Financial abnormal data analysis method and system based on artificial intelligence

Abstract

The invention provides a financial abnormal data analysis method and system based on artificial intelligence, and the method comprises the steps: collecting to-be-recognized financial data, and constructing a dynamic data flow diagram according to the financial data; according to the dynamic data flow diagram; constructing an anomaly detection model and a prediction model; generating an anomaly early warning and decision support report according to output results of the anomaly detection model and the prediction model; and carrying out adaptive optimization on the anomaly detection model and the prediction model, calculating the deviation between the current model and the historical model, automatically triggering model updating when the deviation exceeds a preset threshold value, updating an anomaly early warning and decision support report according to the change of model parameters, and generating a latest anomaly early warning and decision support report. According to the method, a more accurate, real-time and explainable financial anomaly prediction and analysis scheme is provided through self-adaptive causal reasoning and a data flow graph model in combination with an incremental learning mechanism of real-time machine learning.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and in particular relates to an artificial intelligence-based financial abnormal data analysis method and system. Background Art

[0002] Financial data anomaly detection is one of the core tasks in the fields of corporate financial management, risk control and auditing. As the global economic situation and market environment become increasingly complex, the types of financial risks faced by enterprises are becoming more and more diverse and difficult to predict. Traditional financial anomaly detection methods often rely on manually set rules, fluctuation detection based on statistical models, or some basic machine learning models. These methods have played a role in some scenarios, but they have shown obvious limitations when faced with increasingly complex and dynamically changing financial data. Traditional methods usually rely on preset rules or thresholds and cannot adapt to changing financial environments and market conditions. For example, rule-based anomaly detection methods often assume that the change pattern of financial data is static, which makes it unable to adapt to the company's business changes, external economic fluctuations and sudden market events. Although anomaly detection methods based on statistical models, such as standard deviation and moving average, can detect abnormal fluctuations in financial data to a certain extent, they ignore the potential causal relationship between data, cannot effectively explain the deep-seated reasons behind the anomaly, and usually produce more false positives or false negatives. In addition, although anomaly detection methods based on machine learning can dig out nonlinear relationships in data to a certain extent, many models still lack sufficient interpretability, which makes it difficult for financial analysts and decision makers to understand and respond to detected anomalies. Especially when a company's financial data is affected by complex external factors (such as policy changes, market fluctuations, etc.) and internal changes (such as changes in management decisions, adjustments to operating strategies, etc.), existing anomaly detection methods cannot accurately grasp these complex causal relationships, resulting in failure of anomaly prediction or untimely response.

[0003] Therefore, although existing technologies have achieved results in some simple financial anomaly identification, they are often unable to cope with the dynamically changing financial environment and the complex causal relationships between data. Most existing methods rely on static models or rule-based means, lacking the ability to dynamically update and adapt. In the absence of an explanation of the deep-seated causes behind the anomalies, the early warning information provided by many existing methods is often vague, making it difficult to help managers make timely and effective decisions. At the same time, existing financial anomaly analysis systems still lack an in-depth understanding and revelation of the complex causal relationships behind corporate financial data, which limits the company's ability to prevent potential financial risks. Therefore, there is an urgent need for a new solution that can overcome the shortcomings of existing technologies and provide more accurate, real-time, and explainable financial anomaly detection and prediction. Summary of the invention

[0004] The purpose of this invention is to design a financial anomaly data analysis method and system based on artificial intelligence, combined with the incremental learning mechanism of real-time machine learning, successfully overcome the defects of anomaly detection in dynamic financial environments in the existing technology, and provide a more accurate, real-time, and explainable financial anomaly prediction and analysis solution. These innovations not only improve the efficiency of financial anomaly detection, but also enhance the system's ability to respond to complex financial problems, providing more reliable data support for financial decision-making.

[0005] In order to achieve the above object, a first aspect of the present invention provides a method for analyzing abnormal financial data based on artificial intelligence, the method comprising the following steps: S101, collecting financial data to be identified, and constructing a dynamic data flow graph based on the financial data, and introducing a causal reasoning method into the dynamic data flow graph to optimize the edge weights of the dynamic data flow graph. When new financial data flows in, the data in the sliding window is used to update the dynamic data flow graph; S102. Based on the dynamic data flow graph, an adaptive causal reasoning method based on a sliding time window is used to construct a causal reasoning model of the dynamic data flow graph. In the causal reasoning model, regression coefficients are used to quantify the causal relationship between financial indicators, and the regression coefficients are used to form the causal strength. An update strategy based on the regression results is used to update the edge weights of the data flow graph. At the same time, an edge weight threshold is designed. If the current edge weight is lower than the edge weight threshold, the edge is removed. S103, constructing anomaly detection models and prediction models. When there is a real-time data stream input, according to each node corresponding to the key features of the real-time data stream, extracting the causal inference features of the node from the causal influence of the adjacent nodes of the dynamic data flow graph, and representing the causal influence score of the node; The anomaly detection model determines whether an anomaly occurs based on the causal influence score and its historical distribution, and the prediction model predicts future moments based on the current node; S104, generating an abnormality warning and decision support report according to the output results of the abnormality detection model and the prediction model; S105. Adaptively optimize the anomaly detection model and the prediction model, and calculate the deviation between the current anomaly detection model and the current prediction model and the historical anomaly detection model and the historical prediction model. When the deviation exceeds a preset threshold, the model update is triggered, and the anomaly warning and decision support report is updated at the same time to generate the latest anomaly warning and decision support report.

[0006] Furthermore, the dynamic data flow graph is represented as ,in, Represents the node set at time t, each node Represents a financial indicator. Represents the edges connecting these nodes, and the edges represent the dependencies between indicators; The dynamic data flow graph is constructed as follows: Each financial indicator As a node , so the node set , each node represents a specific financial indicator; Initially establish edge sets based on the correlation between financial indicators ; Among them, the correlation between the financial indicators is the node and The Pearson correlation coefficient between them is used to construct preliminary edges based on the Pearson correlation coefficient: If the absolute value of the Pearson correlation coefficient is greater than the set threshold, then at the node and Establish a preliminary boundary between , , indicating a strong correlation between the two financial indicators.

