Financial investment risk value prediction system based on big data group statistical algorithm

Abstract

The invention discloses a financial investment risk value prediction system based on a big data group statistical algorithm, and relates to the technical field of financial investment risk value prediction. The system comprises a multi-source data acquisition module, the output end of the multi-source data acquisition module is electrically connected with a dynamic feature screening module, and the output end of the dynamic feature screening module is electrically connected with a space-time risk conduction model. According to the method, multiple modules of multi-source data acquisition, dynamic feature screening, space-time risk conduction modeling, risk value prediction, intelligent blocking decision making and the like are integrated, all the modules work cooperatively, through deep mining of massive multi-source data and comprehensive application of multiple algorithms, the risk features in the financial market can be accurately recognized, and the risk prediction efficiency of the financial market is improved. Under a complex multi-asset linkage scene, the false alarm rate of the system is reduced by 50%, the prediction accuracy is improved by 25%-30%, investment loss caused by misjudgment is effectively avoided, and a reliable basis is provided for decision making of investors.

Description

Technical Field

[0001] The present invention relates to the technical field of financial investment risk value prediction, and particularly to a financial investment risk value prediction system based on a big data group statistical algorithm. Background Art

[0002] In the field of financial investment, accurately predicting the risk value is crucial for ensuring investment security and improving the income level. However, traditional financial investment risk prediction models rely on fixed feature combinations and are difficult to adapt to the dynamic changes of the financial market. When facing sudden situations such as policy adjustments and black swan events, their limitations are particularly prominent. In the feature screening link, traditional methods usually adopt manual or fixed algorithms, such as feature dimensionality reduction technology based on PCA, which cannot respond to market changes in real time, resulting in a lag in the model's capture of market information.

[0003] Existing systems also have obvious deficiencies in risk conduction analysis. For example, it is difficult to capture risk conduction paths across assets and regions, and lack the ability to provide real-time early warnings for systemic risks. Most existing graph models are static networks with a low frequency of edge weight updates. Taking a static financial knowledge graph as an example, it cannot reflect the dynamic changes of the financial market in a timely manner. In addition, traditional models have poor interpretability, are difficult to clearly show the origin and conduction logic of risks, and cannot meet the requirements of regulatory compliance. In terms of blocking strategies, existing methods mostly rely on rules or historical experience and lack an intelligent decision-making mechanism driven by dynamic features, making it difficult to play an effective role in the complex and ever-changing financial market.

[0004] Therefore, we provide a financial investment risk value prediction system based on a big data group statistical algorithm to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a financial investment risk value prediction system based on a big data group statistical algorithm, which solves the problems that the existing financial investment risk value prediction system cannot respond to market changes in real time, has a low frequency of edge weight updates, and lacks an intelligent decision-making driven by dynamic features through the cooperation of a dynamic feature screening module, a spatio-temporal risk conduction model, a risk value prediction module, and an intelligent blocking decision module.

[0006] To solve the above technical problems, the present invention is realized through the following technical solutions:

[0007] The present invention relates to a financial investment risk value prediction system based on a big data swarm computing algorithm, including a multi-source data acquisition module. The output end of the multi-source data acquisition module is electrically connected to a dynamic feature screening module. The output end of the dynamic feature screening module is electrically connected to a spatio-temporal risk conduction model. The output end of the spatio-temporal risk conduction model is electrically connected to a risk value prediction module. The output end of the risk value prediction module is electrically connected to an intelligent blocking decision module; the output end of the intelligent blocking decision module is electrically connected to the input end of the spatio-temporal risk conduction model, and the output end of the intelligent blocking decision module is electrically connected to the input end of the dynamic feature screening module.

[0008] The present invention is further configured such that the key feature generation method of the dynamic feature screening module includes: optimizing the initial population of the feature subset by a genetic algorithm, and the fitness function is:

[0009] Fitness(S) = α·SharpeRatio(S) + β·VaR(S) + γ·Corr(S)

[0010] Where S is the feature subset, α, β, and γ are weight coefficients, SharpeRatio is the Sharpe ratio, VaR is the value at risk, and Corr is the feature correlation.

