consecutive trading days. The 10 consecutive trading days are based on the trading cycle typically representing half a calendar month in financial trading; this window length can capture recent volatility characteristics while filtering out high-frequency intraday noise. Since the continuous strength factor includes two components at different scales: the moving average of the price difference and the percentage price deviation, the system performs dimensionless processing using the historical scale parameters of each continuous strength factor. For example, the conditional standard deviation corresponding to the moving average of the price difference is divided by the historical moving average scale of the moving average of the price difference, and the conditional standard deviation corresponding to the percentage price deviation is divided by the historical percentage volatility scale of the percentage price deviation, resulting in the dimensionless uncertainty component.

[0119] The computational pipeline performs a cyclic summation operation on the absolute or squared values ​​of each dimensionless uncertainty component within the estimated time window. The result of this cyclic summation is divided by the total number of recorded steps to smooth the feature space migration. The total number of recorded steps is, for example, a constant number of days (10), which is consistent with the historical estimation interval limit and is used to calculate the average fluctuation level within the time window. Then, a weighted summation operation is performed on the average dimensionless uncertainty components of the two continuous intensity factors according to preset non-negative weights, outputting a non-negative parameter uncertainty index. And the parameter uncertainty index will be used as a key parameter uncertainty benchmark for investment in the underlying financial risk control.

[0120] S8. Solve for investment weights by combining uncertainty indicators.

[0121] In an optional embodiment, the return distribution of the assets to be invested is input into the optimization framework, the structural uncertainty index is mapped to a concentration penalty term, the parameter uncertainty index is mapped to a risk tolerance cap of conditional value of risk, and the investment weights are solved based on the concentration penalty term and the risk tolerance cap.

[0122] A portfolio optimization framework is constructed using the CVXPY convex optimization library. Conditional Value at Risk (VaR) is calculated based on the piecewise monotonic quantile function or discrete samples of return distribution constructed earlier, rather than simply equating a single 5th percentile point with VaR.

[0123] Specifically, when using the rate of return When representing the future return of an asset, first convert the rate of return into a loss variable. Regarding confidence levels For example, confidence level Extracting the sample with the largest loss from the discrete sample of the profit distribution. The proportional sample is used, and the mean of the tail samples is calculated to obtain the discrete conditional value of risk. If a piecewise monotonic quantile function is used, the tail conditional value of risk can be approximated by discrete integral. The calculation method for tail conditional value of risk is as follows:

[0124] in, Value at risk for tail conditions; Confidence level; It is a piecewise monotonic quantile function; For integration variables; This is for adjusting the coefficient.

[0125] The parameter uncertainty index is normalized to obtain a non-negative normalized parameter uncertainty value. A risk tolerance scaling factor is then generated based on the reciprocal of the sum of the normalized parameter uncertainty value and a preset positive bias term. The calculation method for the risk tolerance scaling factor is as follows:

[0126] in, This is the risk tolerance scaling factor; Preset positive bias term; This is the normalized value for parameter uncertainty.

[0127] It should be noted that this embodiment constructs the calculation formula for the risk tolerance scaling factor based on the decay characteristics of the inverse proportional function. The value of the preset positive bias term is calibrated by the extreme value deviation in historical risk backtesting data, for example, set to 0.1, to ensure that the denominator is not zero and the mapping curve is smooth. Alternatively, an equivalent monotonically decreasing mapping function can be used to generate the risk tolerance scaling factor, so that the higher the parameter uncertainty, the lower the risk tolerance scaling factor.

[0128] The structural uncertainty index is mapped to the multiplier coefficient of the concentration penalty term using a linear scaling function. The multiplier coefficient is then multiplied by the Herfindahl-Hirschman index, which is calculated from the sum of squared investment weights, to form the concentration penalty term.

[0129] Under the constraints of setting non-negative weights for each asset and a sum of 1, and ensuring that the portfolio's conditional value of risk does not exceed the adjusted tail risk tolerance limit, the objective function is to maximize the expected return minus the concentration penalty term.