[0007] Furthermore, the introduction of the causal reasoning method to optimize the edge weights of the dynamic data flow graph includes: Estimate the causal relationship between two financial indicators through Granger causality analysis, including: For two financial indicators and If the financial indicators Financial indicators Has significant predictive power, then financial indicators Financial indicators has causal influence; Wherein, the significant predictive ability is determined according to Granger causality analysis.

[0008] Furthermore, in S1, when the financial environment changes, the edges and weights in the dynamic data flow graph are dynamically adjusted by calculating new causal relationships each time and combining correlations. The weight changes of the edges in the graph are controlled by dynamic adjustment items, so that the graph structure can adapt to changes in real-time data.

[0009] Furthermore, in the causal reasoning model, when a node enters, a causal chain is generated for each node to describe the path of abnormal propagation: For each node , starting from this node, along the edge in the dynamic data flow graph Trace back and find all the nodes that affect it, forming a causal chain , which describes the chain from other indicators to The causal transmission path of Among them, for each causal path , assign a weight to it , where the weight is the product of the weights of all edges on the path: Sort all causal paths and select the top K strongest causal chains for subsequent anomaly detection; In the causal inference model for constructing a dynamic data flow graph using an adaptive causal inference method based on a sliding time window, the financial data in the time series is divided into sliding windows of length T, and the data in each window is represented as ,in, Representation Node In the time series data at time t, the reasoning goal of setting the causal relationship is to estimate the causal impact between various indicators through the regression model: Assume that within window T, the indicator Indicators The causal impact of is modeled through regression model; The causal relationship is quantified in each regression, and an L1 regularization term is introduced to suppress insignificant regression coefficients, thereby making the model more sparse and removing redundant causal paths.

[0010] Furthermore, before building the anomaly detection model and prediction model, a sliding window mechanism is used to monitor the changes in real-time data and extract key features. , from adjacent nodes Extract the causal influence of , represents the causal influence score of the node.

[0011] Furthermore, the anomaly detection model is constructed as follows: Collection Node Historical data, including nodes The mean and standard deviation of the past W moments, combined with the node The causal effect of the node is calculated by analyzing the standard deviation through the absolute difference between the causal effect and the mean. If the anomaly score is greater than the anomaly threshold, the node is considered There is an anomaly; The prediction model is constructed as follows: The causal strength estimated by the regression model is used as the node For Node The prediction coefficient of the node is used to establish a multivariate time series prediction model based on the causal relationship diagram and historical data. The observations at time t are weighted and corrected using the prediction error to obtain the final node The predicted value of ; where the prediction error follows a normal distribution; At the same time, regularization terms are introduced into the anomaly detection model and the prediction model to prevent overfitting of the model and optimize its adaptability to historical data; Among them, the update mechanism of the anomaly detection model and the prediction model is as follows: Whenever the system detects new data or anomalies, the anomaly detection model and the prediction model automatically adjust the weighting coefficients of causal influences. , prediction coefficient and the threshold for anomaly scoring , so that the prediction and detection effects can adapt to new data patterns.

[0012] Furthermore, the S104 specifically includes: A. Generation of abnormal warning report: Based on the anomaly score of each node, the degree of anomaly of each node is calculated, which is measured based on the ratio of the anomaly score to the historical standard deviation of the node; According to the degree of abnormality of each node, the nodes are divided into different abnormality levels, including mild, moderate and severe abnormalities; Use a sliding window approach to calculate the duration of anomalies: In the past W time steps, if the node If the anomaly score of a node exceeds a certain threshold continuously, the anomaly is considered to be persistent. Based on the causal graph model, the impact of the anomaly on other related nodes is evaluated. By calculating the influence propagation between nodes, the node Whether the abnormality will affect other key financial indicators; B. Decision support report generation: Future values ​​based on forecasts , showing the forecast trend of each financial indicator at a future moment, including the forecast value and the corresponding confidence interval; C. Combine the anomaly score, degree of anomaly, and predicted future value to assess the risk level for the overall financial system through indicator function analysis; D. Provide specific response recommendations to decision makers based on risk levels.

[0013] Furthermore, the deviation calculation is the average absolute deviation of the anomaly scores at time t and t-1; when the average absolute deviation exceeds a preset threshold, the update of the anomaly detection model and the prediction model is automatically triggered.

[0014] In a second aspect of the present invention, a financial abnormal data analysis system based on artificial intelligence is provided, the system comprising the following modules: The financial data collection module is used to collect the financial data to be identified, build a dynamic data flow graph based on the financial data, and introduce a causal reasoning method into the dynamic data flow graph to optimize the edge weights of the dynamic data flow graph. When new financial data flows in, the data in the sliding window is used to update the dynamic data flow graph. Dynamic data flow graph construction module is used to construct a causal reasoning model of the dynamic data flow graph based on the dynamic data flow graph using an adaptive causal reasoning method based on a sliding time window. In the causal reasoning model, regression coefficients are used to quantify the causal relationship between financial indicators, and regression coefficients are used to form causal strength. The edge weights of the data flow graph are updated using an update strategy based on regression results. At the same time, an edge weight threshold is designed. If the current edge weight is lower than the edge weight threshold, the edge is removed. The risk warning module is used to build anomaly detection models and prediction models. When there is real-time data stream input, the causal inference features of the nodes corresponding to the key features of the real-time data stream are extracted from the causal influence of the adjacent nodes in the dynamic data flow graph to represent the causal influence score of the node. The anomaly detection model determines whether an anomaly occurs based on the causal influence score and its historical distribution, and the prediction model predicts future moments based on the current node; A report generation module, used to generate anomaly warning and decision support reports based on the output results of the anomaly detection model and the prediction model; The system optimization module is used to adaptively optimize the anomaly detection model and the prediction model, and calculate the deviation between the current anomaly detection model and the current prediction model and the historical anomaly detection model and the historical prediction model. When the deviation exceeds a preset threshold, the model update is triggered, and the anomaly warning and decision support report is updated at the same time to generate the latest anomaly warning and decision support report.

