Multi-factor dynamic correction investment benefit intelligent evaluation system
By incorporating event-driven factor correction, adaptive weight optimization, anomaly propagation early warning, and strategy execution modules, combined with deep reinforcement learning and graph neural networks, the real-time response and robustness issues of existing investment benefit evaluation systems are resolved, enabling dynamic investment strategy optimization for engineering projects.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUBEI POST TELECOMM PLANNING DESIGN
- Filing Date
- 2026-02-03
- Publication Date
- 2026-05-26
AI Technical Summary
Existing investment benefit assessment systems struggle to respond in real time to unforeseen events in the face of high-frequency dynamic adjustments in engineering projects. This results in delayed model weight adjustments, an inability to accurately depict nonlinear relationships under extreme conditions, high computational complexity, and an inability to balance robustness and real-time performance in predictions. Consequently, investment strategies fail, leading to a decline in benefits.
Employing an event-driven factor correction module, an adaptive weight optimization module, an anomaly propagation early warning module, and a strategy execution module, this system uses deep reinforcement learning and graph neural network technologies to capture macro-level events and project risks in real time, dynamically adjust factor weights, identify anomaly propagation paths, and perform rebalancing under cost and risk control.
It significantly improves the model's response speed and sensitivity to sudden changes in projects, enables adaptive adjustment of investment strategies, optimizes the risk-reward ratio, and generates project allocation schemes that take into account profitability, robustness, and real-time performance.
Smart Images
Figure CN122089480A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of investment evaluation, and more specifically, to an intelligent investment benefit evaluation system with multi-factor dynamic correction. Background Technology
[0002] In the field of project and engineering investment management, multi-factor models are the core analytical framework for investment benefit assessment, risk management, and project portfolio construction. Traditional models select a series of statistical factors, such as cost, schedule, and risk, that can explain differences in investment returns, and determine fixed weights for each factor based on historical data, aiming to achieve stable returns exceeding benchmarks. This approach is widely used in investment decisions for large-scale engineering projects, and in the development of medium- and long-term asset allocation and quantitative strategies for infrastructure funds and insurance institutions. With increased project participation and advancements in data technology, especially in the context of high-frequency dynamic adjustments in mainstream engineering assets such as infrastructure and industrial projects, investors rely on these models to capture anomalies during project execution and guide investment decisions. However, the engineering investment environment is not in static equilibrium; its operation is continuously disturbed by various internal and external factors, including macroeconomic cycles, industry policy adjustments, international geopolitical events, and the collective behavior of project participants.
[0003] While existing technologies attempt to adjust factor weights in stages using methods such as rolling time window regression and Bayesian statistical optimization, significant shortcomings remain. Firstly, these methods inherently rely on the statistical regularities of historical data, resulting in a lag in model weight adjustments and an inability to respond in real-time to structural changes in engineering projects caused by unforeseen events. Secondly, traditional linear or parametric models struggle to accurately depict the complex nonlinear relationships between factors and returns under extreme conditions. For instance, during periods of concentrated project risk, the explanatory power of traditional factors can drastically decrease or even become ineffective. More critically, existing technological solutions often suffer from excessively high computational complexity when implementing dynamic adjustments, failing to balance real-time performance with predictive robustness. This can lead to rapid failure of investment strategies during events such as sudden policy reversals or black swan events, resulting in significant declines in investment returns or even substantial financial losses. This severely restricts the intelligence level and risk response capabilities of engineering investment benefit assessment systems.
[0004] Therefore, there is an urgent need for a lightweight, real-time responsive, and adaptive learning-capable dynamic weight correction technique to effectively integrate event-driven signals and quantitative models, thereby improving the robustness of investment strategies in different project environments. Summary of the Invention
[0005] This invention addresses the technical problems existing in the prior art by providing a multi-factor dynamic correction intelligent investment benefit evaluation system. This system utilizes an event-driven factor correction module, an adaptive weight optimization module, an anomaly propagation early warning module, and a strategy execution module to solve the problems mentioned in the background.
[0006] The technical solution of this invention to solve the above-mentioned technical problems is as follows: it specifically includes: an event-driven factor correction module, an adaptive weight optimization module, an anomaly propagation early warning module, and a strategy execution module;
[0007] Event-driven factor correction module: When the preset macroeconomic event keywords or market sentiment indicators are detected to exceed the threshold, it collects text and unstructured data streams in real time, extracts event entities and sentiment polarity through natural language processing technology, calculates the dynamic impact coefficient for each investment factor based on predefined factor mapping rules, and outputs a set of event vectors containing event timestamps, a list of affected investment factors, and impact intensity.
[0008] The adaptive weight optimization module receives the set of event vectors and the historical multi-factor return sequence output from the event-driven factor correction module. The historical multi-factor return sequence is a set of return data of each investment factor at multiple consecutive historical time points. It adopts a network framework based on deep reinforcement learning, uses the risk-adjusted return within the rolling time window as the reward signal, generates the investment factor weight distribution for the next trading cycle through the strategy network, evaluates the state value corresponding to the current investment factor weight distribution through the value network, and outputs the optimized investment factor weight matrix for the next trading cycle.
