A portfolio decision system and method based on fuzzy stochastic multi-objective optimization

Abstract

This invention relates to the field of financial investment decision-making technology, and discloses a portfolio decision-making system and method based on fuzzy stochastic multi-objective optimization. The portfolio decision-making method based on fuzzy stochastic multi-objective optimization includes: collecting and preprocessing multi-source heterogeneous data; performing cross-modal consistency verification on financial information from different sources and types; constructing two parallel processing paths, an emotional channel and a cognitive channel, to extract the emotional and cognitive features of investor psychological states, respectively; dynamically weighting and fusing features from different information sources to construct an investor psychological state model; introducing investor psychological states as fuzzy variables into the multi-objective optimization model, and using fuzzy stochastic programming to comprehensively consider multiple objectives such as return, risk, and volatility resistance to generate the optimal portfolio decision. This invention effectively integrates the influence of investor emotional and cognitive factors on the market, and solves the problems of heterogeneous financial information processing and multi-objective balance optimization.

Description

Technical Field

[0001] This invention relates to the field of financial investment decision-making technology, and more specifically, to a portfolio decision-making system and method based on fuzzy stochastic multi-objective optimization. Background Technology

[0002] Financial investment decision-making is a complex process that requires processing heterogeneous information from various sources and finding a balance among multiple objectives such as return, risk, and stability. Existing investment decision-making systems mainly suffer from the following technical shortcomings: The system struggles to effectively process and integrate heterogeneous financial information from diverse sources, including text, audio, video, and structured data. It lacks the ability to verify information consistency and filter noise, resulting in incomplete and inaccurate decision-making criteria and an inability to effectively integrate the impact of investor sentiment and cognitive factors on the market. Existing systems typically focus only on quantitative analysis of market data, neglecting the influence of investor psychology on market fluctuations and failing to address market volatility caused by human factors. The investment decision-making process requires simultaneously considering multiple conflicting objectives such as maximizing returns, minimizing risk, and portfolio stability. However, existing technologies often employ simple weighting methods when dealing with these objectives, making it difficult to adapt to the dynamic changes in objective priorities under different market conditions.

[0003] Therefore, there is a need for a portfolio decision-making method that can effectively process heterogeneous financial information, integrate investor sentiment and cognition factors, and achieve a dynamic balance of multiple objectives, in order to improve the accuracy, robustness, and adaptability of investment decisions. Summary of the Invention

[0004] This invention provides a portfolio decision-making system and method based on fuzzy stochastic multi-objective optimization, which solves the technical problems in related technologies such as difficulty in effectively processing heterogeneous financial information, inability to accurately quantify investor psychological states, and difficulty in achieving multi-objective dynamic balance.

[0005] This invention provides a portfolio decision-making method based on fuzzy stochastic multi-objective optimization, comprising the following steps: Collect multi-source heterogeneous data and preprocess it to obtain preprocessed multi-source heterogeneous data, which includes text data, transaction data, behavioral data and multimedia data; Cross-modal consistency verification is performed on the preprocessed multi-source heterogeneous data to identify and remove contradictory or abnormal information, thus obtaining the verified data. Two parallel processing paths, the emotional channel and the cognitive channel, are constructed for the verified data to extract the emotional features and cognitive features of the investor's psychological state, respectively. Based on an adaptive fusion gating network, the extracted emotional and cognitive features are dynamically weighted and fused to construct an investor psychological state model. By introducing the investor psychological state model as a fuzzy variable into the multi-objective optimization model, and by using fuzzy stochastic programming to comprehensively consider multiple objectives such as return, risk and volatility resistance, the optimal portfolio decision is generated.

[0006] Furthermore, the cross-modal consistency verification includes the following steps: For different types of data, corresponding encoders are built to map data of different modalities to a unified semantic space; For different multi-source heterogeneous data of the same financial event, the encoder is used to obtain its representation in a unified semantic space and calculate its similarity. Introduce negative samples to calculate cross-modal consistency scores; A threshold is set based on the calculated cross-modal consistency score. When the cross-modal consistency scores of a certain type of multi-source heterogeneous data and multiple other types of multi-source heterogeneous data are all below the threshold, they are marked as abnormal information and filtered out.

[0007] Furthermore, the extraction of emotional features includes: Using sentiment analysis algorithms in the financial field, we extract the sentiment polarity, intensity, and specific emotion type from text and multimedia data to obtain the probability distribution vector of sentiment categories. Based on the probability distribution vector of the aforementioned emotion categories, time series analysis is used to capture the trend and fluctuation pattern of emotion changes.