[0015] The beneficial technical effects of the present invention are at least as follows: The present invention solves many problems existing in the existing technology in financial anomaly detection by combining causal reasoning, data flow graphs and real-time machine learning technology. First, in view of the limitations of the existing technology in dealing with dynamic financial data, the present invention uses an adaptive causal reasoning method to model and analyze the complex relationships between financial data, and can dynamically capture the changes in the causal relationship between data. For example, by constructing a dynamic data flow graph, the system can flexibly respond to the influence of factors such as market changes and adjustments to the company's internal business strategy, thereby achieving real-time and accurate anomaly detection. This innovation effectively overcomes the shortcomings of traditional methods that cannot adapt to complex financial environments and rapidly changing market conditions.

[0016] Secondly, existing financial anomaly detection methods usually lack an effective explanation of the causes of anomalies, resulting in insufficient financial analysis and decision support. In order to solve this problem, the present invention introduces a causal reasoning model and combines it with a data flow diagram to comprehensively analyze the causal chain behind the anomaly and accurately reveal the root cause of the abnormal data. This not only improves the accuracy of anomaly detection, but also provides transparent decision support for financial managers, helping them understand the specific context of the anomaly and then take effective measures to intervene and adjust. This innovation solves the shortcomings of traditional methods in lacking explanatory power and causal analysis, enabling financial managers to make more informed decisions.

[0017] Finally, in view of the limitations of traditional models in dynamic updating and learning, this paper introduces an incremental learning mechanism based on real-time machine learning to ensure that the system can be updated and adjusted at any time according to new financial data. This adaptive learning mechanism enables the model to reflect new trends and rules in real time in the changes of data streams, and effectively respond to emergencies or market fluctuations. Through incremental learning, the system can not only maintain a high level of prediction accuracy, but also avoid model obsolescence when the data environment changes, thereby greatly improving the real-time and adaptability of anomaly detection.

[0018] Therefore, the present invention successfully overcomes the defects of anomaly detection in dynamic financial environments by combining adaptive causal reasoning and data flow graph models with the incremental learning mechanism of real-time machine learning, and provides a more accurate, real-time, and explainable financial anomaly prediction and analysis solution. These innovations not only improve the efficiency of financial anomaly detection, but also enhance the system's ability to respond to complex financial problems, providing more reliable data support for financial decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0020] Figure 1 The present invention is a flowchart of an abnormal financial data analysis method based on artificial intelligence. DETAILED DESCRIPTION

[0021] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0022] In one or more embodiments, Figure 1As shown, the present invention discloses a method for analyzing abnormal financial data based on artificial intelligence, and the method comprises the following steps: S101, collect the financial data to be identified, and construct a dynamic data flow graph based on the financial data, and introduce a causal reasoning method into the dynamic data flow graph to optimize the edge weights of the dynamic data flow graph. Whenever new financial data flows in, the data in the sliding window is used to update the dynamic data flow graph; wherein, the dynamic data flow graph ,in Represents the node set at time t, each node Represents a financial indicator. Represents the edges connecting these nodes, and the edges represent the dependencies between indicators.

[0023] Specifically, build a dynamic, sequential financial data flow diagram , extract information from multiple dimensions and multiple financial indicators, and clearly describe the temporal dependencies and causal relationships between various financial indicators in the graph. This step provides basic input for subsequent anomaly detection and reasoning analysis.

[0024] The input is a financial dataset ,in represents the ith financial indicator (such as income, expenditure, profit, debt, etc.), which is recorded in time series, such as day, month, quarter. N represents the number of financial indicators.

[0025] Furthermore, to ensure the time consistency of the data, the data of all indicators are firstly processed with unified timestamps to ensure the consistency of the time granularity of different indicators. Missing values ​​are filled by interpolation, and outliers are eliminated by the principle of 3 times the standard deviation (or the upper and lower quartiles are used to eliminate outliers).

[0026] In order to eliminate the dimensional impact between different financial indicators, Perform standardization. The standardization formula is: ; in, For indicators The mean of For indicators The standard deviation of .

[0027] Furthermore, a dynamic data flow graph is constructed based on financial data ,in Represents the node set at time t, each node Represents a financial indicator. Represents the edges connecting these nodes, and the edges represent the dependencies between indicators.

[0028] Node definition: Each financial indicator (After preprocessing) are all treated as a node , so the node set , each node represents a specific financial indicator, such as revenue, profit, liabilities, etc.

[0029] Initial edge establishment: Initially establish edges based on the correlation between financial indicators. .set up for and The Pearson correlation coefficient is: ; like (set threshold), then at the node and Establish a preliminary boundary between , indicating a strong correlation between the two financial indicators.

[0030] Furthermore, in the initial establishment of Based on this, we further introduce causal relationship modeling and optimize edge weights The traditional correlation method cannot directly express the causal relationship between financial indicators, so it is necessary to introduce causal reasoning methods to optimize edge weights.

[0031] Furthermore, the steps of causal relationship modeling are: Granger Causality analysis is used to estimate the causal relationship between financial indicators. and ,if right has significant predictive power (i.e. The historical value can be significantly improved prediction accuracy), then right Has a causal impact.

[0032] Assume that Granger causality strength is , then the weight The initial value of can be set to: ; in, is the indicator function, if the correlation Greater than the set threshold , then the causal strength is introduced as the edge weight, otherwise the weight is 0.

