Method and system for generating an asset allocation strategy for a financial institution

Abstract

The application provides an asset allocation strategy generation method and system of a financial institution, which can be used in the financial field or other fields. The method comprises the following steps: obtaining a state matrix, wherein the state matrix is constructed by multi-dimensional data, and the multi-dimensional data comprises a market fluctuation index, asset historical returns and a customer portrait vector of a customer; obtaining respective corresponding asset weight adjustment vectors according to the state matrix and a plurality of single-agent strategy networks; optimizing network parameters of each single-agent strategy network according to a multi-objective reward mechanism and the asset weight adjustment vectors of each single-agent strategy network to obtain a trained multi-agent model, wherein the trained multi-agent model comprises a plurality of trained single-agent strategy networks; and applying the trained multi-agent model to generate an asset allocation strategy of a target customer in the financial institution. The application can generate an asset allocation strategy in combination with market trends and customer investment preferences, improve the accuracy of the asset allocation strategy and improve customer satisfaction.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for generating asset allocation strategies for financial institutions. Background Technology

[0002] Asset allocation strategies typically involve distributing a client's assets (such as funds) across different asset classes based on their needs and risk profile. Banks often need to recommend asset allocation strategies to their clients, making it crucial to ensure that these strategies align with the client's needs.

[0003] Traditional asset allocation relies primarily on the experience and judgment of financial advisors and static model calculations. This approach makes it difficult to dynamically identify and respond to customer needs in real time, especially when market conditions fluctuate wildly or customer investment preferences change. Traditional systems cannot quickly and adaptively adjust asset allocation strategies, resulting in lagging investment strategies, untimely risk control, and decreased customer satisfaction.

[0004] This section is intended to provide background or context for the embodiments of the invention set forth in the claims. The description herein is not an admission that it is prior art simply because it is included in this section. Summary of the Invention

[0005] To address at least one problem in the prior art, this application proposes a method and system for generating asset allocation strategies for financial institutions. This method can generate asset allocation strategies by combining market conditions and customer investment preferences, thereby improving the accuracy of asset allocation strategies and increasing customer satisfaction.

[0006] To address the aforementioned technical problems, this application provides the following technical solution: Firstly, this application provides a method for generating asset allocation strategies for financial institutions, including: Obtain a state matrix, which is constructed from multidimensional data, including: market volatility index, historical asset returns, and customer profile vectors. Based on the state matrix and multiple single-agent policy networks, the corresponding asset weight adjustment vectors are obtained. Based on the multi-objective reward mechanism and the asset weight adjustment vector of each single agent policy network, the network parameters of each network are optimized to obtain the trained multi-agent model, which includes multiple trained single agent policy networks. The trained multi-agent model is used to generate asset allocation strategies for target customers in financial institutions.

[0007] In one embodiment, obtaining the state matrix includes: Construct multiple virtual financial market environments using the Monte Carlo method; In each of the aforementioned virtual financial market environments, a state matrix is ​​obtained.

[0008] In one embodiment, obtaining the corresponding asset weight adjustment vectors based on the state matrix and multiple single-agent policy networks includes: The state matrix is ​​input into each of the single agent policy networks to obtain the action corresponding to the single agent policy network, and the action is determined as the asset weight adjustment vector.

[0009] In one embodiment, optimizing the network parameters of each single-agent policy network according to the multi-objective reward mechanism and the asset weight adjustment vector of each single-agent policy network to obtain the trained multi-agent model includes: Based on the asset weight adjustment vector of the single-agent policy network, the virtual market feedback is determined, which includes: return rate and risk value; The network parameters of the single-agent policy network are continuously optimized by using the multi-objective reward mechanism, return rate, and risk value for backpropagation training. When the standard deviation of the returns of the single agent policy network is lower than the threshold, the trained single agent policy network corresponding to the single agent policy network is obtained, and the trained single agent policy networks constitute the trained multi-agent model.

[0010] In one embodiment, applying the trained multi-agent model to generate asset allocation strategies for target customers in financial institutions includes:

[0011] Obtain the state matrix corresponding to the target customer; The state matrix corresponding to the target customer is input into the trained multi-agent model to obtain the initial asset allocation strategy output by each trained single-agent policy network. Each initial asset allocation strategy includes: the initial asset weight adjustment vector of the target customer. The target asset weight adjustment vector for the target customer is obtained by using a weighted average fusion algorithm and each of the initial asset allocation strategies. Based on the target customer's target asset weight adjustment vector, the target customer's asset allocation strategy is determined. The asset allocation strategy includes at least one of the following: expected return, risk level, and investment ratio of multiple asset classes.

[0012] In one embodiment, obtaining the state matrix includes: Obtain market volatility index, historical asset returns, and customer risk characteristics, including account balance, asset distribution, transaction history, and risk tolerance questionnaire; The customer profile vector is obtained based on the market volatility index and the risk characteristics. The state matrix is ​​constructed based on the market volatility index, historical asset returns, and customer profile vectors.

[0013] Secondly, this application provides an asset allocation strategy generation system for financial institutions, comprising: The acquisition module is used to acquire a state matrix, which is constructed from multi-dimensional data, including: market volatility index, historical asset returns, and customer profile vectors. The adjustment module is used to obtain the corresponding asset weight adjustment vectors based on the state matrix and multiple single-agent policy networks. The training module is used to optimize the network parameters of each single agent policy network according to the multi-objective reward mechanism and the asset weight adjustment vector of each single agent policy network to obtain the trained multi-agent model. The trained multi-agent model includes: multiple trained single agent policy networks. The generation module is used to apply the trained multi-agent model to generate asset allocation strategies for target customers in financial institutions.

