Financial risk dynamic monitoring system and method based on big data and artificial intelligence

Abstract

The invention discloses a financial risk dynamic monitoring system and method based on big data and artificial intelligence, and relates to the technical field of financial risk monitoring, and the method comprises the following steps: obtaining life cycle data, event data and market data of a financial product; preprocessing the life cycle data, and performing mapping processing on a preprocessing result and the market data to generate weight distribution data; and performing dynamic direction comparison on the preprocessing result and the weight distribution data to generate adjustment judgment data. According to the scheme, the stage-event double-layer index structure is constructed, and the weight adjustment action range and the event triggering window are dynamically associated, so that the weight distribution data can accurately act on the risk characteristics in the specific time interval, and the technical defect that the risk portrait cannot accurately reflect the dynamic evolution in the prior art is effectively overcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of financial risk monitoring, in particular to a financial risk dynamic monitoring system and method based on big data and artificial intelligence. BACKGROUND

[0002] With the continuous development of the financial market, there are various types of financial products and frequent transactions, and financial risk management has become a key technical field to ensure market stability and prevent systemic risks. The rapid development of big data and artificial intelligence technology provides new technical means and methods for dynamic monitoring and accurate early warning of financial risks. Using big data and artificial intelligence technology to realize real-time dynamic analysis and prediction of financial risks has become an important research direction in the field of financial technology.

[0003] However, in the process of implementing the technical solutions of the embodiments of the present application, it is found that the above-mentioned technology at least has the following technical problems:

[0004] The prior art uses life cycle data, market data, and event data to model risk portraits according to fixed weight parameters, and realizes risk warning through rules or simple statistical methods. This method has certain phased risk identification ability and can be applied to the preliminary risk assessment scene of some structured financial products. However, the prior art does not consider the dynamic influence of market data and event data on weights at different stages, and due to the lack of a weighted portrait generation mechanism based on real-time weight distribution, the generated risk portrait cannot accurately reflect the real dynamic evolution of financial product risks, resulting in insufficient accuracy of financial product risk monitoring. SUMMARY

[0005] The purpose of the present application is to provide a financial risk dynamic monitoring system and method based on big data and artificial intelligence to solve the problems raised in the background technology.

[0006] In order to achieve the above-mentioned purpose, the technical solutions of the present application are as follows:

[0007] In a first aspect, the present application discloses a financial risk dynamic monitoring method based on big data and artificial intelligence, applied to the financial risk dynamic monitoring of financial products, comprising the following steps:

[0008] Obtaining life cycle data, event data, and market data of the financial product;

[0009] Preprocessing the life cycle data and mapping the preprocessing result and the market data to generate weight distribution data;

[0010] Dynamically comparing the preprocessing result and the weight distribution data to generate adjustment judgment data;

[0011] Nonlinearly fuse the event data and the preprocessing result, and apply the nonlinear fusion result to the weight distribution data to generate adjustment response data;

[0012] Bidirectionally match the adjustment response data and the adjustment judgment data to generate event adjustment data;

[0013] Image nest the preprocessing result and the adjustment response data under a stage-event double-layer index condition, weight the image nest processing result according to the time period according to the weight distribution data, and generate weighted image data;

[0014] Track map the event adjustment data and the adjustment judgment data on the image nest processing result, and further combine the weighted image data to output risk data through a deep neural network.

[0015] In a second aspect, the present application discloses a financial risk dynamic monitoring system based on big data and artificial intelligence, comprising:

[0016] A data acquisition module is configured to acquire life cycle data, event data and market data of a financial product.

[0017] A weight distribution generation module is configured to preprocess the life cycle data, and map the preprocessing result and the market data to generate weight distribution data.

[0018] An adjustment judgment data generation module is configured to dynamically compare the preprocessing result and the weight distribution data to generate adjustment judgment data.

[0019] An event adjustment data generation module is configured to nonlinearly fuse the event data and the preprocessing result, and apply the nonlinear fusion result to the weight distribution data to generate adjustment response data.

[0020] Bidirectionally match the adjustment response data and the adjustment judgment data to generate event adjustment data.

[0021] A weighted image data generation module is configured to image nest the preprocessing result and the adjustment response data under a stage-event double-layer index condition, weight the image nest processing result according to the time period according to the weight distribution data, and generate weighted image data.

[0022] A risk data output module is configured to track map the event adjustment data and the adjustment judgment data on the image nest processing result, and further combine the weighted image data to output risk data through a deep neural network.

[0023] Compared with the prior art, the present application has the following beneficial effects:

[0024] 1. This solution constructs a phase-event two-layer index structure to dynamically associate the scope of weight adjustment with the event triggering window, enabling the weight distribution data to accurately apply to risk characteristics within a specific time interval. This effectively solves the technical defect in existing technologies where risk profiles cannot accurately reflect dynamic evolution. In addition, by injecting weight change distribution data into structural path nodes, continuous dynamic updates of risk profile data in the time dimension are achieved, overcoming the risk assessment lag problem caused by the solidification of weights in traditional methods.

[0025] 2. This solution transforms dynamically adjusted data into graph structure features through trajectory mapping. Combined with the multi-level feature extraction capabilities of deep neural networks, it enables refined modeling of the risk evolution of financial products. This solves the monitoring bias problem caused by static weight allocation and single data dimension in existing technologies. At the same time, it utilizes risk profile attention area data to achieve precise positioning of key areas, thereby improving the accuracy and interpretability of risk data.

[0026] 3. This solution uses time-series matching and structure mapping to accurately associate the impact of events with the weight adjustment area of ​​a specific stage. It can dynamically adjust the correction magnitude of the affected area in the weight distribution data according to the degree of overlap between the time interval of the event and the stage weight adjustment area, thereby improving the accuracy of risk warning. Attached Figure Description

[0027] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0028] Figure 1 This is a flowchart illustrating the steps of the financial risk dynamic monitoring method based on big data and artificial intelligence according to the present invention.

[0029] Figure 2 This invention provides a schematic diagram of the process for generating weight distribution data.

[0030] Figure 3 A schematic diagram of the process for generating risk profile data for the relevant areas provided by this invention;

[0031] Figure 4 A flowchart illustrating the process of generating risk data at each stage, provided by this invention.

[0032] Figure 5 This is a schematic diagram of the module functions of the financial risk dynamic monitoring system based on big data and artificial intelligence provided by the present invention. Detailed Implementation

[0033] It is easy to understand that according to the technical solution of the application, those skilled in the art can propose a plurality of structure modes and implementation modes which can be replaced with each other without changing the essential spirit of the application. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solution of the application, and should not be regarded as the whole or as a limitation or restriction on the technical solution of the application.

[0034] Summary of the application:

[0035] In the prior art, financial risk management relies on fixed weight parameters for risk modeling, and rules or simple statistical methods are used to realize risk early warning. Although such methods have preliminary risk assessment capabilities, they cannot dynamically respond to the influence of market fluctuations and unexpected events on weight distribution. For example, when a certain financial product is in a stage of severe market fluctuations, the traditional method still uses preset weight parameters, resulting in a risk portrait that cannot reflect real-time changes, causing monitoring lag and misjudgment.

[0036] In order to solve the above problems, a risk monitoring mechanism capable of dynamically sensing market changes and event influences needs to be constructed. First, the problem of dynamic data mapping needs to be solved by dividing stages and calculating differences to extract risk characteristics and establish a weight adjustment mechanism. Second, the problem of nonlinear coupling of events and market data needs to be solved by time alignment and path comparison to realize fusion response. Finally, the problem of dynamic weighting of risk portrait needs to be solved by double-index nesting and trajectory mapping to realize multi-dimensional data integration.