[0033] Furthermore, in order to avoid overfitting and increase the stability of the model, an L1 regularization term is added to the causal strength To constrain: ; in, is the regularization strength coefficient. Through this regularization, redundant edges and weakly related causal relationships are reduced, making the data flow graph more concise and focusing on the most relevant financial indicators.

[0034] Furthermore, to ensure that the data flow graph can adapt to changes in the financial environment, a time sliding window mechanism is adopted. A fixed-length time window T is set (for example, the financial data of the past 12 months), and the window continues to slide forward as time goes by.

[0035] Whenever new financial data comes in, the graph is updated using the data in the sliding window. For example, the graph at the current time t Based on the data in the window, the correlation and causal relationship between all indicators are calculated and the new .

[0036] Due to changes in the financial environment, the causal relationship between certain financial indicators may change. Therefore, the edges and weights in the graph are dynamically adjusted by calculating new causal relationships each time and combining correlations.

[0037] In particular, the introduction of dynamic adjustment terms To update the weights: ; in, To adjust the coefficient, the weight changes of the edges in the graph are controlled by dynamically adjusting the items, so that the graph structure can adapt to the changes in real-time data.

[0038] Furthermore, through the above steps, the dynamic financial data flow graph at time t is output ,This graph contains the temporal dependency and causal relationship between financial indicators. The weight of the edge in the graph represents the causal strength and correlation between financial indicators. As a time-series data flow graph, it will serve as the input for subsequent steps (such as anomaly detection, risk prediction, and causal reasoning), supporting subsequent higher-level analysis.

[0039] S102. According to the dynamic data flow graph, an adaptive causal reasoning method based on a sliding time window is used to construct a causal reasoning model of the dynamic data flow graph. In the causal reasoning model, regression coefficients are used to quantify the causal relationship between financial indicators, and the regression coefficients are used to form the causal strength. An update strategy based on the regression results is used to update the edge weights of the data flow graph. At the same time, an edge weight threshold is designed. If the current edge weight is lower than the edge weight threshold, the edge is removed.

[0040] Specifically, in step S101, the present invention obtains a financial data flow diagram through dynamic modeling ,in is a node set containing various financial indicators. Represents the edge set between nodes, edge weight Represents the preliminary correlation between the two financial indicators. In order to conduct an in-depth analysis of the causal relationship between these indicators, the goal of this step is to build a more accurate causal network based on this preliminary correlation information and dynamically update the financial flow diagram.

[0041] Therefore, the input of step S102 is , each node represents the financial indicator at time t, and each edge Representation Node and The present invention further uses an adaptive causal reasoning method to quantize the edges in , and improve their causality.

[0042] Furthermore, in order to improve the data flow graph In order to improve the causal reasoning ability of the data, the present invention adopts an adaptive causal reasoning method based on a sliding time window. Specifically, the present invention divides the financial data in the time series into sliding windows of length T, and the data in each window is represented as: ,in, Representation Node Time series data at time t. In order to establish a more accurate causal inference model, the present invention sets the reasoning goal of causal relationship as estimating the causal influence between various indicators through a regression model.

[0043] Assume that within window T, the indicator Indicators The causal influence of is modeled by the regression model: ; in, is the regression coefficient, indicating right The effect after a lag of k steps is, is the error term, which is assumed to be normally distributed.

[0044] In order to quantify the causal relationship in each regression, the present invention introduces an L1 regularization term to suppress insignificant regression coefficients, thereby making the model more sparse and removing redundant causal paths: ; in, is the L1 norm regularization term, is the regularization strength, which controls the sparsity of causal influences.

[0045] Furthermore, in the causal inference model, the present invention calculates the regression coefficient To quantify the causal relationship between financial indicators, these regression coefficients constitute the causal strength To update the edge weights in the data flow graph , the present invention adopts an update strategy based on regression results: ; in, is the causal strength estimated from the above regression model, is a smoothing factor that controls the degree of integration of new causal information and historical information. is the edge weight at the previous moment.

[0046] In addition, in order to remove redundant edges and reduce the impact of noise, the present invention sets an edge weight threshold ,like If the value is below this threshold, the edge is removed. , making the data flow diagram more concise and focusing on important cause-effect relationships: ; Furthermore, in order to further enhance the abnormal prediction ability of financial data, the present invention generates a causal chain for each node , describing the path of abnormal propagation. The specific steps are as follows: For each node , the present invention starts from it and follows the edge in the data flow graph Trace back and find all nodes that directly or indirectly affect it This forms a chain of cause and effect. , which describes the chain from other indicators to causal transmission path.

[0047] For each causal path , the present invention assigns a weight , which is the product of the weights of all edges on the path: ; in, Reflects the shift from other indicators to the combined impact.

[0048] All causal paths are sorted and the top K strongest causal paths are selected for subsequent anomaly detection.

[0049] Furthermore, through the above causal reasoning and optimization steps, the present invention obtains an optimized causal data flow graph , which includes the following improvements: is the edge set after causal reasoning optimization, where the edge weight It reflects the strength of causal relationship between financial indicators. In the graph, some irrelevant or weakly related edges are removed, which enhances the simplicity of the graph. For each node Multiple causal propagation paths are generated, providing a reliable basis for anomaly detection and early warning.

[0050] Finally, step S102 outputs a data flow graph based on causal optimization , which can provide key support for subsequent anomaly detection and prediction. This step provides a more detailed and highly interpretable financial data flow diagram through precise causal reasoning and dynamic adjustment.

[0051] S103. Construct an anomaly detection model and a prediction model. When there is a real-time data stream input, according to each node corresponding to the key features of the real-time data stream, the causal inference features of the node are extracted from the causal influence of the adjacent nodes of the dynamic data flow graph to represent the causal influence score of the node. The anomaly detection model determines whether an anomaly occurs based on the causal influence score and its historical distribution, and the prediction model makes predictions for future moments based on the current node.