[0011] The present invention is further configured such that the graph structure dynamic adjustment mechanism of the spatio-temporal risk conduction model includes: when it is detected that the feature fluctuation exceeds a threshold (such as the single-day volatility > 3σ), the associated nodes are expanded by the following formula:

[0012] ΔA t = GCN(X t , A t-1 )·δ·ReLU(σ(f i t ) - θ)

[0013] Where ΔA t is the adjacency matrix increment, δ is the expansion step size, θ is the fluctuation threshold, and σ is the normalization function.

[0014] The present invention is further configured such that the hierarchical architecture of the risk value prediction module includes:

[0015] Basic layer: Using LightGBM and XGBoost to generate an initial prediction;

[0016] Optimization layer: Adjusting the model parameters through a quantum particle swarm algorithm, and the optimization objective function is:

[0017]

[0018] Where is the mean absolute error, λ is the regularization coefficient, For the L2 model weights.

[0019] The present invention is further configured such that the reinforcement learning model of the intelligent blocking decision module includes: The reward function of the Markov decision process (MDP) is designed as:

[0020] R t = ω1·ΔVaR t + ω2·Cost t + ω3·Diversity t

[0021] Wherein, ΔVaR t is the risk reduction amplitude (unit: BP), Cost t is the policy execution cost (unit: yuan), and Diversity t is the portfolio diversity index (calculated based on Shannon entropy).

[0022] The present invention is further configured such that the real-time optimization method of the dynamic feature screening module and the spatio-temporal risk conduction model includes: decoupling design of the feature screening layer and the graph update layer, with a feature screening period of 5 seconds and a graph update period of 30 seconds; the latency optimization formula under edge computing deployment:

[0023] Latency = EdgeProcess(f i t ) + CentralProcess(A t )

[0024] Wherein, EdgeProcess is the edge node processing time (<50 ms), and CentralProcess is the central node processing time (<150 ms).

[0025] The present invention is further configured such that the heat map generation method of the spatio-temporal risk conduction model includes: The mapping relationship between key features and risk conduction paths is calculated by the following formula:

[0026]

[0027] Wherein, AttentionScore is the attention weight of the feature and the edge, and Conductance is the conduction strength of the edge.

[0028] The present invention is further configured such that the data quality assessment method of the multi-source data acquisition module includes: The validation formula for the effectiveness of real estate transaction data:

[0029]

[0030] Among them, y is the unit price of second-hand housing, and Income is the per capita income of this region. If the housing price exceeds twice the annual income, it is marked as abnormal data.

[0031] The present invention is further configured such that the model lightweighting method of the spatio-temporal risk conduction model includes: using TensorRT to optimize the inference speed of the graph neural network, and the optimized computational complexity is:

[0032]

[0033] Among them, N l is the number of nodes in the l-th layer, and k l is the compression factor (typical value 8 - 16).

[0034] The present invention is further configured such that the compliance design of the risk value prediction module includes: the feature importance audit module records the key features and their contribution degrees of each prediction, and the calculation formula is:

[0035]

[0036] Among them, T is the time window length, and α is the gradient of feature f i at time t.

[0037] The present invention has the following beneficial effects:

[0038] 1. The present invention integrates multiple modules such as multi-source data collection, dynamic feature screening, spatio-temporal risk conduction modeling, risk value prediction, and intelligent blocking decision-making. Each module works in coordination. Through in-depth mining of massive multi-source data and comprehensive application of various algorithms, it can accurately identify risk features in the financial market. In complex multi-asset linkage scenarios, the system reduces the false alarm rate by 50%, and at the same time increases the prediction accuracy by 25% - 30%, effectively avoiding investment losses caused by misjudgment and providing a reliable basis for investors' decisions.

