Large model evolution multi-agent investment decision matrix system

Abstract

The embodiment of the invention relates to the technical field of large model data analysis, and discloses a large model evolution multi-agent investment decision matrix system. The investment Transform module is used for acquiring financial data and performing feature extraction and feature fusion on the financial data in combination with a context learning enhancement function and a financial domain knowledge enhancement function; the interpretability analysis module is used for acquiring the feature representation in the investment Transform module and carrying out feature relation mapping and interpretability analysis; the multi-agent decision-making module is used for carrying out collaborative decision-making through multiple agents; the optimization module is used for providing interpretation optimization for the interpretability analysis module and providing strategy optimization for the multi-agent decision-making module; and the knowledge base module is used for providing an investment template for the investment Transformer module, providing causal relationship knowledge for the interpretability analysis module and providing rule constraints for the multi-agent decision-making module. The technical problem of investment decision analysis can be at least solved.

Description

Technical Field

[0001] This application relates to the field of large model data analysis technology, and in particular to a multi-agent investment decision matrix system for large model evolution. Background Technology

[0002] Traditional investment decision-making methods rely heavily on human experience and statistical models, which have significant limitations in practical applications. On the one hand, their data processing capabilities struggle to handle the simultaneous analysis of massive amounts of structured financial data (such as market data and corporate financial reports) and unstructured data (such as news sentiment and social media sentiment), leading to incomplete information mining. On the other hand, linear models or empirical rules are ill-suited to capturing non-linear relationships and sudden fluctuations in financial markets, exhibiting weak market adaptability. Furthermore, the inefficiency of manual analysis and decision-making cannot meet the real-time response requirements of high-frequency trading scenarios. In addition, traditional risk control models over-rely on historical data, lacking sufficient early warning capabilities for extreme market events, and are prone to delays in risk identification.

[0003] In recent years, while the application of machine learning technology in the investment field has been gradually promoted, it still faces multiple technical bottlenecks. Existing intelligent investment systems mostly employ single deep learning models, whose decision-making processes are often considered "black boxes," making it difficult to explain the importance of features and decision-making logic to investors or regulatory agencies. This lack of interpretability severely limits its application in high-risk investment scenarios. Simultaneously, single-agent architectures struggle to simultaneously address multiple objectives such as return optimization, risk control, and liquidity management through collaborative decision-making, and lack mechanisms for strategy complementarity and information sharing among multiple agents. Furthermore, the ability of general machine learning models to integrate with prior knowledge in the financial field (such as industry valuation logic and trading rule constraints) is insufficient, easily leading to decision results deviating from economic fundamentals. Traditional models also require retraining with a large number of samples when the market environment changes drastically, and the strategy iteration speed cannot adapt to dynamic market changes. Summary of the Invention

[0004] One objective of this application is to provide a multi-agent investment decision matrix system with large-scale model evolution, which at least addresses the technical problems of investment decision analysis.

[0005] To achieve the above objectives, some embodiments of this application provide a multi-agent investment decision matrix system for large-scale model evolution, comprising: an investment Transformer module for acquiring financial data and performing feature extraction and feature fusion on the financial data by combining context learning enhancement functions and financial domain knowledge enhancement functions; an interpretability analysis module for acquiring feature representations from the investment Transformer module and performing feature relationship mapping and interpretability analysis; a multi-agent decision module for acquiring feature representations from the investment Transformer module and decision interpretations from the interpretability analysis module, and performing collaborative decision-making through multiple agents; an optimization module for providing parameter optimization for the investment Transformer module, explanation optimization for the interpretability analysis module, and strategy optimization for the multi-agent decision module; and a knowledge base module for providing investment templates for the investment Transformer module, causal relationship knowledge for the interpretability analysis module, and rule constraints for the multi-agent decision module.

[0006] Compared with related technologies, the solution provided in this application improves the accuracy of predicting market trends and asset prices through multimodal feature fusion and domain knowledge enhancement; it constructs transparent decision path analysis and causal relationship explanation to meet regulatory requirements and enhance investor trust; it balances multiple investment objectives such as returns, risks, and liquidity through a multi-agent collaborative mechanism; it reduces the model's dependence on historical data and improves system adaptability by utilizing meta-learning and knowledge base update mechanisms; and it automates feature extraction, strategy optimization, and risk control processes to reduce manual intervention and improve decision-making efficiency. Attached Figure Description

[0007] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0008] Figure 1 This is an architecture diagram of a multi-agent investment decision matrix system based on a large model evolution according to an embodiment of this application;

[0009] Figure 2 This is an architecture diagram of an investment Transformer module provided according to an embodiment of this application;

[0010] Figure 3 This is a diagram illustrating a multi-agent decision-making module architecture according to an embodiment of this application.

[0011] Figure 4 This is an architecture diagram of a system optimization and monitoring system provided according to an embodiment of this application;

[0012] Figure 5 This is a flowchart of a multi-agent investment decision matrix method for large model evolution provided in an embodiment of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] This application relates to a multi-agent investment decision matrix system with large model evolution, such as... Figure 1 As shown, the system includes:

[0015] Financial data includes market data, fundamental data, and technical data. Financial data needs to be preprocessed and feature-engineered before entering the investment Transformer module.

[0016] The InvestFormer module acquires financial data and performs feature extraction and fusion on this data by combining contextual learning enhancement functions and financial domain knowledge enhancement functions. Based on the weight parameters of InvestFormer, contextual learning, and domain knowledge, the feature fusion outputs a state vector to the multi-agent decision-making module.

[0017] The interpretability analysis module is used to obtain the feature representations in the investment Transformer module, perform feature relationship mapping and interpretability analysis, and input the analysis layer results into the knowledge base for knowledge updates.

[0018] The multi-agent decision-making module is used to acquire the feature representations from the investment Transformer module and the decision interpretations from the interpretability analysis module, and to perform collaborative decision-making through multiple agents. Each agent (agent 1, agent 2, agent 3) processes the state vector to make local decisions and then makes collaborative decisions, inputting the collaborative decision-making results into the interpretability analysis module.

[0019] An optimization module is used to provide parameter optimization for the investment Transformer module, interpretation optimization for the interpretability analysis module, and strategy optimization for the multi-agent decision-making module;

[0020] The knowledge base module is used to provide investment templates for the investment Transformer module, causal relationship knowledge for the interpretability analysis module, and rule constraints for the multi-agent decision-making module.

[0021] To ensure the consistency and uniqueness of variable definitions across all modules of the system, the following variable naming conventions are adopted in the following examples: -INV_: Variable prefix for the Investment Transformer module; -EXP_: Variable prefix for the Interpretability Analysis module; -AGT_: Variable prefix for the Multi-Agent Decision-Making module; -OPT_: Variable prefix for the Optimization module; -KB_: Variable prefix for the Knowledge Base module; -CASE_: Variable prefix for the Case Management module; -STRAT_: Variable prefix for the Strategy Optimization module; -MKT_: Variable prefix for the Market Analysis module; -RISK_: Variable prefix for the Risk Warning module; -PERF_: Variable prefix for the Performance Evaluation module.