[0130] The SCS solver in the CVXPY library is called to perform convex optimization calculations, outputting the optimal weight vector that maximizes the objective function, and then the optimal weight vector is used as the investment weight for each asset to be invested in at the current time step.

[0131] Furthermore, the structural uncertainty index is mapped to a concentration penalty term, and the parametric uncertainty index is mapped to the risk tolerance cap of the conditional value of risk. Specifically, the allocation weight of each investable asset in the portfolio is extracted. Under the condition that the sum of the allocation weights is 1, the squares of the allocation weights are calculated and summed to obtain the Herfindahl-Hirschman index, which represents the concentration of the portfolio. The structural uncertainty index is multiplied by the Herfindahl-Hirschman index to construct a concentration penalty instruction for over-allocation of a single asset, and this concentration penalty instruction is used for overall objective constraints. The parametric uncertainty index is normalized to obtain a non-negative normalized parametric uncertainty value. A risk tolerance scaling factor is generated based on the reciprocal of the sum of the normalized parametric uncertainty value and a preset positive bias term, making the risk tolerance scaling factor negatively correlated with the normalized parametric uncertainty value. The risk tolerance scaling factor is multiplied by the initial risk tolerance cap of the conditional value of risk to obtain the adjusted tail risk tolerance cap, so that the optimization framework constrains the conditional value of risk with the adjusted tail risk tolerance cap.

[0132] Just before the portfolio weight calculation workflow is pushed to the terminal target optimization stage, the portfolio control function extracts the current dynamic derivation allocation floating-point value column of all N financial objects mapped to the asset pool from the memory array and compiles it into a basic portfolio planning distribution constraint array. The basic portfolio planning distribution constraint array satisfies the following relation:

[0133] in, Based on the distributed constraint array of the combinatorial planning; The current dynamic projection allocation floating-point value for the first financial object; The current dynamic projection allocation floating-point value for the second financial object; The floating-point value for the current dynamic projection allocation of the Nth financial object.

[0134] The numerical flow must be subject to normalization constraints to satisfy the linear identity gauge boundary condition, i.e., the sum of all weight components equals 1.

[0135] In order to objectively monitor the phenomenon of excessive trading congestion during this process, the solver initiates a power function square transformation operation on the column matrix of the weight space to obtain self-squared terms, and then uses the accumulation network component to obtain the numerical sum, thereby extracting the Herfindahl-Hirschman standard referential cardinality used to define the chip concentration pattern under the metric model.

[0136] Assuming the test configuration only has binary combination ratios of 0.4 and 0.6, the total square product score produced by the evaluator is 0.52. In the subsequent planning generation phase, the Shannon information entropy record value, which indicates the intensity of global information dispersion and is obtained from the previous convergence, is used as a moderating variable. This value is algebraically multiplied with the cardinality obtained from the evaluation to construct a concentration penalty term that restricts the excessively high proportion of a single or few component varieties. This concentration penalty term is then incorporated as an additional constraint into the penalty term of the optimal planning objective function, and overall structure control is implemented. The calculation method for the concentration penalty term is as follows:

[0137] in, This is a concentration penalty item; Record the Shannon information entropy value; The total number of financial objects; For the index number of the financial object; It is a term multiplied by itself by the square.

[0138] It should be noted that this embodiment uses the Herfindahl-Hirschman index, a method for measuring market concentration in industrial economics, and combines it with the information entropy dynamic adjustment coefficient to construct a formula for calculating the concentration penalty term.

