Blockchain-based financial risk early warning method and system

Abstract

The application provides a kind of financial risk early warning method and system based on blockchain, belongs to financial risk early warning technical field, including based on smart contract automatic acquisition blockchain account book real-time data, compare and verify data authenticity with cloud historical database and separate trend and noise component.Application time series decomposition detects abnormal sequence, adopts multivariate causal discovery algorithm to mine causal path to extract core risk indicators.Combined with industry benchmark data flow to build dynamic standard financial trajectory, use random forest model to quantify trajectory deviation to assess risk exposure and divide dynamic threshold.Finally, integrate abnormal sequence, core indicators and threshold, predict future financial risk probability through deep neural network.This method improves the accuracy, timeliness and interpretability of risk early warning through data credibility verification, causal inference and dynamic evaluation.

Description

Technical Field

[0001] This invention relates to the field of financial risk early warning technology, specifically to a blockchain-based financial risk early warning method and system. Background Technology

[0002] Current financial risk early warning systems generally face the fundamental problem of insufficient data source reliability. Traditional methods mainly rely on financial statement data provided by enterprises themselves, which suffers from information lag and difficulty in verifying its authenticity. Existing technologies lack effective automatic verification mechanisms in the data collection stage, making it impossible to effectively verify the authenticity and completeness of financial data. This data quality deficiency directly affects the reliability of subsequent analysis results, causing early warning signals to be generated based on distorted or outdated information.

[0003] At the data analysis level, existing methods have shortcomings in processing financial time-series data. Conventional techniques directly detect anomalies in the raw data, failing to effectively separate long-term trend components from short-term noise components. This makes anomaly identification results susceptible to interference from random fluctuations. After detecting anomalies, existing systems mostly remain at the level of phenomenon description, lacking in-depth exploration of the causal relationships behind the anomalies. They often rely on simple threshold alarms or correlation analysis, failing to reveal the intrinsic connections and driving paths between various financial indicators.

[0004] Current risk assessment models generally employ static benchmarks and fixed thresholds, making them ill-suited to rapidly changing market environments. This rigid assessment system fails to accurately reflect the true risk profile of financial entities under current market conditions. Furthermore, the quantification of risk exposure in these models is often based on linear assumptions, making it difficult to precisely characterize complex nonlinear relationships. Ultimately, risk predictions are often based on limited feature inputs, failing to fully utilize multi-layered risk information. Summary of the Invention

[0005] The purpose of this invention is to provide a blockchain-based financial risk early warning method and system to solve at least one of the technical problems existing in the background art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a blockchain-based financial risk early warning method, comprising:

[0008] The system automatically collects real-time financial transaction data from the blockchain distributed ledger through smart contracts, and compares the real-time financial transaction data with the historical financial database in the cloud to separate the long-term trend components and short-term noise components in the real-time financial transaction data.

[0009] The time series decomposition algorithm is applied to process the long-term trend component and the short-term noise component in order to detect abnormal deviation sequences in the real-time financial transaction data;

[0010] A multivariate causal discovery algorithm is used to perform causal path mining and co-occurrence pattern analysis on the abnormal deviation sequence to extract core risk indicators;

[0011] The core risk indicators are input into the dynamic financial model, combined with industry benchmark data streams, to construct a standard financial trajectory. By quantifying the deviation between the standard financial trajectory and the actual financial trajectory, a random forest model is integrated to assess the risk exposure of the financial entity and classify risk thresholds.

[0012] Based on the abnormal deviation sequence, the core risk indicators, and the risk threshold, a deep neural network model is used to predict the probability of financial risks in future accounting periods and generate risk warning messages.

[0013] Preferably, the step of inputting the core risk indicators into the dynamic financial model, combining industry benchmark data streams to construct a standard financial trajectory, and then using a random forest model to assess the risk exposure of the financial entity and classify risk thresholds by quantifying the deviation between the standard financial trajectory and the actual financial trajectory, includes:

[0014] Using the core risk indicators as input, the data stream is integrated with industry benchmark data. An exponential smoothing algorithm is applied to denoise the multi-source historical financial data to obtain purified financial data. A multinomial regression model is used to fit the purified financial data to generate a standard financial trajectory. The multi-source historical financial data includes: financial reports of companies in the same industry, real transaction flow of the target financial entity in multiple accounting periods, historical default records, and credit rating logs.

[0015] By comparing the actual financial trajectory with the standard financial trajectory, the deviation between the actual financial trajectory and the standard financial trajectory is calculated, and the deviation is measured by the mean squared error index to obtain the deviation score.