[0022] To ensure dimensional consistency of data flow across modules, the core system variable dimensions are defined as follows:

[0023] - Input feature vector x∈R^1024;

[0024] - The Transformer outputs TF(x) ∈ R^1024;

[0025] - Context learning outputs ICL(x)∈R^256;

[0026] - Domain knowledge output Domain(x)∈R^512 (through weighted combination Market(x)∈R^512+

[0027] Risk(x)∈R^256+Value(x)∈R^384 (achieving dimension alignment);

[0028] - The final output is InvestFormer(x) ∈ R^1024;

[0029] - Decision vector d∈R^1024;

[0030] -State representation S(a) i )∈R^512(through weighted combination Obs(a i )∈R^512+Hist(a i )∈R^256+Ctx(a i ()∈R^384 to achieve dimension alignment);

[0031] - Action space A(s) i )∈R^128;

[0032] -Reward function R(a)i )∈R^64.

[0033] All dimensional transformations are achieved through the corresponding transformation matrix W, ensuring mathematical consistency in data transfer between modules.

[0034] The Investment Transformer module includes: providing the interpretability analysis module with quantified data on interpretable objects and feature attention metrics; and generating intermediate features for the multi-agent decision-making module to process, thereby enhancing decision-making information.

[0035] The context learning enhancement function obtains investment template vectors from the knowledge base module, example vectors from the case library, and context vectors from the financial data;

[0036] The financial domain knowledge enhancement function provides market characteristics to the market analysis module, risk characteristics to the risk warning module, and value characteristics to the performance evaluation module.

[0037] like Figure 2 As shown, this embodiment constructs an investment-specific Transformer module, InvestFormer, which provides feature representations for the interpretability analysis module, provides information support for the multi-agent decision-making module, serves as the main optimization target of the optimization module, and extracts analysis results from the knowledge base module.

[0038] InvestFormer(x)=α_INV1·TF(x)+α_INV2·ICL(x)+α_INV3·Domain(x)

[0039] Constraints: α_INV1 + α_INV2 + α_INV3 = 1, α i ≥0;

[0040] Here, TF(x) is an improved Transformer structure for the financial domain: TF(x) = LayerNorm(x + MHA(Q, K, V)) + LayerNorm(FFN(x + MHA(Q, K, V))), where x ∈ R^1024 is the input vector, serving as the interpretation object for Explanation(d) in the interpretability analysis module, containing market data, company financial reports, and macroeconomic indicators; Q (Query matrix), K (Key matrix), and V (Value matrix). The linear transformation yields: Q = x·W_Q, K = x·W_K, V = x·W_V, where W_Q, W_K, W_V ∈ R^1024×1024 are parameter matrices, updated by AutoML(θ) in the optimization module to ensure the model can adapt to market changes; the MHA multi-head attention mechanism supports 8 heads, each with a dimension of 128, specifically calculated as: MHA(Q,K,V) = Concat(head1,head2,...,head8)·W_O, where head i =Attention(Q·W_Q) i ,K·W_K i ,V·W_V i Attention(Q) i ,K i V i ) = softmax(Q i K i T / √d k V i d k =128; FFN uses the GELU activation function, with an intermediate layer dimension of 4096. The generated intermediate features are used for multi-agent state representation to enhance decision-making information.

[0041] ICL(x) = β_ICL1·(W_T·Template(x)⊙W_D·Demo(x)) + β_ICL2·W_C·Context(x), where W_T∈R^256×512, W_D∈R^256×256, and W_C∈R^256×384 are dimensional transformation matrices, ⊙ denotes element-wise multiplication, and the final output is ICL(x)∈R^256; Template(x)∈R^512 is the investment template vector, obtained from the knowledge base module, which includes industry analysis frameworks and valuation models. The model and risk assessment criteria; Demo(x)∈R^256 is the example vector, from the case management module (see below), providing historical similar decision references; Context(x)∈R^384 is the context vector, containing environmental information from the multi-agent collaborative decision-making module, integrating market sentiment, trading volume, and volatility indicators; the weights of β_ICL1 and β_ICL2 are dynamically optimized by the optimization module's MetaLearn(θ), adjusting the influence of each component according to the current market environment, with the constraint: β_ICL1+β_ICL2=1, β i ≥0.

[0042] Domain(x) is a knowledge enhancement function in the financial domain: Domain(x) = γ_INV1·Market(x) + γ_INV2·Risk(x) + γ_INV3·Value(x), where Market(x) ∈ R^512 represents market characteristics, input to the market analysis module (see below), including index trends, sector rotation, and fund flows; Risk(x) ∈ R^256 represents risk characteristics, provided to the risk warning module (see below), integrating volatility, correlation, and extreme event probabilities; Value(x) ∈ R^384 represents value characteristics, used for valuation evaluation in the performance evaluation module (see below), covering fundamental, technical, and sentiment indicators; the weights of γ_INV1, γ_INV2, and γ_INV3 are optimized through MultiTask(θ) of the optimization module to achieve multi-objective balance, with the constraint: γ_INV1 + γ_INV2 + γ_INV3 = 1, γ i ≥0.

[0043] This embodiment achieves an organic integration of financial knowledge and deep learning through an investment-specific Transformer architecture, providing high-quality feature representations and decision-making basis for subsequent modules. This architecture supports parallel computing with a batch size of 32, handles 30 million model parameters, can process 10,000 market data points per second, and has a time complexity of O(n^2). 2 With a space complexity of O(n), it serves as the core feature extraction and representation learning module of the entire system, laying the foundation for intelligent decision-making.

[0044] The interpretability analysis module includes: providing decision explanations for the multi-agent decision-making module, causal analysis for the strategy optimization module (see below), and transparent explanations for the risk warning module (see below), while feeding back the explanation results to the knowledge base module (see below); the interpretability analysis module includes a feature attribution function, a decision tracking function, and a causal analysis function; the feature attribution function is used to quantify the contribution of each feature to the decision, calculate the feature importance and construct a decision path graph, and provide the market analysis module with a feature importance ranking; the decision tracking function is used to track the feature propagation path and analyze the impact of decision feedback, connect the interaction records of the multi-agent decision-making module, and perform decision consequence backtracking; the tracking results are input into the risk warning module for risk tracing; the causal analysis function is used to calculate first-order causal effects based on feature relationships; track the interactions of the multi-agent decision-making module to obtain higher-order causal chains; and input the causal analysis results into the strategy optimization module to guide the direction of strategy adjustment.

[0045] Explainability module: Explanation(d)

[0046] Explanation(d)=α_EXP1·Attribute(d)+α_EXP2·Track(d)+α_EXP3·Causal(d)

[0047] Constraints: α_EXP1 + α_EXP2 + α_EXP3 = 1, α i ≥0;