[0039] 2. The present invention can quickly capture the rapidly changing market dynamics and respond to market changes in a timely manner. For high-frequency trading, the system can meet its strict requirements for timeliness, help investors grasp fleeting trading opportunities, and effectively prevent risks during high-frequency trading.

[0040] 3. The present invention uses a heat map and feature-path two-way mapping technology to clearly and intuitively present the complex risk conduction process. Regulatory authorities and investors can clearly trace the origin of risks, understand the conduction path of risks in each link, and the basis for model decisions. This high degree of interpretability not only helps investors deeply understand investment risks and make more reasonable investment decisions, but also meets the requirements of regulatory agencies for the compliance of financial market supervision. Description of the Drawings

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below.

[0042] Figure 1 It is the overall system schematic diagram of a financial investment risk value prediction system based on a big data group statistical algorithm;

[0043] Figure 2 It is the composition and interaction schematic diagram of the multi-source data acquisition module in a financial investment risk value prediction system based on a big data group statistical algorithm;

[0044] Figure 3 It is the composition and interaction schematic diagram of the dynamic feature screening module in a financial investment risk value prediction system based on a big data group statistical algorithm;

[0045] Figure 4 It is the composition and interaction schematic diagram of the spatio-temporal risk conduction model in a financial investment risk value prediction system based on a big data group statistical algorithm;

[0046] Figure 5 It is the composition and interaction schematic diagram of the risk value prediction module in a financial investment risk value prediction system based on a big data group statistical algorithm.

[0047] Figure 6 It is the composition and interaction schematic diagram of the intelligent blocking decision module in a financial investment risk value prediction system based on a big data group statistical algorithm. Detailed implementation manners

[0048] Next, the technical solutions in the embodiments of the present invention will be described in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0049] Embodiment 1

[0050] Please refer to Figures 1 - 6, the present invention is a financial investment risk value prediction system based on a big data group statistical algorithm, including a multi-source data collection module. The output end of the multi-source data collection module is electrically connected to a dynamic feature screening module. The output end of the dynamic feature screening module is electrically connected to a spatio-temporal risk conduction model. The output end of the spatio-temporal risk conduction model is electrically connected to a risk value prediction module. The output end of the risk value prediction module is electrically connected to an intelligent blocking decision module. The output end of the intelligent blocking decision module is electrically connected to the input end of the spatio-temporal risk conduction model, and the output end of the intelligent blocking decision module is electrically connected to the input end of the dynamic feature screening module. The multi-source data collection module includes a data crawler cluster, a data quality assessment sub-module, and a distributed storage unit. The data crawler cluster includes a securities trading data collector, a real estate trading data API, and a macroeconomic data crawler. The data quality assessment sub-module includes an outlier detection algorithm and a data validity verification formula. The distributed storage unit includes a time-series database and a graph database. The dynamic feature screening module includes a genetic algorithm engine, a reinforcement learning controller, and a feature mapping unit. The genetic algorithm engine includes a feature subset encoding, a crossover and mutation operator, and a fitness function calculation unit. The reinforcement learning controller includes a policy network, a value network, and a dynamic reward function. The feature mapping unit includes an attention weight calculator and a node attribute update interface. The spatio-temporal risk conduction model includes a dynamic knowledge graph construction subsystem, a spatio-temporal graph neural network, and an adversarial training module. The dynamic knowledge graph construction subsystem includes a multi-source data fuser, a time-series edge weight update module, and an abnormal node expansion trigger. The spatio-temporal graph neural network includes a parallel structure of a spatial convolutional layer (GCN) and a temporal convolutional layer (TCN), a spatio-temporal attention mechanism, and a risk heat map generator. The adversarial training module includes a perturbation data generator and a robustness loss function. The risk value prediction module includes a basic model layer, a quantum optimization layer, and a dynamic calibration layer. The basic model layer includes a parallel training framework of LightGBM and XGBoost and a multi-objective loss function. The quantum optimization layer includes a quantum particle swarm optimizer, a hyperparameter search space, and an optimization objective function. The dynamic calibration layer includes a Bayesian posterior correction module and an extreme risk enhancer. The intelligent blocking decision module includes a reinforcement learning decision engine, a policy execution interface, and an interpretability module. The reinforcement learning decision engine includes a state encoder, an action space design, and a reward function calculator. The policy execution interface includes an algorithmic trading API and a circuit breaker trigger. The interpretability module includes a feature contribution calculator and a conduction path visualizer.