[0037] After introducing the basic concept of the application, the embodiments of the application will be specifically introduced with reference to the drawings.

[0038] Embodiment one:

[0039] Please refer to Figure 1 , a financial risk dynamic monitoring method based on big data and artificial intelligence, applied to financial risk dynamic monitoring of financial products, including the following steps:

[0040] Obtain the life cycle data, event data and market data of the financial product;

[0041] Preprocess the life cycle data and map the preprocessing result and the market data to generate weight distribution data;

[0042] Dynamically compare the preprocessing result and the weight distribution data, and generate adjustment judgment data;

[0043] Nonlinearly fuse the event data and the preprocessing result, and apply the nonlinear fusion result to the weight distribution data to generate adjustment response data;

[0044] Bidirectionally match the adjustment response data and the adjustment judgment data to generate event adjustment data;

[0045] The preprocessing result is nested with the adjustment response data under the phase-event double-layer index condition, and the nested processing result is weighted according to the weight distribution data in time periods to generate weighted image data;

[0046] The event adjustment data and the adjustment judgment data are mapped on the nested processing result, and then combined with the weighted image data to output risk data through a deep neural network.

[0047] Among them, the life cycle data refers to the data set with clear time identifier generated or associated in each stage experienced during the existence of the financial product as the analysis object;

[0048] Event data refers to a set of discrete influence factors that disturb or trigger structural changes in the risk state of a financial product during its life cycle. This set is indexed by time stamp, recording the type, intensity, object and dynamic evolution sequence of the influencing event;

[0049] Market data refers to a set of structured data information that can be quantified, time-sequenced, and reflect the financial environment in the external financial market;

[0050] Preprocessing refers to the process of structuring reconstruction, phase division and variation trend modeling of the collected life cycle data on the time axis;

[0051] The preprocessing result refers to the data set generated by sequentially marking, phase dividing, calculating the difference between phases, and mapping risk characteristics of the life cycle data of the financial product according to time order;

[0052] Mapping processing refers to the process of establishing a nonlinear, multiscale, multidimensional correspondence between the preprocessing result of the life cycle data and the structured market data on a unified time axis, and generating weight distribution data reflecting the influence degree of market state on each stage of life cycle by constructing mapping index matrix and weight derivation function group;

[0053] Weight distribution data refers to a relative risk intensity distribution data structure that establishes a relationship between the risk feature distribution extracted from the life cycle data and the market state based on the change trend of the market data in each stage of the life cycle of the financial product;

[0054] Dynamic direction comparison refers to extracting the change trend encoding of the preprocessing result and the weight distribution data, generating direction consistency score through position alignment and bit-by-bit comparison, which can be realized by using trend encoder and cosine similarity calculation;

[0055] Adjustment judgment data refers to a structured data set used to assess whether the life cycle risk trend is highly consistent with the direction of market leading weight adjustment;

[0056] Nonlinear fusion refers to generating event reflection structure by structural mapping of event data and preprocessing results, and comparing paths with weight distribution data. Specifically, it can be achieved by using graph neural network and path similarity analysis;

[0057] Nonlinear fusion result refers to a complex structure data formed after nonlinear function transformation and feedback adjustment of the life cycle structure data expressed in the preprocessing result under the action of event data;

[0058] Adjustment response data refers to the dynamic risk change trend expression data caused by the action of event data on existing weight distribution data and risk characteristic structure in a specific life cycle stage;

[0059] Two-way matching refers to a multi-round data interaction process in which adjustment response data and adjustment judgment data are matched as two parties, based on their respective multi-dimensional structure and time, stage, and event index dimensions, and mutual mapping, feedback, and constraint;

[0060] Event adjustment data refers to a structured data set reflecting the impact of events on dynamic risk weight adjustment, obtained by nonlinear fusion of financial product life cycle data and event data, combined with two-way matching of weight distribution data and adjustment judgment data;

[0061] Portrait nesting processing refers to constructing a primary portrait according to stage-event double-layer index of preprocessing results and adjustment response data, and weighting according to weight distribution data. Specifically, it can be achieved by using multi-dimensional index matrix and time decay function;

[0062] Weighted portrait data refers to a multi-dimensional risk portrait constructed based on financial product life cycle data and event data, which is formed by combining risk feature mapping structure and event trigger response, and adjusting the weight according to weight distribution data in the time dimension;

[0063] Trajectory mapping refers to the process of constructing a time-event path of risk state based on event adjustment data and adjustment judgment data in the portrait structure indexed by stage-event, to form dynamic trajectory data reflecting risk fluctuation evolution;

[0064] Deep neural network refers to a multi-layer nonlinear mapping model, which aims to automatically learn the potential risk feature representation and dynamic evolution law from multi-dimensional, time-series, and complex structure financial risk portrait data;

[0065] The risk data refers to a multi-dimensional financial risk representation result output by a deep neural network based on life cycle data, market data and event data, and processed through multi-layer nonlinear mapping, weighted nesting and trajectory mapping.

[0066] The scheme realizes adaptive adjustment of the weight according to the market state through dynamic direction comparison and weight adjustment scope construction, quantifies the influence of events through structure mapping and path comparison, improves the identification ability of sudden risks, and constructs a multi-dimensional dynamic risk portrait through a stage-event double-layer index and a time weighting mechanism; through the above technical scheme, the present application solves the monitoring lag problem caused by fixed weight in the traditional method, realizes the dynamic evaluation of financial risks according to the market state and the influence of events; through nonlinear fusion and double-layer index nesting, the capture accuracy of risk characteristics in complex scenarios is improved; through dynamic direction collaborative analysis and deep neural network fusion, the prediction ability of risk evolution trend is enhanced, and real-time and accurate risk monitoring support is provided for financial products.

[0067] As introduced above, the complete scheme of the financial risk dynamic monitoring method based on big data and artificial intelligence is introduced, and the acquisition of the life cycle data, event data and market data of the financial product is introduced in detail as follows:

[0068] By accessing the product management system of the financial institution, the life cycle management interface is called, and the life cycle data of each financial product from the issuance date to the current date is obtained according to the product number;

[0069] The life cycle data includes but is not limited to: issuance time, collection period, operation period, clearing time, period dividend record, survival structure adjustment record, etc.; each data is recorded with a time stamp through a system log, and a stage identifier is formed in combination with a product structure change log;

[0070] The event data is pulled in real time through the event publishing interface of the financial information service platform (such as Wind information, Bloomberg terminal, etc.);

[0071] The event data includes but is not limited to: market announcement events (such as central bank interest rate decision), policy adjustment events (such as regulatory rule revision), sudden events (such as geopolitical conflicts, abnormal fluctuations in financial markets), etc.; each event is attached with event time, event type label and impact field identifier;

[0072] The market data is obtained from the accessed real-time financial market platform, and the market index data related to the corresponding financial product is obtained, including but not limited to: benchmark interest rate, volatility index, main market index point, exchange rate level, credit spread curve, etc.; each market data is collected according to minute or hour granularity, and the data synchronization module is used to unify the time line, so that the market data has a time corresponding relationship in different product life cycle stages.

[0073] According to the above technical scheme, the application realizes data standardization input and multi-element driving fusion, various types of data have unified time reference and structure label, so that nonlinear cross calling can be performed based on the data structure itself in the subsequent processing process, and the timeliness of risk identification is improved; it is ensured that all data have clear sources and multi-round transfer capabilities, can participate in multi-stage processing and judgment in the entire monitoring process, and data interruption or data dead angle is avoided.