[0009] Anomaly propagation early warning module: This module obtains historical factor correlation matrices and real-time market trading data from the adaptive weight optimization module and external data interface, respectively. The real-time market trading data consists of the transaction prices and trading volumes of the assets corresponding to each investment factor during the current trading period. It constructs a dynamic directed graph with each investment factor as a node and the correlation strength between factors as edges. The historical factor correlation matrix is a correlation coefficient matrix between each investment factor calculated based on the historical multi-factor return sequence. It uses graph neural network technology to identify abnormal investment factor nodes through node embedding and updating and edge weight propagation algorithms. When an abnormal propagation path exceeding a preset threshold is detected, an early warning signal is generated that includes the investment factor at the source of the risk, the propagation path, and the expected duration of the impact.
[0010] Strategy Execution Module: This module integrates the optimized investment factor weight matrix output by the adaptive weight optimization module with the early warning signal output by the anomaly propagation early warning module. Under the condition of meeting the preset transaction cost constraints and risk control thresholds, it executes the portfolio rebalancing instruction to generate the final asset allocation plan.
[0011] In a preferred embodiment, the macroeconomic event keywords preset in the event-driven factor correction module are extracted from financial news texts, central bank announcements, and regulatory documents. Market sentiment indicators include implied volatility in the options market, social media sentiment index, and search engine attention index. The threshold refers to the change in the value of the market sentiment indicator within two consecutive preset observation periods, which exceeds the standard deviation set multiple calculated based on historical data of the market sentiment indicator.
[0012] In a preferred embodiment, the event-driven factor correction module calculates the dynamic impact coefficient based on a predefined factor mapping rule, specifically as follows:
[0013] First, an attention-based text encoding model is used to transform the extracted event entities and sentiment polarities into event feature vectors. Then, a pre-trained factor sensitivity prediction model is used, taking the event feature vectors as input, to output the sensitivity adjustment parameters corresponding to each investment factor. The event-driven factor correction module generates the dynamic impact coefficient of the investment factor by multiplying the sensitivity adjustment parameters by the current weight value of the corresponding investment factor. The event-driven factor correction module further integrates the dynamic impact coefficient of each investment factor with the investment factor's identification information and event timestamp to finally form a set of event vectors.
[0014] In a preferred embodiment, the adaptive weight optimization module includes a deep reinforcement learning network framework comprising a policy network and a value network. The policy network is used to generate the investment factor weight distribution for the next trading cycle based on the current system state. The value network is used to evaluate the long-term expected return corresponding to different investment factor weight distributions under the current system state. The system state is composed of a set of event vectors, a historical multi-factor return sequence, and a current investment factor weight matrix. The current investment factor weight matrix is formed by the policy execution module after executing the previous rebalancing instruction.
[0015] In a preferred embodiment, the adaptive weight optimization module calculates the risk-adjusted return using an improved Sharpe ratio, and the calculation process includes:
[0016] Within a preset rolling time window, the portfolio return series generated by the investment factor weight distribution is calculated. The excess return series is obtained by subtracting the risk-free rate. The average of the excess return series is divided by its standard deviation, and the transaction cost coefficient calculated based on the proportion of transaction amount is deducted to form the final reward signal value.
[0017] In a preferred embodiment, the process by which the adaptive weight optimization module outputs the optimized investment factor weight matrix for the next trading cycle includes:
[0018] The strategy network outputs an initial weight adjustment vector based on the current system state. After normalization, this weight adjustment vector satisfies the constraint that the sum of the portfolio weights is 1. It is then linearly superimposed with the current investment factor weight matrix to form a candidate optimized investment factor weight matrix for the next trading cycle. After evaluation by the value network, the candidate scheme with the highest state value is selected from multiple candidate optimized investment factor weight matrices as the final output optimized investment factor weight matrix.
[0019] In a preferred embodiment, the construction process of the dynamic directed graph in the anomaly propagation early warning module includes:
[0020] Using each investment factor as a node, the correlation coefficient in the historical factor correlation matrix is used as the initial weight of the edge. The weight of the edge is dynamically updated based on the instantaneous correlation coefficient calculated from real-time market trading data. The instantaneous correlation coefficient is obtained by calculating the rolling correlation coefficient of the asset returns corresponding to each investment factor in the current trading period. The historical factor correlation matrix is directly provided by the adaptive weight optimization module. Real-time market trading data is obtained in real time through an external data interface. The external data interface establishes a connection with the financial market trading database through a preset financial data protocol and receives real-time market data streams released by the exchange at a preset frequency. The real-time market data streams include the transaction price, transaction volume, and timestamp information of the assets corresponding to each investment factor.
[0021] In a preferred embodiment, the process of identifying anomalous investment factor nodes using the graph neural network technology includes:
[0022] The feature representation of each node is updated through a multi-layer graph convolutional network. The initial feature vector of a node is composed of the current return, historical volatility, and impact coefficient from the event vector set of the investment factor. In each layer of graph convolution, the edge weight propagation algorithm aggregates the feature information of neighboring nodes according to the edge weights and outputs the updated node features through a non-linear activation function. Finally, the updated node features are input into an anomaly detection classifier, which outputs the anomaly score of each investment factor node. When the anomaly score of a node exceeds a preset threshold, the investment factor node is marked as an abnormal investment factor node.