[0139] Furthermore, regarding the control logic for asset drawdown distribution, the scheduler further performs dimensional isolation and removal processing on the retrieved parameter uncertainty index representing the uncertainty of asset market parameters. Specifically, by using a predefined envelope amplitude boundary assigned to the historical full sample set, for example, setting the lower limit of volatility to 0.1 and the peak value to 2, the parameter uncertainty index is transformed into a non-negative parameter uncertainty normalized value using a standardization operator. Subsequently, a risk tolerance scaling factor is generated based on the reciprocal of the sum of the parameter uncertainty normalized value and a preset positive bias term. The risk tolerance scaling factor is then truncated at both the upper and lower limits to ensure that its value range falls within a preset safety range, thus preventing the upper limit of risk tolerance from being excessively amplified or compressed. This mapping relationship ensures that the higher the parameter uncertainty, the lower the tail risk tolerance upper limit, and vice versa. The upper and lower limits of the preset safety range are calibrated using the historical maximum drawdown distribution generated by Monte Carlo simulation.

[0140] The initial risk tolerance cap of the conditional value of risk loaded by the configuration engine is used as the baseline cap. An algebraic multiplication update is performed using the risk tolerance scaling factors to obtain the adjusted tail risk tolerance cap. The optimization framework constrains the portfolio's conditional value of risk to no higher than the adjusted tail risk tolerance cap. When the adjusted tail risk tolerance cap decreases, the system tightens the constraint on tail losses, thereby suppressing high-risk allocations under high uncertainty environments.

[0141] The experiment selected daily trading data of 50 core constituent stocks from a financial database over the past 5 years as the test set, with a time window of 252 trading days. Three control groups were constructed for ablation assessment. The basic group used a standard single-agent long short-term memory network combined with a traditional mean-variance optimization model. The partial ablation group used a multi-agent dual-forgetting-gated recurrent neural network model and generated a hybrid hidden market state vector, but removed the constraint mappings of structural and parameter uncertainties. The complete group adopted all the technical features of this embodiment, including a communication topology architecture generated based on marginal probability, and a complete optimization module combining Herfindahl-Hirschman index concentration penalty and a risk tolerance cap scaling mechanism based on conditional value of risk.

[0142] Reference Figure 2During the cumulative net asset value (NAV) evolution of the portfolio, the complete group maintained a high NAV throughout the testing period and achieved significant growth in the later stages. The cumulative NAV trend of the partially ablated group fell between that of the complete group and the base group. The cumulative NAV of the base group experienced a significant decline in the early stages and fluctuated the most drastically throughout the period, ultimately resulting in the lowest NAV level. This demonstrates that the segmentation mechanism based on the heterogeneous perspective of the multi-agent system and the dual forgetting network identified the nonlinear characteristics of the hidden market state, thereby achieving an initial performance improvement of the partially ablated group relative to the base group. The more crucial performance leap stems from the uncertainty penalty mechanism utilized by the complete solution, which dynamically adjusts constraints to achieve a contraction of the asset allocation defense line, reducing NAV volatility and demonstrating excellent disaster recovery capabilities during periods of rapid market style shifts.

[0143] The base group achieved an annualized return of 8.5%, with a maximum drawdown of 22.4% and a Sharpe ratio of 0.65. The partially liquidated group saw its annualized return increase to 11.2%, its maximum drawdown decrease to 18.1%, and its Sharpe ratio reach 0.92. The full group's annualized return rose to 14.6%, its maximum drawdown shrank to 9.5%, and its Sharpe ratio reached a high of 1.45. Furthermore, the full group maintained a stable daily maximum concentration index below 0.12 during periods of extreme market volatility, lower than the crowding levels of the other two groups (above 0.3).

[0144] Reference Figure 3 The concentration indices of the basic group and the partially ablated group are at relatively high and similar levels, indicating that the basic group and the partially ablated group are more concentrated in the allocation of resources or chips, and are prone to overcrowding. The concentration index of the complete group is significantly lower than that of the basic group and the partially ablated group. This proves that mapping structural uncertainty to a concentration penalty term can effectively avoid the risk of excessive chip concentration in a one-sided market, and shows the difference in chip concentration of the portfolio under different mechanisms.