[0052] Specifically, the input comes from step S102: is the data flow graph after causal optimization. For Node The causal inference characteristics.

[0053] The goal is to build a model that can perform anomaly detection and trend prediction based on this causal inference feature and historical data. To this end, the present invention first defines a time window W, which represents the length used to process historical data. Represents a historical data set within a window, where each element Represents the financial indicator at time tk Observed value of .

[0054] Furthermore, in this step, based on the cause-effect diagram and the input data stream , the present invention adopts a sliding window mechanism to monitor the changes of real-time data and extract key features. , the present invention is from the adjacent node Extract the causal influence of , represents the causal influence score of the node.

[0055] In order to capture the dynamic relationship in the data flow, the present invention uses the following formula to calculate the causal influence of each node: ; in, is the weight coefficient, indicating that the node right The degree of influence, is the edge weight of the causal relationship, is the node at time t-1 This formula is based on the transmission of causal flow, which transfers the causal relationship between nodes through Translated into the impact on changes in financial data at the current moment.

[0056] Furthermore, the anomaly detection model is constructed: Anomaly detection is the key goal of this step. In order to identify anomalies in real-time data, the present invention compares the causal influence of each node. and its historical distribution to determine whether an abnormality occurs. Assume that node The mean of the historical data at the past W time points is , the standard deviation is , the present invention calculates the abnormal score of each node by the following formula: ; in, and Respectively represent nodes The mean and standard deviation within the historical window. If the score represents a node The degree to which current causal effects deviate from the historical normal range. , then the node Considered to be abnormal.

[0057] This abnormal score It reflects whether the behavior of each node is inconsistent with its historical pattern, thereby judging whether there are abnormal fluctuations in financial indicators. This scoring model can effectively identify abnormal situations caused by economic fluctuations or operational errors through dynamic detection of causal effects.

[0058] Furthermore, the prediction model is constructed: In addition to anomaly detection, the financial system also needs to have accurate future trend prediction capabilities. Here, the present invention establishes a multivariate time series prediction model based on the causal relationship graph and historical data to predict the future time t+1 node In order to make full use of the causal influence between nodes, the present invention designs the following causal prediction formula: ; in, Is a node For Node The prediction coefficient of Is a node The observed value at time t, is the prediction error, which is assumed to be normally distributed. In order to ensure the accuracy of the prediction results, the present invention will gradually update the prediction coefficients through the regression model. , and adjusts its adaptive weights through the training process of historical data.

[0059] Through this prediction formula, the model can combine historical data and causal structure to predict the value of financial indicators at future moments, providing data support for decision makers. The prediction results will help financial personnel to warn of potential problems in advance and take countermeasures.

[0060] Furthermore, in order to improve the stability and robustness of the model, the present invention introduces a regularization term , which is used to prevent the model from overfitting and optimize its adaptability to historical data. The form of this regularization term is: ; in, is a regularization parameter that controls the balance between model complexity and fitting accuracy. By introducing the regularization term, the present invention can avoid overfitting caused by data noise while ensuring that the model can flexibly adapt to complex financial data.

[0061] In addition, the present invention designs an abnormal enhancement mechanism by scoring the abnormality of the model. Weighting is performed to increase the weight of outliers so that the model can respond more sensitively to major anomalies. Specifically, for nodes whose anomaly scores exceed the threshold, , the present invention enhances their influence on prediction and anomaly scoring, reflecting the impact of these abnormal nodes on the entire financial data system.

[0062] Furthermore, in the end, the model outputs two results: Anomaly detection results: For each node , output anomaly score ,if , then mark the node as an abnormal node.

[0063] Future prediction results: For each node , output the predicted value at the future time t+1 , so that decision makers can predict future trends.

[0064] Through these outputs, the system can monitor financial data in real time, identify possible abnormal fluctuations in advance, and predict future financial trends based on causal models. This feedback mechanism can help financial personnel take timely remedial measures to reduce risks.

[0065] Understandably, in real applications, financial data is constantly changing, and the model needs to have adaptive capabilities to adjust its prediction and anomaly detection mechanisms. Whenever the system detects new data or anomalies, the model automatically adjusts the weighting coefficients of causal influences. , prediction coefficient and the threshold for anomaly scoring , so that the prediction and detection effects can adapt to new data patterns. The update mechanism takes the following form: ; ; in, is the smoothing factor for causal effects, is the learning rate of the prediction coefficient, For Node Through this dynamic adjustment mechanism, the system can adapt to the changing financial environment and maintain the effectiveness of real-time monitoring and forecasting.

[0066] Through the design of step S103, the system can realize anomaly detection and trend prediction in real-time financial data streams through causal relationships. Combining innovative causal impact analysis, anomaly scoring mechanism, and adaptive prediction model, financial decision makers can timely identify abnormal fluctuations, predict future trends, and make self-adjustments based on the feedback mechanism to improve the accuracy and stability of financial data analysis.

[0067] S104: Generate an abnormality warning and decision support report according to the output results of the abnormality detection model and the prediction model.

[0068] Specifically, the purpose of the abnormal warning report is to enable decision makers to quickly identify current abnormal financial indicators and make a detailed analysis of the nature of the abnormalities (degree, duration, etc.) to help them take effective management measures.

[0069] Furthermore, the abnormality score obtained from step S103 of the present invention is Representation Node Abnormal level at time t. To determine the severity of the abnormality, the present invention designs the following calculation method: Calculate the degree of anomaly for each node based on the anomaly score and the node's historical standard deviation Measured by the ratio: ; in, For Node The abnormal score of is the historical standard deviation, reflecting the volatility of the node under normal circumstances.

[0070] According to the calculated , nodes can be divided into different abnormal levels (mild, moderate, severe). The present invention uses a grading function To classify the degree of abnormality: ; in, and The pre-set thresholds distinguish between mild, moderate and severe abnormalities.

