Multi-dimensional investment decision support system and dynamic income prediction method

Abstract

The invention relates to the field of finance, and discloses a multi-dimensional investment decision support system and a dynamic income prediction method, and the system comprises a multi-source heterogeneous data integration layer, a dynamic feature engineering module, a hybrid machine learning model group and a multi-dimensional decision engine which are connected in sequence. The multi-source heterogeneous data integration layer comprises a financial quantification database, an unstructured data processing engine and an industrial chain knowledge graph which are arranged in parallel, and the dynamic feature engineering module integrates a sliding time window dynamic segmentation algorithm and a feature selector based on an Attention mechanism. Structured financial data and unstructured data (such as public opinions and industrial chain supply and demand relationships) are synchronously collected through a multi-source heterogeneous data integration layer, a dynamic industrial chain knowledge graph is constructed, and the defect that a traditional system only depends on financial data is overcome through multi-dimensional data integration, so that investors can more comprehensively capture market signals, and the market quality is improved. Decision deviation caused by one-sided information is reduced, and the objectivity and accuracy of judgment are improved.

Description

Technical Field

[0001] The present invention relates to the field of financial technology, and in particular to a multi-dimensional investment decision support system and a dynamic return prediction method. Background Art

[0002] In today's financial markets, the scientific nature and accuracy of investment decisions are crucial for investors. However, existing investment decision support systems and return forecasting methods have significant flaws:

[0003] Traditional investment decision-making systems primarily rely on structured financial data, such as revenue and profit information from company financial reports. However, in real financial markets, unstructured data, such as social media sentiment, supply and demand fluctuations across the entire supply chain, and macroeconomic policy adjustments, have a significant impact on investment decisions. For example, a sudden burst of negative public opinion can cause a company's stock price to plummet in the short term, but traditional systems are unable to capture this information in a timely manner and incorporate it into their analysis.

[0004] Most traditional forecasting models are trained using fixed time windows, making them ill-suited to the dynamic nature of financial markets. Market conditions are constantly shifting, from stable to volatile to trending, and their underlying patterns change accordingly. Models trained using fixed time windows struggle to dynamically capture these changes, leading to significant deviations in forecast results as market conditions shift.

[0005] Current investment decision-making systems on the market often operate independently for modules such as risk control, return forecasting, and asset allocation. This results in a lack of effective multi-dimensional parameter linkage analysis between these modules. For example, when allocating assets, the dynamic returns of different assets at varying risk levels are not fully considered, making it impossible to optimize the overall investment portfolio. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention provides a multi-dimensional investment decision support system and a dynamic return prediction method, which solves the problem that traditional investment decision systems mainly rely on structured financial data and are not objective enough.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: a multi-dimensional investment decision support system and a dynamic return prediction method, including a multi-source heterogeneous data integration layer, a dynamic feature engineering module, a hybrid machine learning model group and a multi-dimensional decision engine connected in sequence; the multi-source heterogeneous data integration layer includes a financial quantitative database, an unstructured data processing engine and an industrial chain knowledge graph arranged in parallel; the dynamic feature engineering module integrates a sliding time window dynamic segmentation algorithm and a feature selector based on the Attention mechanism; the hybrid machine learning model group adopts a three-layer model architecture including a primary prediction layer, a dynamic correction layer and a meta-learning fusion layer; the multi-dimensional decision engine constructs a three-dimensional decision space based on risk level, expected return and asset liquidity.

[0008] Preferably, the financial quantitative database is connected to the financial data platform in real time through the API interface to collect stock price, trading volume, and bond yield data; the unstructured data processing engine uses NLP technology to build a semantic network to convert news and social media texts into public opinion indexes; the industrial chain knowledge graph dynamically updates the production, sales data and supply and demand relationship weights of upstream and downstream companies.

[0009] Preferably, the sliding time window dynamic segmentation algorithm:

[0010] formula:

[0011]

[0012] Character meaning: : The adjusted time window length. : The time window length before adjustment. : Current market volatility can usually be measured by calculating the standard deviation of asset price returns over a period of time, such as for asset price series , yield rate , volatility ,in is the mean return rate, is the sample size. : Low volatility threshold. : High volatility threshold. : Time window expansion coefficient, . : time window reduction coefficient, .