[0074] As introduced above, the life cycle data, event data and market data of the financial product are obtained, and the preprocessing of the life cycle data is introduced below, which specifically includes:

[0075] The life cycle data is divided into stages, and stage partition data of each stage is generated;

[0076] The stage partition data is mapped, risk feature mapping data of each stage is generated, and the risk feature mapping data is taken as a preprocessing result.

[0077] The stage division processing refers to dividing the life cycle data of the financial product into different stages according to time or business logic, which can be realized by using a clustering algorithm based on a time window or a preset business stage rule;

[0078] The stage partition data refers to a plurality of continuous time sequence label data generated by structured processing based on a data set with time sequence attribute in the financial life cycle data (such as account activity, fund flow frequency, product holding period, customer behavior record, etc.);

[0079] The mapping processing refers to the process of constructing the structural correspondence between the life cycle stage and the risk feature based on the life cycle data that has been divided into stages;

[0080] The risk feature mapping data refers to a data structure formed by associating a set of feature indicators with risk indicating significance in each stage to the corresponding life cycle stage in the form of a mapping relationship, after the life cycle data is divided into a plurality of stages.

[0081] The above content will be described in detail as follows:

[0082] Firstly, the life cycle data is divided into stages; the stage division is based on the structural characteristics such as business process change, behavior mode switching, index transition trend reflected in the life cycle data, and the entire life cycle data is divided into a plurality of stage intervals with time or state continuity by using a preset stage division rule or a data-driven clustering algorithm; a set of stage partition data with unique identification is generated for each stage interval, and the stage partition data is used to describe the structural boundary and content framework of the life cycle data in each stage;

[0083] Secondly, for any two adjacent stages in the stage partition data, difference calculation is performed based on the difference of the structural characteristics, to obtain stage transition data;

[0084] The specific method is as follows:

[0085] The life cycle data of each pair of adjacent stages is used to construct a stage feature vector group, and then the difference measurement index between the feature vector groups, such as Euclidean distance, cosine similarity, KL divergence, correlation coefficient difference, etc., is calculated, and a group of stage transition data for describing the change structure of stage transition is generated;

[0086] Then, the stage partition data and the corresponding stage transition data are cross-mapped, which specifically includes associating and fusing the structural information of the stage partition and the change information of the stage transition, constructing the mapping relationship of each stage by a multi-dimensional mapping algorithm, and forming risk feature mapping data reflecting the risk features of each stage;

[0087] Finally, the risk feature mapping data is used as the preprocessing result of the life cycle data.

[0088] Through the above technical solution, the present application can effectively extract the implicit intra-stage risk features and inter-stage dynamic change features in the life cycle data, generate risk feature mapping data with multi-dimensional correlation, provide more accurate input data basis for subsequent dynamic weight adjustment and risk portrait generation, and thus improve the accuracy of financial product risk dynamic monitoring.

[0089] As introduced above, the preprocessing of the life cycle data is performed, and the mapping processing of the preprocessing result and the market data to generate weight distribution data is introduced below, please refer to Figure 2 , Figure 2 The flowchart for generating weight distribution data provided by the embodiment of the present application is shown in the figure, and the generation of weight distribution data specifically includes:

[0090] The preprocessing result is analyzed by a sliding window to generate risk feature stability data of the current stage;

[0091] The stage partition data and the market data are time-series aligned to generate time-series cross-index data;

[0092] According to the risk feature stability data, the corresponding time section is located in the time-series cross-index data to generate weight adjustment scope data of the current stage;

[0093] The weight adjustment scope data is used as input, and a preset dynamic adjustment rule set is called to generate weight distribution data of the current stage.

[0094] The sliding window analysis refers to a process of continuously sampling and calculating statistical features of the collected data structure in a window by using a fixed-length time window that slides by a step on the preprocessed results.

[0095] The risk feature stability data refers to a structured data set formed by dynamically analyzing the change law of the financial risk features in a specific time window, and used to depict the volatility and trend stability of the risk features.

[0096] The time series alignment operation refers to a process of establishing a corresponding relationship between different time series data, and realizing accurate mapping and synchronization of multiple heterogeneous data sequences in the time dimension by matching time labels and time windows.

[0097] The time series cross-index data refers to a data structure used to realize accurate correspondence and synchronization alignment between multi-dimensional time series data.

[0098] The weight adjustment scope data refers to dynamically positioning and calibrating the specific time section and its corresponding spatial range of weight adjustment in the current risk monitoring stage based on the risk feature stability data and the time series cross-index data.

[0099] The preset dynamic adjustment rule set refers to a structured rule base containing multiple groups of decision rules and adjustment strategies for weight distribution generation.

[0100] The above content will be described in detail as follows:

[0101] The preprocessed results are subjected to structural path similarity clustering, and the path similarity scores between the current stage and the previous and subsequent stages are calculated, with the specific calculation formula as follows:

[0102]

[0103] In the formula, represents the path similarity score of stage , represents the path structure similarity result between stage and stage , represents the path structure similarity calculation function, such as cosine similarity, represents the structural path vector of stage , and the aggregation calculation result is obtained to get the mapping stability index sequence of the current stage;

[0104] The mapping stability index sequence is subjected to the sliding window analysis method, the window size and step are set, and the sequence is segmented and counted.

[0105] determining whether the mapping stability index sequence continuously rises or continuously falls within the continuous N sliding windows, and if so, marking the continuous N sliding windows as a characteristic trend section of the current stage;

[0106] structuring and combining the characteristic trend section and the corresponding determination result to generate risk characteristic stability data of the current stage; the risk characteristic stability data includes the following fields: characteristic trend section start time point, characteristic trend section end time point, determination result, and characteristic trend section length;

[0107] performing time sequence alignment operation on the stage partition data and the market data to generate time sequence cross-index data;

[0108] The specific process is as follows:

[0109] standardizing the timestamps of the stage partition data and the market dynamic data to a unified time format and resolution;

[0110] For the case where the time points do not completely coincide, an interpolation method (such as linear interpolation or spline interpolation) is used to fill the time difference to generate the aligned time sequence;

[0111] Based on the aligned time sequence, time sequence cross-index data is constructed; the time sequence cross-index data takes the time point as the key and associates the corresponding indexes of the stage partition data and the market dynamic data.

[0112] According to the start and end points of the characteristic trend section in the risk characteristic stability data, a sliding window expansion operation is performed (for example, expanding one sliding window forward and backward respectively), and the corresponding time section is located in the time sequence cross-index data, the time sequence cross-index data of the corresponding time section is extracted, and the risk adjustment time domain data of the current stage is constructed; the risk adjustment time domain data not only includes the expanded time range, but also retains the corresponding time sequence cross-index data as the structural basis for subsequent mask operation;

[0113] Using the expanded time interval indicated by the risk adjustment time domain data as a mask time window, the time sequence cross-index data is subjected to window covering operation, and all data segments with data cross events, path characteristic labels or stage mapping relationships within the time window are extracted to generate weight adjustment scope data of the current stage;

[0114] Taking the weight adjustment scope data as input, a preset dynamic adjustment rule set is called to dynamically adjust the weight through the condition-action rules in the rule library, and weight change distribution data of the current stage is generated;

[0115] The preset dynamic adjustment rule set is obtained through expert knowledge modeling, which is a set of structured condition-action mapping rules.