[0023] In a preferred embodiment, the fusion process in the policy execution module is as follows:
[0024] When an early warning signal is received from the abnormal transmission early warning module, the weight allocation of the corresponding investment factor is reduced in the optimized investment factor weight matrix according to the risk source investment factor marked in the signal, and the weight of other investment factors not marked as abnormal is increased accordingly. The transmission path and expected impact duration in the early warning signal are used to determine the magnitude and priority of the weight adjustment.
[0025] In a preferred embodiment, the transaction cost constraints in the strategy execution module include a fixed commission rate, a stamp duty rate calculated as a percentage of the transaction amount, and a market impact cost rate estimated based on the order size; the risk control threshold conditions include the maximum exposure limit of a single investment factor, the overall value at risk limit of the portfolio, and the maximum drawdown control line; before generating the portfolio rebalancing instruction, it is necessary to ensure that the adjusted asset allocation scheme simultaneously meets all transaction cost constraints and risk control threshold conditions, and the adjusted asset allocation scheme is obtained by performing a quadratic programming solution on the optimized investment factor weight matrix using a constrained optimization algorithm.
[0026] The beneficial effects of this invention are as follows: The event-driven factor correction module captures the impact of macro events and project risk indicators in real time and dynamically generates factor impact coefficients, which significantly improves the model's response speed and sensitivity to sudden changes in projects. The adaptive weight optimization module uses a deep reinforcement learning framework to continuously optimize the factor weight distribution based on risk-adjusted returns within a rolling time window, realizing adaptive adjustment of investment strategies and optimization of the risk-reward ratio. The anomaly transmission early warning module constructs a dynamic directed graph and combines it with graph neural network technology to proactively identify abnormal factor transmission paths and effectively warn of systemic engineering risks. The strategy execution module integrates the weight optimization results and risk warning signals, and performs rebalancing under strict investment cost and risk control constraints, ultimately generating a project allocation scheme that takes into account profitability, robustness, and real-time performance. Attached Figure Description
[0027] Figure 1 This is a flowchart of the method of the present invention;
[0028] Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation
[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0030] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.
[0031] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.
[0032] Example 1
[0033] This embodiment provides, for example Figure 1-2 The present invention is a multi-factor dynamic correction intelligent evaluation system for investment benefits, which specifically includes: an event-driven factor correction module, an adaptive weight optimization module, an anomaly transmission early warning module, and a strategy execution module.
[0034] Event-driven factor correction module: When the preset macroeconomic event keywords or market sentiment indicators are detected to exceed the threshold, it collects text and unstructured data streams in real time, extracts event entities and sentiment polarity through natural language processing technology, calculates the dynamic impact coefficient for each investment factor based on predefined factor mapping rules, and outputs a set of event vectors containing event timestamps, a list of affected investment factors, and impact intensity.
[0035] The adaptive weight optimization module receives the set of event vectors and the historical multi-factor return sequence output from the event-driven factor correction module. The historical multi-factor return sequence is a set of return data of each investment factor at multiple consecutive historical time points. It adopts a network framework based on deep reinforcement learning, uses the risk-adjusted return within the rolling time window as the reward signal, generates the investment factor weight distribution for the next trading cycle through the strategy network, evaluates the state value corresponding to the current investment factor weight distribution through the value network, and outputs the optimized investment factor weight matrix for the next trading cycle.
[0036] Anomaly propagation early warning module: This module obtains historical factor correlation matrices and real-time market trading data from the adaptive weight optimization module and external data interface, respectively. The real-time market trading data consists of the transaction prices and trading volumes of the assets corresponding to each investment factor during the current trading period. It constructs a dynamic directed graph with each investment factor as a node and the correlation strength between factors as edges. The historical factor correlation matrix is a correlation coefficient matrix between each investment factor calculated based on the historical multi-factor return sequence. It uses graph neural network technology to identify abnormal investment factor nodes through node embedding and updating and edge weight propagation algorithms. When an abnormal propagation path exceeding a preset threshold is detected, an early warning signal is generated that includes the investment factor at the source of the risk, the propagation path, and the expected duration of the impact.
[0037] Strategy Execution Module: This module integrates the optimized investment factor weight matrix output by the adaptive weight optimization module with the early warning signal output by the anomaly propagation early warning module. Under the condition of meeting the preset transaction cost constraints and risk control thresholds, it executes the portfolio rebalancing instruction to generate the final asset allocation plan.
[0038] In this embodiment, it is specifically necessary to explain the preset macroeconomic event keywords in the event-driven factor correction module. These keywords are extracted from financial news texts, central bank announcements, and regulatory documents. The formation process of macroeconomic event keywords includes: first, establishing a professional vocabulary dictionary covering financial policies, macroeconomic data releases, and industry regulatory changes; then, using text mining technology to extract frequently occurring keyword combinations from historical financial documents; and finally, determining the final list of macroeconomic event keywords through expert review. This list supports regular updates to cover new policy terms and market hotspots. Market sentiment indicators include implied volatility in the options market, social media sentiment index, and search engine attention index. The threshold refers to the change in the value of the market sentiment indicator within two consecutive preset observation periods exceeding a set multiple of the standard deviation calculated based on historical data of the market sentiment indicator. The set multiple of the standard deviation can be configured from 1.5 times to 2.5 times, and the specific multiple is dynamically adjusted according to the market conditions. In the case of insufficient historical data, a preset default multiple of 1.8 times can be used.