[0014] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the asset allocation strategy generation method of the financial institution.

[0015] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the asset allocation strategy generation method for financial institutions.

[0016] Fifthly, this application provides a computer program product, which includes a computer program that, when executed by a processor, implements the asset allocation strategy generation method for financial institutions.

[0017] As can be seen from the above technical solution, this application provides a method and system for generating asset allocation strategies for financial institutions. The method includes: obtaining a state matrix, which is constructed from multi-dimensional data, including market volatility index, historical asset returns, and customer profile vectors; obtaining corresponding asset weight adjustment vectors for each single-agent strategy network based on the state matrix and multiple single-agent strategy networks; optimizing the network parameters of each single-agent strategy network based on a multi-objective reward mechanism and the asset weight adjustment vectors of each single-agent strategy network to obtain a trained multi-agent model, which includes multiple trained single-agent strategy networks; and applying the trained multi-agent model to generate asset allocation strategies for target customers in financial institutions. This method combines market conditions and customer investment preferences to generate asset allocation strategies, improving the accuracy of asset allocation strategies and increasing customer satisfaction. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a first flowchart illustrating the method for generating asset allocation strategies for financial institutions in this application embodiment; Figure 2 This is a second flowchart illustrating the method for generating asset allocation strategies for financial institutions in this application embodiment; Figure 3 This is a schematic diagram of the third process of the asset allocation strategy generation method for financial institutions in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of the asset allocation strategy generation system of a financial institution in the embodiments of this application; Figure 5 This is a schematic block diagram of the system configuration of an electronic device according to an embodiment of this application. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in 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, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0020] Asset allocation, representing the professionalism and high-end nature of the financial industry, provides specialized investment and financial services to high-end clients, enhancing the level of customized investment services for private banking clients. To address the technical issues of insufficient intelligence, inaccurate client profiling, and delayed response to portfolio adjustments in existing asset allocation services, this application proposes a method and system for generating asset allocation strategies for financial institutions. Utilizing intelligent agent technology, it establishes a professional and skillful private banking service chain, starting from the process of understanding, guiding, persuading, and reaching consensus with clients, thereby improving client satisfaction. Specifically, by introducing intelligent agent models and multi-dimensional data analysis technology, it achieves intelligent asset allocation integrating data collection, client profiling generation, strategy evolution, and dynamic optimization. It can simulate client investment behavior under different market scenarios and continuously optimize model parameters through continuous learning, thereby achieving self-evolution and personalized recommendations for asset allocation strategies. The intelligent agent model acts as the "core of dynamic decision-making," adapting in real-time to changes in client risk preferences and market trend fluctuations through information perception, strategy generation, and feedback adjustment, significantly improving the accuracy and flexibility of asset allocation strategies.

[0021] Therefore, the asset allocation strategy generation method and system for financial institutions proposed in this application can break through the reliance of traditional asset allocation systems on human experience and static algorithms from a technical perspective. It can solve the key technical problem of "how to use intelligent agent models to achieve dynamic optimization and intelligent decision-making of customer asset allocation" and provide financial institutions with a sustainable learning and adaptive optimization intelligent asset allocation technology framework.

[0022] Specifically, the asset allocation strategy generation method and system for financial institutions provided in this application can solve technically challenging problems that cannot be easily overcome by traditional algorithms or system architectures, but are achieved through innovative technical design.

[0023] First, intelligent data collection and customer profiling can address the technical challenge of accurately identifying investment characteristics and sentiment preferences in unstructured data. Traditional data collection and analysis often only perform simple statistics on structured data, failing to accurately capture users' potential investment intentions in text, voice, and behavioral data. This solution introduces Natural Language Processing (NLP) and sentiment analysis technologies to semantically fuse multi-source heterogeneous data such as customer social media behavior, historical transaction records, and risk questionnaire results. Through an adaptive semantic clustering model, a multi-dimensional profile is constructed, enabling the system to automatically identify and quantify customer risk preference characteristics, providing a precise foundation for subsequent strategy generation.

[0024] Secondly, in terms of strategy generation and dynamic optimization, existing technologies face the technical challenge of "how to achieve adaptive optimization of asset portfolios in a dynamic market environment." Traditional asset allocation algorithms are mostly based on static risk-return models, which cannot respond to market fluctuations and changes in customer behavior in real time. This solution innovatively introduces a reinforcement learning-driven agent algorithm, which can self-evolve strategies through continuous market feedback. By simulating customer investment behavior and market scenarios, the agent evaluates the risk-reward ratio of different asset weight adjustments in real time and automatically corrects the allocation strategy, realizing a truly "self-learning and self-optimizing" asset allocation mechanism.

[0025] Furthermore, the technical bottleneck of "logical association of multi-source data and synchronization of real-time calculation results" needs to be addressed during the coordination between the data retrieval and strategy execution modules. Due to the complexity and high update frequency of financial data sources, a lack of efficient logical mapping mechanisms can easily lead to data delays or strategy mismatches. This solution can achieve rapid data matching and real-time retrieval by establishing a multi-dimensional index database and a parallel computing architecture, utilizing the logical relationship between data identifiers and asset attribute fields. Simultaneously, a timestamp consistency mechanism is introduced to ensure the synchronization and accuracy of calculation results in a multi-threaded environment.