[0048] Where Attribute(d) is the feature attribution function: Attr(d)=∑(i=1 to n)w i Shapley(f i )+PathExplain(d), where d is the decision vector with dimension d_base = 1024, and is the output feature representation from the InvestFormer function, containing all the feature representations required for investment decisions; w i The feature importance weights are calculated based on the MHA attention scores in TF(x), quantifying the contribution of each feature to the decision, satisfying ∑w i =1; f i As the key feature, it is defined as f i =d i (The i-th component of d), the first 256-dimensional features extracted from d, corresponding to the state space representation of the multi-agent collaborative decision-making module, containing key decision-making elements; PathExplain(d) constructs a complete decision path graph with a size of 128×128, including operational reasons and impact analysis, supporting end-to-end tracing from features to decisions; Shapley(fi Monte Carlo sampling was used with a sample size of 1000. The calculation results were used in the market analysis module to provide feature importance ranking. Track(d) is the decision tracking function: Track(d) = β_EXP1·Forward(d) + β_EXP2·Backward(d), with the constraints: β_EXP1 + β_EXP2 = 1, β i ≥0; where Forward(d)∈R^256 tracks the feature propagation path, records the entire information transmission process based on the attention flow in TF(x); Backward(d)∈R^256 analyzes the impact of decision feedback, connects the agent interaction records of the multi-agent collaborative decision-making module, and realizes the backtracking of decision consequences; the weights of β_EXP1 and β_EXP2 are dynamically adjusted by the optimizer of the optimization module, and resources are allocated according to the importance of tracking; the tracking results are in the form of a directed graph with 64 nodes and 256 edges, and are input into the risk warning module to support risk tracing. Causal(d) is the causal analysis function: Causal(d)=γ_EXP1·ATE(X→Y)+γ_EXP2·∑ i ATE(X→Z i →Y), where ATE is the average treatment effect, X→Y represents the direct causal path, and X→Z i →Y represents the result through the mediator variable Z i The indirect causal path; Direct(d)=ATE(X→Y)∈R^128 analyzes the direct impact, and calculates the first-order causal effect based on the Domain(x) characteristic relationship of InvestFormer(x); Indirect(d)=∑ i ATE(X→Z i →Y)∈R^256 studies indirect effects, tracks agent interactions in a multi-agent collaborative decision-making module, and reveals hidden higher-order causal chains; γ_EXP1 and γ_EXP2 adaptively adjust the weights of direct and indirect causality through MetaLearn(θ) learning in the optimization module, with the constraint: γ_EXP1+γ_EXP2=1, γ i ≥0; The causal analysis results are represented by a directed acyclic graph (DAG) structure, containing 128 nodes and 256 edges, which are used by the strategy optimization module to guide the direction of strategy adjustment.

[0049] This embodiment constructs a complete interpretability analysis framework, achieving comprehensive and transparent explanation of the investment decision-making process through feature attribution, decision tracing, and causal analysis. This framework boasts high computational efficiency, with a single explanation generation time of less than 100 milliseconds and an explanation accuracy of 90%. It can handle feature spaces up to 256 dimensions, providing the necessary transparency and credibility for system decision-making while meeting regulatory compliance requirements, enabling end-users to understand and verify the basis of their decisions.

[0050] The multi-agent decision-making module includes: based on the feature representation and the decision interpretation, the multi-agent decision-making module provides the execution subject for the strategy optimization module (see below), reports the status to the early warning risk module, and feeds back the decision experience to the strategy base (see below); the multi-agent decision-making module includes a state representation function, a decision generation function, and a reward calculation function; the state representation function is used to receive market observations, store historical records, and provide decision background; the market observations come from the investment Transformer module, the historical records come from the case base and establish time-series memory, and the decision background comes from the interpretability analysis module; the decision generation function is used to generate investment decisions, execute rule constraints, and evaluate feedback dynamic adjustments; the investment decisions are based on the features of the investment Transformer module, the rule constraints come from the rule constraints of the knowledge base, and the feedback dynamic adjustments are based on the market analysis module; the reward calculation function is used to calculate returns and evaluate risks; the calculated returns are compared with the valuation of the performance evaluation module, and the evaluated risks are connected to the risk early warning module for evaluation. Figure 3 As shown:

[0051] MASystem:S×A→S×R

[0052] Where S is the state space, A is the action space, and R is the reward space, the system generates new states and reward feedback through state-action mapping.

[0053] State(A) is the state representation function: S(a i )=α_AGT1·Obs(a i )+α_AGT2·Hist(a i )+α_AGT3·Ctx(a i ), constraints: α_AGT1+α_AGT2+α_AGT3=1, α i ≥0; Obs(a i )∈R^512 receives market observations from Market(x), Risk(x), and Value(x) in Domain(x) of InvestFormer, containing real-time market data; Hist(a i)∈R^256 stores historical records, including Success(C) and Failure(C) from the case library in the case management module, establishing a time-series memory; Ctx(a i The state space ∈ R^384 contains environmental information, and the results are explained using the Explanation(d) function of the interpretability analysis module, providing a decision-making context. α_AGT1, α_AGT2, and α_AGT3 are adjusted by the optimizer of the optimization module, adaptively adjusting the weights of each information source according to the market environment. The state space supports parallel processing by up to 32 agents, with a state update frequency of 10 times per second and a state resolution of 99.5%.

[0054] Action(A) is the decision generation function: A(s) i )=β_AGT1·Policy(s i )+β_AGT2·Rule(s i ), constraint condition: β_AGT1+β_AGT2=1, β i ≥0; where Policy(s) i Investment decisions are generated based on the TF(x) feature of InvestFormer, including nine basic operations such as buy, sell, and hold; Rule(s) ∈ R^128. i The rules ∈ R^64 are used to enforce risk control restrictions, derived from the rule set in the knowledge base Graph(K) of the knowledge base module. β_AGT1 and β_AGT2 are dynamically adjusted based on feedback from the Analysis(M) module of the market analysis module to ensure a balance between policy and rules. The decision accuracy reaches 95%, the latency is less than 10 milliseconds, and the processing complexity is O(n^64). 2 The decision space of ). Reward(A) is the reward calculation function: R(a) i )=γ_AGT1·Return(a i )+γ_AGT2·Risk(a i ), constraint condition: γ_AGT1+γ_AGT2=1, γ i ≥0; where Return(a i The return is calculated using R^64, including three dimensions: absolute return, relative return, and risk-adjusted return, and compared with the Evaluate(P) valuation from the performance evaluation module; Risk(a iRisk is assessed using the ∈R^128 method, which includes various risk measures such as volatility risk, liquidity risk, and systemic risk. It is integrated with the Monitor(R) risk control indicator of the risk warning module. γ_AGT1 and γ_AGT2 are dynamically determined based on the MultiTask(θ) multi-objective optimization of the optimization module and are adaptively adjusted according to the investor's risk preferences. The reward signal is updated in real time after each decision, and the historical reward decay factor is 0.95 to ensure the maximization of long-term value.

[0055] This embodiment implements a distributed investment decision-making mechanism, which transforms the feature representation and decision interpretation of the preceding modules into specific investment actions. It supports simultaneous processing of 8 investment categories, with a decision latency of <5ms in high-frequency strategies and more than 1,000 feature indicators in low-frequency strategies. The decision accuracy is improved by 30% compared to traditional methods, and the agent collaboration efficiency is improved by 50%. The system can be expanded to 128 parallel agents to form a complete multi-agent investment decision-making ecosystem.

[0056] The optimization module includes: parameter optimization for the investment Transformer module, interpretation optimization for the interpretability analysis module, policy optimization for the multi-agent decision-making module, and global parameter tuning for the system; the optimization module includes an automatic machine learning function, a multi-task learning function, and a meta-learning function; the automatic machine learning function is used to search for network structures and adjust parameters; the network structure search includes optimizing the number of layers, heads, and dimensions, and the parameter adjustment searches for solutions in the parameter space; the multi-task learning function is used for task management and knowledge sharing; the task management integrates various sub-tasks of the system, and the knowledge sharing establishes knowledge bridges between modules in the system for information and parameter sharing; the meta-learning function is used for learning initialization and rapid adaptation; the learning initialization provides a learning starting point for the context learning enhancement function, and the rapid adaptation is an adaptive adjustment of the causal analysis function.

[0057] like Figure 4 As shown:

[0058] OptimSystem(θ)=arg min_θL_total(θ)

[0059] Where: L_total(θ)=λ1·L_pred+λ2·L_explain+λ3·L_decision+λ4·L_reg, constraint condition: ∑λ i =1,λ i ≥0;

[0060] L_pred is the prediction loss, L_explain is the explanation loss, L_decision is the decision loss, and L_reg is the regularization term.