[0039] In a specific example, the formula for determining whether the market sentiment indicator breaks through the threshold is as follows: In the formula: For the first Market sentiment index values for each observation period; For the first Market sentiment index values for each observation period; The variation range over two consecutive observation periods; The standard deviation of historical data for the market sentiment indicator is calculated using the following formula: ;in The number of historical observation periods. This is the average of historical market sentiment indicators; Set a factor for the standard deviation, ranging from 1.5 to 2.5, with a default value of 1.8;
[0040] The event-driven factor correction module calculates the dynamic impact coefficient based on predefined factor mapping rules. The specific operation is as follows:
[0041] First, an attention-based text encoding model is used to transform the extracted event entities and sentiment polarities into event feature vectors. This text encoding model employs a structure combining a bidirectional long short-term memory network and an attention mechanism, enabling it to automatically identify key information in the text and extract semantic features. Its training data is derived from the correspondence between historical financial news and market data. Subsequently, a pre-trained factor sensitivity prediction model is used, taking the event feature vectors as input and outputting the sensitivity adjustment parameters for each investment factor. This factor sensitivity prediction model is a multilayer perceptron neural network, trained using historical event data and corresponding investment factor return rate change data, establishing the event feature vectors through supervised learning. The model maps the sensitivity parameters of investment factors to the nonlinear laws governing the impact of different event types on different investment factors. The event-driven factor correction module generates the dynamic impact coefficient of the investment factor by multiplying the sensitivity adjustment parameter by the current weight value of the corresponding investment factor. The event-driven factor correction module further integrates the dynamic impact coefficient of each investment factor with the investment factor's identification information and the event timestamp to form an event vector set. The data structure of the event vector set is stored in dictionary format, including a unique event identifier, a timestamp field, a list of affected investment factors, and an impact coefficient value field for each investment factor, which facilitates direct calling and processing by subsequent modules.
[0042] In a specific example, the attention-based text encoding process is represented as follows:
[0043] ;
[0044] ;
[0045] ;
[0046] in: The word embedding sequence of the event text, with dimension . , For text length, Word embedding dimension; The hidden state sequence is the output of a bidirectional LSTM. ,in The dimension of the hidden layer (values range from 128 to 512); This is the weight matrix for the attention mechanism. For attention vectors, For attention dimension (ranging from 64 to 256); For the first The attention weight of each word satisfies ; This is the weighted event context feature vector;
[0047] The sentiment polarity score is calculated using a fully connected layer and a softmax classifier: ;in , , The probability vector is a three-category sentiment vector (negative, neutral, positive).
[0048] The final event feature vector is obtained by concatenation: ;in This represents a vector concatenation operation;
[0049] The factor sensitivity prediction model employs a multilayer perceptron structure:
[0050] ;
[0051] ;
[0052] ;
[0053] in: , , The weight matrix for each layer, and the dimension of the hidden layer. The value range is from 64 to 256; , , These are the bias vectors for each layer; The total number of investment factors; for The sensitivity adjustment parameter vector for each investment factor is ensured to take values in the range of (0,1) by the Sigmoid activation function.
[0054] No. The formula for calculating the dynamic impact coefficient of each investment factor is as follows:
[0055] ;
[0056] in: Let be the dynamic impact coefficient of the k-th investment factor; This is the sensitivity adjustment parameter for the k-th investment factor; This represents the current weight value of the k-th investment factor.
[0057] The event vector set is represented as:
[0058] ;
[0059] in: For event timestamps; List the m affected investment factors; This is the corresponding impact intensity vector.
[0060] In this embodiment, it is specifically noted that the deep reinforcement learning network framework in the adaptive weight optimization module includes a policy network and a value network. The policy network is used to generate the investment factor weight distribution for the next trading cycle based on the current system state. The value network is used to evaluate the long-term expected return corresponding to different investment factor weight distributions in the current system state. The system state is composed of a set of event vectors, a historical multi-factor return sequence, and a current investment factor weight matrix. The current investment factor weight matrix is formed by the policy execution module after executing the previous rebalancing instruction. The event vector set is the real-time output data from the event-driven factor correction module. The historical multi-factor return sequence refers to the logarithmic return sequence calculated from the closing prices of each investment factor in the past several preset trading cycles. The policy network and the value network continuously optimize the network parameters by combining offline historical data training and online incremental learning.
[0061] In a specific example, the deep reinforcement learning network framework is as follows:
[0062] System status At time t, it is defined as:
[0063] ;
[0064] in: The feature representation of the event vector set at time t is obtained by performing a time-decay-weighted average of the impact coefficients of all event vectors within the most recent time window:
[0065] in The set of events within the valid time window before time t. The time decay coefficient, For event timestamps; This is a historical multi-factor return sequence matrix. For a moment The vector of returns for each investment factor, where T is the historical backtracking length; Let be the current investment factor weight vector, satisfying and ;
[0066] Policy Network The output weight adjustment vector is used in the network structure, which includes two fully connected layers.