[0026] In summary, this solution innovatively addresses core technical challenges such as semantic recognition of unstructured data, dynamic optimization of asset allocation, and real-time correlation of multi-source data, forming a highly intelligent, automated, and adaptive asset allocation system. These technological breakthroughs not only improve the system's operational efficiency and decision-making accuracy but also significantly enhance its practical value and technological originality in complex financial environments.

[0027] The method and system for generating asset allocation strategies for financial institutions proposed in this application have at least the following technical effects: It enables intelligent collection, analysis, prediction, and dynamic optimization of financial data, achieving efficient, accurate, and personalized asset management. It can acquire multi-dimensional data in real time, including market conditions, macroeconomic data, industry trends, and client investment behavior. Through natural language processing and sentiment analysis, it performs in-depth analysis of unstructured data, automatically extracting core features that influence asset allocation decisions.

[0028] During the data analysis phase, the system leverages reinforcement learning and multi-factor models to achieve adaptive dynamic optimization of the investment portfolio. Through continuous market feedback and risk-reward assessment, the intelligent agent model can autonomously learn the optimal allocation strategy under different market conditions, thereby maintaining the portfolio's stability and returns even in highly volatile markets. It possesses strong self-evolution capabilities, continuously adjusting its allocation schemes in response to changes in client needs and external market conditions, ensuring the investment strategy remains optimal at all times.

[0029] By establishing a multi-dimensional logical mapping database, real-time access and synchronous updates of data can be achieved at each stage of data acquisition, processing, and decision output. Information flows within the system without human intervention, automatically transforming raw data into executable strategies, thus significantly improving computational efficiency and data consistency.

[0030] The asset allocation process can shift from "human experience-based decision-making" to "intelligent algorithm-driven" approaches. The technological benefits include: more comprehensive data collection, more precise information processing, more flexible investment strategies, higher execution efficiency, and more intelligent risk control. Through this system, financial institutions can provide customized investment solutions to clients at lower costs and higher efficiency, significantly improving the scientific rigor and return stability of asset allocation, and achieving an intelligent upgrade in the asset management field.

[0031] By generating and dynamically optimizing strategies through intelligent agents, self-learning and real-time optimization of asset allocation strategies can be achieved in complex financial market environments, reaching dynamic adaptability and predictive accuracy that traditional algorithms cannot achieve. An intelligent agent decision network can be constructed based on reinforcement learning algorithms, continuously receiving market data, client trading behavior, and risk preference feedback during system operation, and automatically updating strategy parameters. The intelligent agent simulates asset weight adjustments under different market conditions, performing multiple rounds of iterative calculations on indicators such as returns, volatility, and correlation, ultimately forming the optimal strategy matrix, thereby significantly improving the robustness and risk hedging capabilities of the investment portfolio.

[0032] By leveraging unstructured data semantic recognition, deep semantic mining of multi-source customer data (including social text, investment comments, behavioral logs, etc.) can be achieved. Utilizing natural language processing and sentiment analysis technologies, implicit risk propensities and psychological expectations of customers can be quantified into calculable model parameters. This allows for accurate customer profiling without relying on traditional questionnaires or human interviews, thereby improving the personalization and suitability of asset allocation and significantly enhancing the matching accuracy of recommendation strategies and customer satisfaction.

[0033] By designing a multi-dimensional logical mapping database and a parallel computing architecture, data from different sources and with different structures can be correlated, retrieved, and synchronously updated within milliseconds. By establishing a logical mapping relationship between data index tables and asset attribute fields, automatic correlation and real-time updates can be achieved when calling macroeconomic data, industry indices, and individual stock data, avoiding decision-making biases caused by data latency in traditional solutions. Its parallel computing mechanism can further improve the computational efficiency of this solution in large-scale financial data processing, enabling dynamic adjustments to asset portfolios to respond to market changes within milliseconds, significantly improving the real-time performance and accuracy of asset allocation.

[0034] As described above, this solution achieves groundbreaking technical results in intelligent learning, accurate matching, and high-speed decision-making through the synergy of core technologies such as intelligent agent dynamic optimization, semantic recognition, and logical mapping database, providing more efficient, intelligent, and secure technical support for financial asset allocation.

[0035] The following examples illustrate this in detail.

[0036] To improve the accuracy and customer satisfaction of asset allocation strategies by combining market conditions and customer investment preferences, this embodiment provides a method for generating asset allocation strategies for financial institutions, where the executing entity is a financial institution's asset allocation strategy generation system. This system includes, but is not limited to, a server. Figure 1 As shown, this method specifically includes the following: Step 100: Obtain the state matrix, which is constructed from multidimensional data, including: market volatility index, historical asset returns, and customer profile vectors.

[0037] Specifically, the state matrix can be constructed from multidimensional data, including market volatility index, historical asset returns, and customer profile vectors. The market volatility index measures the expected volatility of the financial market over a future period. Historical asset returns represent the rate of return generated by price changes of a specific financial asset (such as stocks, bonds, funds, gold, etc.) over a specific past period. For example, a stock return of 8%, a bond return of 4%, a commodity return of 3%, and a cash return of 0.5% could be represented by the following columns: first column: unique customer identifier; second column: customer age; third column: account balance; ...; column i: market volatility index; and columns i+1 to n: returns of various asset classes over a period of time.