[0051] Specifically: The data crawler cluster collects multi-domain data through the securities trading data collector, real estate trading data API, and macroeconomic data crawler. The data quality assessment sub-module uses outlier detection algorithms and data validity verification formulas to clean and verify the collected data. The cleaned data is stored in the time-series database and graph database. The genetic algorithm engine uses feature subset encoding, crossover and mutation operators, and fitness function calculation units to preliminarily screen the data and form an initial feature subset. The policy network and value network of the reinforcement learning controller further optimize the feature subset based on the dynamic reward function. The feature mapping unit maps the selected key features to the knowledge graph node attributes through the attention weight calculator and node attribute update interface. The multi-source data fusion device of the dynamic knowledge graph construction subsystem integrates multi-source data. The time-series edge weight update module updates the edge weights according to the algorithm. The abnormal node expansion trigger expands the graph nodes when detecting abnormal feature fluctuations. The spatial convolutional layer (GCN) and temporal convolutional layer (TCN) of the spatio-temporal graph neural network extract spatio-temporal features in parallel, capture the risk conduction path by combining the spatio-temporal attention mechanism, and visualize the conduction path through the risk heat map generator. The perturbation data generator of the adversarial training module generates adversarial samples to train the model to enhance its robustness. The LightGBM and XGBoost parallel training framework in the basic model layer generates initial risk prediction values based on the multi-objective loss function. The quantum optimization layer uses the quantum particle swarm optimizer to adjust the model parameters in the hyperparameter search space according to the optimization objective function. The dynamic calibration layer calibrates the prediction results through the Bayesian posterior correction module and extreme risk enhancer. The state encoder of the reinforcement learning decision engine encodes the graph state and feature information, and generates the optimal intervention strategy by combining the action space design and reward function calculator. The policy execution interface executes the trading strategy through the algorithm trading API and circuit breaker mechanism trigger. The interpretability module provides a basis for decision-making through the feature contribution calculator and conduction path visualizer, and feeds the policy back to the dynamic feature screening module and spatio-temporal risk conduction model to achieve the dynamic optimization of the system.

[0052] Embodiment 2

[0053] Please refer to Figures 2 - 6 , the key feature generation method of the dynamic feature screening module includes: the genetic algorithm optimizes the initial population of the feature subset, and the fitness function is:

[0054] Fitness(S) = α·SharpeRatio(S) + β·VaR(S) + γ·Corr(S)

[0055] where S is the feature subset, α, β, and γ are weight coefficients, SharpeRatio is the Sharpe ratio, VaR is the value at risk, and Corr is the feature correlation.

[0056] Optimize the initial population of feature subsets through a genetic algorithm, guided by a fitness function that includes the Sharpe ratio, value at risk, and feature correlation. On the one hand, considering both returns (Sharpe ratio), risks (value at risk), and the relationships between features, ensure that the selected feature set can comprehensively reflect the risk and return characteristics of financial investments; on the other hand, perform intelligent optimization on the feature set from the initial stage to improve the quality of the input data for the subsequent risk prediction model, thereby enhancing the accuracy and reliability of risk value prediction.

[0057] The graph structure dynamic adjustment mechanism of the spatio-temporal risk conduction model includes: when it is detected that the feature fluctuation exceeds the threshold (such as the single-day volatility > 3σ), expand the associated nodes through the following formula:

[0058] ΔA t = GCN(X t , A t-1 )·δ·ReLU(σ(f i t ) - θ)

[0059] where ΔA t is the adjacency matrix increment, δ is the expansion step size, θ is the fluctuation threshold, and σ is the normalization function.