[0008] Furthermore, the cognitive feature extraction includes: Identify patterns of cognitive biases in investor trading behavior; The identified cognitive bias patterns are quantified into numerical features for subsequent decision-making.

[0009] Furthermore, the adaptive fusion gating network is represented as: ; in, This represents the output of the adaptive fusion gating network; , They represent The first and nth feature vectors, where n represents the index of the input feature vector; For the i-th feature vector of the input The nonlinear transformation function; for Corresponding adaptive weights; Calculated in the following way: ; in Representation of features Reliability score; express Represents the product of reliability and correlation; This represents the feature index used when calculating the normalized denominator; This represents the sum of the importance indices of all features, used to normalize the weights so that their sum equals 1. Representation of features Reliability rating, express Relevance score, express The j-th eigenvector.

[0010] Furthermore, the dynamic weighted fusion also includes constructing a temporal psychological state evolution model: ; in, Indicates time The final investor psychological state; Indicates time This indicates the psychological state of investors; Indicates time The current mental state representation obtained through an adaptive fusion gating network; It is a time series smoothing factor.

[0011] Furthermore, the investor psychological state model includes factors influencing investor psychological state. Defined as a fuzzy variable, represented using triangular fuzzy numbers: ; in, A fuzzy representation of factors influencing investor psychology. The central value of the calculated investor psychological state influencing factor; This represents the lower bound of the factor influencing investor psychology. This represents the upper bound of the factor influencing investor psychology.

[0012] Furthermore, the multi-objective function in the multi-objective optimization model includes: Maximizing expected return: ; Minimize investment risk: ; Maximizing resistance to volatility: ; in, This represents the objective function for expected returns; This represents the objective function for investment risk. Represent the objective function for resisting volatility; Indicates the maximum value; Indicates the minimum value; Indicates the assets in the portfolio The weights; This indicates the total number of assets in the portfolio; Assets Expected rate of return; Indicates the psychological state of investors towards assets Additional impact coefficient of revenue; The covariance matrix representing the rate of return on assets; Representing vectors Transpose of; Assets Stability indicators; Assets Sensitivity coefficient to fluctuations in investor psychology.

[0013] Furthermore, the multi-objective function is transformed into a single objective function: ; in, This represents the transformed single objective function. Indicates the maximum value; This represents the objective function for expected returns; This represents the objective function for investment risk. Represent the objective function for resisting volatility; , , These represent the relative importance of returns, risks, and volatility resistance in decision-making, respectively. , The targets are respectively The maximum and minimum values ​​are obtained by searching within the feasible region using an optimization algorithm; , , These represent the normalized values ​​of the objective functions for return, risk, and volatility resistance, respectively, with the normalized range being [0,1].

[0014] This invention provides a portfolio decision-making system based on fuzzy stochastic multi-objective optimization, used to execute the aforementioned portfolio decision-making method based on fuzzy stochastic multi-objective optimization, comprising: The data acquisition and preprocessing module is used to acquire multi-source heterogeneous data and preprocess it to obtain preprocessed multi-source heterogeneous data. The cross-modal consistency verification module is used to perform cross-modal consistency verification on preprocessed multi-source heterogeneous data, identify and remove contradictory or abnormal information, and obtain verified data. The emotional and cognitive feature extraction module constructs two parallel processing paths for the verified data: an emotional channel and a cognitive channel, to extract emotional and cognitive features of investors' psychological states, respectively. An adaptive fusion gating network module is used to dynamically weight and fuse extracted emotional and cognitive features to construct an investor psychological state model. A fuzzy stochastic multi-objective optimization module is used to generate optimal portfolio decisions.

[0015] The beneficial effects of this invention are as follows: With enhanced information processing capabilities, this invention, based on a cross-modal consistency verification method, can effectively identify and filter contradictions and anomalies between different information sources, thereby improving the processing accuracy of multi-source heterogeneous financial information.

[0016] The accuracy of decision-making is enhanced by introducing investor psychological state as a fuzzy variable into a multi-objective optimization model and using an adaptive fusion gating network to dynamically adjust the weights of different information sources. This invention can effectively cope with irrational market fluctuations.

[0017] With enhanced risk control capabilities, this invention, based on dual-channel feature extraction of emotion and cognition and fuzzy random multi-objective optimization, can accurately identify and quantify market volatility risks caused by human factors, thereby improving the robustness of investment portfolios.