[0038] Step 200: Based on the state matrix and multiple single-agent policy networks, obtain the corresponding asset weight adjustment vectors for each.

[0039] Step 300: Optimize the network parameters of each single agent policy network according to the multi-objective reward mechanism and the asset weight adjustment vector of each single agent policy network to obtain the trained multi-agent model. The trained multi-agent model includes: multiple trained single agent policy networks.

[0040] Specifically, the multi-objective reward mechanism can be represented by the application of the reward function R = α × return rate - β × risk value. α and β can be set according to the actual situation, and this application does not impose any restrictions on them.

[0041] Step 400: Apply the trained multi-agent model to generate asset allocation strategies for target customers in financial institutions.

[0042] Specifically, the financial institutions mentioned can be banks, securities companies, and insurance companies, etc.

[0043] To simulate customer investment behavior under different financial market environments and thus improve the reliability of the trained multi-agent model, such as... Figure 2 As shown, in one embodiment, step 100 includes: Step 101: Construct multiple virtual financial market environments using the Monte Carlo method.

[0044] Step 102: Obtain the state matrix in each of the virtual financial market environments.

[0045] Specifically, for each single-agent policy network, the asset weight adjustment vector in each virtual financial market environment can be obtained based on the state matrix in each virtual financial market environment; the network parameters of the single-agent policy network are optimized according to the multi-objective reward mechanism and the asset weight adjustment vector in each virtual financial market environment until the standard deviation of the returns in each virtual financial market environment is lower than the threshold, thus obtaining the trained single-agent policy network.

[0046] To improve the reliability of obtaining the asset weight adjustment vector, in one embodiment, step 200 includes: Step 201: Input the state matrix into each of the single agent policy networks to obtain the action corresponding to the single agent policy network, and determine the action as the asset weight adjustment vector.

[0047] To improve the reliability of multi-agent model training and thus the accuracy of subsequent multi-agent model applications, such as... Figure 3 As shown, in one embodiment, step 300 includes: Step 301: Determine the virtual market feedback based on the asset weight adjustment vector of the single agent policy network. The virtual market feedback includes: return rate and risk value.

[0048] Specifically, virtual market feedback can be generated based on the asset weight adjustment vector of the single-agent policy network and historical market data. Historical market data may include the price and trading volume of assets (such as stocks, indices, futures contracts, etc.), and the risk value may be the maximum drawdown rate or volatility, etc.

[0049] Step 302: Use the multi-objective reward mechanism, return rate and risk value to perform backpropagation training, and continuously optimize the network parameters of the single agent policy network.

[0050] Step 303: When the standard deviation of the reward of the single agent policy network is lower than the threshold, the trained single agent policy network corresponding to the single agent policy network is obtained, and the trained single agent policy networks constitute the trained multi-agent model.

[0051] To improve the reliability of the asset allocation strategy, in one embodiment, step 400 includes: Step 401: Obtain the state matrix corresponding to the target customer.

[0052] Specifically, the state matrix corresponding to the target customer can be constructed from the current market volatility index, the current historical returns of assets, and the customer profile vector of the target customer.

[0053] Step 402: Input the state matrix corresponding to the target customer into the trained multi-agent model to obtain the initial asset allocation strategy output by each trained single-agent policy network. Each initial asset allocation strategy includes: the initial asset weight adjustment vector of the target customer.

[0054] Step 403: Using the weighted average fusion algorithm and each of the initial asset allocation strategies, the target asset weight adjustment vector for the target customer is obtained.

[0055] For example, initial asset allocation strategy 1 could be [0.4, 0.3, 0.2, 0.1], meaning 40% allocated to stocks, 30% to bonds, 20% to commodities, and 10% to cash; initial asset allocation strategy 2 could be [0.3, 0.4, 0.2, 0.1], meaning 30% allocated to stocks, 40% to bonds, 20% to commodities, and 10% to cash. Assuming the weights of initial asset allocation strategies 1 and 2 are 0.3 and 0.7 respectively, then the target asset weight adjustment vector could be [0.4×0.3+0.3×0.7, 0.3×0.3+0.4×0.7, 0.2, 0.1], meaning the investment proportions of stocks, bonds, commodities, and cash are 33%, 37%, 20%, and 10% respectively.

[0056] Step 404: Determine the asset allocation strategy of the target customer based on the target asset weight adjustment vector. The asset allocation strategy includes at least one of the following: expected return, risk level, and investment ratio of multiple asset classes.

[0057] Specifically, the expected returns and risk levels of the target customers can be determined based on their investment ratios.

[0058] Specifically, the system can analyze the current state of a client's overall portfolio to generate a portfolio summary. Starting with the first step of the asset allocation model—asset review—it moves beyond simply displaying the client's current asset allocation percentages, portfolio details, return details, and asset review score. Leveraging the AI's writing capabilities, the asset allocation recommendation report provides a complete and clear description of the client's current assets, summarizing issues in the current asset structure and explaining its alignment with the previous asset allocation goals. Personalized product portfolio recommendations can be generated. The asset allocation model will provide a complete asset allocation strategy based on market conditions and fluctuations. The AI ​​will follow up on the model's content, summarizing the asset allocation strategy and explaining the reasons for recommended products in the asset allocation recommendation report, further enhancing the asset allocation service experience. The system can continuously track client asset allocation strategies. The AI ​​will establish an intelligent portfolio analysis and monitoring mechanism, reading daily client asset allocation portfolio holdings, return information, transaction changes, and market financial indices to perform portfolio performance attribution and portfolio change monitoring, providing early warnings and attribution analysis explanations to clients and private banking account managers.