[0016] Based on the offset score, the financial entity's risk exposure is assessed using a random forest model to generate an initial risk assessment value. An evolutionary strategy algorithm is then used to optimize the initial risk assessment value to obtain the financial entity's risk score.

[0017] Using the aforementioned risk score and a seasonal autoregressive model, the risk dynamics of financial entities within a predetermined time window are simulated to obtain risk simulation results. These results are then interpreted, and risk thresholds are defined.

[0018] Preferably, the step of assessing the risk exposure of the financial entity based on the offset score using a random forest model to generate an initial risk assessment value, and then optimizing the initial risk assessment value using an evolutionary strategy algorithm to obtain the risk score of the financial entity, includes:

[0019] The offset score is processed by deep feature extraction using a convolutional neural network to generate a basic risk assessment value.

[0020] Based on the aforementioned basic risk assessment values, a hybrid frog-leap algorithm and an artificial bee colony algorithm are used to jointly optimize the random forest model to obtain the optimized risk model.

[0021] The offset score is input into the optimized risk model, the model evaluation value is output, and the Grey Wolf Optimization Algorithm and Flame Algorithm are applied to refine the model evaluation value to generate a refined evaluation value.

[0022] By using the Bayesian averaging method, the basic risk assessment value, the optimized risk model, and the refined assessment value are integrated to generate a risk score for the financial entity.

[0023] Preferably, the step of inputting the offset score into the optimized risk model, outputting a model evaluation value, and applying the Grey Wolf Optimization Algorithm and the Flame Algorithm to refine the model evaluation value to generate a refined evaluation value includes:

[0024] The optimized risk model is invoked, and the offset score is used as the input feature to perform risk inference calculation and output a preliminary risk score.

[0025] The initial risk score is searched using the Grey Wolf optimization algorithm to locate the optimal solution. The initial risk score is then adjusted based on the optimal solution to obtain the Grey Wolf optimization result.

[0026] Based on the gray wolf optimization results, combined with the flame algorithm, a local refined search is performed to further optimize the preliminary risk score and generate a refined post-evaluation value.

[0027] Preferably, the multivariate causal discovery algorithm is used to perform causal path mining and co-occurrence pattern analysis on the abnormal deviation sequence to extract core risk indicators, including:

[0028] A data matrix is ​​constructed based on the abnormal deviation sequence. Independent component analysis is used to perform feature dimensionality reduction on the data matrix to obtain a compressed data matrix. A multivariate causal discovery algorithm is applied to traverse the compressed data matrix to discover frequent causal patterns.

[0029] The frequent causal patterns are verified by structural equation model causality test, the causal relationships of the financial transaction data are determined, and co-occurrence rules are screened from the frequent causal patterns. The causal relationships and the co-occurrence rules are then combined to form a key rule set.

[0030] Information gain, Pearson correlation coefficient, and Manhattan distance are used as evaluation metrics to assess the causal relationships and co-occurrence rules of the key rule set and generate aggregate scores.

[0031] Based on the aggregated score, candidate core risk indicators are selected from the frequent causal patterns. The support vector machine model is used to perform weight allocation processing on the candidate core risk indicators to extract the core risk indicators.

[0032] Preferably, the evaluation of the causal relationships and co-occurrence rules of the key rule set using information gain, Pearson correlation coefficient, and Manhattan distance as evaluation metrics, and the generation of an aggregate score, includes:

[0033] Calculate the information gain of causal relationships and co-occurrence rules in the key rule set to obtain an association score;

[0034] Using the Pearson correlation coefficient, a linear correlation analysis is performed on the causal relationships and co-occurrence rules in the key rule set, and the correlation coefficients of the causal relationships and co-occurrence rules in the key rule set are calculated to obtain a correlation score;

[0035] The Manhattan distance is used to evaluate the dispersion of the causal relationship and the co-occurrence rule, and the discrete values ​​of the causal relationship and the co-occurrence rule are calculated to obtain the dispersion score;

[0036] By combining the correlation score, the relevance score, and the dispersion score, a weighted score is applied to each causal relationship and co-occurrence rule in the key rule set to generate an aggregate score.

[0037] Preferably, the application of a time series decomposition algorithm to process the long-term trend component and short-term noise component to detect abnormal deviation sequences in the real-time financial transaction data includes:

[0038] Based on the long-term trend components and short-term noise components, a time series decomposition algorithm is used to perform pattern separation processing on the real-time financial transaction data to identify normal fluctuation patterns and abnormal fluctuation patterns. Based on the normal fluctuation patterns and the abnormal fluctuation patterns, the decomposition results are obtained.