[0116] The technical scheme can capture the influence of market changes on weight distribution in real time, realize intelligent optimization of weight parameters by combining dynamic adjustment rules, and dynamically adjust the weight distribution of risk characteristics in each stage according to real-time changes in market data, thereby effectively eliminating evaluation deviation caused by market fluctuations and product stage misalignment, accurately limiting the weight action range by constructing a risk adjustment time domain, avoiding the influence of noise data on distribution calculation, and improving the accuracy of dynamic monitoring of financial product risks.

[0117] The above is mapping the preprocessing result and market data to generate weight distribution data. The following introduces dynamic direction comparison of the preprocessing result and weight distribution data to generate adjustment judgment data, which specifically includes:

[0118] The change trend of the preprocessing result and the weight distribution data in the same stage is extracted and coded to obtain a risk direction coding sequence and a weight direction coding sequence;

[0119] In the time sequence range limited by the weight adjustment scope data in the same stage, the risk direction coding sequence and the weight direction coding sequence are positionally aligned to generate each cooperative vector pair;

[0120] Each cooperative vector pair is compared bit by bit, and a direction consistency score is calculated according to the comparison result;

[0121] The adjustment judgment data is generated by integrating the risk direction coding sequence, the weight direction coding sequence, and the direction consistency score.

[0122] The change trend refers to the change direction of data in the time dimension, which can be realized by difference calculation or sliding window statistical method;

[0123] The coding operation refers to the process of converting original structural change trend data into a discrete symbol sequence with directionality, rhythm, and stage characteristics;

[0124] The risk direction coding sequence refers to the conversion of the change trend of the risk characteristics into a sequence composed of discrete numerical values, which can be realized by binary coding or piecewise linear coding method;

[0125] The weight direction coding sequence refers to the conversion of the change trend of the weight distribution into a sequence composed of discrete numerical values, which can be realized by the same coding rule as the risk direction coding;

[0126] The position alignment refers to the structure mapping process of one-to-one correspondence between the time stamp and the coding index in the risk direction coding sequence and the weight direction coding sequence in the same stage;

[0127] The synergistic vector pair refers to the combination of risk direction encoding and weight direction encoding in the same time interval, which can be generated by timestamp alignment or sliding window matching method;

[0128] Bit-by-bit comparison refers to the operation of matching the encoding elements at the same position in the risk direction encoding sequence and the weight direction encoding sequence in the synergistic vector pair for structural direction consistency;

[0129] Direction consistency score refers to the similarity measurement of the encoding values at the corresponding positions in the synergistic vector pair, which can be generated by cosine similarity or Hamming distance calculation method.

[0130] The above contents are described in detail as follows:

[0131] The change trend of the pre-processing result and the weight distribution data in the same stage is extracted respectively, and the change trend is encoded to obtain the risk direction encoding sequence and the weight direction encoding sequence;

[0132] The change trend extraction and encoding processing of the pre-processing result are exemplarily described to obtain the risk direction encoding sequence:

[0133] Based on the pre-processing result, the risk value structure change trend between stages of the life cycle is extracted, and three change directions are used for encoding: rising (↑), falling (↓), and flat (→), thereby generating the risk direction encoding sequence. For example, if the pre-processing result of a stage of the life cycle is: [0.3, 0.5, 0.4, 0.6, 0.5], the risk direction encoding sequence is "↑↓↑↓";

[0134] In the time sequence range defined by the weight adjustment scope data in the same stage, the risk direction encoding sequence and the weight direction encoding sequence are positionally aligned to generate each synergistic vector pair;

[0135] Bit-by-bit comparison is performed on each synergistic vector pair, and the direction consistency score is calculated according to the comparison result;

[0136] The following rules are used for direction similarity judgment:

[0137] If the corresponding positions are consistent in direction (such as ↑ and ↑, or ↓ and ↓), the score is 1;

[0138] If the directions are opposite (such as ↑ and ↓), the score is -1;

[0139] If one side is flat (→), the score is 0.5;

[0140] Direction consistency score = sum of all position scores / total length of synergistic vector pair;

[0141] determining whether the direction consistency score is greater than a preset discrimination threshold, if yes, marking as direction consistent, otherwise marking as direction deviation;

[0142] Comprehensive judgment result, risk direction encoding sequence, weight direction encoding sequence, direction consistency score, structured combination, generate adjustment judgment data.

[0143] The above technical scheme, the present application can solve the problem of weight adjustment lag caused by lack of dynamic direction comparison in the prior art, through encoding sequence generation and collaborative vector analysis, realize the direction consistency quantitative evaluation of risk characteristics and weight distribution, so that the adjustment judgment data can accurately reflect the dynamic matching state of the two, and further improve the adaptability of the financial risk monitoring model to real-time market changes.

[0144] In order to better understand the above content, an application scenario example is given as follows:

[0145] Applied to the financial risk monitoring of a certain financial product;

[0146] Get the risk preprocessing result sequence of a certain financial product of an enterprise at different stages of the life cycle from the risk life cycle monitoring system, and the specific data is as follows:

[0147] Preprocessing result (sampling point in a certain stage): [0.3, 0.5, 0.4, 0.6, 0.5];

[0148] At the same time, get the weight distribution data corresponding to the same stage from the risk weight adjustment module: [0.4, 0.4, 0.45, 0.42, 0.47];

[0149] Collect the time sequence range limited by the weight adjustment scope data of this stage, and determine the time sequence interval boundary of the subsequent trend alignment;

[0150] Based on the preprocessing result, the risk direction encoding sequence is extracted by using the three-state trend encoding method:

[0151] Rule definition:

[0152] Upward ↑: the current value is higher than the previous sampling point;

[0153] Downward ↓: the current value is lower than the previous sampling point;

[0154] Flat →: the difference between the current value and the previous sampling point is within the preset threshold range (±0.02);

[0155] Application:

[0156] 0.3→0.5: up ↑;

[0157] 0.5→0.4: down ↓;

[0158] 0.4→0.6: up ↑;

[0159] 0.6→0.5: down ↓;

[0160] Result: risk direction encoding sequence = "↑↓↑↓";

[0161] Weight direction encoding sequence generation:

[0162] Similarly, the same encoding rule is applied to the weight change data to extract the weight change trend:

[0163] 0.4→0.4: flat →;

[0164] 0.4→0.45: up ↑;

[0165] 0.45→0.42: down ↓;

[0166] 0.42→0.47: up ↑;

[0167] Result: weight direction encoding sequence = "→↑↓↑"

[0168] According to the weight adjustment scope data to determine the timing range, the risk direction encoding sequence and the weight direction encoding sequence are aligned according to the position to form a pair of collaborative vectors:

[0169] Collaborative vector pair sequence = [(↑, →), (↓, ↑), (↑, ↓), (↓, ↑)];

[0170] The direction score of each pair of collaborative vectors is calculated by using a preset rule:

[0171] Rule:

[0172] Consistent (↑ and ↑, ↓ and ↓) gets 1 point;

[0173] Opposite (↑ and ↓, ↓ and ↑) gets -1 point;

[0174] One side flat (→) gets 0.5 points;

[0175] Calculation process:

[0176] (↑, →): 0.5 points;

[0177] (↓, ↑): -1;

[0178] (↑, ↓): -1;

[0179] (↓, ↑): -1;

[0180] Cumulative score: 0.5-1-1-1=-2.5;

[0181] Direction consistency score=-2.5 / 4=-0.625;

[0182] The preset threshold is 0.3, and the score-0.625<0.3, which is determined as direction deviation;

[0183] The structured combination of risk direction encoding sequence, weight direction encoding sequence, direction consistency score and determination result generates adjustment judgment data, and the format example is as follows:

[0184] {"risk direction encoding sequence": "↑↓↑↓", "weight direction encoding sequence": "→↑↓↑", "direction consistency score": -0.625, "direction consistency determination": "deviation"}.