[0061] The system includes three core optimization components:

[0062] AutoML(θ) is the automatic machine learning function: AutoML(θ) = α_OPT1·Structure(θ) + α_OPT2·Param(θ), with the constraints: α_OPT1 + α_OPT2 = 1, α i ≥0; Structure(θ)∈R^256 optimizes the structure, automatically searches for the optimal network structure, and optimizes hyperparameters such as the number of layers, heads, and dimensions for InvestFormer's TF(x); Param(θ)∈R^384 adjusts the parameters, searching for the optimal solution in a 30 million parameter space, supporting the Attribute(d) of the interpretability analysis module and the S(a) of the multi-agent collaborative decision-making module. i The training process involves α_OPT1 and α_OPT2, whose performance is dynamically adjusted through cross-validation. The validation set size is 20% of the training set, and the validation cycle is every 1000 iterations. The structure search space contains 10... 5 The possible combinations are analyzed, and the parameter optimization uses a second-order optimization algorithm, which improves the convergence speed by 3 times compared with the first-order method.

[0063] MultiTask(θ) is a multi-task learning function: MultiTask(θ) = β_OPT1·Task(θ) + β_OPT2·Share(θ), with constraints: β_OPT1 + β_OPT2 = 1, β i ≥0; Task(θ)∈R^256 manages tasks, integrates the core tasks of the system, including 10 sub-tasks such as feature extraction, interpretation generation, and decision execution; Share(θ)∈R^384 shares knowledge, establishes knowledge bridges between modules of the system, and promotes information and parameter sharing between modules; β_OPT1 and β_OPT2 are dynamically allocated based on task importance and difficulty, and adopt an adaptive weighting mechanism with task weights ranging from [0.01 to 10]; supports parallel processing of 16 tasks, with a knowledge transfer efficiency of 80% between tasks and a 50% reduction in parameter redundancy.

[0064] MetaLearn(θ) is the meta-learning function: MetaLearn(θ) = γ_OPT1·Init(θ) + γ_OPT2·Adapt(θ), with the constraints: γ_OPT1 + γ_OPT2 = 1, γ i≥0; where Init(θ)∈R^256 is the learning initialization, providing a good starting point for β_ICL1 and β_ICL2 in InvestFormer's ICL(x); Adapt(θ)∈R^384 is the fast adaptation, supporting γ_EXP1 and γ_EXP2 in Causal(d) of the interpretability analysis module to quickly adapt to market changes; γ_OPT1 and γ_OPT2 are dynamically balanced through the meta-validation set, and the meta-learning rate is adaptively adjusted in the range of [0.001, 0.1]; the adaptation speed is 5 times faster than traditional methods, and only 10 samples are needed to adapt to the new market environment.

[0065] This embodiment establishes a system optimization mechanism to improve overall performance through automatic machine learning, multi-task optimization, and meta-learning. This optimization framework automatically updates parameters daily, supports incremental and online learning, achieves a parallelism of 32, maintains GPU utilization above 95%, and improves overall model performance by 40% and parameter efficiency by 65% ​​compared to the baseline, providing a driving force for continuous iterative optimization of the system.

[0066] The knowledge base module includes: a knowledge graph function, a knowledge update function, and a knowledge query function; the knowledge graph function stores financial entities and representational relationships; the financial entities provide template materials for the context learning enhancement function, and the representational relationships define the relationship types between financial entities, constructing a knowledge network for the causal analysis function; the knowledge update function acquires new knowledge and verifies knowledge; the new knowledge comes from the decision-making experience of the multi-agent decision-making module and the learning results of the case management module, and the verified knowledge verifies the authenticity and timeliness of the knowledge through the interpretation results of the feature attribution function; the knowledge query function is used for precise and fuzzy queries; the precise query provides index queries and extracts deterministic rules from the decision generation function, and the fuzzy query is based on semantic similarity matching, providing relevant templates for the investment template vector. It provides Template(x) for InvestFormer's ICL(x), causal relationship knowledge for the Causal(d) of the interpretability analysis module, and Rule(s) for the multi-agent collaborative decision-making module. i It provides rule constraints and also receives experience summaries from the Report(P) module of the performance evaluation module:

[0067] KnowSystem(K)={Graph(K), Update(K), Query(K)}

[0068] Where Graph(K) is the knowledge graph function: Graph(K) = α_KB1·Entity(K) + α_KB2·Relation(K), with the constraints: α_KB1 + α_KB2 = 1, αi ≥0; Entity(K)∈R^512 stores entities, including 100,000 financial entities such as stocks, industries, and indices, providing template materials for InvestFormer's ICL(x); Relation(K)∈R^384 represents relations, defining 50 types of relations between entities, supporting Causal(d) causal analysis in the interpretability analysis module, and constructing a complete knowledge network; α_KB1 and α_KB2 are determined by MultiTask(θ) optimization in the optimization module, dynamically balanced according to the importance of entities and relations; the graph density reaches an average of 15 relations per entity, the query latency is <5ms, and it supports multi-hop inference with a maximum depth of 8.

[0069] Update(K) is the knowledge update function: Update(K) = β_KB1·New(K) + β_KB2·Verify(K), with the constraints: β_KB1 + β_KB2 = 1, β i ≥0; New(K)∈R^256 acquires new knowledge from the decision-making experience of the MASystem(A) module of the multi-agent collaborative decision-making module and the learning results of the Learn(C) module of the case management module; Verify(K)∈R^128 verifies knowledge by using the interpretation results of the Attribute(d) module of the interpretability analysis module to verify the authenticity and timeliness of knowledge; β_KB1 and β_KB2 are dynamically adjusted within the range of [0,1] based on the knowledge reliability and novelty scores to ensure the quality of knowledge; the knowledge update cycle is a full update once a day, an incremental update every hour, and an emergency knowledge update in real time to keep the knowledge base up-to-date.

[0070] Query(K) is the knowledge query function: Query(K) = γ_KB1·Exact(K) + γ_KB2·Fuzzy(K), with constraints: γ_KB1 + γ_KB2 = 1, γ i ≥0; Exact(K)∈R^128 exact query, providing O(1) complexity index query, supporting Rule(s) of multi-agent collaborative decision-making module iThe system rapidly extracts deterministic rules; Fuzzy(K)∈R^256 fuzzy queries, based on semantic similarity matching, provide relevant template recommendations for InvestFormer's Template(x); γ_KB1 and γ_KB2 dynamically balance query precision requirements [0.5, 1.0], finding the optimal point between speed and relevance; the query engine supports 1000 concurrent QPS, with an average response time of <10ms, a precision rate of 95%, and a recall rate of 90%. This embodiment constructs a dynamic knowledge management system that supports the storage, use, and updating of professional financial knowledge. The knowledge base has a capacity of 100GB, covering 15 major categories in the financial field, containing 3 million structured knowledge entries and 10 million relationships, supporting real-time incremental updates and version control, with a knowledge timeliness verification accuracy rate of 95%, providing reliable professional knowledge support for the system and becoming the cornerstone of decision-making wisdom.