[0060] The graph structure of the spatio-temporal risk conduction model has a dynamic adjustment mechanism. When the feature fluctuation exceeds the threshold, expand the associated nodes according to a specific formula. This mechanism can promptly capture market abnormal fluctuations, enabling the model structure to adaptively adjust with the dynamic changes in the market, enhancing the model's ability to capture risk conduction paths, better adapting to the complex and ever-changing financial market environment, and improving the accuracy and timeliness of risk conduction analysis.

[0061] The hierarchical architecture of the risk value prediction module includes:

[0062] Basic layer: Use LightGBM and XGBoost to generate initial predictions;

[0063] Optimization layer: Adjust the model parameters through the quantum particle swarm algorithm, and the optimization objective function is:

[0064]

[0065] where is the mean absolute error, λ is the regularization coefficient, is the L2 model weight.

[0066] The hierarchical architecture of the risk value prediction module is reasonable. The basic layer uses LightGBM and XGBoost to generate initial predictions, and uses their high efficiency and accuracy to initially capture risk characteristics; the optimization layer adjusts the model parameters through the quantum particle swarm optimization algorithm, and takes the objective function considering the mean absolute error and L2 regularization as the optimization direction, which can not only reduce the prediction error but also prevent overfitting, optimize the model from different levels, and improve the accuracy and stability of risk value prediction.

[0067] The reinforcement learning model of the intelligent blocking decision module includes: The reward function of the Markov decision process (MDP) is designed as:

[0068] R t = ω1·ΔVaR t + ω2·Cost t + ω3·Diversity t

[0069] Where, ΔVaR t is the risk reduction amplitude (unit: BP), Cost t is the policy execution cost (unit: yuan), and Diversity t is the portfolio diversity index (calculated based on Shannon entropy).

[0070] The reward function of the reinforcement learning model of the intelligent blocking decision module is scientifically designed. Considering the risk reduction amplitude, policy execution cost, and portfolio diversity comprehensively, it encourages the decision-making model to balance cost control and portfolio rationality while reducing risks. It prompts the decision-making module to make decisions that are more in line with actual investment needs and have better comprehensive benefits, improves the scientificity and effectiveness of investment decisions, and enhances the risk resistance and sustainability of the investment portfolio.

[0071] The real-time optimization methods of the dynamic feature screening module and the spatio-temporal risk conduction model include: the decoupled design of the feature screening layer and the graph update layer, the feature screening period is 5 seconds, and the graph update period is 30 seconds; the latency optimization formula under edge computing deployment:

[0072] Latency = EdgeProcess(f i t ) + CentralProcess(A t )

[0073] Where, EdgeProcess is the edge node processing time (<50ms), and CentralProcess is the central node processing time (<150ms).

[0074] The dynamic feature screening module and the spatio-temporal risk conduction model adopt a decoupled design and set different update cycles, enabling the two to be independently optimized according to their own characteristics without interference, thereby improving the operation efficiency of the module. The edge computing deployment combines with the latency optimization formula to clarify the processing time limits of edge nodes and central nodes, significantly reducing the system processing latency, meeting the stringent requirements for real-time in financial investment risk prediction, and being able to capture market changes more timely.

[0075] The method for generating the heat map of the spatio-temporal risk conduction model includes: The mapping relationship between key features and the risk conduction path is calculated by the following formula:

[0076]

[0077] Among them, AttentionScore is the attention weight of the feature and the edge, and Conductance is the conduction intensity of the edge.

[0078] The method for generating the heat map of the spatio-temporal risk conduction model combines the attention weight of the feature and the edge and the conduction intensity of the edge through the mapping formula of key features and the risk conduction path, accurately quantifying the relationship between the two, enabling the heat map to intuitively and accurately display the risk conduction situation, and providing a clear and effective visualization basis for risk analysis and decision-making.