[0185] As introduced above, the dynamic direction comparison of preprocessing result and weight distribution data is performed to generate adjustment judgment data. The nonlinear fusion of event data and preprocessing result is introduced below, and the nonlinear fusion result is applied to the weight distribution data to generate adjustment response data, which specifically includes:

[0186] The event data and the weight adjustment scope data of each stage are time sequence matched to generate event coverage domain data of each stage;

[0187] The event coverage domain data and the preprocessing result of the same stage are structurally mapped to generate event reflection structure data, which is taken as the nonlinear fusion result;

[0188] The nonlinear fusion result and the weight distribution data are mapped path compared to generate adjustment response data.

[0189] Among them, the time sequence matching refers to the process of structurally mapping and associating the time stamp of each event in the event data with the time boundary of the weight adjustment scope corresponding to each stage based on the discrete time points and time intervals on the time axis;

[0190] Event coverage domain data refers to the associated data set obtained by aligning the event data and the weight adjustment scope data in the time dimension through time sequence matching operation, which can be realized by using sliding window algorithm or dynamic time warping algorithm;

[0191] Structural mapping processing refers to the operation of feature fusion of event coverage domain data and preprocessing result through nonlinear transformation, which can be realized by using graph convolution network or attention mechanism;

[0192] Event reflection structure data refers to the composite data structure obtained by structurally mapping the event coverage domain data and the preprocessing result in the same risk stage;

[0193] The mapping path comparison refers to a process of similarity measurement between the event reflection structure data and the weight distribution data on the adjustment path, which can be realized by using a dynamic time warping algorithm or a path matching algorithm.

[0194] The above content will be described in detail as follows:

[0195] Firstly, the event data and the weight adjustment scope data corresponding to each stage are aligned according to the time sequence;

[0196] Specifically, for each risk stage, the time label of the event data in the time range of the stage is extracted, all events in the time range are filtered out to form an event subset corresponding to the stage, and the time interval information of the weight adjustment scope data is combined to establish a time sequence mapping relationship between the event and the weight adjustment scope, and finally event coverage domain data reflecting the event time distribution and the action range of each stage is generated;

[0197] Subsequently, for the event coverage domain data generated in the same life cycle stage, the pre-processing result of the stage is combined, and a structure mapping method is used to match the two;

[0198] The specific operation includes mapping the time and space nodes in the event coverage domain data to the risk structure path and nodes, calculating the disturbance intensity and influence path of the event on the risk structure by a nonlinear mapping function, and finally forming event reflection structure data reflecting the event influence trajectory and its nonlinear action characteristics, and the specific calculation formula is as follows:

[0199]

[0200]

[0201] In the formula, represents the event reflection structure data, represents the event coverage domain data, represents a single event The nonlinear mapping function of the pre-processing result , , represents an adjustment parameter, represents the distance measurement between the time point of a single event and the related node in the pre-processing result , represents the weight of the pre-processing result .

[0202] Finally, the event reflection structure data is mapped and compared with the weight change distribution data, specifically including calculating the correlation coefficient of the event reflection structure data and the weight change distribution data by the Pearson correlation coefficient method, and generating the adjustment response data based on the comprehensive calculation results.

[0203] As introduced above, the event data and the preprocessing results are nonlinearly fused, and the nonlinear fusion results are applied to the weight distribution data to generate the adjustment response data. The adjustment response data and the adjustment judgment data are bidirectionally matched to generate the event adjustment data, specifically including:

[0204] The adjustment response data and the adjustment judgment data are respectively time-synchronized and feature vectorized;

[0205] The feature vectors of the adjustment response data and the adjustment judgment data are bidirectionally matched by using a bidirectional matching algorithm;

[0206] Based on the matching results, the feature information of the adjustment response data and the adjustment judgment data is fused to generate the event adjustment data;

[0207] The above content will be described in detail as follows:

[0208] The adjustment response data and the adjustment judgment data are respectively time series preprocessed to correct their time stamps and time resolutions, so as to ensure that the two data are completely aligned in the time dimension;

[0209] Numerical indicators representing the risk dynamic fluctuation characteristics of events, such as fluctuation amplitude, duration, frequency component, etc., are extracted from the adjustment response data;

[0210] Adjustment consistency indicators are extracted from the adjustment judgment data, including adjustment direction consistency measure and adjustment intensity difference value, etc.

[0211] The above features are converted into a unified dimensional vector form;

[0212] The two sets of vectorized features are matched and analyzed by using a bidirectional matching algorithm (such as bidirectional nearest neighbor matching, cross similarity calculation, etc.);

[0213] Based on the matching results, the feature information of the adjustment response and the adjustment judgment is fused to construct the event adjustment data.

[0214] Through the above technical solutions, the application realizes deep correlation analysis of event influence and risk adjustment consistency, improves the accuracy and timeliness of the event adjustment data, thereby enhancing the precision and response speed of financial risk dynamic monitoring, and effectively supporting the real-time decision and early warning capability of risk management.

[0215] As introduced above, the adjustment response data and the adjustment judgment data are bidirectionally matched to generate event adjustment data. The following introduces that after the event adjustment data is generated, portrait index structure data is generated according to stage partition data and event coverage domain data, which specifically includes:

[0216] Interval mapping processing is performed on the event coverage domain data to construct an event mapping function.

[0217] A two-dimensional index matrix is constructed according to the stage partition data and the event mapping function.

[0218] A bidirectional mapping structure is generated based on the two-dimensional index matrix; the bidirectional mapping structure includes a stage-to-event mapping relationship and an event-to-stage mapping relationship.

[0219] The portrait index structure data is generated by comprehensively considering the event coverage domain data, the two-dimensional index matrix, and the bidirectional mapping structure.

[0220] Interval mapping processing refers to a mathematical mapping process of mapping the event influence time intervals defined in the event coverage domain data to the corresponding stage partition time periods;

[0221] The event mapping function refers to a correlation function generated by calculating the overlap degree of the event time interval and the stage time label; specifically, interval overlap coefficient calculation or time window matching algorithm can be used to implement it.

[0222] The two-dimensional index matrix refers to a matrix structure constructed by taking the stage time label as the row index and the event time interval as the column index; specifically, sparse matrix storage technology can be used to implement it.

[0223] The bidirectional mapping structure refers to a composite data structure that simultaneously contains a stage-to-event forward index and an event-to-stage reverse index; specifically, a bidirectional hash table or a graph database can be used to implement it.

[0224] The portrait index structure data refers to a multi-dimensional index data structure constructed based on the time and event double-axis mapping.

[0225] The following describes the above content in detail:

[0226] First, the stage partition data is discretized to obtain a stage time label sequence; discretization is achieved by dividing the continuous time interval into several equal or unequal time units.

[0227] Subsequently, according to the time interval corresponding to each event in the event coverage domain data, the time tag sequence in the above stage is matched by overlapping, and an event mapping function is constructed; the event mapping function determines the influence weight of the event on each stage time unit by calculating the overlapping degree between the event time interval and the time unit represented by the stage time tag, and then realizes the mapping of the event time interval to the stage time tag sequence;

[0228] Next, based on the stage partition data and the event mapping function, a two-dimensional index matrix is constructed; the rows of the two-dimensional index matrix represent each time unit in the stage time tag sequence, the columns represent each event in the event coverage domain, and the matrix elements correspond to the mapping weight or correlation strength between the event and the stage time unit;

[0229] Then, based on the two-dimensional index matrix, a bidirectional mapping structure is generated, which includes stage-to-event mapping relationship and event-to-stage mapping relationship, and is used to describe the mapping from the stage time unit to the event and the mapping from the event to the stage time unit, respectively;

[0230] Finally, the stage time tag sequence, the event coverage domain data, the two-dimensional index matrix and the bidirectional mapping structure are integrated to form the portrait index structure data.