[0071] Furthermore, a sliding window method is used to calculate the duration of abnormal events. In the past W time steps, if the node Anomaly score Continuously exceeding a certain threshold , then the anomaly is considered to be persistent. According to the causal graph model , the present invention evaluates the impact of anomalies on other related nodes. By calculating the influence propagation between nodes, the node is further evaluated. Will the abnormality of the company affect other key financial indicators? The impact assessment formula is: ; in, Representation Node The abnormality of its adjacent nodes The impact of is the causal influence weight between adjacent nodes, The above anomaly evaluation helps to build a complete anomaly warning information, so that the report can show the severity, duration and possible chain reaction of the anomaly.

[0072] Furthermore, decision support reports should not only identify current abnormal situations, but also combine future forecast results to help decision makers understand possible future trends and potential risks, thereby guiding them in formulating strategies.

[0073] Based on the future value predicted in step S103 , the present invention shows the forecast trend of each financial indicator at a future moment, including the forecast value and the corresponding confidence interval. To this end, the present invention uses the following formula to calculate the confidence interval: ; where z is the critical value of the standard normal distribution (for example, z = 1.96 corresponds to a 95% confidence level), For Node Standard deviation of the forecast value. By showing the future trend forecast and its confidence interval for each node, the report can provide decision makers with specific financial trend forecasts and help them identify potential risks in the future (such as cash flow shortages or reduced revenue).

[0074] Furthermore, combined with the anomaly score , abnormality 、Future prediction and impact assessment , the present invention assesses the risk level of the overall financial system. The present invention proposes a new comprehensive risk score , which combines the severity of the anomaly and the risk of future trends: ; in, For Node The importance weight reflects the relative importance of the node in the entire financial system. It is an adjustment coefficient used to adjust the impact of future forecast results deviating from the confidence interval. is an indicator function, when the predicted value Not in the confidence interval The value is 1 when it is inside, otherwise it is 0. This formula takes into account the current severity of anomalies and future uncertainties, helping decision makers to assess the overall risk and make countermeasures.

[0075] Furthermore, based on the above risk assessment results , the report will provide decision makers with specific response suggestions. For example, if the abnormal score of certain nodes is high and the future trend is unstable, the report may suggest increasing risk reserves, adjusting investment directions, or strengthening monitoring of certain financial projects.

[0076] Furthermore, in order to ensure that decision makers can intuitively understand the report content, the present invention designs multiple visualization elements: Anomaly heat map: displays the anomaly score of each node , using color intensity to identify the severity of the anomaly.

[0077] Trend Forecast Chart: Displays forecast values ​​for future moments and its confidence interval, intuitively presenting future financial trends.

[0078] Risk distribution map: showing the overall risk score distribution of data, helping decision makers identify which areas present higher risks.

[0079] Furthermore, the report generation process is automated, and the system will automatically generate and push reports periodically or when an anomaly is detected in real time. The report content includes anomaly warning, future prediction, risk assessment and decision-making recommendations.

[0080] Regular updates: Updated reports are generated after each financial period based on new real-time data.

[0081] Dynamic feedback: When key anomalies or trend changes are detected, the system automatically updates the report and pushes it to decision makers.

[0082] Through the design of step S104, the present invention can not only provide real-time abnormal warning, but also provide comprehensive decision support for decision makers in combination with future trend forecasting. This innovative report generation method combines abnormal detection, causal relationship diagram analysis, future trend forecasting and comprehensive risk assessment, providing an efficient and systematic decision-making tool for financial management, which helps to identify problems in a timely manner and take appropriate measures.

[0083] S105. Adaptively optimize the anomaly detection model and the prediction model, and calculate the deviation between the current model and the historical model. When the deviation exceeds a preset threshold, the model update is automatically triggered. According to the changes in the model parameters, the anomaly warning and decision support report is updated to generate the latest anomaly warning and decision support report.

[0084] Specifically, to ensure that the entire system can quickly respond to changes in the market or financial environment, the present invention designs an adaptive feedback mechanism. Large fluctuations, or risk scores When a certain predetermined threshold is exceeded, the system automatically triggers the model retraining process and updates the forecast results based on the latest financial data.

[0085] Furthermore, feedback mechanism: In the adaptive mechanism of the system, the present invention calculates the deviation of the current model To evaluate the accuracy of the model: ; Where N is the total number of nodes; and are the anomaly scores at time t and t-1, respectively.

[0086] when Exceeds a preset threshold , the system will automatically trigger a model update: ; This indicator function This ensures that the model's adaptive update mechanism is only activated when the abnormal score fluctuation exceeds the set threshold, ensuring computational efficiency.

[0087] Furthermore, as the model parameters and predicted values ​​change dynamically, the system will update the report and generate the latest risk assessment and decision recommendations. The adaptively optimized model not only provides more accurate abnormal warnings, but also adjusts decision recommendations based on the latest financial data. Strategy adjustment: Based on the results of adaptive optimization, the system will automatically generate new decision suggestions. The adjusted strategy will take the following aspects into consideration: Risk Threshold Adjustment: If the risk score Beyond a certain threshold, decision recommendations may require additional risk control measures (such as increasing financial reserves).

[0088] Forecast trend revision: If the future forecast value If a large fluctuation occurs, the system will modify the strategy recommendations based on the parameters optimized by the model.

[0089] Furthermore, the system will automatically perform the following operations after each financial period: Model adaptive optimization: Automatically adjust model parameters based on real-time data and prediction results.

[0090] Report generation and optimization: Generate new decision support reports to provide the latest risk assessment and decision recommendations based on optimized models.

[0091] Through this automated optimization mechanism, the system is able to ensure a high degree of flexibility and accuracy in an ever-changing financial environment.