[0067] ;
[0068] ;
[0069] ;
[0070] in: For the set of policy network parameters; $ ,in For the state vector dimension, For the hidden layer dimension of the policy network; , ; This is a hidden layer representation of the policy network; The logits vector is adjusted for weights; The normalized weight adjustment vector is ensured by the Softmax operation. and ;
[0071] Value Network To assess state value, the network architecture also includes two fully connected layers:
[0072] ;
[0073] ;
[0074] in: For the set of value network parameters; , ,in For the hidden layer dimension of the value network; , ; This represents the hidden layer of the value network. For state The estimated value;
[0075] In the adaptive weight optimization module, the risk-adjusted return is calculated using an improved Sharpe ratio, and the calculation process includes:
[0076] Within a preset rolling time window, the portfolio return series generated by the investment factor weight distribution is calculated. The excess return series is obtained by subtracting the risk-free rate. The average of the excess return series is divided by its standard deviation, and a transaction cost coefficient calculated based on the proportion of transaction amount is deducted to form the final reward signal value. The length of the rolling time window can be set to 20 to 60 trading days, and the specific length is configured according to the length of the investment period. The transaction cost coefficient includes a fixed commission rate, a stamp duty rate, and a market impact cost rate estimated based on the order size. This coefficient can be preset and adjusted according to different markets and trading instruments.
[0077] In a specific example, the improved Sharpe ratio is calculated as follows:
[0078] The portfolio's return at time t is: ;in Let be the return (logarithmic return) of the k-th investment factor at time t.
[0079] In the rolling time window Inside( (The window length is 20 to 60 trading days). The excess return series is as follows: ;in This is the risk-free rate (annualized), which needs to be converted to the interest rate for the corresponding trading period;
[0080] The mean and standard deviation of the excess returns are as follows:
[0081] ;
[0082] ;
[0083] in The average return of the portfolio;
[0084] The transaction cost coefficient is calculated as follows:
[0085] ;
[0086] in: This represents the change in the weight of the k-th factor. A fixed commission rate (e.g., 0.03%). The stamp duty rate (charged only on sales transactions, e.g., 0.1%). The market shock cost coefficient, raised to the power of 1.5, reflects the non-linear relationship between market shock and transaction size. This represents the absolute value of the weight change. This indicates that stamp duty will only be levied on the portion of the shares sold.
[0087] The improved Sharpe ratio (reward signal) is:
[0088] ;
[0089] in To prevent division by zero of small constants;
[0090] The process by which the adaptive weight optimization module outputs the optimized investment factor weight matrix for the next trading cycle includes:
[0091] The strategy network outputs an initial weight adjustment vector based on the current system state. After normalization, this weight adjustment vector satisfies the constraint that the sum of the portfolio weights is 1. It is then linearly superimposed with the current investment factor weight matrix to form a candidate optimized investment factor weight matrix for the next trading cycle. After evaluation by the value network, the candidate scheme with the highest state value is selected from multiple candidate optimized investment factor weight matrices as the final output optimized investment factor weight matrix. Multiple candidate optimized investment factor weight matrices are generated by inputting system states with different random noises into the strategy network to increase the exploratory nature of weight decisions. The evaluation of state value is based on the expected value of long-term cumulative reward predicted by the value network, which is calculated recursively using the Bellman equation.
[0092] In a specific example, M perturbation states are generated by adding Gaussian noise to the system state:
[0093] ;
[0094] Where m = 1, 2, ..., M, The standard deviation of noise. It is the identity matrix;
[0095] Generate a weight adjustment vector for each perturbation state:
[0096] ;
[0097] And calculate the candidate weight vector:
[0098] ;
[0099] in Update the step size for weights; ensure through normalization ;
[0100] Calculate the state value for each candidate solution:
[0101] ;
[0102] Select the candidate with the highest value as the final output:
[0103] ;
[0104] The Actor-Critic algorithm was used for training, and the time difference error (TD error) was:
[0105] ;
[0106] in Discount factor;
[0107] Policy network loss function (policy gradient):
[0108] ;
[0109] Value network loss function (mean squared error):
[0110] ;
[0111] Parameters are updated via gradient descent:
[0112] ;
[0113] ;
[0114] in These are the learning rates for the policy network and the value network, respectively.
[0115] In this embodiment, it is specifically necessary to explain the construction process of the dynamic directed graph in the anomaly propagation early warning module, which includes:
[0116] Using each investment factor as a node, the correlation coefficient in the historical factor correlation matrix is used as the initial weight of the edge. The weight of the edge is dynamically updated based on the instantaneous correlation coefficient calculated from real-time market trading data. The instantaneous correlation coefficient is obtained by calculating the rolling correlation coefficient of the return of the corresponding asset of each investment factor in the current trading period. The length of the rolling window can be set to 5 to 20 trading hours, and the specific length is configured according to the market volatility characteristics. The historical factor correlation matrix is directly provided by the adaptive weight optimization module. Real-time market trading data is obtained in real time through an external data interface. The external data interface establishes a connection with the financial market trading database through a preset financial data protocol and receives the real-time market data stream released by the exchange at a preset frequency. The real-time market data stream contains the transaction price, transaction volume and timestamp information of the corresponding asset of each investment factor. The preset frequency can be configured to the second level, minute level or higher frequency according to the monitoring needs. The financial data protocol includes financial information exchange protocol, securities trading system dedicated interface protocol or real-time market streaming transmission protocol.