[0059] To improve the dimensions and accuracy of customer profiles, in one embodiment, step 100 includes: Step 111: Obtain market volatility index, historical asset returns, and customer risk characteristics, including account balance, asset distribution, transaction records, and risk tolerance questionnaire.

[0060] Step 112: Obtain the customer profile vector based on the market volatility index and the risk characteristics.

[0061] Step 113: Construct the state matrix based on the market volatility index, historical asset returns, and customer profile vectors.

[0062] To further illustrate this solution, this application provides an application example of a method for generating asset allocation strategies for financial institutions, as described in detail below: Step 1: The user data collection server collects user data. Step 1 includes: Step 11, Customer Profile Generation: By accessing the financial institution's customer database, trading system, and external public market data interfaces, extract customer account balances, asset distribution, transaction records, risk tolerance questionnaires, and real-time market fluctuation data. Using a semantic analysis algorithm based on natural language processing, extract keywords from customer communication records (including financial consultation text and voice-transcribed text) to generate a semantic risk preference vector for the customer.

[0063] Specifically, the account balance refers to the customer's total assets in financial institutions, including but not limited to bank savings accounts, investment accounts, and third-party platform balances, reflecting the scale of funds; for example, customer A's savings account balance is 500,000 yuan, and the investment account balance is 1 million yuan. The asset distribution refers to the allocation ratio of customer assets in different categories (such as stocks, bonds, and cash), used to analyze investment preferences; for example, customer B's asset distribution is 60% stocks, 30% bonds, and 10% cash. The transaction records include historical buying and selling operations, frequency, and amount, used to identify behavioral patterns; for example, customer C's transaction frequency in the past year is an average of 5 times per month, with an average transaction amount of 100,000 yuan. The risk tolerance questionnaire uses quantitative scoring to classify customers into conservative, balanced, or aggressive types; for example, a questionnaire score of 70 corresponds to a balanced profile. The real-time market fluctuation data includes dynamic indicators such as stock indices and exchange rates, used for environmental awareness. These data collectively form the basis of the customer profile, but the semantic risk preference vector is mainly extracted from communication records, with other data used for subsequent risk classification.

[0064] The multi-source heterogeneous data includes structured data from financial institution databases (such as account balances and asset distribution) and external unstructured data (such as social text and voice recordings), which are converted into computable vector representations through deep feature embedding algorithms. For example, word embedding (Word2Vec) is used to convert financial consultation text into semantic vectors, or normalization techniques are used to map account balances into numerical vectors. These vector representations are used for customer profile generation, enabling multi-dimensional feature fusion and improving profile accuracy.

[0065] Step 12, Risk Preference Identification: Using clustering algorithms (such as K-means++), the risk characteristics of the customer group are classified into three risk profiles: conservative, balanced, and aggressive. The initial strategy parameters of the agent are dynamically adjusted based on these risk profiles.

[0066] Step 13, Dynamic Data Acquisition: Collect market conditions, policy changes, industry index fluctuations, and other information in real time, and store them in a data buffer in time series form for subsequent use by the multi-agent model.

[0067] Furthermore, distributed data crawling technology can be used to transform and align the formats of heterogeneous data from multiple sources, such as bank accounts, third-party wealth management platforms, securities trading systems, and social media sentiment indices. Through deep feature embedding algorithms, unstructured data (such as text comments and voice tags) can be transformed into computable vector representations, improving the dimensionality and accuracy of customer profiles.

[0068] Furthermore, isolation forest and time-series residual analysis methods can be used to automatically identify abnormal data in customer investment behavior (such as extreme transactions and missing data), and data repair or interpolation can be performed through time backtracking mechanisms.

[0069] Furthermore, a data encryption mechanism based on differential privacy can be used to perturb sensitive customer data (such as account balances, asset distribution, and transaction records); the transport layer uses a TLS+AES dual-layer encryption protocol to ensure data security and compliance.

[0070] Based on the data extracted in step 11, including account balance, asset distribution, transaction records, and risk tolerance questionnaires, the system first calculates a risk characteristic vector for each customer. For example, through normalization, the account balance is mapped to a size index in the 0-1 range, and the asset distribution is transformed into an asset class weight vector. Subsequently, the K-means++ clustering algorithm is used to classify the risk characteristic vectors, dividing the customer group into three risk profiles: conservative, balanced, and aggressive. These risk profiles are used to dynamically adjust the agent's initial strategy parameters: for example, for conservative customers, the agent's initial parameters favor low-volatility asset weights; for aggressive customers, the parameters allow for higher risk exposure.

[0071] Step 2: Agent Modeling and Learning. The server performs agent modeling and learning. Step 2 includes: Step 21, Environmental Feedback Simulation: Simulate the random fluctuation trajectory of asset prices using the Monte Carlo method to construct multiple virtual financial market environments E(t).