[0013] Preferably, the feature selector based on the Attention mechanism:

[0014]

[0015]

[0016]

[0017] , : are the i-th and j-th feature vector representations in the input feature vector respectively; , : learnable weight matrix, : represents the correlation score between the i-th eigenvector and the j-th eigenvector, Attention weights normalized by the Softmax function; : The context vector obtained by weighted summation of all feature vectors based on the attention weights.

[0018] Preferably, the long short-term memory network of the primary prediction layer controls the update of cell state and hidden state through input gate, forget gate and output gate, and the random forest performs feature selection and node splitting through the Gini index, and the two are combined to output preliminary prediction results.

[0019] Preferably, the Bayesian optimization algorithm of the dynamic correction layer models the objective function through Gaussian process, combines the acquisition function with balanced exploration and utilization, and finds the optimal parameter combination within a relatively small number of iterations.

[0020] Preferably, the meta-learning fusion layer takes the output of the primary prediction layer and the dynamic correction layer as input, trains the meta-model to learn the model weights, and the final prediction result is the weighted sum of the prediction results of each model.

[0021] Preferably, the multi-dimensional decision engine dynamically embeds the investor risk preference coefficient α∈[0,1] into the decision model, and recommends a low-risk, high-liquidity combination when α=0, and recommends a high-risk, high-return combination when α=1.

[0022] A multi-dynamic profit forecasting method comprises the following steps:

[0023] S1. Multi-source heterogeneous data integration: collecting financial quantitative data, processing unstructured text data, and building an industrial chain knowledge graph; S2. Dynamic feature engineering processing: adaptively adjusting the time window length, screening key features and weighting them based on the Attention mechanism; S3. Hybrid machine learning model prediction: generating the final prediction results through the primary prediction layer, dynamic correction layer, and meta-learning fusion layer; S4. Multi-dimensional decision engine analysis: constructing a three-dimensional decision space, combining investor portraits to generate personalized investment advice; S5. System dynamic iteration: optimizing each system module based on real-time feedback data.

[0024] Preferably, the meta-learning fusion layer uses the Stacking ensemble learning method to take the prediction results of the primary prediction layer and the dynamic correction layer as input, trains the meta-model learning model weights, and realizes weighted fusion of multi-model prediction results.

[0025] The present invention provides a multi-dimensional investment decision support system and a dynamic return forecasting method. It has the following beneficial effects:

[0026] 1. This invention uses a multi-source heterogeneous data integration layer to synchronously collect structured financial data and unstructured data (such as public opinion and industrial chain supply and demand relationships) and construct a dynamic industrial chain knowledge graph. This multi-dimensional data integration makes up for the shortcomings of traditional systems that rely solely on financial data, enabling investors to capture market signals more comprehensively, reduce decision-making biases caused by one-sided information, and improve the objectivity and accuracy of judgments.

[0027] 2. The present invention utilizes the STWDS algorithm, a dynamic feature engineering module, to automatically adjust the time window based on market volatility: extending the time window to capture long-term trends during low volatility and shortening it to respond to short-term fluctuations during high volatility. This addresses the problem of fixed time window models being unable to adapt to market state transitions. Furthermore, a feature selector based on the Attention mechanism filters and weights key factors from over 200 dimensional features in real time. This, combined with Bayesian optimization and stacking ensemble learning within a hybrid machine learning model group, dynamically optimizes parameters and integrates the advantages of multiple models. This mechanism enables the system to rapidly respond to market changes, significantly improving forecast accuracy under varying market conditions and reducing errors during trend transitions.