[0071] The system also includes a case management module: This module provides example materials for the context learning enhancement function, historical memory for the state representation function, and experiential references for the strategy optimization module, and feeds back the learning results to the knowledge base module. The case management module includes a case storage function, an experiential learning function, and a case retrieval function. The case storage function stores successful cases and records failed cases. Storing successful cases provides successful examples for the context learning enhancement function, while recording failed cases is used to receive risk aversion data from market observations. The experiential learning function is used to mine commonalities and summarize rules. Mining commonalities includes mining common features and patterns to input into the interpretability analysis module, and summarizing rules is used to refine patterns into executable rules and update the knowledge base module. The case retrieval function performs similarity searches and insight extraction. The similarity search, based on the current market feature vector, finds the similar cases with the highest historical rewards in the multi-agent decision-making module. The insight extraction extracts key insights from the retrieved cases, providing strategy generation references for the strategy optimization module. A case management framework is constructed to provide sample materials for Demo(x) in InvestFormer's ICL(x) and for Hist(a) in the multi-agent collaborative decision-making module. i It provides historical memory, offering experience references for the Generate(S) module of the strategy optimization module, accepting decision results from each module to update cases, and simultaneously feeding back learning results to the Update(K) module of the knowledge base:

[0072] CaseSystem(C)={Store(C), Learn(C), Retrieve(C)}

[0073] Where Store(C) is the case storage function: Store(C) = α_CASE1·Success(C) + α_CASE2·Failure(C), with the constraint: α_CASE1 + α_CASE2 = 1, α i ≥0; Success(C)∈R^512 stores successful cases, including information such as rate of return, holding period, and key decision points, providing an example for Demo(x) in InvestFormer's ICL(x); Failure(C)∈R^384 records failed cases, including reasons for loss, risk signals, and error analysis, used for the S(a) of the multi-agent collaborative decision-making module. i Risk avoidance in ); α_CASE1 and α_CASE2 are dynamically adjusted through the AutoML(θ) evaluation of the optimization module, and their weights are determined based on the representativeness and teaching value of the cases; the storage capacity supports 1 million cases, adopts a hierarchical storage architecture, and popular cases are cached in memory with an access latency of <1ms. Learn(C) is the experience learning function: Learn(C)=β_CASE1·Pattern(C)+β_CASE2·Rule(C), with the constraints: β_CASE1+β_CASE2=1, β i ≥0; Pattern(C)∈R^256 extracts patterns, mines common features and rules from the case set, and inputs them into the Track(d) interpretation system of the interpretability analysis module; Rule(C)∈R^128 summarizes rules, refines patterns into executable rules, updates the Graph(K) knowledge base of the knowledge base module, and enriches entity relationships; β_CASE1 and β_CASE2 are dynamically determined based on the pattern reliability range [0.7, 1.0] to ensure learning quality; supports 10 pattern mining algorithms, with a rule summarization accuracy of 92% and rule redundancy controlled within 10%.

[0074] Retrieve(C) is the case retrieval function:

[0075] Retrieve(C)=γ_CASE1·Similar(C)+γ_CASE2·Insight(C), constraints: γ_CASE1+γ_CASE2=1, γ i ≥0; Similar(C)∈R^256 Similarity retrieval, based on the current market feature vector, finds the R(a) of the multi-agent collaborative decision-making module. iThe highest-rewarded similar cases in history; Insight(C)∈R^128 insight extraction, extracting key insights from the retrieved cases, for reference in the Base(S) strategy generation of the strategy optimization module; γ_CASE1, γ_CASE2 are dynamically adjusted through market similarity [0,1], reducing similarity weight and increasing insight weight when market differences are large; the retrieval algorithm adopts approximate nearest neighbor search, supporting second-level retrieval of a case library of hundreds of millions, with a Top-10 accuracy of 96%. This embodiment realizes the systematic management and learning application of investment experience, builds a distributed case library with a capacity of 10TB, supports incremental updates and retrieval, covers 15 major asset classes, contains more than 500,000 successful cases and 300,000 failed cases, improves experience learning efficiency by 60% compared with traditional methods, can process 10,000 new cases per day, provides the system with rich practical experience reference, and ensures that decisions are based on historical experience and lessons.

[0076] The system also includes a strategy optimization module: based on the feature representation, the strategy optimization module uses the causal analysis of the interpretability analysis module to guide the decision execution of the multi-agent decision-making module, and feeds back the evaluation feedback results to the risk warning module. This embodiment constructs a strategy optimization framework based on the InvestFormer(x) feature representation, using the Causal(d) causal analysis of the interpretability analysis module to guide the A(s) of the multi-agent collaborative decision-making module. i The decision is executed, and the evaluation results are fed back to the Alert(R) module of the risk warning module:

[0077] StrategySystem(S)={Generate(S),Evaluate(S),Adapt(S)}

[0078] Where Generate(S) is the policy generation function: Generate(S) = α_STRAT1·Base(S) + α_STRAT2·Adapt(S), with the constraint: α_STRAT1 + α_STRAT2 = 1, α i ≥0; Base(S)∈R^256 generates the basic strategy, using InvestFormer's TF(x) feature representation, creating core strategy components including asset allocation, market timing, and stock selection; Adapt(S)∈R^384 adjusts the adaptive strategy, dynamically adjusting strategy parameters and weights based on the Causal(d) causal analysis results of the interpretability analysis module; α_STRAT1 and α_STRAT2 are dynamically adjusted through the MetaLearn(θ) optimizer of the optimization module, and weights are allocated according to market stability [0.1, 1.0]; strategy generation speed <50ms, supports simultaneous maintenance of 128 alternative strategies, covering 9 common market environments.

[0079] Evaluate(S) is the policy evaluation function: Evaluate(S) = β_STRAT1·Perform(S) + β_STRAT2·Risk(S), with the constraints: β_STRAT1 + β_STRAT2 = 1, β i Performance is evaluated using ≥0; Perform(S)∈R^128, utilizing the R(a) of the multi-agent collaborative decision-making module. i The reward signal and performance evaluation module's Return(P) indicator comprehensively evaluates strategy performance; Risk(S)∈R^256 assesses risk and connects to the Risk Warning module's Monitor(R) risk control indicator to quantify strategy risk in multiple dimensions; β_STRAT1 and β_STRAT2 are dynamically adjusted based on market environment and investor risk preferences [0.2, 5.0] to balance returns and risks in different market stages; the evaluation dimensions include 12 key indicators such as absolute return, relative return, maximum drawdown, and Sharpe ratio, and the evaluation period supports multiple frequencies such as daily, weekly, and monthly.

[0080] Adapt(S) is the policy adaptation function: Adapt(S) = γ_STRAT1·Online(S) + γ_STRAT2·Robust(S), with constraints: γ_STRAT1 + γ_STRAT2 = 1, γ i ≥0; Online(S)∈R^256 is updated online, based on the latest state S(a) of the multi-agent collaborative decision-making module. i The strategy parameters are adjusted in real time; Robust(S)∈R^128 robustness enhancement is achieved by using Failure(C) failure cases in the case management module to enhance the stability of the strategy in extreme markets; γ_STRAT1 and γ_STRAT2 are dynamically balanced according to the market change rate [0.01, 10], and the weights are updated online for rapidly changing markets; the adaptation frequency supports three modes: high frequency (millisecond level), medium frequency (minute level), and low frequency (daily level), with an adaptation accuracy of 94%.