[0079] The data quality evaluation method of the multi-source data collection module includes: The validation formula for the effectiveness of real estate transaction data:

[0080]

[0081] Among them, y is the unit price of second-hand houses, and Income is the per capita income of the local area. If the housing price exceeds twice the annual income, it is marked as abnormal data.

[0082] The data quality evaluation method of the multi-source data collection module gives a validation formula for the effectiveness of real estate transaction data, uses the comparison between the unit price of second-hand houses and the per capita income of the local area to judge whether the data is abnormal, is simple and direct and conforms to the logic of the real estate market, can effectively filter out unreasonable data, improve the data quality, and provide a reliable data basis for subsequent risk prediction.

[0083] The model lightweight method of the spatio-temporal risk conduction model includes: Using TensorRT to optimize the inference speed of the graph neural network, and the optimized computational complexity is:

[0084]

[0085] Among them, N l is the number of nodes in the l-th layer, and k l is the compression factor (typical value 8 - 16).

[0086] The spatio-temporal risk conduction model uses TensorRT to optimize the inference speed of the graph neural network. Combining with the computational complexity formula, while lightweighting the model, it significantly improves the inference efficiency, reduces the consumption of computing resources, enabling the model to process a large amount of data in a shorter time and meeting the requirements for the efficient operation of the model in financial investment risk prediction.

[0087] The compliance design of the risk value prediction module includes: The feature importance audit module records the key features and their contribution degrees for each prediction. The calculation formula is:

[0088]

[0089] where T is the time window length, and α is the gradient of feature f i at time t.

[0090] The compliance design of the risk value prediction module records the key features and their contribution degrees through the feature importance audit module. The calculation formula is based on the feature gradients within the time window, which can effectively trace the role of features in each prediction, enhancing the interpretability of the model prediction and meeting the requirements of financial supervision for the compliance and transparency of the risk prediction model.

[0091] Example 3

[0092] High-frequency stock trading scenario

[0093] Scenario background

[0094] A quantitative investment company focuses on high-frequency stock trading and needs to respond to the fluctuations in the stock market within an extremely short time to capture trading opportunities and control risks. Due to the rapid changes in the stock market, traditional risk prediction methods are difficult to meet its requirements for timeliness and accuracy.

[0095] System operation process

[0096] Multi-source data collection: The securities trading data collector of the data crawler cluster obtains real-time stock trading data at a millisecond frequency, including information such as stock price, trading volume, bid and ask quotes, etc. The data quality assessment sub-module performs real-time cleaning on the collected data to ensure the accuracy and integrity of the data, and then stores the data in the time-series database.

[0097] Dynamic feature screening: The genetic algorithm engine quickly conducts a preliminary screening of the massive trading data, extracts features closely related to stock price fluctuations, such as short-term price volatility, trading volume change rate, etc., to form an initial feature subset. The reinforcement learning controller dynamically optimizes the feature subset according to real-time market changes. The feature mapping unit maps the selected key features to the corresponding nodes of the knowledge graph to provide support for subsequent analysis.

[0098] Spatiotemporal Risk Propagation Modeling: The dynamic knowledge graph construction subsystem integrates multi-source data and updates the edge weights of the knowledge graph based on the correlation between stocks and market dynamics. The spatiotemporal graph neural network uses the parallel structure of GCN and TCN, combined with the spatiotemporal attention mechanism, to capture the risk propagation path in the stock market. The adversarial training module enhances the robustness of the model to market abnormal fluctuations by generating adversarial samples.

[0099] Risk Value Prediction: The parallel training framework of LightGBM and XGBoost in the basic model layer, based on the multi-objective loss function, quickly generates the initial prediction value of stock investment risk. The quantum optimization layer uses the quantum particle swarm optimizer to quickly adjust the model parameters within the hyperparameter search space to improve the prediction accuracy. The dynamic calibration layer calibrates the prediction results through the Bayesian posterior correction module and the extreme risk enhancer.