[0028] 3. The present invention constructs a three-dimensional decision space of "risk, return, and liquidity" through a multi-dimensional decision engine, linking the risk control, return forecast, and asset allocation modules for analysis rather than operating independently. At the same time, the system generates personalized recommendations based on the investor's risk preference coefficient α (01): when α=0, it recommends low-risk, high-liquidity assets; when α=1, it focuses on high-risk, high-return targets, and builds a profile based on investment objectives and timeframes. This multi-dimensional linkage and personalized mechanism breaks the modular separation of traditional systems, achieving a dynamic balance between risk and return while meeting the needs of different investors, improving the scientific nature and return potential of portfolio allocation. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a perspective view of the present invention. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0031] Example:

[0032] Please see the attached Figure 1 An embodiment of the present invention provides a multi-dimensional investment decision support system and dynamic return forecasting method, comprising a multi-source heterogeneous data integration layer, comprising a financial quantitative database, an unstructured data processing engine, and an industry chain knowledge graph. The financial quantitative database accesses APIs from authoritative financial data platforms such as Bloomberg and Wind to obtain comprehensive and real-time financial quantitative data, including key data indicators for various financial products such as stock prices, trading volumes, and bond yields, providing a solid data foundation for subsequent analysis. The unstructured data processing engine utilizes natural language processing (NLP) technology to construct a semantic network, converting unstructured text data such as news reports and social media comments into a quantifiable public opinion index. For example, a sentiment analysis algorithm can be used to determine whether the market sentiment towards a particular industry or company is positive, negative, or neutral, thereby providing a public opinion reference for investment decisions. The industry chain knowledge graph is used to collect production, sales, inventory, and other information from upstream and downstream enterprises in the industry chain in real time, constructing a dynamically updated industry chain knowledge graph. By analyzing the weights of nodes (enterprises) and edges (supply and demand relationships) in the graph, the mutual influence relationships among various links in the industrial chain can be clearly displayed. For example, the impact of changes in the production capacity of a key raw material supplier on multiple downstream industry companies can be used to help investors grasp investment opportunities and risks at the industrial chain level.

[0033] The multi-source heterogeneous data integration layer is connected to a dynamic feature engineering module, which utilizes the STWDS algorithm and adaptively adjusts the data sampling period based on market volatility. When market volatility is low and relatively stable, the time window is appropriately widened to capture longer-term trend information. When market volatility rises sharply and volatile conditions occur, the time window is narrowed to quickly capture short-term market fluctuations. For example, in the stock market, if stock price fluctuations are minimal over a period of time, the algorithm automatically extends the time window to 30 days to analyze long-term trends. If stock price fluctuations are significant, the time window is shortened to 3 days to promptly respond to short-term market fluctuations. This includes a feature selector based on the attention mechanism, which calculates the importance scores of over 200 features in real time. These features cover multiple dimensions, including financial data, public opinion indexes, and supply and demand relationships in the industry chain. This attention mechanism enables the model to focus on the most critical features for the current market state and ignore redundant information, improving model training efficiency and prediction accuracy. For example, when analyzing investment opportunities in an emerging industry, if the current market is more focused on industry policy guidance and technological innovation progress, the feature selector will assign higher weights to features such as relevant policy indicators and technological R&D investment.

[0034] The sliding time window dynamic segmentation algorithm:

[0035] formula:

[0036]

[0037] : The adjusted time window length. : The time window length before adjustment. : Current market volatility can usually be measured by calculating the standard deviation of asset price returns over a period of time, such as for asset price series , yield rate , volatility ,in is the mean return rate, is the sample size. : Low volatility threshold. : High volatility threshold. : Time window expansion coefficient, . : time window reduction coefficient, .

[0038] First, calculate the current market volatility σ based on historical asset price data. Then, compare the calculated volatility σ with the pre-set thresholds σlow and σhigh: If σ<σlow, it means that the market is in a low-volatility, stable state. At this time, multiply the original time window length Told by the expansion coefficient k1 to obtain a new time window length Tnew in order to obtain longer-term trend information. If σlow≤σ≤σhigh, the market volatility is within the normal range, and the time window length remains unchanged, that is, Tnew=Told. If σ>σhigh, the market volatility is severe, and the original time window length Told is divided by the reduction coefficient k2 to obtain a new time window length Tnew, so as to quickly capture short-term market changes.