[0081] This embodiment establishes a strategy optimization system to achieve continuous evolution of investment strategies. The system supports 50 basic strategy components, which can be combined to form 10... 6 The system combines various strategies, with the fastest strategy optimization cycle being completed in 10 minutes. The strategy iteration efficiency is 75% higher than traditional methods, with an alpha increase of 42% in the backtesting environment and a win rate increase of 25% in the live trading environment. This provides the system with continuously optimized decision-making strategies, ensuring that investment decisions always remain market-adaptable and competitive.

[0082] The system also includes a market analysis module: this module processes market data from the Investment Transformer module, utilizes the feature importance of the interpretability analysis module to provide market observations for the multi-agent decision-making module, and inputs the prediction results into the risk warning module. This embodiment constructs a market analysis framework, processes Market(x) data in the Domain(x) of InvestFormer, and utilizes the Shapley(f) property of the interpretability analysis module. i The importance of the feature is the Obs(a) of the multi-agent collaborative decision-making module. i It provides market observations and inputs the forecast results into the Monitor(R) module of the risk warning module:

[0083] MarketSystem(M)={Analysis(M), Predict(M), Scenario(M)}

[0084] Where Analysis(M) is the market analysis function: Analysis(M) = α_MKT1·Macro(M) + α_MKT2·Micro(M), with the constraints: α_MKT1 + α_MKT2 = 1, α i ≥0; Macro(M)∈R^512 analyzes the macroeconomic aspect, processing macroeconomic data from InvestFormer's Market(x), including indicators such as GDP, interest rates, and inflation; Micro(M)∈R^384 analyzes the microeconomic aspect, combining enterprise and industry data from the Graph(K) knowledge graph in the knowledge base module to study the microeconomic structure; α_MKT1 and α_MKT2 are dynamically adjusted according to the importance of the market stage [0.1, 10]. Systematic market conditions enhance the weight of macroeconomic analysis, while structural market conditions enhance the weight of microeconomic analysis; The analysis dimensions include four major categories: fundamentals, technicals, capital flows, and sentiment, totaling 56 key indicators, with an analysis accuracy of 92%.

[0085] Predict(M) is the prediction function: Predict(M) = β_MKT1·Trend(M) + β_MKT2·Risk(M), with the constraints: β_MKT1 + β_MKT2 = 1, β i≥0; Trend(M)∈R^256 predicts trends, based on the State(A) representation of the multi-agent collaborative decision-making module, and applies time series analysis and machine learning models to predict future trends; Risk(M)∈R^128 assesses risks, identifies potential market risk factors, and inputs them into the Alert(R) warning system of the risk warning module to provide risk prediction; β_MKT1 and β_MKT2 are learned through the AutoML(θ) optimizer of the optimization module and dynamically adjusted according to the prediction accuracy [0.5, 1.0] and the urgency of risk warning [0, 5]; the prediction period covers three time scales: short-term (1-5 days), medium-term (1-4 weeks), and long-term (1-12 months), with prediction accuracies of 85%, 76%, and 68%, respectively.

[0086] Scenario(M) is the scenario analysis function: Scenario(M) = γ_MKT1·Base(M) + γ_MKT2·Stress(M) + γ_MKT3·Extreme(M), with constraints: γ_MKT1 + γ_MKT2 + γ_MKT3 = 1, γ i ≥0; Base(M)∈R^256 analyzes the baseline scenario, constructs the normal development path of the market, and provides a reference for Base(S) of the strategy optimization module; Stress(M)∈R^192 tests the stress scenario, simulates market stress test, and provides a risk assessment benchmark for Risk(R) of the risk warning module; Extreme(M)∈R^128 analyzes the extreme scenario, evaluates the market performance under extreme events, and strengthens the robustness of the Robust(S) strategy of the strategy optimization module; γ_MKT1, γ_MKT2, and γ_MKT3 are dynamically adjusted according to the market volatility level [0.5, 3.0], and the weight of stress and extreme scenarios is increased during periods of high volatility; the scenario analysis supports 10,000 Monte Carlo simulations, covering a 95% confidence interval, with 32 types of scenarios, including historical crisis recurrence and new risk simulations.

[0087] This embodiment constructs a market analysis system, providing comprehensive market insight support. The system can process 1 million market data points per second, analyzing 42 major global markets, 500 industry categories, and 10,000 trading instruments. The macroeconomic indicator library contains 2,000 economic indicators, improving prediction accuracy by 35% compared to traditional methods, and achieving a market anomaly detection accuracy rate of 93%. This provides a solid market research foundation for investment decisions, ensuring that decisions are based on a deep understanding of the market.

[0088] The system also includes a risk warning module and a performance evaluation module: the risk warning module is used to monitor risk data in the Investment Transformer module, track risk sources using the interpretability analysis module, analyze the decision risk of the multi-agent decision-making module, and receive the prediction results from the market analysis module; this embodiment constructs a risk warning framework, monitors Risk(x) data in the Domain(x) of InvestFormer, tracks risk sources using Track(d) of the interpretability analysis module, analyzes the decision risk of the MASystem(A) of the multi-agent collaborative decision-making module, and simultaneously receives the prediction results of Risk(M) from the market analysis module.

[0089] RiskSystem(R)={Monitor(R),Alert(R),Control(R)}

[0090] Where Monitor(R) is the risk monitoring function: Monitor(R) = α_RISK1·Market(R) + α_RISK2·Position(R) + α_RISK3·Liquidity(R), with the constraints: α_RISK1 + α_RISK2 + α_RISK3 = 1, α i ≥0; Market(R)∈R^256 monitors market risk, using the Analysis(M) analysis results of the market analysis module to track 16 risk indicators such as volatility and correlation anomalies in real time; Position(R)∈R^384 monitors position risk, based on the Action(A) decision status of the multi-agent collaborative decision-making module, assessing 12 position risks such as concentration, leverage, and duration; Liquidity(R)∈R^192 monitors liquidity risk, analyzing 10 liquidity indicators such as trading depth, slippage, and bid-ask spread to ensure the feasibility of trading; α_RISK1, α_RISK2, and α_RISK3 are dynamically optimized by the MultiTask(θ) of the optimization module, allocating monitoring resources according to the importance of risk type [0.1, 10]; the monitoring frequency supports three modes: millisecond-level high-frequency monitoring, second-level regular monitoring, and minute-level deep monitoring, with a coverage of 99.9%.

[0091] Alert(R) is the warning function: Alert(R) = β_RISK1·Early(R) + β_RISK2·Action(R) + β_RISK3·Escalate(R), with the constraints: β_RISK1 + β_RISK2 + β_RISK3 = 1, β i≥0; where Early(R)∈R^128 generates early warnings, based on 17 early warning models, combined with the explanation results of Explanation(d) in the interpretability analysis module, providing risk warnings and explanations; Action(R)∈R^256 provides suggestions, generates risk response plans, guides the strategy(S) strategy adjustment in the strategy optimization module, and mitigates or avoids risks; Escalate(R)∈R^128 risk escalation, according to the severity of the risk [1-5 levels], initiates different levels of emergency response, and triggers automatic risk control measures when necessary; β_RISK1, β_RISK2, β_RISK3 are dynamically determined based on the urgency of the risk [0,1], with emergency risks increasing the weight of actions and escalation; early warning delay <10ms, supports 24-hour uninterrupted monitoring, early warning accuracy reaches 96%, and false alarm rate is controlled below 2%.