[0072] Specifically, in each virtual financial market environment, the agent's action output (i.e., asset weight adjustment vector) is evaluated. Multiple virtual financial market environments (i.e., multi-dimensional scenarios such as bull market, sideways market, and bear market) can be automatically generated for the agent to learn the optimal allocation strategy under different market conditions.

[0073] Step 22, Multi-Agent Model Training: Based on the Deep Reinforcement Learning (DRL) algorithm, each agent is treated as an autonomous learning decision-making unit. Each agent acquires "state-action-reward" triplet data through interaction with the environment. The steps are as follows: State initialization: Construct a state matrix S0 based on variables such as customer profile vector, market volatility index, and historical asset returns.

[0074] Action decision generation: Through the policy network (i.e. the single agent policy network mentioned above) π(a|s;θ) and the state matrix, the agent generates action a, i.e., the asset weight adjustment vector, in a given state.

[0075] Environmental feedback sampling: A virtual market feedback r(t) is generated based on historical market data. This virtual market feedback includes multi-dimensional indicators such as return, maximum drawdown, and volatility.

[0076] Policy parameter update: The agent uses the reward function R=α×return rate-β×risk value for backpropagation training to continuously optimize the network parameters θ. This reward function is a multi-objective reward mechanism that combines return rate and risk value.

[0077] Model convergence criterion: When the standard deviation of the reward (i.e., the output value of the reward function) in consecutive iterations is lower than the threshold ε, the policy network is considered to have converged, and the policy parameter set {θ1, θ2, ..., θ} is output. n Once all the single-agent policy networks have converged, the trained multi-agent model is obtained.

[0078] Furthermore, after the multi-agent model is trained, the system calculates the policy complementarity among agents through a cooperative game mechanism; for example, the Shapley value is used to evaluate the contribution rate of each agent to the overall return. A weighted average fusion algorithm is used to output the comprehensive policy matrix W: specifically, let the policy weight of agent i be ωi (calculated based on its historical return standard deviation), then the comprehensive policy W = ∑(ωi) Wi), where Wi is the asset weight vector output by agent i. This process forms the optimal asset allocation strategy of swarm intelligence, enhancing the robustness of the strategy.

[0079] Specifically, assuming the constructed virtual financial market environments include A1, A2, and A3, a state matrix for virtual financial market environment A1 can be built based on variables such as the customer profile vector, market volatility index, and historical asset returns. Using the policy network π(a|s;θ) and the state matrix in virtual financial market environment A1, action 'a', i.e., the asset weight adjustment vector, is generated in virtual financial market environment A1 under a given state. Based on the asset weight adjustment vector in virtual financial market environment A1, virtual market feedback is generated, including multi-dimensional indicators such as return rate, maximum drawdown rate, and volatility. Backpropagation training is performed using the reward function and the return rate of virtual financial market environment A1 to continuously optimize the network parameters θ. Similarly, the reward function, the return and risk values ​​of the virtual financial market environment A2 can be used for backpropagation training to continuously optimize the network parameters θ. The network parameters θ can be continuously optimized until the standard deviation of returns in different virtual financial market environments is lower than the threshold ε. The current network parameters are then used as the optimized network parameters, and the policy network converges.

[0080] In this application example, the training process of the multi-agent model introduces a multi-objective reward mechanism, which not only considers maximizing returns but also achieves optimal risk constraints through dynamic weight adjustment, thereby enhancing the model's market adaptability.

[0081] Furthermore, meta-learning algorithms can be introduced during the training of multi-agent models to enable the models to quickly adapt to changes in the market environment under conditions of limited new data, thereby improving the generalization ability of the algorithms.

[0082] Furthermore, through transfer learning mechanisms, the network parameters of a single-agent policy network that has converged in a specific market (such as A-shares) can be transferred to another market (such as US stocks or Hong Kong stocks) for retraining, thereby achieving cross-market policy reuse and reducing model training costs.

[0083] Step 23, Multi-agent Cooperative Optimization: After the training of multiple single-agent policy networks is completed, the policy complementarity among agents is calculated through a cooperative game mechanism, and the comprehensive policy matrix W is output using a weighted average fusion algorithm. This leads to the formation of an optimal asset allocation strategy based on collective wisdom.

[0084] Furthermore, it can generate decision-making path diagrams and profit contribution rate analysis diagrams for intelligent agents, and use the SHAP (Shapley Additive Explanation) algorithm to explain model behavior, providing investment advisors with understandable intelligent recommendation basis and improving system transparency and trust.

[0085] Step 3: The portfolio optimization server performs portfolio optimization. Step 3 includes: Step 31: Calculate asset weights by receiving the policy parameter set {θ1, θ2, ..., θ} output by the agent. n} and the comprehensive strategy matrix W Perform a return-risk balance calculation.

[0086] Step 32: Use the mean-variance model to solve for the expected returns and covariance of different assets.

[0087] Step 33: Based on time series analysis and Bayesian prediction, anticipate market volatility trends and generate rebalancing signals. The final asset allocation recommendations (i.e., asset allocation strategies) include: asset class, investment ratio, expected return, risk level, and other information.

[0088] Furthermore, a risk transmission matrix between assets can be constructed based on complex network theory to provide early warning of potential systemic risks. When the volatility of a certain asset class exceeds a threshold, the system automatically triggers a weight adjustment strategy.

[0089] Furthermore, sentiment analysis algorithms can monitor the sentiment index of news, social platforms, and analyst comments in real time, incorporate sentiment factors into portfolio optimization models, and dynamically adjust market expectation parameters.