[0231] The present scheme can accurately capture the dynamic association relationship between stages and events by constructing a bidirectional mapping structure containing an overlapping matching mechanism. Through the above technical scheme, the present application effectively solves the problem of inaccurate association between events and stages in the prior art, which leads to distortion of the risk portrait. By establishing a bidirectional index mechanism based on time overlapping matching, the dynamic coupling of stage division and event influence is realized, so that the risk portrait can accurately reflect the combined influence of cross-stage events.

[0232] As introduced above, after generating the event adjustment data, the portrait index structure data is generated according to the stage partition data and the event coverage domain data. The generation of the portrait index structure data is introduced as follows: please refer to Figure 3 , Figure 3 The flowchart for generating the risk portrait focus area data provided by the embodiment of the present application, which specifically includes:

[0233] According to the adjustment response data and the risk feature stability data, a structure comparison mapping matrix is constructed;

[0234] The structure continuity score between each structure pair in the structure comparison mapping matrix is calculated;

[0235] It is judged whether the structure continuity score is greater than a preset disturbance threshold. If yes, it is determined as a continuous state, otherwise it is determined as a structure disturbance, and a structure disturbance labeling result is generated.

[0236] The adjustment stability judgment data is generated by integrating the adjustment response data and the risk characteristic stability data.

[0237] The risk image stability label data is generated by embedding the risk characteristic stability data and the adjustment stability judgment data into the nodes of the portrait index structure data.

[0238] The risk portrait focus area data of each stage is generated by comparing the risk image stability label data with the portrait index structure data.

[0239] The structure comparison mapping matrix is a two-dimensional matrix formed by cross-mapping the structure characteristics of the adjustment response data and the risk characteristic stability data. It can be realized by using multi-dimensional tensor decomposition combined with feature alignment algorithm.

[0240] The structure continuity score is a quantitative index obtained by calculating the trend similarity of the adjustment response data and the risk characteristic stability data in the time series. It can be realized by using dynamic time warping algorithm combined with cosine similarity calculation.

[0241] The disturbance threshold is a critical value for distinguishing normal fluctuations from abnormal disturbances. It can be determined by historical data analysis combined with sliding window statistical method.

[0242] The structure disturbance annotation result is the label information of the continuity state of each structural unit (node or substructure) in the risk portrait, which is determined based on the structure comparison mapping matrix calculated from the adjustment response data and the risk characteristic stability data.

[0243] The adjustment stability judgment data is a multi-dimensional structural annotation data constructed based on the structure continuity matching and dynamic disturbance judgment results of the adjustment response data and the risk characteristic stability data.

[0244] The risk image stability label data is a multi-dimensional label set formed by integrating the adjustment stability judgment result and the risk characteristic stability data. It can be realized by using graph embedding technology combined with feature splicing method.

[0245] The structure comparison is a process of quantitatively evaluating the spatial relationship, time sequence continuity and topological association between elements in the high-dimensional and multi-dimensional mapping matrix constructed for different data structures in the risk dynamic monitoring method.

[0246] The risk image focus area data refers to a set of key time-event space regions in the multi-stage risk image that show significant structural disturbance, insufficient risk stability, and strong adjustment response, determined based on the structural comparison result of the risk image stability label data and the image index structure data.

[0247] The above will be described in detail as follows:

[0248] First, based on the obtained adjustment response data and risk feature stability data, a node correspondence mapping and feature vector extraction technique is used to construct a structural comparison mapping matrix.

[0249] Specifically, the dynamic risk fluctuation features in the adjustment response data and the stability indicators in the risk feature stability data are mapped in a multi-dimensional space, and for the same time stage and event index node, a structural comparison mapping matrix is formed, and each element in the structural comparison mapping matrix represents the matching degree or similarity between two structure nodes.

[0250] For each structure pair in the structural comparison mapping matrix, a weighted similarity algorithm is used to calculate the structural continuity score.

[0251] A disturbance threshold is preset, and the structural state is determined according to the structural continuity score: when the structural continuity score is greater than the disturbance threshold, it is determined that the structure pair is in a continuous state; when the structural continuity score is lower than the disturbance threshold, it is determined that the structure pair has occurred structural disturbance, and the corresponding structural disturbance annotation result is generated.

[0252] The adjustment stability judgment data is generated by integrating the structural disturbance annotation result, the adjustment response data, and the risk feature stability data.

[0253] The risk feature stability data and the generated adjustment stability judgment data are embedded in each node in the image index structure data to form the risk image stability label data; the embedding process is realized through node attribute expansion and multi-level association mapping.

[0254] Finally, the risk image stability label data and the image index structure data are structurally compared, and the focus area in the multi-stage risk image is identified based on the label changes and spatial distribution characteristics between nodes; this process uses a nonlinear clustering and region division algorithm, combined with label change amplitude and adjacency relationship, to generate risk image focus area data for each stage.

[0255] The scheme realizes real-time dynamic monitoring of the adjustment process by constructing a structure comparison mapping matrix and a continuity score calculation mechanism. Meanwhile, by embedding stability labels into the portrait structure nodes, the identification accuracy of the risk focus area is expanded from a single dimension to a time-space multi-dimensional correlation dimension. Through the above technical solution, the present application effectively solves the adjustment lag problem caused by the lack of dynamic stability evaluation in the prior art, can capture the matching deviation between the adjustment strategy and the risk characteristics in real time during the risk evolution process of financial products, thereby improving the sensitivity of the risk portrait to abnormal fluctuations. By constructing a multi-dimensional stability label and a focus area identification mechanism, the risk monitoring system can accurately locate the potential disturbance source and avoid global misjudgment caused by local fluctuations.

[0256] As introduced above, after generating the portrait index structure data, the adjustment response data and the risk characteristic stability data are matched. The following introduces portrait nesting processing of the preprocessing result and the adjustment response data, and weighting of the portrait nesting processing result according to the weight distribution data to generate weighted portrait data, which specifically includes:

[0257] The preprocessing result is nested under the phase time period index of the portrait index structure data, and the adjustment response data is nested under the event window index, so as to construct primary portrait construction data under the condition of phase-event double-layer index, and take it as the portrait nesting processing result;

[0258] According to the portrait nesting processing result, the weight distribution data is injected into the structure path node according to the phase time period, and the weight adjustment scope data is taken as a boundary limitation condition to construct weighted portrait generation data.

[0259] The phase-event double-layer index condition refers to constructing a nested structure by taking the phase time period and the event window as independent and associated index dimensions, which can be realized by using the hierarchical index mechanism of the time series database;

[0260] The primary portrait construction data refers to the structured data set for expressing the dynamic flow state of the risk structure in the time-event coordinate system, which is formed by nesting and mapping the preprocessing result and the adjustment response data according to the phase time period and the event window double dimensions based on the multi-dimensional time-space index condition.