[0092] Control(R) is the risk control execution function: Control(R) = γ_RISK1·Limit(R) + γ_RISK2·Hedge(R) + γ_RISK3·Stop(R), with constraints: γ_RISK1 + γ_RISK2 + γ_RISK3 = 1, γ i ≥0; where Limit(R)∈R^192 sets the limit, based on the Rule(s) of the knowledge base module. i The system employs a rule base to implement protective measures such as risk limits and transaction thresholds; Hedge(R)∈R^128 executes hedging, automatically generating hedging strategies to balance portfolio risk exposure when risk exceeds the threshold; Stop(R)∈R^64 executes stop-loss, triggering an automatic stop-loss mechanism in extreme cases to cut off further losses and protect capital; γ_RISK1, γ_RISK2, and γ_RISK3 dynamically balance risk control strategy effectiveness [0.5, 1.0] and market liquidity [0.1, 1.0] to ensure accurate and effective risk control execution; control latency <5ms, risk control accuracy reaches 99.5%, reliability reaches 99.99%, and it operates 24 / 7.

[0093] This embodiment implements a comprehensive risk management mechanism to ensure investment security. The system supports multi-layered risk protection, including 128 risk monitoring indicators, 64 early warning types, and 32 risk control measures. The average lead time for risk identification is 2.5 trading days, and the accuracy rate of extreme event early warning reaches 87%. The overall risk control effect is 60% better than traditional methods, effectively controlling the maximum drawdown within a preset range. It provides comprehensive risk protection for investment decisions, ensuring that the system strictly controls risks while pursuing returns.

[0094] The performance evaluation module provides valuation references based on the investment Transformer module, uses the interpretability analysis module to explain the reasons for performance, analyzes the revenue data of the multi-agent decision-making module, and feeds back the evaluation results to the knowledge base module, case management module, and strategy optimization module.

[0095] This embodiment constructs a performance evaluation framework, providing a valuation reference based on InvestFormer's Value(x), using the Explanation(d) of the Explainability Analysis module to explain the reasons for performance, analyzing the Reward(A) return data of the multi-agent collaborative decision-making module, and feeding back the evaluation results to the knowledge base module, case management module, and strategy optimization module:

[0096] PerformSystem(P)={Evaluate(P), Report(P), Optimize(P)}

[0097] Where Evaluate(P) is the evaluation function: Evaluate(P) = α_PERF1·Return(P) + α_PERF2·Risk(P) + α_PERF3·Benchmark(P), with constraints: α_PERF1 + α_PERF2 + α_PERF3 = 1, α i ≥0; Return(P)∈R^256 calculates the profit, based on the multi-agent collaborative decision-making module's R(a i The system includes a reward signal, calculating 15 indicators such as absolute return, relative return, and risk-adjusted return; Risk(P)∈R^384 assesses risk using the Monitor(R) risk indicator from the risk warning module, quantifying 12 risk measures such as volatility, maximum drawdown, and VaR; Benchmark(P)∈R^192 performs benchmarking analysis, comparing with the market, peers, and custom benchmarks to assess excess returns and risk exposure; α_PERF1, α_PERF2, and α_PERF3 are determined through multi-objective optimization using the MultiTask(θ) function of the optimization module, dynamically balancing return and risk dimensions according to the assessment objective; the assessment accuracy reaches 99.9%, supporting six timeframes: real-time, daily, weekly, monthly, quarterly, and yearly, covering 25 core performance indicators.

[0098] Report(P) is the report generation function: Report(P) = β_PERF1·Analysis(P) + β_PERF2·Suggest(P) + β_PERF3·Visual(P), with constraints: β_PERF1 + β_PERF2 + β_PERF3 = 1, β i≥0; Analysis(P)∈R^256 analyzes performance, combining the interpretation results of Attribute(d) in the interpretability analysis module to deeply analyze the causes and influencing factors of performance; Suggest(P)∈R^128 provides suggestions to guide the strategy optimization of Adapt(S) in the strategy optimization module and proposes specific improvement directions; Visual(P)∈R^192 provides visualization, generating intuitive charts, including 18 visualization components such as profit curves, attribution analysis, and risk distribution; β_PERF1, β_PERF2, and β_PERF3 are dynamically adjusted based on the purpose and importance of the report, and the depth and professionalism are adjusted according to the audience type [technical, management, customer]; the report generation speed is <1 second, supports multiple output formats such as PDF, HTML, and API, and the automation level reaches 98%. Optimize(P) is the optimization suggestion function: Optimize(P) = γ_PERF1·Strategy(P) + γ_PERF2·Parameter(P) + γ_PERF3·Resource(P), with constraints: γ_PERF1 + γ_PERF2 + γ_PERF3 = 1, γ i ≥0; where, Strategy(P)∈R^256 strategy optimization, proposes strategy-level improvement suggestions for Strategy(S) of the strategy optimization module, including asset allocation, timing strategy, etc.; Parameter(P)∈R^192 parameter tuning, provides parameter optimization direction for OptimSystem(θ) of the optimization module, and finely adjusts model parameters; Resource(P)∈R^128 resource allocation, guides the optimal configuration of computing resources, data resources and model complexity, and improves system efficiency; γ_PERF1, γ_PERF2, γ_PERF3 are dynamically balanced according to optimization urgency [0.1, 1.0] and expected effect [0.5, 2.0], and concentrate resources on high-return optimization points; the accuracy rate of optimization suggestions reaches 90%, and the average performance improvement is 12% and the risk reduction is 15% after implementation.

[0099] This embodiment establishes a complete performance evaluation system, forming a closed-loop optimization of investment decisions. The system automatically generates a comprehensive evaluation report daily, supporting multi-dimensional, multi-period, and multi-benchmark comprehensive evaluation. Performance attribution is accurate to three levels: asset class, industry, and individual bond, with interpretability reaching 95%. Evaluation results are directly fed back to each module to form a closed-loop optimization, continuously improving the overall system performance. The performance evaluation module forms an organic whole with the preceding nine modules, constructing a complete intelligent investment closed loop from feature extraction and decision generation to risk control and performance evaluation. This enables the system to continuously learn, evolve, and optimize, achieving a spiral increase in investment decision-making capabilities.

[0100] It is worth mentioning that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problem proposed in this application; however, this does not mean that other units are absent from this embodiment.

[0101] This application relates to a multi-agent investment decision matrix method for large model evolution, such as... Figure 5 As shown, the method includes:

[0102] S101: Acquire financial data and perform feature extraction and fusion by combining contextual learning and domain knowledge enhancement functions;

[0103] S102, perform relation mapping and attribution analysis based on the extracted features to generate decision explanations;

[0104] S103, combining feature representation and decision interpretation, generates investment decisions through multi-agent interaction;

[0105] S104 optimizes parameters and strategies for the feature extraction, interpretation analysis, and decision generation processes;

[0106] S105, build and update the knowledge base to provide prior knowledge and rule constraints for feature extraction, interpretation analysis and decision generation.

[0107] The steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this application. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but without changing the core design of the algorithm and process, are also within the scope of protection of this application.

[0108] It is not difficult to see that this embodiment is a method embodiment corresponding to the system embodiment, and this embodiment can be implemented in conjunction with the system embodiment. The relevant technical details mentioned in the system embodiment are still valid in this embodiment, and will not be repeated here to reduce repetition. Correspondingly, the relevant technical details mentioned in this embodiment can also be applied to the system embodiment.

[0109] The flowcharts or block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of devices, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-specific system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0110] The scope of this application is defined by the appended claims rather than the foregoing description, and is therefore intended to encompass all variations falling within the meaning and scope of equivalents of the claims. No reference numerals in the claims should be construed as limiting the scope of the claims. Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a device claim may also be implemented by a single unit or device in software or hardware. Terms such as "first," "second," etc., are used only for distinguishing descriptions and do not indicate any particular order, nor should they be construed as indicating or implying relative importance.