[0092] Through the design of step S105, the entire system implements an adaptive update mechanism based on real-time feedback, ensuring that the prediction model and anomaly detection algorithm can be automatically optimized according to the latest financial data. This not only improves the accuracy of anomaly detection, but also provides decision makers with more accurate risk assessment and decision support. Ultimately, the system can cope with the rapidly changing financial environment and continue to provide effective decision support for enterprises.

[0093] In another embodiment of the present invention, a financial abnormal data analysis system based on artificial intelligence is provided, and the system includes the following modules: The financial data collection module is used to collect the financial data to be identified, and to construct a dynamic data flow graph based on the financial data. The causal reasoning method is introduced into the dynamic data flow graph to optimize the edge weights of the dynamic data flow graph. Whenever new financial data flows in, the data in the sliding window is used to update the dynamic data flow graph; wherein, the dynamic data flow graph ,in Represents the node set at time t, each node Represents a financial indicator. Represents the edges connecting these nodes, and the edges represent the dependencies between indicators; The dynamic data flow graph construction module is used to construct a causal reasoning model of the dynamic data flow graph based on the dynamic data flow graph using an adaptive causal reasoning method based on a sliding time window. In the causal reasoning model, regression coefficients are used to quantify the causal relationship between financial indicators, and regression coefficients are used to form causal strength. The edge weights of the data flow graph are updated using an update strategy based on regression results. At the same time, an edge weight threshold is designed. If the current edge weight is lower than the edge weight threshold, the edge is removed. Among them, in the causal reasoning model, when a node enters, a causal chain is generated for each node to describe the path of abnormal propagation: For each node The present invention starts from this node and follows the edge in the dynamic data flow graph. Trace back and find all nodes that directly or indirectly affect it , forming a causal chain , which describes the chain from other indicators to The causal transmission path of Among them, for each causal path , assign a weight to it , which is the product of the weights of all edges on the path: ; in, Reflects the shift from other indicators to The combined impact of For the edge, is the edge weight; Sort all causal paths and select the top K strongest causal paths for subsequent anomaly detection; The risk warning module is used to build anomaly detection models and prediction models. When there is a real-time data stream input, the causal derivation features of the node are extracted from the causal influence of the adjacent nodes of the dynamic data flow graph according to each node corresponding to the key features of the real-time data stream, and the causal influence score of the node is expressed. The anomaly detection model determines whether an anomaly occurs based on the causal influence score and its historical distribution, and the prediction model predicts the future time based on the current node; A report generation module, used to generate anomaly warning and decision support reports based on the output results of the anomaly detection model and the prediction model; The system optimization module is used to adaptively optimize the anomaly detection model and prediction model, and calculate the deviation between the current model and the historical model. When the deviation exceeds a preset threshold, the model update is automatically triggered. According to the changes in model parameters, the anomaly warning and decision support report is updated to generate the latest anomaly warning and decision support report.

[0094] These are just some preferred embodiments of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of implementing the above embodiments and making equivalent changes according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A financial abnormal data analysis method based on artificial intelligence, characterized in that: The method comprises the following steps: S101, collecting financial data to be identified, and constructing a dynamic data flow graph based on the financial data, and introducing a causal reasoning method into the dynamic data flow graph to optimize the edge weights of the dynamic data flow graph. When new financial data flows in, the data in the sliding window is used to update the dynamic data flow graph; S102. Based on the dynamic data flow graph, an adaptive causal reasoning method based on a sliding time window is used to construct a causal reasoning model of the dynamic data flow graph. In the causal reasoning model, regression coefficients are used to quantify the causal relationship between financial indicators, and the regression coefficients are used to form the causal strength. An update strategy based on the regression results is used to update the edge weights of the data flow graph. At the same time, an edge weight threshold is designed. If the current edge weight is lower than the edge weight threshold, the edge is removed. S103, constructing anomaly detection models and prediction models. When there is a real-time data stream input, according to each node corresponding to the key features of the real-time data stream, extracting the causal inference features of the node from the causal influence of the adjacent nodes of the dynamic data flow graph, and representing the causal influence score of the node; The anomaly detection model determines whether an anomaly occurs based on the causal influence score and its historical distribution, and the prediction model predicts future moments based on the current node; S104, generating an abnormality warning and decision support report according to the output results of the abnormality detection model and the prediction model; S105. Adaptively optimize the anomaly detection model and the prediction model, and calculate the deviation between the current anomaly detection model and the current prediction model and the historical anomaly detection model and the historical prediction model. When the deviation exceeds a preset threshold, the model update is triggered, and the anomaly warning and decision support report is updated at the same time to generate the latest anomaly warning and decision support report.

2. The method for analyzing abnormal financial data based on artificial intelligence according to claim 1, characterized in that: The dynamic data flow graph is represented as ,in, Represents the node set at time t, each node Represents a financial indicator. Represents the edges connecting these nodes, and the edges represent the dependencies between indicators; The dynamic data flow graph is constructed as follows: Each financial indicator As a node , so the node set , each node represents a specific financial indicator; Initially establish edge sets based on the correlation between financial indicators ; Among them, the correlation between the financial indicators is the node and The Pearson correlation coefficient between them is used to construct preliminary edges based on the Pearson correlation coefficient: If the absolute value of the Pearson correlation coefficient is greater than the set threshold, then at the node and Establish a preliminary boundary between , ,side Indicates a strong correlation between these two financial indicators.

3. The method for analyzing abnormal financial data based on artificial intelligence according to claim 2, characterized in that: The causal reasoning method is introduced to optimize the edge weights of the dynamic data flow graph, including: Estimate the causal relationship between two financial indicators through Granger causality analysis, including: For two financial indicators and If the financial indicators Financial indicators Has significant predictive power, then financial indicators Financial indicators has causal influence; Wherein, the significant predictive ability is determined according to Granger causality analysis.

4. The method for analyzing abnormal financial data based on artificial intelligence according to claim 1, characterized in that: In S1, when the financial environment changes, the edges and weights in the dynamic data flow graph are dynamically adjusted by calculating new causal relationships each time and combining correlations. The weight changes of the edges in the graph are controlled by dynamic adjustment items, so that the graph structure can adapt to changes in real-time data.