[0018] Enhanced system adaptability: Through dynamic adjustment of the importance of different information sources and adaptive optimization of multi-objective weights, this invention can automatically adjust decision-making strategies according to changes in the market environment, improving the system's adaptability and universality. It demonstrates superior stability and effectiveness compared to traditional methods across different market cycles and asset classes. Attached Figure Description

[0019] Figure 1 This is a flowchart of a portfolio decision-making method based on fuzzy stochastic multi-objective optimization according to the present invention; Figure 2 This is a flowchart of the present invention for collecting and preprocessing multi-source heterogeneous data; Figure 3 This is a detailed flowchart of the present invention for cross-modal consistency verification of preprocessed multi-source heterogeneous data; Figure 4 This is a detailed flowchart of the emotion feature extraction and cognitive feature extraction of the present invention; Figure 5 This is a detailed flowchart of the present invention for constructing an investor psychological state model; Figure 6 This is a detailed flowchart of the process for generating optimal portfolio decisions according to the present invention. Detailed Implementation

[0020] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0021] At least one embodiment of the present invention discloses a portfolio decision-making method based on fuzzy stochastic multi-objective optimization, such as... Figures 1 to 6 As shown, it includes the following steps: Step 1: Collect multi-source heterogeneous data and perform preprocessing; Multi-source heterogeneous data includes text data, transaction data, behavioral data, and multimedia data; Specifically, the following steps are included: Step 1.1, Data Acquisition; Collect the following types of multi-source heterogeneous data from various data sources: Text data includes social media comments, financial news reports, and research institution analysis reports. Transaction data: including structured data such as market transaction records, trading volume, and price changes; Behavioral data: including investor online behavior data, application usage data, etc. Multimedia data: including financial videos, audio programs, and other related content.

[0022] Step 1.2, Data Cleaning; The collected heterogeneous data from multiple sources is cleaned to remove obvious erroneous data, duplicate data, and incomplete data.

[0023] Step 1.3, Data Standardization; For multi-source heterogeneous data of different types and sources, appropriate standardization methods are applied to convert the data into a unified format and scale for easier subsequent processing. For numerical data, the Z-score standardization method is used. ; in, Indicates multi-source heterogeneous data values; Represents the arithmetic mean of multi-source heterogeneous data values; This represents the standard deviation of values ​​from multi-source heterogeneous data. This represents the standardized data values. The transformed data will exhibit a distribution with a mean of 0 and a standard deviation of 1, which is beneficial for comparing data of different dimensions and for model training.

[0024] Step 1.4, preliminary noise filtering; Statistical methods and domain knowledge are used to perform preliminary noise filtering on multi-source heterogeneous data, removing obvious outliers and irrelevant information. For time series data, moving average filtering can be applied. ; in, Indicates time Multi-source heterogeneous data points; Indicates the size of the sliding window; Indicates the index position within the window; This represents the filtered data points.

[0025] The output of this step is pre-processed multi-source heterogeneous data, which lays the foundation for subsequent in-depth analysis and feature extraction.

[0026] Step 2: Perform cross-modal consistency verification on the preprocessed multi-source heterogeneous data, and identify and remove contradictory or abnormal information; Specifically, the following steps are included: Step 2.1, Modal encoder construction; Build corresponding encoders for different types of data. This maps data from different modalities to a unified semantic space.

[0027] For text data, a pre-trained language model is used as the encoder; For transaction data, a time-series feature extraction network is used; For multimedia data, a corresponding multimedia feature extraction network is used.

[0028] Step 2.2, cross-modal similarity calculation; For the same financial event or target, different multi-source heterogeneous data and via encoder and We obtain their representations in a unified semantic space, and then calculate their similarity: ; in, Represents the cosine similarity function; , Representing modes and modality Multi-source heterogeneous data; , These represent the modes used for processing. and modality Encoder for multi-source heterogeneous data; express via encoder The resulting representation vector; express via encoder The resulting representation vector; These represent vectors. and The Euclidean norm.

[0029] Step 2.3, calculate the cross-modal consistency score; Introducing negative samples Calculate the cross-modal consistency score: ; in, express and Cross-modal consistency score between This represents a negative sample, which is a modality unrelated to the current financial event or target. Multi-source heterogeneous data; Represents the cosine similarity function; This represents the maximum value.

[0030] Step 2.4, Abnormal Information Identification and Filtering; Based on the calculated cross-modal consistency score, a threshold is set. When the cross-modal consistency scores of a certain type of multi-source heterogeneous data and multiple other types of multi-source heterogeneous data are all below the threshold, they are marked as abnormal information and filtered out: ; in, Indicates the filtering function; express and Cross-modal consistency score between; The threshold representing the cross-modal consistency score; Indicates the total number of modes; This indicates traversing the modal. All other modalities besides; For indicator functions, when When the value is 1, it is 1; otherwise, it is 0.