[0117] In a specific example, the historical factor correlation matrix Calculated using the Pearson correlation coefficient:
[0118] ;
[0119] in: Let be the historical correlation coefficient between the i-th and j-th investment factors. ; For the i-th investment factor at time... The rate of return (logarithmic return); Let T be the historical average return of the i-th investment factor; T is the length of the historical lookback period. It is a symmetric matrix;
[0120] In the scrolling window of the current trading session Within, the instantaneous correlation coefficient is:
[0121] ;
[0122] in: The length of the scrolling window (ranging from 5 to 20 transaction moments); The average rate of return within the window;
[0123] The edge weights of a dynamic directed graph are weighted by a combination:
[0124] ;
[0125] in: Let be the weight of the directed edge from node i to node j at time t; This is the historical correlation weighting coefficient (default value is 0.7); if Then the edge is removed, where The edge weight threshold is set to 0.1; the dynamic directed graph is represented as follows. ,in: Let K be the set of nodes, corresponding to K investment factors. Let be the set of edges at time t; This is the edge weight matrix;
[0126] The process of identifying anomalous investment factor nodes using graph neural network technology includes:
[0127] The feature representation of each node is updated through a multi-layer graph convolutional network. The initial feature vector of a node is composed of the current return, historical volatility, and impact coefficient from the event vector set of the investment factor. The current return uses the logarithmic return calculated using the latest closing price, and the historical volatility uses the standard deviation of the returns over the past 20 trading days. In each layer of graph convolution, the edge weight propagation algorithm aggregates the feature information of neighboring nodes according to the edge weights and outputs the updated node features through a non-linear activation function. The non-linear activation function includes, but is not limited to, the ReLU function, the Sigmoid function, or the Tanh function. Finally, the updated node features are input into an anomaly detection classifier, which outputs the anomaly score of each investment factor node. The anomaly detection classifier is a support vector machine or an isolated forest algorithm. After training with historical data, it can identify anomaly patterns. When the node anomaly score exceeds a preset threshold, the investment factor node is marked as an anomalous investment factor node. The preset threshold is determined by calculating the statistical distribution of the node anomaly scores under historical normal conditions, and is usually set at 2.5 to 3.5 times the standard deviation of the historical average anomaly score.
[0128] In a specific example, the initial feature vector of a node is constructed as follows:
[0129] ;
[0130] in: Let be the current return (logarithmic return) of the i-th investment factor at time t. Historical volatility One trading day; The impact coefficient comes from the set of event vectors; , ;
[0131] No. Node feature update rules for layer graph convolution ( (where L is the number of graph convolutional layers, ranging from 2 to 4).
[0132] ;
[0133] in: Let i be the set of incoming neighbor nodes; Let be the edge weight from node j to node i; This represents vector concatenation, which combines the vector's own features with the weighted average features of its neighbors. For the first Layer weight matrix, hidden layer dimension The value ranges from 64 to 256; For the first Layer bias vector; The initial input serves as the input for the next layer. Initial features of the nodes;
[0134] After L layers of graph convolution, the final embedding of node i is represented as ;
[0135] A two-layer fully connected network is used as the anomaly detection classifier:
[0136] ;
[0137] ;
[0138] in: , , Let be the dimension of the hidden layer of the classifier; , ; This represents the hidden layer of the classifier; For the outlier score of node i, the Sigmoid activation function ensures that the output is between 0 and 1;
[0139] Criteria for determining abnormal nodes:
[0140] ;
[0141] Among them, the abnormal threshold Determined based on the historical anomaly score distribution:
[0142] ;
[0143] in: This represents the mean of the abnormal scores under historical normal conditions. The standard deviation of outlier scores under historical normal conditions; This represents the number of historical samples. This is the threshold multiple, with a default value of 3.0;
[0144] Define the conduction intensity matrix:
[0145] ;
[0146] From the abnormal source node (satisfy Starting from n, a breadth-first search is used to identify the conduction path. The conduction strength of a path of length n is defined as:
[0147] ;
[0148] in , , The maximum conduction path length (values from 3 to 5);
[0149] The path is marked as an abnormal conduction path when the conduction strength meets the following conditions:
[0150] ;
[0151] Wherein the conduction threshold: ;
[0152] The expected duration of the impact is estimated using the average duration of similar historical events:
[0153] ;
[0154] in A set of historically similar anomalous events. The duration of the h-th historical event (in transaction days).
[0155] In this embodiment, it is specifically necessary to explain that the fusion process in the policy execution module is as follows:
[0156] When an early warning signal is received from the abnormal transmission early warning module, the weight allocation of the corresponding investment factor is reduced in the optimized investment factor weight matrix according to the risk source investment factor marked in the signal, and the weight of other investment factors not marked as abnormal is increased accordingly. The transmission path and expected impact duration in the early warning signal are used to determine the magnitude and priority of the weight adjustment. The magnitude of the weight adjustment is inversely proportional to the length of the transmission path in the early warning signal and directly proportional to the expected impact duration. The priority is sorted according to the number of investment factors involved in the transmission path and the severity of the expected impact. High-risk transmission paths enjoy higher adjustment priority.