[0145] This invention also discloses an intelligent investment process system based on multi-agent collaboration, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the intelligent investment process method based on multi-agent collaboration according to this invention is implemented.

[0146] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0147] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for intelligent investment process based on multi-agent collaboration, characterized in that, include: S1. Obtain the financial time series data of the assets to be invested and assign observation perspectives to each agent; S2. Construct a hybrid hidden market state vector composed of discrete state codes and continuous intensity factors; S3. Each agent uses a sequence transformation model to generate a local prior probability distribution for the hybrid hidden market state vector; S4. Generate a communication topology based on the marginal probabilities of the discrete state codes in the global posterior probability distribution of the previous time step; S5. Each agent performs dual-channel belief propagation, transmitting the local prior probability distribution and the Kullback-Leibler divergence between its own beliefs and those of its neighbors as primary and secondary channel messages, respectively. The Kullback-Leibler divergence is used to adjust the neighbor influence weights and update the beliefs to obtain the global posterior probability distribution for the current time step; S6. Drive the prediction network based on the global posterior probability distribution to generate prediction quantile points for the assets to be invested, forming a return distribution representation. S7. Calculate the structural uncertainty index of the discrete state code in the global posterior probability distribution, and the parameter uncertainty index of the continuous strength factor; S8. Input the return distribution expression into the optimization framework, map the structural uncertainty index to the concentration penalty term, map the parameter uncertainty index to the risk tolerance upper limit of the conditional value of risk, and solve the investment weight by combining the concentration penalty term and the risk tolerance upper limit.

2. The intelligent investment process method based on multi-agent collaboration according to claim 1, characterized in that, The process of acquiring financial time-series data of the assets to be invested in and assigning observation perspectives to each agent includes: receiving time-series information containing specific investment targets from a financial database through an application programming interface; arranging the extracted data sequences in memory in ascending order of time as independent time-series data tensor record matrices to form a baseline three-dimensional tensor, and using the independently arranged time-series data tensor record matrix as a baseline observation source to segment and call to each agent; and performing slicing and mapping operations on the baseline three-dimensional tensor in terms of spatial and feature dimensions based on the pre-set heterogeneous feature preference configurations of each agent to achieve independent assignment of observation perspectives.

3. The intelligent investment process method based on multi-agent collaboration according to claim 2, characterized in that, The construction of the hybrid hidden market state vector, which consists of discrete state encoding and continuous strength factor, includes: generating a one-dimensional array in the storage space, presetting the element at index 0 of the array as an integer format variable limited to 1 to 3, the values ​​of which represent three macro-structure patterns respectively, to obtain the discrete state encoding; extracting the absolute moving average of the difference between the closing price and the opening price within a set historical time window, and the percentage deviation of the closing price relative to its historical moving average, as two floating-point variables, which are stored in index 1 and index 2 spaces respectively, and the continuous strength factor is composed of the two floating-point variables.

4. The intelligent investment process method based on multi-agent collaboration according to claim 1, characterized in that, Each agent employs a gated recurrent neural network model with a double forget gate as its sequence transformation model. The generation of the local prior probability distribution includes: feeding the received data tensor into the gated recurrent neural network model with a double forget gate; performing a normalized exponential transformation on the first mapping branch of the gated hidden output state matrix to output three edge prediction probability values ​​corresponding to different discrete state codes; simultaneously guiding the second mapping branch of the gated hidden output state matrix to a conditional mean square evaluation predictor with a linear fully connected mapping layer to generate the expected value of the predicted target and the distribution variance; and encapsulating the edge prediction probability values, the expected value of the predicted target, and the distribution variance as a local prior probability distribution.