[0031] Based on the distribution characteristics of cross-modal consistency scores using historical data, the mean and standard deviation of all cross-modal consistency scores are calculated, and a strict threshold is set. The threshold is the mean minus 1-2 times the standard deviation, used to control the stringency of anomaly detection; the higher the stringency threshold, the more stringent the filtering. express The proportion that contradicts other modal data; This is the preset filtering ratio threshold.

[0032] The output of this step is a high-quality dataset that has undergone cross-modal consistency verification and anomaly filtering, which significantly reduces the impact of errors and noise in subsequent analyses.

[0033] Step 3: Construct two parallel processing paths, the emotional channel and the cognitive channel, for the verified data, and extract the emotional features and cognitive features of the investor's psychological state, respectively. Specifically, the following steps are included: Step 3.1, Heterogeneous information projection; Using heterogeneous information projection function This maps different types of heterogeneous data from multiple sources to a unified semantic space. ; in, Indicates the type of data Feature extraction function; Represents the projection matrix; Represents the projection bias vector; express Representation vectors in a unified semantic space.

[0034] Step 3.2, Extraction of Emotional Channel Features; Extracting sentiment features from text and multimedia data, including: Fine-grained sentiment analysis: Utilizing sentiment analysis algorithms from the financial field, this method extracts the sentiment polarity, intensity, and specific emotion type from text and multimedia data, yielding a probability distribution vector for sentiment categories. ; in, express The representation vector in the unified semantic space after heterogeneous information projection; The weight matrix representing sentiment analysis; Represents the bias vector for sentiment analysis; This represents the softmax activation function; express The probability distribution vector of sentiment categories.

[0035] Sentiment fluctuation tracking: Based on the sentiment category probability distribution vectors obtained above, time series analysis is used to capture sentiment change trends and fluctuation patterns. ; in, Indicates the current time point; Indicates the previous point in time; express time The sampling data; express time The sampling data; , They represent and The distribution of sentiment categories; It represents the absolute difference in sentiment distribution between two points in time; This represents the total number of modalities used in the computation of multi-source heterogeneous data; This represents the average of the sentiment changes across all samples. Indicates time The overall sentiment fluctuation index.

[0036] Step 3.3, Cognitive Channel Feature Extraction; Extracting cognitive bias features from transaction data, including: Behavioral pattern analysis: Identifying cognitive bias patterns in investor trading behavior: ; in, Indicates current transaction data The results of cognitive bias pattern recognition in the text; It is a cognitive bias identification function; Indicates current transaction data The representation vector in the unified semantic space after heterogeneous information projection; Represents historical transaction data Representation vectors in the same semantic space.

[0037] Cognitive bias quantification: Quantifying identified cognitive bias patterns into numerical features that can be used for subsequent decision-making. ; in, A comprehensive quantitative result representing cognitive bias; This represents the total number of cognitive bias patterns defined in the system; An index representing cognitive bias patterns; Indicates the first The identification results of various cognitive bias patterns; It is the first Weighting coefficients corresponding to different cognitive bias patterns; This represents a weighted summation of all cognitive bias patterns.

[0038] Step 3.4, Feature vector generation; Integrating emotional and cognitive features, a complete feature vector of investor psychological state is generated: ; in, This represents the final generated feature vector of investor psychological state. This represents the probability distribution vector of sentiment categories obtained from sentiment analysis. Indicators representing the trend and pattern of emotional changes obtained from emotional fluctuation tracking; This represents the vector of cognitive bias pattern recognition results; This represents the comprehensive quantitative result of cognitive bias.

[0039] The output of this step is a feature vector of investor psychological state, which includes both sentiment analysis results and cognitive bias characteristics, providing a comprehensive representation of irrational factors for subsequent investor psychological state models.

[0040] Step 4: Based on an adaptive fusion gating network, the extracted emotional and cognitive features are dynamically weighted and fused to construct an investor psychological state model; Specifically, the following steps are included: Step 4.1, Construction of Adaptive Fusion Gated Network; The adaptive fusion gating network is represented as: ; in, This represents the output of the adaptive fusion gating network; , They represent The first and nth feature vectors, where n represents the index of the input feature vector; For the i-th feature vector of the input The nonlinear transformation function; for Corresponding adaptive weights; Adaptive weights Calculated in the following way: ; where exp Representation of features Reliability score; express Represents the product of reliability and correlation; This represents the feature index used when calculating the normalized denominator; This represents the sum of the importance indices of all features, used to normalize the weights so that their sum equals 1. Representation of features Reliability rating, express Relevance score, express The j-th eigenvector; Implementation of the adaptive fusion gating network: The adaptive fusion gating network adopts a multi-input single-output feedforward neural network structure, which includes a feature transformation layer, a weight generation layer and a weighted fusion layer.