[0157] In the strategy execution module, transaction cost constraints include a fixed commission rate, a stamp duty rate calculated as a percentage of the transaction amount, and a market impact cost rate estimated based on order size. Risk control threshold conditions include the maximum exposure limit for a single investment factor, the overall portfolio value at risk limit, and the maximum drawdown control line. The maximum exposure limit for a single investment factor is typically set at 5% to 15% of the total portfolio assets, with the specific value adjusted according to the risk appetite of the investment strategy. Before generating a portfolio rebalancing instruction, it is necessary to ensure that the adjusted asset allocation plan simultaneously meets all transaction cost constraints and risk control threshold conditions. The adjusted asset allocation plan is obtained by solving the optimized investment factor weight matrix using a constrained optimization algorithm with quadratic programming. The constrained optimization algorithm treats transaction costs and risk control thresholds as hard constraints during the solution process, and iteratively optimizes to find a feasible solution that satisfies all constraints and is closest to the optimized investment factor weight matrix.
[0158] In a specific example, the risk adjustment factor for the k-th investment factor is defined as follows:
[0159] ;
[0160] in: This is the risk adjustment coefficient for the k-th investment factor; The risk adjustment intensity parameter is set to 0.5 by default. The shortest transmission path length, which includes factor k; Maximum conduction path length limit; The expected duration of the effect of factor k; To influence the maximum duration (e.g., 30 trading days); the calculation of the risk adjustment factor ensures that: the shorter the path ( The smaller the size (or the longer the duration of the effect) (The larger the value), the higher the risk adjustment factor. The smaller the value, the stronger the weight compression of the factor;
[0161] The fused target weight vector:
[0162] ;
[0163] in: The optimized weight vector output by the adaptive weight optimization module; This is a vector of risk adjustment coefficients; This represents element-wise multiplication (Hadamard product); the denominator is used for normalization to ensure that the sum of the weights is 1;
[0164] The final asset allocation is obtained by solving the following constrained optimization problem:
[0165] ;
[0166] subject to:
[0167] ;
[0168] in: Let be the final asset allocation weight vector to be solved; This is a penalty term for deviating from the target weight; This is the transaction cost weighting coefficient; From Adjust to Total transaction costs:
[0169] ;
[0170] in ; This represents the maximum exposure limit for a single factor (e.g., 20% or 30%). Confidence level The Value at Risk (VAT) is calculated using either historical simulation or variance-covariance methods.
[0171] ;
[0172] in Take 0.95 or 0.99. For portfolio return; For the limit of value at risk; The maximum drawdown estimate based on historical backtesting:
[0173] ;
[0174] in For a moment The combined value, This refers to the length of the backtesting period; The maximum drawdown control line (e.g., 10% or 15%).
[0175] The aforementioned quadratic programming problem can be solved efficiently using standard convex optimization solvers (such as CVXOPT and Gurobi).
[0176] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0177] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0178] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0179] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0180] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0181] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0182] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A multi-factor dynamic correction intelligent evaluation system for investment benefits, characterized in that, Specifically, it includes: The module includes an event-driven factor correction module, an adaptive weight optimization module, an anomaly propagation and early warning module, and a strategy execution module. Event-driven factor correction module: When the preset macroeconomic event keywords or market sentiment indicators are detected to exceed the threshold, it collects text and unstructured data streams in real time, extracts event entities and sentiment polarity through natural language processing technology, calculates the dynamic impact coefficient for each investment factor based on predefined factor mapping rules, and outputs a set of event vectors containing event timestamps, a list of affected investment factors, and impact intensity. The adaptive weight optimization module receives the set of event vectors and the historical multi-factor return sequence output from the event-driven factor correction module. The historical multi-factor return sequence is a set of return data of each investment factor at multiple consecutive historical time points. It adopts a network framework based on deep reinforcement learning, uses the risk-adjusted return within the rolling time window as the reward signal, generates the investment factor weight distribution for the next trading cycle through the strategy network, evaluates the state value corresponding to the current investment factor weight distribution through the value network, and outputs the optimized investment factor weight matrix for the next trading cycle. Anomaly propagation early warning module: This module obtains historical factor correlation matrices and real-time market trading data from the adaptive weight optimization module and external data interface, respectively. The real-time market trading data consists of the transaction prices and trading volumes of the assets corresponding to each investment factor during the current trading period. It constructs a dynamic directed graph with each investment factor as a node and the correlation strength between factors as edges. The historical factor correlation matrix is a correlation coefficient matrix between each investment factor calculated based on the historical multi-factor return sequence. It uses graph neural network technology to identify abnormal investment factor nodes through node embedding and updating and edge weight propagation algorithms. When an abnormal propagation path exceeding a preset threshold is detected, an early warning signal is generated that includes the investment factor at the source of the risk, the propagation path, and the expected duration of the impact. Strategy Execution Module: This module integrates the optimized investment factor weight matrix output by the adaptive weight optimization module with the early warning signal output by the anomaly propagation early warning module. Under the condition of meeting the preset transaction cost constraints and risk control thresholds, it executes the portfolio rebalancing instruction to generate the final asset allocation plan.
2. The intelligent investment benefit evaluation system with multi-factor dynamic correction according to claim 1, characterized in that: The macroeconomic event keywords preset in the event-driven factor correction module are extracted from financial news texts, central bank announcements, and regulatory documents. Market sentiment indicators include implied volatility in the options market, social media sentiment index, and search engine attention index. The threshold is defined as the change in the value of the market sentiment indicator within two consecutive preset observation periods exceeding the standard deviation set multiple calculated based on historical data of the market sentiment indicator.