5. The method for analyzing abnormal financial data based on artificial intelligence according to claim 1, characterized in that: In the causal reasoning model, when a node enters, a causal chain is generated for each node to describe the path of abnormal propagation: For each node , starting from this node, along the edge in the dynamic data flow graph Trace back and find all the nodes that affect it, forming a causal chain , which describes the chain from other financial indicators to The causal transmission path of For each causal path , assign a weight to it , where the weight is the product of the weights of all edges on the path: Sort all causal chains and select the top K strongest causal chains for subsequent anomaly detection; In the causal inference model for constructing a dynamic data flow graph using an adaptive causal inference method based on a sliding time window, the financial data in the time series is divided into sliding windows of length T, and the data in each window is represented as ,in, Representation Node In the time series data at time t, the reasoning goal of setting the causal relationship is to estimate the causal impact between various financial indicators through the regression model: Assume that within window T, the financial indicator Subject to financial indicators The causal impact of is modeled through regression model; The causal relationship is quantified in each regression, and an L1 regularization term is introduced to suppress insignificant regression coefficients, thereby making the model more sparse and removing redundant causal chains.

6. The method for analyzing abnormal financial data based on artificial intelligence according to claim 1, characterized in that: Before building the anomaly detection model and prediction model, a sliding window mechanism is used to monitor the changes in real-time data and extract key features. , from adjacent nodes Extract the causal influence of , represents the causal influence score of the node.

7. The method for analyzing abnormal financial data based on artificial intelligence according to any one of claims 5 or 6, characterized in that: The anomaly detection model is constructed as follows: Collection Node Historical data, including nodes The mean and standard deviation of the past W moments, combined with the node The causal effect of the node is calculated by analyzing the standard deviation through the absolute difference between the causal effect and the mean. If the anomaly score is greater than the anomaly threshold, the node is considered There is an anomaly; The prediction model is constructed as follows: The causal strength estimated by the regression model is used as the node For Node The prediction coefficient of the node is used to establish a multivariate time series prediction model based on the causal relationship diagram and historical data. The observations at time t are weighted and corrected using the prediction error to obtain the final node The predicted value of ; where the prediction error follows a normal distribution; At the same time, regularization terms are introduced into the anomaly detection model and the prediction model to prevent overfitting of the model and optimize its adaptability to historical data; Among them, the update mechanism of the anomaly detection model and the prediction model is as follows: Whenever the system detects new data or anomalies, the anomaly detection model and the prediction model automatically adjust the weighting coefficients of causal influences. , prediction coefficient and the threshold for anomaly scoring , so that the prediction and detection effects can adapt to new data patterns.

8. The method for analyzing abnormal financial data based on artificial intelligence according to claim 7, characterized in that: The S104 specifically includes: A. Generation of abnormal warning report: Based on the anomaly score of each node, the degree of anomaly of each node is calculated, which is measured based on the ratio of the anomaly score to the historical standard deviation of the node; According to the degree of abnormality of each node, the nodes are divided into different abnormality levels, including mild, moderate and severe abnormalities; Use a sliding window approach to calculate the duration of anomalies: In the past W time steps, if the node If the anomaly score of a node exceeds the threshold continuously, the anomaly is considered to be persistent. Based on the causal graph model, the impact of the anomaly on other related nodes is evaluated. By calculating the influence propagation between nodes, the node Whether the abnormality will affect other key financial indicators; B. Decision support report generation: Future values ​​based on forecasts , showing the forecast trend of each financial indicator at a future moment, including the forecast value and the corresponding confidence interval; C. Combine the anomaly score, degree of anomaly, and predicted future value to assess the risk level for the overall financial system through indicator function analysis; D. Provide specific response recommendations to decision makers based on risk levels.

9. The method for analyzing abnormal financial data based on artificial intelligence according to claim 1, characterized in that: The deviation calculation is the average absolute deviation of the anomaly scores at time t and t-1; when the average absolute deviation exceeds a preset threshold, the update of the anomaly detection model and the prediction model is automatically triggered.

10. An artificial intelligence-based financial abnormal data analysis system, characterized in that: The system includes the following modules: The financial data collection module is used to collect the financial data to be identified, build a dynamic data flow graph based on the financial data, and introduce a causal reasoning method into the dynamic data flow graph to optimize the edge weights of the dynamic data flow graph. When new financial data flows in, the data in the sliding window is used to update the dynamic data flow graph. Dynamic data flow graph construction module is used to construct a causal reasoning model of the dynamic data flow graph based on the dynamic data flow graph using an adaptive causal reasoning method based on a sliding time window. In the causal reasoning model, regression coefficients are used to quantify the causal relationship between financial indicators, and regression coefficients are used to form causal strength. The edge weights of the data flow graph are updated using an update strategy based on regression results. At the same time, an edge weight threshold is designed. If the current edge weight is lower than the edge weight threshold, the edge is removed. The risk warning module is used to build anomaly detection models and prediction models. When there is real-time data stream input, the causal inference features of the nodes corresponding to the key features of the real-time data stream are extracted from the causal influence of the adjacent nodes in the dynamic data flow graph to represent the causal influence score of the node. The anomaly detection model determines whether an anomaly occurs based on the causal influence score and its historical distribution, and the prediction model predicts future moments based on the current node; A report generation module, used to generate anomaly warning and decision support reports based on the output results of the anomaly detection model and the prediction model; The system optimization module is used to adaptively optimize the anomaly detection model and the prediction model, and calculate the deviation between the current anomaly detection model and the current prediction model and the historical anomaly detection model and the historical prediction model. When the deviation exceeds a preset threshold, the model update is triggered, and the anomaly warning and decision support report is updated at the same time to generate the latest anomaly warning and decision support report.

Images