[0039] The feature selector based on the Attention mechanism:

[0040]

[0041]

[0042]

[0043] , : They are respectively represented by the i-th and j-th eigenvectors in the input eigenvector. Each eigenvector can contain the encoding of various information such as financial indicators, public opinion index, industry chain related data, etc. , : A learnable weight matrix used to perform linear transformation on the feature vector to extract more valuable information for calculating the correlation between features. : represents the correlation score between the i-th eigenvector and the j-th eigenvector, The attention weight after normalization by the Softmax function reflects the The eigenvector for The importance of the eigenvectors, and , is the total number of eigenvectors. : The context vector obtained by weighted summation of all feature vectors based on attention weights, which integrates all feature information and highlights the Important features related to each feature.

[0044] First, for the input feature vector set ,First, through linear transformation, using the weight matrix and Transform each eigenvector separately, calculate the dot product between the eigenvectors, and get the correlation score .

[0045] Next, use the Softmax function to score the correlation Perform normalization to obtain attention weight , so that the importance of different features is comparable and the sum is 1.

[0046] Finally, according to the calculated attention weight , perform weighted summation on all feature vectors to obtain the context vector corresponding to each feature These context vectors comprehensively consider the information of all features and weight them according to their importance. They can then be used for model training and prediction, focusing on the features that are most critical to the current market status.

[0047] The dynamic feature engineering module is connected to a hybrid machine learning model group, which adopts a three-layer model architecture and includes:

[0048] The primary prediction layer uses a combination of long short-term memory (LSTM) and random forest models. LSTM excels at processing time series data and can capture long-term dependencies in financial data, such as predicting stock price trends over time. Random forests, through a voting mechanism among multiple decision trees, comprehensively analyze multiple features, enhancing the model's generalization capabilities and reducing the risk of overfitting. Together, these two methods provide preliminary predictions of financial data from multiple perspectives.

[0049] Specifically: Primary prediction layer

[0050]

[0051]

[0052]

[0053]

[0054]

[0055] : The input vector at each moment, such as data on financial market prices and trading volumes. : The hidden state vector at each moment is used to convey the information of the time series. : The cell state vector at each moment is used to memorize long-term information. : Input gate, which controls the degree to which the current input information enters the cell state. σ is the Sigmoid function, which maps the result to between 0 and 1. : Forget gate, controls the degree of retention of cell state information at the previous moment. : Output gate, which controls the degree to which cell state information is output to the hidden state. , , , : Weight matrices input to various gates and cell states. , , , : The weight matrix from hidden state to each gate and cell state. , , : Weight matrix from cell state to each gate; , , , Bias vectors for each gate and cell state; Element-wise multiplication operation.

[0056] First, based on the current input and the hidden state at the previous moment , cell status , calculate the input gate through the Sigmoid function and the Forget Gate The value of determines the inflow of current input information and the degree of retention of cell state information at the previous moment.

[0057] Then, the candidate cell states are calculated

[0058]

[0059] Combined with the results of the forget gate and input gate, update the cell state

[0060] .

[0061] Next, the output gate is calculated by the Sigmoid function The value of the cell state after the update Calculate the hidden state at the current moment

[0062] , hidden state Can be used as input or output for subsequent predictions.

[0063] Random forest part: Random forest consists of multiple decision trees. The decision tree construction process involves feature selection and node splitting. Taking the Gini index as an example of feature selection metric,

[0064]

[0065] : The probability distribution vector of the sample belonging to a certain category, , is the total number of categories. : The sample belongs to The probability of the class. : The Gini index of a node is used to measure the impurity of the node. The smaller the value, the purer the node sample. , : are the Gini indexes of the left and right child nodes after splitting. , : are the number of samples of the left child node and the right child node respectively. : The number of samples of the parent node. : The Gini index after splitting is used to evaluate the effect of feature splitting. The smallest feature and splitting point are used to split the node.

[0066] First, for each node, calculate the Gini index of the current node .