3. The intelligent investment benefit evaluation system with multi-factor dynamic correction according to claim 2, characterized in that: The event-driven factor correction module calculates the dynamic impact coefficient based on predefined factor mapping rules. The specific operation is as follows: First, an attention-based text encoding model is used to transform the extracted event entities and sentiment polarities into event feature vectors. Then, a pre-trained factor sensitivity prediction model is used, taking the event feature vectors as input, to output the sensitivity adjustment parameters corresponding to each investment factor. The event-driven factor correction module generates the dynamic impact coefficient of the investment factor by multiplying the sensitivity adjustment parameters by the current weight value of the corresponding investment factor. The event-driven factor correction module further integrates the dynamic impact coefficient of each investment factor with the investment factor's identification information and event timestamp to finally form a set of event vectors.
4. The intelligent investment benefit evaluation system with multi-factor dynamic correction according to claim 3, characterized in that: In the adaptive weight optimization module, the deep reinforcement learning network framework includes a policy network and a value network; The strategy network is used to generate the investment factor weight distribution for the next trading cycle based on the current system state. The value network is used to evaluate the long-term expected return corresponding to different investment factor weight distributions under the current system state. The system state is composed of a set of event vectors, a historical multi-factor return sequence, and the current investment factor weight matrix. The current investment factor weight matrix is formed by the strategy execution module after executing the previous rebalancing instruction.
5. The intelligent investment benefit evaluation system with multi-factor dynamic correction according to claim 4, characterized in that: In the adaptive weight optimization module, the risk-adjusted return is calculated using an improved Sharpe ratio, and the calculation process includes: Within a preset rolling time window, the portfolio return series generated by the investment factor weight distribution is calculated. The excess return series is obtained by subtracting the risk-free rate. The average of the excess return series is divided by its standard deviation, and the transaction cost coefficient calculated based on the proportion of transaction amount is deducted to form the final reward signal value.
6. The intelligent investment benefit evaluation system with multi-factor dynamic correction according to claim 5, characterized in that: The process by which the adaptive weight optimization module outputs the optimized investment factor weight matrix for the next trading cycle includes: The strategy network outputs an initial weight adjustment vector based on the current system state. After normalization, this weight adjustment vector satisfies the constraint that the sum of the portfolio weights is 1. It is then linearly superimposed with the current investment factor weight matrix to form a candidate optimized investment factor weight matrix for the next trading cycle. After evaluation by the value network, the candidate scheme with the highest state value is selected from multiple candidate optimized investment factor weight matrices as the final output optimized investment factor weight matrix.
7. The intelligent investment benefit evaluation system with multi-factor dynamic correction according to claim 6, characterized in that: In the anomaly propagation early warning module, the construction process of the dynamic directed graph includes: Using each investment factor as a node, the correlation coefficient in the historical factor correlation matrix is used as the initial weight of the edge. The weight of the edge is dynamically updated based on the instantaneous correlation coefficient calculated from real-time market trading data. The instantaneous correlation coefficient is obtained by calculating the rolling correlation coefficient of the asset returns corresponding to each investment factor in the current trading period. The historical factor correlation matrix is directly provided by the adaptive weight optimization module. Real-time market trading data is obtained in real time through an external data interface. The external data interface establishes a connection with the financial market trading database through a preset financial data protocol and receives real-time market data streams released by the exchange at a preset frequency. The real-time market data streams include the transaction price, transaction volume, and timestamp information of the assets corresponding to each investment factor.
8. The intelligent investment benefit evaluation system with multi-factor dynamic correction according to claim 7, characterized in that: The process of identifying anomalous investment factor nodes using graph neural network technology includes: The feature representation of each node is updated through a multi-layer graph convolutional network. The initial feature vector of a node is composed of the current return, historical volatility, and impact coefficient from the event vector set of the investment factor. In each layer of graph convolution, the edge weight propagation algorithm aggregates the feature information of neighboring nodes according to the edge weights and outputs the updated node features through a non-linear activation function. Finally, the updated node features are input into an anomaly detection classifier, which outputs the anomaly score of each investment factor node. When the anomaly score of a node exceeds a preset threshold, the investment factor node is marked as an abnormal investment factor node.
9. The intelligent investment benefit evaluation system with multi-factor dynamic correction according to claim 8, characterized in that: In the strategy execution module, the fusion process is as follows: When an early warning signal is received from the abnormal transmission early warning module, the weight allocation of the corresponding investment factor is reduced in the optimized investment factor weight matrix according to the risk source investment factor marked in the signal, and the weight of other investment factors not marked as abnormal is increased accordingly. The transmission path and expected impact duration in the early warning signal are used to determine the magnitude and priority of the weight adjustment.
10. The intelligent investment benefit evaluation system with multi-factor dynamic correction according to claim 9, characterized in that: In the strategy execution module, transaction cost constraints include a fixed commission rate, a stamp duty rate calculated as a percentage of the transaction amount, and a market impact cost rate estimated based on the order size. Risk control threshold conditions include the maximum exposure limit of a single investment factor, the overall value at risk limit of the portfolio, and the maximum drawdown control line. Before generating the portfolio rebalancing instruction, it is necessary to ensure that the adjusted asset allocation plan simultaneously meets all transaction cost constraints and risk control threshold conditions. The adjusted asset allocation plan is obtained by solving the quadratic programming problem of the optimized investment factor weight matrix using a constrained optimization algorithm.