[0041] Feature transformation layer for each Apply independent nonlinear transformation The weight generation layer is based on reliability scores. and correlation score Calculate dynamic weights The weighted fusion layer combines the transformed features according to their weights to generate the final output.

[0042] In practical applications, this adaptive fusion gating network can be used in financial market sentiment analysis scenarios. For example, during periods of market panic, it can automatically increase the weight of social media sentiment features, while during periods of stability, it can increase the weight of fundamental data features, thus achieving dynamic adaptation to different market environments.

[0043] Step 4.2, Generation of Emotional Cognition Fusion Features; Using an adaptive fusion gating network, emotional channel features are incorporated. and cognitive channel characteristics This can be integrated into a unified investor psychological state: ; in, This represents the fused feature vector of investor psychological state. This represents the feature vector extracted from the emotional channel; This represents the feature vector extracted from the cognitive channel; This represents a nonlinear transformation function applied to emotional features; This represents a nonlinear transformation function applied to cognitive features; Represents the adaptive weights of the emotional channel; Represents the adaptive weights of cognitive channels; This represents the adaptive fusion gating network function.

[0044] ; ; Step 4.3, Calculation of nonlinear emotion influence factor; Based on the integrated investor psychological characteristics, calculate its nonlinear impact factor on the market: ; in, Factors influencing investor psychology; This represents the fused feature vector of investor psychological state. Represents the weight matrix; Represents the bias vector; This represents the sigmoid activation function; This represents the influence coefficient.

[0045] Step 4.4: Construct a temporal psychological state evolution model; Considering the time evolution characteristics of investor psychological states, a time-series psychological state evolution model is constructed: ; in, Indicates time The final investor psychological state; Indicates time This indicates the psychological state of investors; Indicates time The current real-time mental state representation obtained through an adaptive fusion gating network; It is a time series smoothing factor.

[0046] This time-series psychological state evolution model achieves a balance between the continuity and timeliness of investors' psychological states, and can effectively capture the gradual and sudden changes in market sentiment.

[0047] Implementation of the temporal psychological state evolution model: This model employs an exponentially weighted moving average structure, taking into account both historical cumulative psychological states and the current immediate psychological state. Temporal smoothing factor. It's not a fixed value, but rather adaptively determined by market conditions: during periods of high market volatility, it increases... This makes the time-series psychological state evolution model more sensitive to changes in the current state; during periods of market stability, it reduces... This is to maintain the continuity of one's psychological state.

[0048] In practical applications, this time-series psychological state evolution model can effectively handle abnormal market fluctuations caused by sudden changes in investor sentiment, such as after the release of major economic policies, by adjusting... It can capture rapid changes in sentiment, provide early warnings of potential herding effects, and offer risk hedging signals for investment portfolios.

[0049] The output of this step is a representation of investor psychological state that integrates multi-source information and its impact on the market, providing key inputs of irrational factors for the final investment decision.

[0050] Step 5: Introduce the investor psychological state model as a fuzzy variable into the multi-objective optimization model, and generate the optimal portfolio decision by comprehensively considering multiple objectives such as return, risk and volatility resistance through fuzzy stochastic programming. Specifically, the following steps are included: Step 5.1, Define fuzzy variables; Factors influencing investor psychology Defined as a fuzzy variable, represented using triangular fuzzy numbers: ; in, A fuzzy representation of factors influencing investor psychology; The central value of the factor influencing investor psychological state; This represents the lower bound of the factor influencing investor psychology. This represents the upper bound of the factors influencing investor psychology. These three values, determined based on historical data analysis and the current market volatility range, together form a triangular fuzzy number used to represent the uncertainty of the impact of investor psychology.

[0051] Step 5.2, Construction of multi-objective function; Define the following three optimization objectives: Maximizing expected return: ; in, This represents the expected return function of the investment portfolio. Indicates the maximum value; Indicates the assets in the portfolio Configuration weights; Assets The expected basic rate of return; This indicates the total number of assets included in the investment portfolio; Indicates the psychological state of investors towards assets Additional impact coefficient of revenue; Fuzzy influencing factors representing investor psychological state; This indicates a summation operation on all assets.