[0067] Then, for each possible split point of each feature, calculate the Gini index of the left and right child nodes after the split and , and calculate the Gini index after splitting based on the number of child node samples .

[0068] Continue to select The smallest feature and splitting point are used to split the node, and this process is repeated until the stopping condition is met (such as the number of node samples is less than the threshold, the Gini index is less than the threshold, etc.), and a decision tree is constructed.

[0069] Repeat the above steps to build multiple decision trees to form a random forest. During prediction, the final prediction result is obtained by voting (classification problems) or averaging (regression problems) the multiple decision trees.

[0070] Dynamic Correction Layer: This layer uses a Bayesian optimization algorithm to adjust model parameters in real time. As market data continuously updates, model parameters need to be dynamically optimized to adapt to new market conditions. Bayesian optimization constructs a probabilistic model of the objective function to find the optimal parameter combination within a limited number of trials, ensuring accurate model predictions. For example, when new policy changes or major market events occur, the dynamic correction layer rapidly adjusts model parameters, enabling the model to adapt promptly.

[0071] Bayesian optimization finds the optimal parameters by constructing a surrogate model of the objective function (usually a Gaussian process) and balancing exploration (finding new possible areas) and exploitation (using known good areas):

[0072]

[0073]

[0074] : The output value of the objective function, in this system, is the error or benefit of the model prediction.

[0075] : Gaussian process is a probabilistic model used to model the objective function.

[0076] : The mean function of the Gaussian process, which means that when the input is The mean prediction of the objective function when .

[0077] : The kernel function of the Gaussian process is used to measure the input and The similarities between them.

[0078] : Acquisition function, used to guide the selection of the next parameter sampling point.

[0079] : Exploration utilizes a balance parameter that changes gradually with iterations to control the degree of exploration of new areas.

[0080] :When input is The mean prediction of the objective function is .

[0081] When calculating, first, according to the existing parameter sampling points and its corresponding objective function value , build a Gaussian process model and determine the mean function and kernel function .

[0082] Then, through the acquisition function Calculation at different parameter points The acquisition function value at , selects the parameter point with the largest acquisition function value as the next sampling point for testing, and obtains the new objective function value.

[0083] Add new sampling points and objective function values ​​to the existing data and update the Gaussian process model.

[0084] Repeat the above steps and iterate continuously until the stopping conditions are met (such as reaching the maximum number of iterations, the objective function value converges, etc.) to find the optimal model parameters

[0085] Meta-learning fusion layer: Dynamically generates weights for different models through the stacking ensemble learning method. Taking the prediction results of multiple models from the primary prediction layer and the dynamic correction layer as input, the meta-learning fusion layer automatically assigns the optimal weight to each model based on historical data and current market conditions, and combines the advantages of multiple models to obtain the final accurate prediction result. Specifically:

[0086]

[0087] : The final prediction result.

[0088] : The number of models involved in the integration, which is the total number of models in the primary prediction layer and dynamic correction layer in this system.

[0089] :No. The prediction results of the model.

[0090] :No. The weights of the models, and .

[0091] First, multiple models of the primary prediction layer and the dynamic correction layer are used to predict the training data and obtain the prediction results of each model. .

[0092] Then, using these prediction results As input features, the true label is used as output, and a meta-model (such as a linear regression model) is trained to learn how to combine the prediction results of these models to obtain the weight of each model. .

[0093] In the prediction stage, the prediction results of each model for new data are , according to the learned weights Perform weighted summation to obtain the final prediction result

[0094] The hybrid machine learning model group is connected to a multidimensional decision-making engine, which constructs a comprehensive three-dimensional decision space that comprehensively considers the portfolio's risk level, expected return, and asset liquidity. Within this space, different investment portfolios correspond to different coordinate points, allowing investors to intuitively visualize the trade-offs between risk, return, and liquidity across various investment options. For example, high-risk, high-return portfolios are located farther away from low-risk, low-return portfolios in the space, and liquidity may also vary. An investor profile is constructed based on information such as the investor's risk appetite, investment objectives, and investment horizon. By dynamically embedding a risk appetite coefficient α∈[0,1] into the decision model, personalized investment decision recommendations are provided for each investor. A risk appetite coefficient α of 0 indicates an investor's extreme risk aversion and a preference for low-risk, low-return, but highly liquid portfolios. A risk appetite coefficient of 1 indicates a high risk tolerance and a preference for high-risk, high-return investment opportunities.