[0039] By matching the decomposition results with a preset normal fluctuation template, financial transaction data that does not conform to the normal fluctuation template are detected and labeled as abnormal deviation sequences.

[0040] Preferably, the step of automatically collecting real-time financial transaction data from the blockchain distributed ledger via smart contracts and comparing the real-time financial transaction data with a historical financial database in the cloud to separate the long-term trend component and short-term noise component from the real-time financial transaction data includes:

[0041] Deploy smart contracts on blockchain nodes to periodically capture real-time financial transaction data and verify the authenticity of the real-time financial transaction data;

[0042] The real-time financial transaction data is synchronized to the cloud-based historical financial database, time series alignment is performed, and a moving average algorithm is applied to smooth data fluctuations, separating long-term trend components and short-term noise components.

[0043] Hash functions are used to verify the data integrity of the long-term trend component and the short-term noise component, ensuring that the transmission process is tamper-proof.

[0044] Preferably, the step of predicting the probability of financial risk in future accounting periods using a deep neural network model based on the abnormal deviation sequence, the core risk indicators, and the risk threshold, and generating a risk warning message, includes:

[0045] The abnormal deviation sequence, the core risk indicators, and the risk threshold are integrated to form an input feature vector, and a deep neural network model is trained to capture risk patterns.

[0046] Using a trained deep neural network model, predict the probability of financial risk in future accounting periods and output the probability distribution;

[0047] Based on the probability distribution, risk warning messages are generated and distributed through a blockchain network.

[0048] Secondly, the present invention provides a blockchain-based financial risk early warning system, comprising:

[0049] The acquisition module is used to automatically collect real-time financial transaction data from the blockchain distributed ledger through smart contracts, and compare the real-time financial transaction data with the cloud historical financial database to separate the long-term trend component and short-term noise component in the real-time financial transaction data.

[0050] The extraction module is used to process the long-term trend components and short-term noise components using a time series decomposition algorithm to detect abnormal deviation sequences in the real-time financial transaction data; and to use a multivariate causal discovery algorithm to perform causal path mining and co-occurrence pattern analysis on the abnormal deviation sequences to extract core risk indicators.

[0051] The segmentation module is used to input the core risk indicators into the dynamic financial model, combine industry benchmark data streams to construct a standard financial trajectory, and integrate a random forest model to assess the risk exposure of financial entities and segment risk thresholds by quantifying the deviation between the standard financial trajectory and the actual financial trajectory.

[0052] The prediction module is used to predict the probability of financial risk in future accounting periods based on the abnormal deviation sequence, the core risk indicators, and the risk threshold using a deep neural network model, and generate risk warning messages.

[0053] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the blockchain-based financial risk early warning method as described in the first aspect.

[0054] Fourthly, the present invention provides a computer device including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor invokes the program instructions to execute the blockchain-based financial risk early warning method as described in the first aspect.

[0055] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the blockchain-based financial risk warning method as described in the first aspect.

[0056] The beneficial effects of this invention are: it effectively separates long-term trends from short-term noise in the data, and the improved data quality reduces the sensitivity to false signals and noise in the subsequent anomaly detection stage; it applies a multivariate causal discovery algorithm to mine causal paths and analyze co-occurrence patterns of detected abnormal deviations, inferring the potential causal direction between financial indicators, and identifying fundamental core risk-driving indicators from intertwined anomalies; by quantifying the deviation between this dynamic trajectory and the actual financial trajectory, the risk assessment benchmark can be adjusted in real time with the market environment, overcoming the lag of static benchmarks.

[0057] The advantages of additional aspects of the invention will be set forth more clearly in the following description or will be learned by practice of the invention. Attached Figure Description

[0058] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a flowchart of a blockchain-based financial risk early warning method according to an embodiment of the present invention.

[0060] Figure 2 This is a flowchart illustrating the model evaluation value refinement process described in an embodiment of the present invention.

[0061] Figure 3 This is a flowchart illustrating the real-time financial data acquisition and component separation process described in an embodiment of the present invention.

[0062] Figure 4 This is a causal relationship strength matrix analysis diagram of the blockchain financial risk indicators described in this embodiment of the invention.