[0052] The expected return maximization function consists of two parts: the first part is the basic expected return, and the second part is the additional return impact after considering the investor's psychological state.

[0053] Minimize investment risk: ; in, This represents the risk function of a portfolio. Indicates the minimum value; Represents the portfolio weight vector; Representing vectors Transpose of; The covariance matrix representing the rate of return on assets.

[0054] Maximizing resistance to volatility: ; in, The volatility resistance function of a portfolio; Indicates the maximum value; Assets Stability indicators; Assets Sensitivity coefficient to fluctuations in investor psychology; Fuzzy influencing factors representing investor psychological state; This indicates a summation operation on all assets.

[0055] The volatility resistance maximization function consists of two parts: the first part is the contribution of the asset's fundamental stability, and the second part is the impact of volatility sensitivity after considering the investor's psychological state. The difference between the two parts is the comprehensive volatility resistance score.

[0056] Step 5.3, Define constraints; Set the following constraints: The weights sum to 1: ; in, Indicates the assets in the portfolio Configuration weights, This represents the total number of assets included in the portfolio. This constraint ensures that the sum of the weights of all assets equals 1, meaning that all funds are allocated.

[0057] Nonnegativity constraint: ; in, Indicates the assets in the portfolio Configuration weights, This indicates the total number of assets included in the portfolio. This constraint ensures that the allocation ratio of each asset is not negative, i.e., short selling is not allowed.

[0058] Asset class constraints: ; in, Indicates the first A collection of asset classes; Indicates the assets in the portfolio Configuration weights; This indicates the upper limit of this type of asset; This indicates the total number of asset classes; this constraint ensures that the portfolio's allocation across all asset classes does not exceed a preset limit, guaranteeing investment diversification.

[0059] Step 5.4, solving the fuzzy random multi-objective optimization problem; Transform the above three objective functions into a single objective function: ; in, This represents the transformed single objective function. Indicates the maximum value; This represents the objective function for expected returns; This represents the objective function for investment risk. Represent the objective function for resisting volatility; , , These represent the relative importance of returns, risks, and volatility resistance in decision-making, respectively. , The targets are respectively The maximum and minimum values ​​obtained by the optimization algorithm within the feasible region are determined as follows: under the premise of satisfying all constraints, each objective function is optimized separately to obtain the extreme value range of each objective function; , , These represent the normalized values ​​of the objective functions for return, risk, and volatility resistance, respectively. The normalized values ​​have a range of [0,1], eliminating the influence of differences in the units and numerical ranges of different objective functions. The minus sign before the target indicates that due to risk objectives Originally a minimization objective, when transformed into a unified maximization objective function, a negative value needs to be taken, so that the combined objective function is minimized. The larger the value, the better.

[0060] For those containing fuzzy variables The objective function is transformed into a deterministic objective using the expected value transformation method: ; ; in, This represents the expected value of the objective function for expected returns. This represents the expected value of the anti-volatility objective function; The expected value of the fuzzy variable representing the influencing factor of investor psychological state; Assets The expected basic rate of return; Indicates the psychological state of investors towards assets Additional impact coefficient of revenue; Assets Stability indicators; Assets Sensitivity to fluctuations in investor psychology; Indicates the assets in the portfolio Configuration weights; This indicates the total number of assets included in the portfolio.

[0061] Triangular fuzzy number The expected value is calculated as follows: ; in, Represents triangular fuzzy numbers Expected value This represents the lower bound of the factor influencing investor psychology. This represents the central value of the factor influencing investor psychology, i.e., the most likely degree of influence. This represents the upper bound of the factor influencing investor psychology. The formula for calculating the expected value of a triangular fuzzy number uses the arithmetic mean of three eigenvalues ​​as the expected value of the fuzzy variable.

[0062] Finally, the transformed single-objective optimization problem is solved using an improved particle swarm optimization algorithm to obtain the optimal portfolio weight vector. ,in This represents the asset allocation scheme obtained by the optimization solution, which includes the optimal allocation ratio of each asset.

[0063] Implementation of fuzzy stochastic multi-objective optimization models: The fuzzy stochastic multi-objective optimization model adopts an improved non-dominated sorting particle swarm optimization algorithm structure, which includes a fuzzy variable processing module, an objective function standardization module, and a dynamic weight adjustment module.

[0064] The fuzzy variable processing module quantifies and transforms uncertain factors; The objective function standardization module ensures the comparability of objectives with different dimensions; The dynamic weight adjustment module adjusts the weights of each target based on the characteristics of the market cycle.

[0065] During the model training phase, historical backtesting data is used to optimize algorithm parameters, such as particle count, inertia weight, and learning factor.