[0261] The above contents will be described in detail as follows:

[0262] The risk characteristic mapping subset corresponding to each phase time period in the preprocessing result is accurately nested under each time period index node according to the phase time period index in the portrait index structure data;

[0263] The dynamic disturbance information corresponding to each event window in the adjustment response data is nested under each event window index node according to the event window index in the portrait index structure data;

[0264] In combination with the two types of nested data, relying on the two-dimensional index system composed of stage time period index and event window index, the preprocessing results and the adjustment response data are nested and fused in the stage-event dual dimension to form preliminary portrait construction data;

[0265] The preliminary portrait construction data constructed above is taken as the portrait nesting processing result;

[0266] For the portrait nesting processing result, the weight distribution data is accurately injected into the path nodes of the portrait structure according to the stage time period index; the injection process maps the weight information to the specific node position of the risk path according to the time period division of the weight distribution data;

[0267] At the same time, the weight adjustment scope data is used to set the boundary limit condition of the weighting processing; the boundary limit condition is used to limit the effective range of weight injection, prevent the weight information from exceeding the reasonable action area, and ensure the spatial and logical consistency of the weighted portrait;

[0268] By applying the weight distribution data and the weight adjustment scope data to the portrait nesting processing result respectively, the final weighted portrait generation data is constructed.

[0269] The present scheme dynamically associates the weight adjustment scope with the event trigger window by constructing a stage-event double-layer index structure, and further realizes continuous dynamic update of risk portrait data in the time dimension by injecting weight change distribution data into the structure path nodes; through the above technical scheme, the present application effectively solves the technical defect that the risk portrait cannot accurately reflect the dynamic evolution in the prior art, realizes the spatio-temporal association of risk feature data and event response data through the stage-event double-layer index structure, and makes the generated weighted portrait data able to reflect the influence of market changes and sudden events on risk assessment in real time by combining the dynamic weight injection mechanism.

[0270] As introduced above, the preprocessing result and the adjustment response data are nested and processed, and the portrait nesting processing result is weighted according to the weight distribution data to generate weighted portrait data, and the event adjustment data and the adjustment judgment data are mapped on the portrait nesting processing result to output risk data through a deep neural network, please refer to Figure 4 , Figure 4 The flowchart for generating risk data of each stage provided by the embodiment of the present application, which generates risk data of each stage, specifically includes:

[0271] The event adjustment data and the adjustment judgment data are taken as inputs and are respectively mapped to the graph structure path nodes and the path flow direction scalar in the portrait nesting processing result, and risk adjustment trend data is output through structure trend analysis.

[0272] According to the weighted portrait data and the risk adjustment trend data, output the risk level data of each stage by a deep neural network;

[0273] Integrate the risk level data and the risk portrait focus area data to obtain the risk data of each stage.

[0274] Among them, the graph structure path node refers to the basic unit node of the risk characteristic state on the specific time-event coordinate axis in the multi-dimensional nested structure of the portrait nesting processing result;

[0275] The path flow direction scalar refers to the quantitative value of the directionality and amplitude of the change of the risk factors along a certain risk path in the graph structure of the portrait nesting processing result;

[0276] The structure trend analysis refers to the quantitative calculation of the connection relationship of the graph structure path node and the change direction of the path flow direction scalar, which can be realized by using a graph convolution network or a time series propagation algorithm;

[0277] The risk adjustment trend data refers to the trend quantitative index generated by analyzing the change mode of the graph structure path node and the path flow direction scalar, which can specifically include trend strength, direction consistency, and propagation rate;

[0278] The risk level data refers to the comprehensive quantitative description of the financial risk under the influence of multiple time stages and events.

[0279] The above contents will be described in detail as follows:

[0280] Map the event adjustment data to the graph structure path node in the portrait nesting processing result;

[0281] Specifically, according to the node time sequence coordinates in the portrait index structure data, the risk adjustment amplitude of each time point in the event adjustment data is correspondingly mapped to the corresponding path node to form a node weight sequence. In the mapping process, a time synchronization algorithm is used to ensure the strict correspondence between the event adjustment data timestamp and the path node time point, and to avoid time sequence misplacement;

[0282] At the same time, the adjustment judgment data is mapped to the path flow direction scalar in the portrait nesting processing result. The mapping obtains the strength and direction information of the path flow direction by quantitatively processing the time sequence characteristics and cooperativity indexes in the adjustment judgment data;

[0283] After the above mapping is completed, the structure trend analysis is performed, which includes the following specific operations:

[0284] Calculate the time sequence change rate of the path node weight, and the specific calculation formula is as follows:

[0285]

[0286] wherein, denotes the node , denotes the node , denotes the path node weight of the node at time , denotes the path node weight of the node at time ,

[0287] denotes the time difference window;

[0288]

[0289] wherein, denotes the spatial gradient of the path segment , denotes the path flow direction scalar of the path segment , denotes the distance metric of the path segment .

[0290] According to the structural trend analysis result, the risk adjustment trend data is calculated by using a nonlinear regression method, and the specific calculation formula is as follows:

[0291]

[0292] wherein, denotes the risk adjustment trend data of the node , denotes the fusion factor of the time trend and the spatial gradient, denotes the adjacent node set of the node , denotes the spatial gradient of the path segment .

[0293] A deep neural network is constructed; the deep neural network includes an input layer, a graph structure nested coding layer, a time sequence evolution discrimination layer, and an output layer.

[0294] The input layer receives the input weighted portrait data and the risk adjustment trend data.

[0295] The graph structure nested coding layer encodes the structure of the stage node and the event influence edge in the weighted portrait data by using a graph convolution mechanism.

[0296] The risk adjustment trend data is nested in the graph convolution process to obtain a graph structure nested tensor.

[0297] The time evolution discrimination layer introduces a time-aware gated recurrent structure to perform multi-stage dynamic modeling on the graph structure nested tensor;

[0298] In the modeling, the risk adjustment trend data is dynamically introduced to control the weight, and a stage risk evolution vector sequence is obtained;

[0299] The output layer uses a fully connected layer to project the stage risk evolution vector sequence to a fixed-dimensional risk level space and output the risk level data of each stage;

[0300] The risk level data and the risk portrait attention area data are combined to obtain the risk data of each stage.

[0301] The above technical scheme can realize fine modeling of the risk evolution of the financial product based on the synergistic effect of dynamic trajectory mapping and deep learning, solve the monitoring deviation problem caused by static weight distribution and single data dimension in the prior art, and make the generated risk data more consistent with the dynamic fluctuation characteristics in the actual market environment through structure trend analysis and region attention mechanism.

[0302] Embodiment two:

[0303] Please refer to Figure 5 , the financial risk dynamic monitoring system based on big data and artificial intelligence, comprising:

[0304] A data acquisition module is configured to acquire life cycle data, event data, and market data of a financial product.

[0305] A weight distribution generation module is configured to preprocess the life cycle data and map the preprocessing result and the market data to generate weight distribution data.

[0306] An adjustment judgment data generation module is configured to dynamically compare the preprocessing result and the weight distribution data to generate adjustment judgment data.

[0307] An event adjustment data generation module is configured to nonlinearly fuse the event data and the preprocessing result, and apply the nonlinear fusion result to the weight distribution data to generate adjustment response data.

[0308] The adjustment response data and the adjustment judgment data are bidirectionally matched to generate event adjustment data.

[0309] A weighted portrait data generation module is configured to portrait-nest the preprocessing result and the adjustment response data under the stage-event double-index condition, and weight the portrait-nested processing result according to the time period based on the weight distribution data to generate weighted portrait data.

[0310] The risk data output module is configured to perform track mapping on the portrait nesting processing result by combining the event adjustment data and the adjustment judgment data, and further combining the weighted portrait data, and outputting risk data through a deep neural network.