[0111] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily made by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims, and the above embodiments should be regarded as exemplary and non-limiting.

Claims

1. A multi-agent investment decision matrix system with large-scale model evolution, characterized in that, The system includes: The Transformer module is used to acquire financial data and perform feature extraction and feature fusion on the financial data by combining context learning enhancement functions and financial domain knowledge enhancement functions. The interpretability analysis module is used to obtain the feature representations in the investment Transformer module, perform feature relationship mapping and interpretability analysis; A multi-agent decision-making module is used to obtain the feature representations in the investment Transformer module and the decision interpretations in the interpretability analysis module, and to make collaborative decisions through multiple agents; An optimization module is used to provide parameter optimization for the investment Transformer module, interpretation optimization for the interpretability analysis module, and strategy optimization for the multi-agent decision-making module; The knowledge base module is used to provide investment templates for the investment Transformer module, causal relationship knowledge for the interpretability analysis module, and rule constraints for the multi-agent decision-making module.

2. The system according to claim 1, characterized in that, The investment Transformer module includes: The Investment Transformer module provides the interpretability analysis module with quantified data on interpretable objects and feature attention metrics; the intermediate features generated by the Investment Transformer module are processed by the multi-agent decision-making module to enhance decision-making information. The context learning enhancement function obtains investment template vectors from the knowledge base module, example vectors from the case library, and context vectors from the financial data; The financial domain knowledge enhancement function provides market characteristics to the market analysis module, risk characteristics to the risk warning module, and value characteristics to the performance evaluation module.

3. The system according to claim 2, characterized in that, The interpretability analysis module includes: The interpretability analysis module provides decision explanations for the multi-agent decision-making module, causal analysis for the strategy optimization module, and transparent explanations for the risk warning module, while feeding back the explanation results to the knowledge base module. The interpretability analysis module includes a feature attribution function, a decision tracking function, and a causal analysis function; The feature attribution function is used to quantify the contribution of each feature to the decision, calculate the feature importance and construct a decision path graph, and provide the feature importance ranking to the market analysis module; The decision tracking function is used to track feature propagation paths and analyze the impact of decision feedback, connect the interaction records of the multi-agent decision-making module, and perform decision consequence backtracking; the tracking results are input into the risk warning module for risk source tracing; The causal analysis function is used to calculate first-order causal effects based on feature relationships; track the interactions of the multi-agent decision-making module to obtain higher-order causal chains; and input the causal analysis results into the policy optimization module to guide the policy adjustment direction.

4. The system according to claim 3, characterized in that, The multi-agent decision-making module includes: Based on the feature representation and the decision interpretation, the multi-agent decision-making module provides the execution subject for the strategy optimization module, reports the status to the early warning risk module, and feeds back the decision experience to the strategy library. The multi-agent decision-making module includes a state representation function, a decision generation function, and a reward calculation function; The state representation function is used to receive market observations, store historical records, and provide decision-making context; the market observations come from the investment Transformer module, the historical records come from the case library and establish time-series memory, and the decision-making context comes from the interpretability analysis module; The decision generation function is used to generate investment decisions, enforce rule constraints, and evaluate feedback dynamic adjustments; the investment decisions are based on the features of the investment Transformer module, the rule constraints are derived from the rule constraints of the knowledge base, and the feedback dynamic adjustments are based on the market analysis module; The reward calculation function is used to calculate returns and assess risks; the calculated returns are compared with the valuation of the performance evaluation module, and the assessed risks are evaluated by the risk warning module.

5. The system according to claim 4, characterized in that, The optimization module includes: The optimization module is used to provide parameter optimization for the investment Transformer module, interpretation optimization for the interpretability analysis module, strategy optimization for the multi-agent decision-making module, and global parameter tuning for the system. The optimization module includes automatic machine learning functions, multi-task learning functions, and meta-learning functions; The automatic machine learning function is used to search for network structure and adjust parameters; the network structure search includes optimizing the number of layers, the number of heads, and the dimension, and the parameter adjustment seeks solutions in the parameter space. The multi-task learning function is used to manage tasks and share knowledge; the management task integrates the various sub-tasks of the system, and the shared knowledge establishes knowledge bridges between the various modules of the system to share information and parameters. The meta-learning function is used for learning initialization and rapid adaptation; the learning initialization provides a learning starting point for the context learning enhancement function, and the rapid adaptation is an adaptive adjustment of the causal analysis function.

6. The system according to claim 5, characterized in that, The knowledge base module includes: The knowledge base module includes knowledge graph functions, knowledge update functions, and knowledge query functions; The knowledge graph function is used to store financial entities and representation relationships; the financial entities provide template materials for the context learning enhancement function, and the representation relationships define the relationship types between financial entities, thus constructing a knowledge network for the causal analysis function. The knowledge update function is used to acquire new knowledge and verify knowledge; the new knowledge comes from the decision-making experience of the multi-agent decision-making module and the learning results of the case management module, and the knowledge verification verifies the authenticity and timeliness of the knowledge through the interpretation results of the feature attribution function. The knowledge query function is used for precise and fuzzy queries; the precise query provides indexed queries to extract deterministic rules from the decision generation function, and the fuzzy query provides relevant templates for the investment template vector based on semantic similarity matching.

7. The system according to claim 6, characterized in that, The system also includes a case management module: The case management module is used to provide example materials for the context learning enhancement function, provide historical memory for the state representation function, provide experience reference for the policy optimization module, and feed the learning results back to the knowledge base module. The case management module includes case storage functions, experience learning functions, and case retrieval functions; The case storage function is used to store successful cases and record failed cases; the stored successful cases provide successful instances for the context learning enhancement function, and the recorded failed cases are used to receive risk aversion in market observations. The experience learning function is used to discover commonalities and inductive rules; the discovery of commonalities includes discovering common features and patterns and inputting them into the interpretability analysis module; the inductive rules are used to extract patterns into executable rules and update the knowledge base module. The case retrieval function is used for similarity retrieval and insight extraction; the similarity retrieval is based on the current market feature vector to find similar cases with the highest historical rewards in the multi-agent decision-making module; the insight extraction is used to extract key insights from the retrieved cases to provide a reference for strategy generation for the strategy optimization module.

8. The system according to claim 7, characterized in that, The system also includes a strategy optimization module: The strategy optimization module, based on the feature representation and utilizing the causal analysis of the interpretability analysis module, guides the decision execution of the multi-agent decision-making module and feeds back the evaluation feedback results to the risk warning module.

9. The system according to claim 8, characterized in that, The system also includes a market analysis module: The market analysis module processes market data from the investment Transformer module, utilizes the feature importance of the interpretability analysis module to provide market observations for the multi-agent decision-making module, and inputs the prediction results into the risk warning module.

10. The system according to claim 9, characterized in that, The system also includes a risk warning module and a performance evaluation module: The risk warning module is used to monitor risk data in the investment Transformer module, use the interpretability analysis module to track risk sources, analyze the decision risk of the multi-agent decision-making module, and receive the prediction results of the market analysis module. The performance evaluation module provides valuation references based on the investment Transformer module, uses the interpretability analysis module to explain the reasons for performance, analyzes the revenue data of the multi-agent decision-making module, and feeds back the evaluation results to the knowledge base module, case management module, and strategy optimization module.

Images