[0090] Furthermore, LSTM (Long Short-Term Memory) networks can be used to predict the return sequence of a portfolio, generating future return ranges and confidence intervals to assist investors in making forward-looking decisions.

[0091] Step 4: Visual output and interactive terminal execution of visual output.

[0092] The investment strategy visualization interface displays the final asset allocation recommendations through dynamic dashboards and candlestick charts, allowing users to adjust risk preference parameters or investment horizons. Asset allocation strategies can also be regenerated.

[0093] Specifically, a closed-loop intelligent asset allocation decision-making system can be constructed through data flow and logical collaboration among the aforementioned user data acquisition server, agent modeling and learning server, portfolio optimization server, and visualization output and interactive terminal. The user data acquisition server can provide structured customer and market data input; the agent modeling and learning server can train multi-agent models based on the input data, output optimized strategies, and perform strategy learning, model iteration, and behavior optimization in a simulated market environment; the portfolio optimization server can quantitatively optimize the agent strategy results and predict returns; and the visualization output and interactive terminal can realize user interaction and strategy visualization display, forming a closed loop: customer feedback → new data acquisition → agent relearning → strategy iteration and optimization.

[0094] From a software perspective, in order to generate asset allocation strategies by combining market conditions and customer investment preferences, thereby improving the accuracy of asset allocation strategies and increasing customer satisfaction, this application provides an embodiment of an asset allocation strategy generation system for financial institutions that implements all or part of the aforementioned asset allocation strategy generation method. See [link to relevant documentation]. Figure 4 The asset allocation strategy generation system of the financial institution specifically includes the following: Acquisition module 01 is used to acquire a state matrix, which is constructed from multi-dimensional data, including: market volatility index, historical asset returns, and customer profile vectors. The adjustment module 02 is used to obtain the corresponding asset weight adjustment vectors based on the state matrix and multiple single-agent policy networks. Training module 03 is used to optimize the network parameters of each single agent policy network according to the multi-objective reward mechanism and the asset weight adjustment vector of each single agent policy network to obtain the trained multi-agent model. The trained multi-agent model includes: multiple trained single agent policy networks. The generation module 04 is used to apply the trained multi-agent model to generate asset allocation strategies for target customers in financial institutions.

[0095] As described above, the asset allocation strategy generation method and system for financial institutions provided in this application can be specifically applied to the scenario of "application of asset allocation based on intelligent agent models." Centered on the asset allocation business of financial institutions, the system achieves intelligent generation of customer profiles, deep learning of investment preferences, and adaptive optimization of asset portfolio strategies through the embedding and dynamic evolution mechanism of intelligent agent models. In practice, this system first collects multi-dimensional information from users through customer service terminals, including data on risk preferences, liquidity needs, asset size, investment cycle, and historical trading behavior. This data is aggregated into the system's central database. Through a pre-trained intelligent agent model, customer behavior patterns are dynamically learned and clustered to form personalized customer profiles. An intelligent decision engine can be applied to interactively model different customer profiles with dynamic market information to generate customized asset allocation recommendations. For example, when market volatility increases, the intelligent agent automatically adjusts the weighting of customer risk tolerance and reallocates asset proportions to achieve a new balance between stability and returns in the investment portfolio. The entire process is automated, requiring no human intervention. Simultaneously, the system provides optimization suggestions to investment advisors through a visual interface, enabling them to have more precise strategic basis when communicating with clients. Furthermore, the system can evolve investment strategies through continuous data learning and model updates. When new trends emerge in the investment behavior of a particular client group, the system automatically identifies and updates model parameters, ensuring the real-time and forward-looking nature of asset allocation strategies. In practice, this means that banks or investment institutions can respond to market changes more quickly, providing high-net-worth clients with more intelligent and personalized investment services. In this way, the intelligent agent model is no longer merely an algorithmic tool, but becomes the intelligent engine of asset allocation business. It transforms the traditional wealth management model, which relies on human experience and static analysis, into a dynamic, learnable, and evolvable intelligent decision-making system, significantly improving the professionalism of financial institutions in customer service, investment advisory, and risk control. The core value of this solution lies in transforming the traditional financial service chain of "understanding the client—guiding the client—persuading the client—reaching a consensus" into a quantifiable, optimizable, and traceable intelligent system through technological means, fundamentally reshaping the operational logic and service model of asset allocation business.

[0096] The embodiments of the asset allocation strategy generation system for financial institutions provided in this specification can be used to execute the processing flow of the embodiments of the asset allocation strategy generation method for financial institutions described above. Its functions will not be repeated here, but can be referred to the detailed description of the embodiments of the asset allocation strategy generation method for financial institutions described above.

[0097] Figure 5 This is a schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention, such as... Figure 5As shown, the electronic device includes: a memory 501, a processor 502, and a computer program stored in the memory 501 and executable on the processor 502. When the processor 502 executes the computer program, it implements the following method: Obtain a state matrix, which is constructed from multidimensional data, including: market volatility index, historical asset returns, and customer profile vectors. Based on the state matrix and multiple single-agent policy networks, the corresponding asset weight adjustment vectors are obtained. Based on the multi-objective reward mechanism and the asset weight adjustment vector of each single agent policy network, the network parameters of each network are optimized to obtain the trained multi-agent model, which includes multiple trained single agent policy networks. The trained multi-agent model is used to generate asset allocation strategies for target customers in financial institutions.