[0095] A dynamic return prediction method for a multi-dimensional investment decision support system includes the following steps:

[0096] S1. Multi-source heterogeneous data integration

[0097] S101. Financial quantitative data collection

[0098] By accessing the API of an authoritative financial data platform, we can obtain quantitative data such as stock prices, trading volumes, and bond yields in real time and build a financial quantitative database.

[0099] S102. Unstructured Data Processing

[0100] Using natural language processing technology, we transform unstructured texts such as news reports and social media comments into quantifiable public opinion indices, and use sentiment analysis to determine market sentiment trends (positive / negative / neutral).

[0101] S103. Construction of industrial chain knowledge graph

[0102] Real-time collection of production, sales, inventory and other information from upstream and downstream enterprises in the industrial chain, construction of a dynamically updated knowledge graph, analysis of the weights of nodes (enterprises) and edges (supply and demand relationships), and the disclosure of the mutual influence between links in the industrial chain.

[0103] S2. Dynamic feature engineering processing

[0104] S201. Adaptive time window adjustment

[0105] Dynamically adjust the data sampling period according to market volatility: expand the time window when volatility is low to capture long-term trends; narrow the time window when volatility is high to focus on short-term changes; keep the time window length unchanged during normal fluctuations.

[0106] S202. Multi-dimensional feature selection and weighting

[0107] Based on the attention mechanism, the importance scores of more than 200-dimensional features (financial data, public opinion index, industrial chain relationships, etc.) are calculated, key features are automatically screened and weighted, and redundant information interference is reduced.

[0108] S3. Hybrid Machine Learning Model Prediction

[0109] S301. Primary prediction layer modeling

[0110] Fusion of the Long Short-Term Memory (LSTM) and Random Forest models: LSTM captures the time series dependencies of financial data; Random Forest comprehensively analyzes features through a voting mechanism of multiple decision trees to output preliminary prediction results.

[0111] S302. Dynamic Correction Layer Optimization

[0112] Using the Bayesian optimization algorithm, model parameters are dynamically adjusted according to real-time market data. Through Gaussian process modeling and acquisition function optimization, the optimal parameter combination is found within a relatively small number of iterations.

[0113] S303. Meta-learning fusion layer integration

[0114] The stacking ensemble learning method is adopted, with the output of the primary prediction layer and the dynamic correction layer as input, and the model weights are automatically assigned through a meta-model (such as linear regression) to generate the final prediction results.

[0115] S4. Multi-dimensional decision engine analysis

[0116] S401. Construction of three-dimensional decision space

[0117] By comprehensively considering the investment portfolio's risk level, expected return, and asset liquidity, a three-dimensional decision-making space is constructed to intuitively display the balance relationship between different investment options.

[0118] S402. Personalized investment advice generation

[0119] Build a profile based on the investor's risk preference coefficient α (α∈[0,1]), investment goals, term, and other information, dynamically embed α into the decision-making model, and output an adapted investment portfolio recommendation:

[0120] When α=0, a low-risk, high-liquidity portfolio is recommended;

[0121] When α=1, a high-risk, high-return combination is recommended.

[0122] S5. System dynamic iteration

[0123] Based on real-time feedback data, we cyclically optimize multi-source data integration logic, feature engineering parameters, model architecture, and decision-making strategies to continuously improve the system's prediction accuracy and decision-making support capabilities.