[0100] Intelligent Blocking Decision-making: The reinforcement learning decision-making engine generates the optimal trading strategy based on the risk prediction value and market conditions. The strategy execution interface quickly executes trading instructions through the algorithmic trading API to achieve the buy or sell operations of stocks. When the risk value exceeds the preset threshold, the circuit breaker mechanism trigger immediately starts to stop trading to control risks. The interpretability module provides an intuitive basis for investment decisions and feeds back the trading results to the dynamic feature screening module and the spatiotemporal risk propagation model to achieve the continuous optimization of the system.

[0101] Implementation Effect

[0102] Through the application of this system, in the high-frequency stock trading scenario, this quantitative investment company has successfully increased the accuracy rate of risk prediction by 30%, shortened the trading response time to within 100ms, effectively improved the investment return, and reduced the investment risk.

[0103] Example 4

[0104] Diversified Asset Allocation Scenario

[0105] Scenario Background

[0106] A large asset management company manages various types of assets, including stocks, bonds, real estate, etc. To achieve the optimal allocation of assets and reduce the risk of the investment portfolio, a system that can comprehensively evaluate the risks of different assets is needed.

[0107] System Operation Process

[0108] Multi-source Data Collection: The data crawler cluster collects stock, bond, real estate market, and macroeconomic data through the securities trading data collector, real estate trading data API, and macroeconomic data crawler. After the data quality assessment sub-module cleans and validates the data, the data is stored in the time-series database and the graph database.

[0109] Dynamic feature screening: The genetic algorithm engine screens out features related to the risks of various assets from multi-source data, such as changes in the credit ratings of bonds and the supply-demand relationship in the real estate market. The reinforcement learning controller optimizes the feature subset according to market dynamics. The feature mapping unit maps key features to the nodes of the knowledge graph to construct a knowledge graph of multi-asset associations.

[0110] Spatio-temporal risk conduction modeling: The dynamic knowledge graph construction subsystem integrates multi-source data and updates the edge weights of the knowledge graph to reflect the risk conduction relationship between different assets. The spatio-temporal graph neural network captures the spatio-temporal risk conduction path of the multi-asset market and visualizes the risk conduction process through a risk heat map generator. The adversarial training module enhances the model's adaptability to complex market environments.

[0111] Risk value prediction: The LightGBM and XGBoost parallel training frameworks in the basic model layer generate risk prediction values for various assets based on a multi-objective loss function. The quantum optimization layer optimizes the model parameters to improve the prediction accuracy. The dynamic calibration layer corrects the prediction results, taking into account the correlation between different assets and market uncertainties.

[0112] Intelligent blocking decision-making: The reinforcement learning decision-making engine generates an optimal asset allocation strategy based on the risk prediction values of multi-assets. The policy execution interface adjusts the proportions of various assets in the investment portfolio through an algorithmic trading API. When the risk value of a certain asset is too high, the circuit breaker mechanism trigger restricts the investment scale of that asset. The interpretability module provides a clear basis for asset management decisions and feeds back the policy execution results to the dynamic feature screening module and the spatio-temporal risk conduction model to achieve the dynamic optimization of the system.

[0113] Implementation effects

[0114] In a diversified asset allocation scenario, by applying this system, the asset management company has successfully reduced the risk of the investment portfolio by 25%, while improving the efficiency and returns of asset allocation. The interpretability of the system provides strong support for asset management decisions, enhancing the scientificity and transparency of decisions.

[0115] The above-described preferred embodiments of the present invention disclosed are only used to help explain the present invention. The preferred embodiments do not describe all details in detail, nor limit the invention to the specific embodiments described. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the present invention, so that those skilled in the relevant technical fields can well understand and utilize the present invention.

Claims

1. A financial investment risk value prediction system based on a big data group statistical algorithm, including a multi-source data acquisition module, characterized in that: The output end of the multi-source data acquisition module is electrically connected to a dynamic feature screening module, the output end of the dynamic feature screening module is electrically connected to a spatio-temporal risk conduction model, the output end of the spatio-temporal risk conduction model is electrically connected to a risk value prediction module, and the output end of the risk value prediction module is electrically connected to an intelligent blocking decision-making module; The output end of the intelligent blocking decision-making module is electrically connected to the input end of the spatio-temporal risk conduction model, and the output end of the intelligent blocking decision-making module is electrically connected to the input end of the dynamic feature screening module.