[0066] The output of this step is the optimal portfolio decision, which includes the allocation ratio of each asset. It takes into account both traditional risk and return factors and the influence of investor psychology, achieving a balanced optimization of multiple objectives.

[0067] Application example of this implementation method: To verify the effectiveness and practicality of this implementation method, the following application examples in actual investment decision-making scenarios are shown.

[0068] Multi-source heterogeneous data acquisition and preprocessing: Table 1: Sample data collection sources and quantities;

[0069] The collected data were standardized. Text data was standardized using Z-score, and transaction data was filtered using moving average to obtain a preliminarily cleaned dataset.

[0070] Cross-modal consistency verification and anomaly filtering: Table 2: Encoder configuration and cross-modal consistency score matrix for different modal data;

[0071] Set a consistency threshold Filtering ratio threshold Anomaly identification and filtering were performed on all data sources. In this example, the consistency scores of all modalities exceeded the threshold, and no obvious anomalies were found.

[0072] Emotional cognition dual-channel feature extraction: Table 3: Examples of feature extraction results for emotional and cognitive channels;

[0073] Final quantitative cognitive bias total score This indicates that there is significant irrational behavior in the market.

[0074] Multi-source heterogeneous data fusion and investor psychology modeling: Table 4: Weight allocation for adaptive fusion gating network (March 23, 2020, the day of market low).

[0075] Influencing factors of investor psychology after integration This indicates that investor psychology has a strong influence on the market.

[0076] The time-series psychological state evolution model is used to track changes in investor psychological state, and a time-series smoothing factor is selected. (Considering the drastic market fluctuations), the final temporal psychological state representation is obtained.

[0077] Investment decision generation based on fuzzy stochastic multi-objective optimization: Table 5: Comparison of Multi-Objective Optimization Parameters and Portfolio Decision Results;

[0078] The multi-objective optimization problem is solved using an improved non-dominated sorting particle swarm optimization algorithm. Parameter settings: population size = 100, maximum number of iterations = 500, inertia weight = 0.8, cognitive factor... Social factors .

[0079] Technical effectiveness verification: Table 6: Comparison of the performance of different models during the market panic to recovery period from March to June 2020;

[0080] Table 6 shows that this implementation significantly outperforms traditional methods and general quantitative models in key performance indicators during periods of severe market volatility. In particular, this implementation demonstrates extremely strong risk control capabilities in terms of Sharpe ratio and maximum drawdown.

[0081] Table 7: Relationship between cross-modal consistency verification and information quality;

[0082] Table 7 verifies the effect of the cross-modal consistency verification mechanism on improving information quality. When the consistency threshold is set to 0.40, the system achieves a good balance between information accuracy and decision accuracy, significantly improving decision accuracy without losing too much effective information.

[0083] The experimental results above demonstrate that this implementation method has significant advantages in processing heterogeneous financial information, integrating investor sentiment and cognitive factors, and multi-objective optimization.

[0084] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A portfolio decision-making method based on fuzzy stochastic multi-objective optimization, characterized in that, Includes the following steps: Collect multi-source heterogeneous data and preprocess it to obtain preprocessed multi-source heterogeneous data, which includes text data, transaction data, behavioral data and multimedia data; Cross-modal consistency verification is performed on the preprocessed multi-source heterogeneous data to identify and remove contradictory or abnormal information, thus obtaining the verified data. Two parallel processing paths, the emotional channel and the cognitive channel, are constructed for the verified data to extract the emotional features and cognitive features of the investor's psychological state, respectively. Based on an adaptive fusion gating network, the extracted emotional and cognitive features are dynamically weighted and fused to construct an investor psychological state model. By introducing the investor psychological state model as a fuzzy variable into the multi-objective optimization model, and by using fuzzy stochastic programming to comprehensively consider multiple objectives such as return, risk and volatility resistance, the optimal portfolio decision is generated.

2. The portfolio decision-making method based on fuzzy stochastic multi-objective optimization according to claim 1, characterized in that, The cross-modal consistency verification includes the following steps: For different types of data, corresponding encoders are built to map data of different modalities to a unified semantic space; For different multi-source heterogeneous data of the same financial event, the encoder is used to obtain its representation in a unified semantic space and calculate its similarity. Introduce negative samples to calculate cross-modal consistency scores; A threshold is set based on the calculated cross-modal consistency score. When the cross-modal consistency scores of a certain type of multi-source heterogeneous data and multiple other types of multi-source heterogeneous data are all below the threshold, they are marked as abnormal information and filtered out.