[0124] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A multi-dimensional investment decision support system, characterized by: It includes a multi-source heterogeneous data integration layer, a dynamic feature engineering module, a hybrid machine learning model group and a multi-dimensional decision engine connected in sequence; the multi-source heterogeneous data integration layer includes a parallel financial quantitative database, an unstructured data processing engine and an industrial chain knowledge graph; the dynamic feature engineering module integrates a sliding time window dynamic segmentation algorithm and a feature selector based on the Attention mechanism; the hybrid machine learning model group adopts a three-layer model architecture including a primary prediction layer, a dynamic correction layer and a meta-learning fusion layer; the multi-dimensional decision engine constructs a three-dimensional decision space based on risk level, expected return and asset liquidity.

2. A multi-dimensional investment decision support system according to claim 1, characterized in that: The financial quantitative database is connected to the financial data platform in real time through the API interface to collect stock price, trading volume, and bond yield data; the unstructured data processing engine uses NLP technology to build a semantic network to convert news and social media texts into public opinion indexes; the industrial chain knowledge graph dynamically updates the production and sales data and supply and demand relationship weights of upstream and downstream companies.

3. A multi-dimensional investment decision support system according to claim 1, characterized in that: The sliding time window dynamic segmentation algorithm: formula: Character meaning: : The adjusted time window length; : Time window length before adjustment; : Current market volatility can usually be measured by calculating the standard deviation of asset price returns over a period of time, such as for asset price series , yield rate , volatility ,in is the mean return rate, is the sample size; : low volatility threshold; :High volatility threshold; : Time window expansion coefficient, ; : time window reduction coefficient, .

4. A multi-dimensional investment decision support system according to claim 1, characterized in that: The feature selector based on the Attention mechanism: , : are the i-th and j-th feature vector representations in the input feature vector respectively; , : learnable weight matrix, : represents the correlation score between the i-th eigenvector and the j-th eigenvector, Attention weights normalized by the Softmax function; : The context vector obtained by weighted summation of all feature vectors based on the attention weights.

5. A multi-dimensional investment decision support system according to claim 1, characterized in that: The long short-term memory network of the primary prediction layer controls the update of cell state and hidden state through input gate, forget gate and output gate, and the random forest performs feature selection and node splitting through the Gini index, and the two are combined to output preliminary prediction results.

6. A multi-dimensional investment decision support system according to claim 1, characterized in that: The Bayesian optimization algorithm of the dynamic correction layer uses Gaussian process modeling to model the objective function, combines the acquisition function to balance exploration and utilization, and finds the optimal parameter combination within a relatively small number of iterations.

7. A multi-dimensional investment decision support system according to claim 1, characterized in that: The meta-learning fusion layer takes the output of the primary prediction layer and the dynamic correction layer as input, trains the meta-model to learn the model weights, and the final prediction result is the weighted sum of the prediction results of each model.

8. A multi-dimensional investment decision support system according to claim 1, characterized in that: The multi-dimensional decision engine dynamically embeds the investor risk preference coefficient α∈[0,1] into the decision model, and recommends a low-risk, high-liquidity portfolio when α=0, and recommends a high-risk, high-return portfolio when α=1.

9. A multi-dynamic return forecasting method, comprising a multi-dimensional investment decision support system according to any one of claims 18, characterized in that: The following steps are involved: S1. Multi-source heterogeneous data integration: collecting financial quantitative data, processing unstructured text data, and building an industry chain knowledge graph; S2. Dynamic feature engineering: adaptively adjusting the time window length, screening key features based on the attention mechanism, and weighting them; S3. Hybrid machine learning model prediction: Generates the final prediction result through the primary prediction layer, dynamic correction layer, and meta-learning fusion layer; S4. Multi-dimensional decision engine analysis: Constructing a three-dimensional decision space and generating personalized investment recommendations based on investor profiles; S5. Dynamic system iteration: Optimize each system module based on real-time feedback data.

10. The dynamic income prediction method of a multi-dimensional investment decision support system according to claim 9, characterized in that: The meta-learning fusion layer uses the Stacking ensemble learning method to take the prediction results of the primary prediction layer and the dynamic correction layer as input, trains the meta-model learning model weights, and realizes the weighted fusion of the multi-model prediction results.

Images