[0063] Figure 5 This is a graph showing the separation analysis of long-term trends and short-term noise in blockchain financial transaction data as described in an embodiment of the present invention. Detailed Implementation

[0064] 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 denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0065] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0066] Example 1

[0067] In this embodiment 1, a blockchain-based financial risk early warning system is provided, including:

[0068] The acquisition module is used to automatically collect real-time financial transaction data from the blockchain distributed ledger through smart contracts, and compare the real-time financial transaction data with the cloud historical financial database to separate the long-term trend component and short-term noise component in the real-time financial transaction data.

[0069] The extraction module is used to process the long-term trend components and short-term noise components using a time series decomposition algorithm to detect abnormal deviation sequences in the real-time financial transaction data; and to use a multivariate causal discovery algorithm to perform causal path mining and co-occurrence pattern analysis on the abnormal deviation sequences to extract core risk indicators.

[0070] The segmentation module is used to input the core risk indicators into the dynamic financial model, combine industry benchmark data streams to construct a standard financial trajectory, and integrate a random forest model to assess the risk exposure of financial entities and segment risk thresholds by quantifying the deviation between the standard financial trajectory and the actual financial trajectory.

[0071] The prediction module is used to predict the probability of financial risk in future accounting periods based on the abnormal deviation sequence, the core risk indicators, and the risk threshold using a deep neural network model, and generate risk warning messages.

[0072] like Figure 1 As shown, in this embodiment, the above-described system is used to implement a blockchain-based financial risk early warning method, which includes:

[0073] Through smart contracts deployed on blockchain network nodes, real-time financial transaction data recorded in the distributed ledger is automatically and continuously collected. The data verification logic built into the smart contracts ensures the authenticity and immutability of the collected data. The collected real-time financial transaction data is then compared with historical financial databases stored in the cloud. This comparison process involves time alignment and standardization of the data, and data decomposition techniques, such as seasonal decomposition or moving averages, are applied to effectively separate long-term trend components from short-term random noise components in the real-time data. Time series decomposition algorithms are applied to conduct in-depth analysis of the separated long-term trend components and short-term noise components, identifying outliers or sequences that deviate from normal fluctuation patterns, thereby detecting abnormal deviation sequences. Multivariate causal discovery algorithms are used to mine causal paths and analyze co-occurrence patterns in the detected abnormal deviation sequences to reveal potential causal relationships and linkages between different financial variables, and to extract the core risk indicators most sensitive to risk responses. Next, these core risk indicators are input into a dynamic financial model. This model combines real-time industry benchmark data streams to construct a standard financial trajectory reflecting the industry's health level. By quantitatively calculating the deviation between the financial entity's actual financial trajectory and this standard trajectory, and integrating ensemble learning models such as random forests, the model comprehensively assesses the financial entity's risk exposure, thereby classifying different levels of risk thresholds. Based on the abnormal deviation sequences, core risk indicators, and determined risk thresholds obtained from the aforementioned steps, a deep neural network model is used to learn the nonlinear mapping relationship between complex risk characteristics and the probability of future risk occurrence. This predicts the probability of financial risk in a specific future accounting period and automatically generates specific risk warning messages, which are recorded on the blockchain to ensure traceability.

[0074] See Figure 2In one specific embodiment, the implementation process begins with the input of core risk indicators, which are derived from the results of a prior multivariate causal discovery analysis of abnormal deviation sequences. In practice, these core risk indicators serve as the primary input features of the dynamic financial model, which also needs to integrate continuously updated industry benchmark data streams from external sources. The industry benchmark data stream is a multi-source dataset, comprising publicly available financial reports from companies in the same industry, actual transaction records of the target financial entity over multiple accounting periods, historical default event records, and credit rating change logs from third-party credit rating agencies. An exponential smoothing algorithm is applied to preprocess the multi-source historical financial data, filtering out random fluctuations and abnormal spikes to obtain a purified set of financial data. A multinomial regression model is then used to fit the purified financial data to a curve, generating a standard financial trajectory that represents the general development pattern of the industry by finding the best-fit curve.

[0075] In practice, a standard financial trajectory serves as the evaluation benchmark, compared to the actual financial trajectory of the financial entity. The actual financial trajectory is derived from a processed sequence of real transaction data recorded in the blockchain distributed ledger. By calculating the difference between each data point on the actual financial trajectory and its corresponding point on the standard financial trajectory, the degree of deviation can be quantified. The mean squared error (MSE) metric is selected as the standard, calculating the average of the squares of all differences to output a comprehensive offset score. This offset score reflects the degree of deviation of the financial entity's actual performance from industry standards. In practice, the offset score is then fed into a random forest model for preliminary analysis. The random forest model consists of multiple decision trees, which can integrate the results of multiple weak classifiers to assess the financial entity's risk exposure and generate an initial risk assessment value. This initial risk assessment value is a preliminary risk quantification result.