5. The intelligent investment process method based on multi-agent collaboration according to claim 1 or 4, characterized in that, The method of generating a communication topology based on the marginal probability of discrete state codes in the global posterior probability distribution of the previous time step includes: extracting the marginal probability distribution matrix of discrete state codes retained from the previous operation cycle of all multi-agents in the global communication domain; selecting any two agents to calculate the absolute difference in their marginal probability distributions and establishing connection distance values ​​based on Chebyshev distance; comparing each connection distance value with a set distance limit parameter, setting the communication permission status flag to 1 if it is lower than or equal to the distance limit parameter, and setting the communication permission status flag to 0 if it exceeds the distance limit parameter; and filling the adjacency matrix of an undirected graph based on all the communication permission status flags to complete the generation of the communication topology.

6. The intelligent investment process method based on multi-agent collaboration according to claim 1, characterized in that, The step of transmitting the Kullback-Leibler divergence between one's own beliefs and those of neighbors as a secondary channel message includes: obtaining the discrete classification belief probability maintained by oneself and the discrete classification belief probability transmitted by each permitted neighboring agent; adding a preset smoothing term to the discrete classification belief probability maintained by oneself and the discrete classification belief probability transmitted by neighboring agents and re-normalizing them; calculating the Kullback-Leibler divergence based on the normalized self-belief probability and the belief probability of neighboring agents to obtain the real value of the deviation base point result; using the real value of the deviation base point result as an inconsistency feedback index to measure the mismatch distance of local neighboring nodes, and using the inconsistency feedback index to constitute the secondary channel message.

7. The intelligent investment process method based on multi-agent collaboration according to claim 1, characterized in that, The calculation of structural uncertainty and parameter uncertainty indices includes: calculating the product of the marginal probabilities of each class and the logarithm of each class in the global posterior probability distribution, summing all the product results and inverting them to obtain the Shannon entropy expression value as the structural uncertainty index; extracting the conditional variance or conditional standard deviation corresponding to the continuous intensity factor, and performing dimensionless processing according to the historical scale parameter of the continuous intensity factor to obtain the dimensionless uncertainty component of each continuous intensity factor; averaging the absolute value or square value of each dimensionless uncertainty component within the estimation time window, and performing a weighted summation operation to obtain the non-negative parameter uncertainty index.

8. The intelligent investment process method based on multi-agent collaboration according to claim 1, characterized in that, The method of solving for investment weights by combining concentration penalty term and risk tolerance cap includes: extracting the allocation weight ratio of the assets to be invested in the portfolio; calculating and summing the squares of the allocation weight ratios under the condition that the sum of the allocation weight ratios is 1 to obtain the Herfindahl-Hirschman index; multiplying the structural uncertainty index with the Herfindahl-Hirschman index to construct a concentration penalty term instruction for overall objective constraint; normalizing the parameter uncertainty index to obtain a parameter uncertainty normalized value; generating a risk tolerance scaling factor based on the reciprocal of the sum of the parameter uncertainty normalized value and a preset positive bias term; multiplying the risk tolerance scaling factor with the initial risk tolerance cap to obtain the adjusted tail risk tolerance cap, so that the optimization framework constrains the value at risk with the adjusted tail risk tolerance cap.

9. The intelligent investment process method based on multi-agent collaboration according to claim 1, characterized in that, The prediction network driven by the global posterior probability distribution generates predicted quantile points to form a return distribution expression for the investable asset. This includes: extracting discrete state marginal probabilities, continuous intensity factor conditional mean vectors, and conditional variance vectors from the global posterior probability distribution as context variables and inputting them into the prediction network decoder; connecting a quantile regression layer at the end of the prediction network decoder to predict the key quantile points of the investable asset at a future time step in parallel; using the key quantile points as anchor points for the return distribution expression, and constructing a piecewise monotonic quantile function by combining historical return residual samples of the same state, generating a monotonically arranged discrete sample of the return distribution; and constructing a non-parametric empirical cumulative distribution function approximating the future return of the investable asset as the return distribution expression by using the piecewise monotonic quantile function or the set of discrete sample return distributions.

10. An intelligent investment process system based on multi-agent collaboration, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the intelligent investment process method based on multi-agent collaboration according to any one of claims 1-9.

Images