2. The financial investment risk value prediction system based on the big data group statistical algorithm according to claim 1, characterized in that: The key feature generation method of the dynamic feature screening module includes: using a genetic algorithm to optimize the initial population of feature subsets, and the fitness function is: Fitness(S)=α·SharpeRatio(S)+β·VaR(S)+γ·Corr(S) where S is the feature subset, α, β, and γ are weight coefficients, SharpeRatio is the Sharpe ratio, VaR is the value at risk, and Corr is the feature correlation.

3. A financial investment risk value prediction system based on a big data group statistical algorithm according to claim 1, characterized in that: The graph structure dynamic adjustment mechanism of the spatio-temporal risk conduction model includes: when it is detected that the feature fluctuation exceeds the threshold, the associated nodes are expanded through the following formula: ΔA t = GCN(X t , A t-1 )·δ·ReLU(σ(f i t ) - θ) where, ΔA t is the adjacency matrix increment, δ is the expansion step, θ is the fluctuation threshold, and σ is the normalization function.

4. A financial investment risk value prediction system based on a big data group statistical algorithm according to claim 1, characterized in that: The hierarchical architecture of the risk value prediction module includes: Basic layer: Using LightGBM and XGBoost to generate initial predictions; Optimization layer: Adjusting the model parameters through the quantum particle swarm optimization algorithm, and the optimization objective function is: wherein, is the mean absolute error, λ is the regularization coefficient, is the L2 model weight.

5. A financial investment risk value prediction system based on a big data group statistical algorithm according to claim 1, characterized in that: The reinforcement learning model of the intelligent blocking decision-making module includes: The reward function of the Markov decision process (MDP) is designed as: R t = ω1·ΔVaR t + ω2·Cost t + ω3·Diversity t Among them, ΔVaR t is the risk reduction amplitude, Cost t is the strategy execution cost, and Diversity t is the portfolio diversity index.

6. The financial investment risk value prediction system based on the big data swarm computing algorithm according to claim 1, wherein: The real-time optimization method of the dynamic feature screening module and the spatio-temporal risk conduction model includes: decoupling design of the feature screening layer and the graph update layer, the feature screening period is 5 seconds, and the graph update period is 30 seconds; the delay optimization formula under edge computing deployment: Latency=EdgeProcess(f i t )+CentralProcess(A t ) where EdgeProcess is the edge node processing time and CentralProcess is the central node processing time.

7. A financial investment risk value prediction system based on a big data group statistical algorithm according to claim 1, characterized in that: The heat map generation method of the spatio-temporal risk conduction model includes: The mapping relationship between the key features and the risk conduction path is calculated through the following formula: where AttentionScore is the attention weight of the feature and the edge, and Conductance is the conduction intensity of the edge.

8. A financial investment risk value prediction system based on a big data group statistical algorithm according to claim 1, characterized in that: The data quality assessment method of the multi-source data acquisition module includes: The validity verification formula of real estate transaction data: where y is the unit price of second-hand houses and Income is the per capita income of the local area. If the housing price exceeds twice the annual income, it is marked as abnormal data.

9. A financial investment risk value prediction system based on a big data group statistical algorithm according to claim 1, characterized in that: The model lightweight method of the spatio-temporal risk conduction model includes: Using TensorRT to optimize the inference speed of the graph neural network, and the optimized computational complexity is: Among them, N l is the number of nodes in the l-th layer, and k l is the compression factor.

10. A financial investment risk value prediction system based on a big data group statistical algorithm according to claim 1, characterized in that: The compliance design of the risk value prediction module includes: The feature importance audit module records the key features and their contribution degrees of each prediction, and the calculation formula is: where T is the time window length and α is the gradient of feature f i at time t.

Images