[0098] This embodiment discloses a computer program product, which includes a computer program that, when executed by a processor, implements the following method: Obtain a state matrix, which is constructed from multidimensional data, including: market volatility index, historical asset returns, and customer profile vectors. Based on the state matrix and multiple single-agent policy networks, the corresponding asset weight adjustment vectors are obtained. Based on the multi-objective reward mechanism and the asset weight adjustment vector of each single agent policy network, the network parameters of each network are optimized to obtain the trained multi-agent model, which includes multiple trained single agent policy networks. The trained multi-agent model is used to generate asset allocation strategies for target customers in financial institutions.

[0099] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the following method: Obtain a state matrix, which is constructed from multidimensional data, including: market volatility index, historical asset returns, and customer profile vectors. Based on the state matrix and multiple single-agent policy networks, the corresponding asset weight adjustment vectors are obtained. Based on the multi-objective reward mechanism and the asset weight adjustment vector of each single agent policy network, the network parameters of each network are optimized to obtain the trained multi-agent model, which includes multiple trained single agent policy networks. The trained multi-agent model is used to generate asset allocation strategies for target customers in financial institutions.

[0100] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0101] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0102] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0103] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0104] In the description of this specification, the references to terms such as "an embodiment," "a specific embodiment," "some embodiments," "for example," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0105] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating asset allocation strategies for financial institutions, characterized in that, include: Obtain a state matrix, which is constructed from multidimensional data, including: market volatility index, historical asset returns, and customer profile vectors. Based on the state matrix and multiple single-agent policy networks, the corresponding asset weight adjustment vectors are obtained. Based on the multi-objective reward mechanism and the asset weight adjustment vector of each single agent policy network, the network parameters of each network are optimized to obtain the trained multi-agent model, which includes multiple trained single agent policy networks. The trained multi-agent model is used to generate asset allocation strategies for target customers in financial institutions.

2. The method for generating asset allocation strategies for financial institutions according to claim 1, characterized in that, The acquisition of the state matrix includes: Construct multiple virtual financial market environments using the Monte Carlo method; In each of the aforementioned virtual financial market environments, a state matrix is ​​obtained.

3. The method for generating asset allocation strategies for financial institutions according to claim 1, characterized in that, The step of obtaining the corresponding asset weight adjustment vectors based on the state matrix and multiple single-agent policy networks includes: The state matrix is ​​input into each of the single agent policy networks to obtain the action corresponding to the single agent policy network, and the action is determined as the asset weight adjustment vector.

4. The method for generating asset allocation strategies for financial institutions according to claim 1, characterized in that, The step of optimizing the network parameters of each single-agent policy network according to the multi-objective reward mechanism and the asset weight adjustment vector of each single-agent policy network to obtain the trained multi-agent model includes: Based on the asset weight adjustment vector of the single-agent policy network, the virtual market feedback is determined, which includes: return rate and risk value; The network parameters of the single-agent policy network are continuously optimized by using the multi-objective reward mechanism, return rate, and risk value for backpropagation training. When the standard deviation of the returns of the single agent policy network is lower than the threshold, the trained single agent policy network corresponding to the single agent policy network is obtained, and the trained single agent policy networks constitute the trained multi-agent model.

5. The method for generating asset allocation strategies for financial institutions according to claim 1, characterized in that, The application of the trained multi-agent model to generate asset allocation strategies for target customers in financial institutions includes: Obtain the state matrix corresponding to the target customer; The state matrix corresponding to the target customer is input into the trained multi-agent model to obtain the initial asset allocation strategy output by each trained single-agent policy network. Each initial asset allocation strategy includes: the initial asset weight adjustment vector of the target customer. The target asset weight adjustment vector for the target customer is obtained by using a weighted average fusion algorithm and each of the initial asset allocation strategies. Based on the target customer's target asset weight adjustment vector, the target customer's asset allocation strategy is determined. The asset allocation strategy includes at least one of the following: expected return, risk level, and investment ratio of multiple asset classes.

6. The method for generating asset allocation strategies for financial institutions according to claim 1, characterized in that, The acquisition of the state matrix includes: Obtain market volatility index, historical asset returns, and customer risk characteristics, including account balance, asset distribution, transaction history, and risk tolerance questionnaire; The customer profile vector is obtained based on the market volatility index and the risk characteristics. The state matrix is ​​constructed based on the market volatility index, historical asset returns, and customer profile vectors.

7. An asset allocation strategy generation system for financial institutions, characterized in that, include: The acquisition module is used to acquire a state matrix, which is constructed from multi-dimensional data, including: market volatility index, historical asset returns, and customer profile vectors. The adjustment module is used to obtain the corresponding asset weight adjustment vectors based on the state matrix and multiple single-agent policy networks. The training module is used to optimize the network parameters of each single agent policy network according to the multi-objective reward mechanism and the asset weight adjustment vector of each single agent policy network to obtain the trained multi-agent model. The trained multi-agent model includes: multiple trained single agent policy networks. The generation module is used to apply the trained multi-agent model to generate asset allocation strategies for target customers in financial institutions.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the asset allocation strategy generation method for financial institutions as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the asset allocation strategy generation method for financial institutions as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the asset allocation strategy generation method for financial institutions as described in any one of claims 1 to 6.

Images