[0311] The embodiment has the same technical effects as those of the first embodiment.

[0312] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. The data mentioned in the present application is normalized and dimensionally unified before performing xx calculation.

[0313] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A dynamic monitoring method for financial risks based on big data and artificial intelligence, applied to the dynamic monitoring of financial risks in financial products, characterized in that... Includes the following steps: Acquire lifecycle data, event data, and market data for financial products; The lifecycle data is preprocessed, and the preprocessing results are mapped to the market data to generate weight distribution data; The preprocessing results and the weight distribution data are dynamically compared in direction to generate adjustment judgment data; The event data and preprocessing results are nonlinearly fused, and the nonlinear fusion result is applied to the weight distribution data to generate adjustment response data. The adjustment response data and the adjustment judgment data are matched bidirectionally to generate event adjustment data; The preprocessing results and the adjustment response data are used to perform a nested profile processing with stage-event two-level index conditions, and the profile nesting processing results are weighted according to the weight distribution data according to the time period to generate weighted profile data. The event adjustment data and adjustment judgment data are mapped onto the profile nesting processing result, and then combined with the weighted profile data to output risk data through a deep neural network.

2. The method for dynamic monitoring of financial risks based on big data and artificial intelligence according to claim 1, characterized in that: Preprocessing the lifecycle data specifically includes: The lifecycle data is divided into stages to generate stage partition data for each stage; The stage partition data is mapped to generate risk feature mapping data for each stage, and this data is used as the preprocessing result.

3. The method for dynamic monitoring of financial risks based on big data and artificial intelligence according to claim 2, characterized in that: The preprocessing results are mapped to the market data to generate weight distribution data, specifically including: A sliding window analysis is performed on the preprocessing results to generate risk characteristic stability data for the current stage; Perform time-series alignment on the stage partition data and the market data to generate time-series cross-index data; Based on the risk characteristic stability data, the corresponding time segment is located in the time series cross-index data to generate the weight adjustment scope data for the current stage. The weight adjustment scope data is used as input, and a preset set of dynamic adjustment rules is invoked to generate the weight distribution data for the current stage.

4. The method for dynamic monitoring of financial risks based on big data and artificial intelligence according to claim 3, characterized in that: The dynamic directional comparison between the preprocessing results and the weight distribution data to generate adjustment judgment data specifically includes: The change trends of the preprocessing results and the weight distribution data in the same stage are extracted and encoded to obtain the risk direction encoding sequence and the weight direction encoding sequence. Within the time frame defined by the weight adjustment scope data in the same stage, the risk direction encoding sequence and the weight direction encoding sequence are aligned to generate various cooperative vector pairs; For each pair of cooperative vectors, perform a bit-by-bit comparison and calculate the directional consistency score based on the comparison results; By combining the risk direction coding sequence, the weight direction coding sequence, and the direction consistency score, adjustment judgment data is generated.

5. The method for dynamic monitoring of financial risks based on big data and artificial intelligence according to claim 3, characterized in that: The process of performing nonlinear fusion of the event data and preprocessing results, and applying the nonlinear fusion result to the weighted distribution data to generate adjusted response data, specifically includes: The event data is matched temporally with the weight adjustment scope data of each stage to generate event coverage data for each stage. The event coverage domain data and the preprocessing results at the same stage are subjected to structure mapping processing to generate event-reflecting structure data, which is then used as a nonlinear fusion result. The nonlinear fusion results are compared with the weight distribution data through a mapping path to generate the adjustment response data.

6. The method for dynamic monitoring of financial risks based on big data and artificial intelligence according to claim 5, characterized in that: After generating the event adjustment data, the process further includes generating profile index structure data based on the stage partition data and the event coverage domain data, specifically including: Perform interval mapping processing on the event coverage data to construct an event mapping function; Construct a two-dimensional index matrix based on the stage partition data and the event mapping function; A bidirectional mapping structure is generated based on the two-dimensional index matrix; the bidirectional mapping structure includes a stage-to-event mapping relationship and an event-to-stage mapping relationship. By combining the event coverage data, the two-dimensional index matrix, and the bidirectional mapping structure, a profile index structure data is generated.

7. The method for dynamic monitoring of financial risks based on big data and artificial intelligence according to claim 6, characterized in that: After generating the profile index structure data, the process further includes matching the adjustment response data and the risk characteristic stability data, specifically including: Based on the adjustment response data and the risk characteristic stability data, a structural alignment mapping matrix is ​​constructed; Calculate the structural continuity score between each structural pair in the structural alignment mapping matrix; Determine whether the structural continuity score is greater than a preset disturbance threshold. If it is, the structure is determined to be in a continuous state; otherwise, it is determined that a structural disturbance has occurred, and a structural disturbance labeling result is generated. By combining the structural disturbance annotation results, the corresponding adjustment response data, and the risk characteristic stability data, adjustment stability judgment data is generated. The risk feature stability data and the adjustment stability judgment data are jointly embedded into the nodes of the portrait index structure data to construct risk image stability label data; By structurally comparing the risk image stability label data with the profile index structure data, risk profile attention area data for each stage is generated.

8. The method for dynamic monitoring of financial risks based on big data and artificial intelligence according to claim 7, characterized in that: The preprocessing results are combined with the adjusted response data to perform a profile nesting process, and the profile nesting process results are weighted according to the weight distribution data to generate weighted profile data. Specifically, this includes: The preprocessing results are nested under the stage time period index of the portrait index structure data, and the adjustment response data is nested under the event window index. The primary portrait construction data is constructed using the stage-event dual-level index conditions, and it is used as the portrait nesting processing result. Based on the image nesting processing results, the weight distribution data is injected into the structural path nodes according to the stage time period, and the weight adjustment scope data is used as the boundary constraint condition to construct the weighted image generation data.

9. The method for dynamic monitoring of financial risks based on big data and artificial intelligence according to claim 8, characterized in that: The event adjustment data and adjustment judgment data are mapped onto the profile nesting processing result, and then combined with weighted profile data to output risk data through a deep neural network. Specifically, this includes: The event adjustment data and adjustment judgment data are used as inputs and mapped to the graph structure path nodes and path flow scalars in the profile nesting processing results, respectively. Through structural trend analysis, risk adjustment trend data is output. Based on the weighted profile data and risk adjustment trend data, risk level data for each stage is output through a deep neural network. By combining the risk level data and the risk profile attention area data, risk data for each stage is obtained.

10. A dynamic monitoring system for financial risks based on big data and artificial intelligence, characterized in that: include: The data acquisition module is used to acquire lifecycle data, event data, and market data of financial products. The weight distribution generation module is used to preprocess the life cycle data and map the preprocessing results to the market data to generate weight distribution data. The adjustment judgment data generation module is used to dynamically compare the preprocessing results and the weight distribution data to generate adjustment judgment data. The event adjustment data generation module is used to perform nonlinear fusion of the event data and the preprocessing result, and apply the nonlinear fusion result to the weight distribution data to generate adjustment response data. The adjustment response data and the adjustment judgment data are matched bidirectionally to generate event adjustment data; The weighted portrait data generation module is used to perform portrait nesting processing on the preprocessing results and the adjustment response data with stage-event two-level index conditions, and to weight the portrait nesting processing results according to the weight distribution data according to the time period to generate weighted portrait data. The risk data output module is used to perform trajectory mapping between the event adjustment data and the adjustment judgment data on the profile nesting processing result, and then combine the weighted profile data to output risk data through a deep neural network.

Images