3. The portfolio decision-making method based on fuzzy stochastic multi-objective optimization according to claim 1, characterized in that, The emotional feature extraction includes: Using sentiment analysis algorithms in the financial field, we extract the sentiment polarity, intensity, and specific emotion type from text and multimedia data to obtain the probability distribution vector of sentiment categories. Based on the probability distribution vector of the aforementioned emotion categories, time series analysis is used to capture the trend and fluctuation pattern of emotion changes.

4. The portfolio decision-making method based on fuzzy stochastic multi-objective optimization according to claim 1, characterized in that, The cognitive feature extraction includes: Identify patterns of cognitive biases in investor trading behavior; The identified cognitive bias patterns are quantified into numerical features for subsequent decision-making.

5. The portfolio decision-making method based on fuzzy stochastic multi-objective optimization according to claim 1, characterized in that, The adaptive fusion gating network is represented as follows: ； in, This represents the output of the adaptive fusion gating network; , They represent The first and nth feature vectors, where n represents the index of the input feature vector; For the i-th feature vector of the input The nonlinear transformation function; for Corresponding adaptive weights; Calculated in the following way: ； in, Represents the natural exponential function; Representation of features Reliability score; express Represents the product of reliability and correlation; This represents the feature index used when calculating the normalized denominator; This represents the sum of the importance indices of all features, used to normalize the weights so that their sum equals 1. Representation of features Reliability rating, express Relevance score, express The j-th eigenvector.

6. The portfolio decision-making method based on fuzzy stochastic multi-objective optimization according to claim 1, characterized in that, The dynamic weighted fusion also includes constructing a temporal psychological state evolution model: ； in, Indicates time The final investor psychological state; Indicates time This indicates the psychological state of investors; Indicates time The current mental state representation obtained through an adaptive fusion gating network; It is a time series smoothing factor.

7. The portfolio decision-making method based on fuzzy stochastic multi-objective optimization according to claim 1, characterized in that, The investor psychological state model includes factors influencing investor psychological state. Defined as a fuzzy variable, represented using triangular fuzzy numbers: ； in, A fuzzy representation of factors influencing investor psychology. The central value of the calculated investor psychological state influencing factor; This represents the lower bound of the factor influencing investor psychology. This represents the upper bound of the factor influencing investor psychology.

8. The portfolio decision-making method based on fuzzy stochastic multi-objective optimization according to claim 1, characterized in that, The multi-objective function in the multi-objective optimization model includes: Maximizing expected return: ； Minimize investment risk: ； Maximizing resistance to volatility: ； in, This represents the objective function for expected returns; This represents the objective function for investment risk. Represent the objective function for resisting volatility; Indicates the maximum value; Indicates the minimum value; Indicates the assets in the portfolio The weights; This indicates the total number of assets in the portfolio; Assets Expected rate of return; Indicates the psychological state of investors towards assets Additional impact coefficient of revenue; The covariance matrix representing the rate of return on assets; Representing vectors Transpose of; Assets Stability indicators; Assets Sensitivity coefficient to fluctuations in investor psychology.

9. The portfolio decision-making method based on fuzzy stochastic multi-objective optimization according to claim 8, characterized in that, Transform a multi-objective function into a single-objective function: ； in, This represents the transformed single objective function. Indicates the maximum value; This represents the objective function for expected returns; This represents the objective function for investment risk. Represent the objective function for resisting volatility; , , These represent the relative importance of returns, risks, and volatility resistance in decision-making, respectively. , The targets are respectively The maximum and minimum values ​​are obtained by searching within the feasible region using an optimization algorithm; , , These represent the normalized values ​​of the objective functions for return, risk, and volatility resistance, respectively, with the normalized range being [0,1].

10. A portfolio decision-making system based on fuzzy stochastic multi-objective optimization, used to execute the portfolio decision-making method based on fuzzy stochastic multi-objective optimization as described in any one of claims 1-9, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source heterogeneous data and preprocess it to obtain preprocessed multi-source heterogeneous data. The cross-modal consistency verification module is used to perform cross-modal consistency verification on preprocessed multi-source heterogeneous data, identify and remove contradictory or abnormal information, and obtain verified data. The emotional and cognitive feature extraction module constructs two parallel processing paths for the verified data: an emotional channel and a cognitive channel, to extract emotional and cognitive features of investors' psychological states, respectively. An adaptive fusion gating network module is used to dynamically weight and fuse extracted emotional and cognitive features to construct an investor psychological state model. A fuzzy stochastic multi-objective optimization module is used to generate optimal portfolio decisions.

Images