[0076] Evolutionary strategy algorithms are introduced to optimize initial risk assessment values. Evolutionary strategy algorithms are population-based optimization algorithms that search the solution space by simulating mutation, recombination, and selection processes in biological evolution. In practice, the evolutionary strategy algorithm uses the initial risk assessment value as a starting point, iteratively generates new candidate solutions and evaluates their fitness, ultimately finding a better risk score. The risk score is a refined numerical representation of a financial entity's risk exposure.

[0077] In practical implementation, to further improve the depth and accuracy of risk assessment, convolutional neural networks (CNNs) are used for deep feature extraction of offset scores. Through its multiple convolutional and pooling layers, the CNN automatically learns the complex, non-linear patterns inherent in the offset scores and generates a base risk assessment value. This base risk assessment value contains deep features mined from the original offset scores. In practice, the performance optimization of the random forest model is achieved through the collaborative work of a hybrid frog-jumping algorithm and an artificial bee colony algorithm. The hybrid frog-jumping algorithm is responsible for global exploration, simulating the foraging behavior of a frog population to find promising areas within a broad solution space. The artificial bee colony algorithm is responsible for fine-tuning local areas, simulating the process of bees searching for nectar sources, performing a deep search within the area located by the hybrid frog-jumping algorithm. The hybrid frog-jumping algorithm and the artificial bee colony algorithm jointly adjust the hyperparameters of the random forest model, thereby obtaining an optimized risk model with stronger predictive performance.

[0078] The offset score is input into the optimized risk model, which performs forward propagation calculations and outputs a model evaluation value. The model evaluation value is the optimized risk model's single-step risk assessment. In practice, the Grey Wolf Optimization Algorithm is used to refine the model evaluation value. Simulating the social hierarchy and hunting mechanisms of a wolf pack, the Grey Wolf Optimization Algorithm uses iterative population search to locate the optimal solution region and adjusts the model evaluation value based on the found optimal solution, producing the Grey Wolf optimized result. The Flame Algorithm is then activated, performing a local refinement search based on the Grey Wolf optimization result. Simulating the phenomenon of sparks flying during flame flickering, the Flame Algorithm performs denser sampling near the Grey Wolf optimization result, further refining the model evaluation value and ultimately producing a refined evaluation value. The refined evaluation value is the result after processing by two levels of optimization algorithms.

[0079] The Bayesian averaging method is used to integrate basic risk assessment values, model assessment values ​​output from optimized risk models, and refined assessment values. Bayesian averaging is a statistical method that generates a more robust and reliable risk score for financial entities by weighted averaging these values, taking into account the uncertainties of information from different sources. The risk score is a comprehensive quantitative output of risk assessment. In practice, a seasonal autoregressive (SAR) model is used to analyze the time-series characteristics of the risk score. The SAR model can capture seasonal fluctuations and autocorrelation in the data. Using the risk score as input, the SAR model simulates the dynamic risk of financial entities within a predetermined time window. The simulation generates a series of possible risk evolution paths, yielding the risk simulation results. The risk simulation results demonstrate the possible trends and ranges of risk changes over a future period. By interpreting and analyzing the risk simulation results, such as by setting quantile or extreme value thresholds, different levels of risk thresholds can be identified. Risk thresholds are used to define safe zones, areas of concern, and risk zones, providing a basis for risk warning.

[0080] Understandably, core risk indicators form the foundation for dynamic financial modeling. Industry benchmark data streams provide the necessary context and comparison benchmarks for the model. In practical implementation, the parameter selection of the exponential smoothing algorithm affects the denoising effect; the smoothing coefficient needs to be adjusted according to the data volatility characteristics. The order of the multinomial regression model is crucial; too low an order leads to underfitting, while too high an order leads to overfitting. The calculation of the mean squared error index needs to ensure that the actual financial trajectory and the standard financial trajectory are strictly aligned in time. The number and depth of decision trees in the random forest model are key hyperparameters affecting model performance. The population size and mutation intensity of the evolutionary strategy algorithm need to be carefully set. The network structure design of the convolutional neural network needs to match the feature dimensions of the offset scoring. The collaborative mechanism between the hybrid frog leaping algorithm and the artificial bee colony algorithm needs to be clearly defined, such as whether it is executed sequentially or in parallel. The social hierarchy parameter in the gray wolf optimization algorithm needs to be reasonably initialized. The search radius of the flame algorithm controls the fineness of the local search. The Bayesian averaging method needs to assign appropriate prior weights to each value to be fused. The lag order and seasonality of a seasonal autoregressive model need to be determined based on the characteristics of the financial data. Risk thresholds typically require consideration of industry experience and regulatory requirements.

[0081] In practice, the construction of dynamic financial models and the calculation of risk scores are iterative and optimization processes. Models may require periodic retraining and calibration with new data to maintain predictive accuracy. The quality of cleaned financial data directly impacts the reliability of standard financial trajectories. The calculation method for offset scores can be adjusted based on specific risk concerns, such as focusing on deviations in liquidity risk, debt repayment risk, or profitability. When assessing the risk exposure of financial entities, random forest models can also output feature importance rankings, which help understand which core risk indicators contribute most to the assessment results. The optimization objective of evolutionary strategy algorithms can be to minimize prediction error or maximize correlation with real risk events. Convolutional neural networks, through deep feature extraction, can automatically learn complex patterns that are difficult for the human eye to recognize. The performance of the optimized risk model can be evaluated through cross-validation. The combination of the gray wolf optimization algorithm and the flame algorithm forms a coarse-to-fine optimization strategy. Bayesian averaging effectively reduces the risk of overfitting that may arise from a single model. Seasonal autoregressive models simulate the dynamic risk of financial entities within a predetermined time window, generating multiple simulation scenarios to assess the distribution of risk rather than just point estimates. Defining risk thresholds is a step that requires business knowledge, and the determination of risk thresholds needs to balance the sensitivity of early warnings with the false alarm rate.

[0082] In one specific embodiment, industry benchmark data streams can be obtained in real time from authorized financial data service providers via an application programming interface (API). Optionally, the multinomial regression model can try different orders and select the optimal order through cross-validation. In specific implementations, the initial risk assessment value generated by the random forest model can be stored and recorded as an important intermediate result. The number of iterations of the evolutionary strategy algorithm can be set according to computational resources and accuracy requirements. Training convolutional neural networks requires a large amount of labeled data, i.e., historical data and its corresponding real risk results. The collaborative optimization process of the hybrid frog leaping algorithm and the artificial bee colony algorithm may require a long computation time. The optimal solution located by the gray wolf optimization algorithm can serve as an important reference for adjusting the model evaluation value. The number of sparks generated by the flame algorithm affects the breadth of the local search. The setting of prior weights in the Bayesian averaging method can be based on the performance of each model on the validation set. Parameter estimation of the seasonal autoregressive model can be completed using the maximum likelihood estimation method. Visualization of risk simulation results helps to intuitively understand risk dynamics.

[0083] It is understandable that a standard financial trajectory represents the expected development path under ideal market conditions. It is also understandable that the deviation between the actual financial trajectory and the standard financial trajectory is a direct signal for risk identification. In practice, the random forest model's assessment of a financial entity's risk exposure is a comprehensive judgment process. Evolutionary strategy algorithms optimize the initial risk assessment values, helping to improve the accuracy and stability of risk scores. Convolutional neural networks perform deep feature extraction on the deviation scores, uncovering risk features that shallow models cannot capture. The hybrid frog-leaping algorithm and artificial bee colony algorithm collaboratively optimize the random forest model to improve its generalization ability. The gray wolf optimization algorithm and flame algorithm refine the model assessment values, a key step in refining risk scores. The Bayesian averaging method fuses assessment values ​​from different sources, contributing to more reliable risk scores. Seasonal autoregressive models simulate risk dynamics, providing a basis for forward-looking risk management. Defining risk thresholds is a necessary operation to transform continuous risk scores into discrete warning levels.

[0084] In one specific embodiment, a multidimensional data matrix is ​​constructed based on the anomalous deviation sequence. The rows of the multidimensional data matrix typically correspond to different time points or observation samples, while the columns correspond to different financial variables or anomalous measurement dimensions. In practice, independent component analysis (ICA) is applied to perform dimensionality reduction on the multidimensional data matrix. ICA aims to decompose the observed multidimensional data matrix into a linear combination of statistically independent non-Gaussian components, thereby eliminating redundant information between variables and revealing potential independent source signals, ultimately obtaining a compressed data matrix with lower dimensionality and more independent features. A multivariate causal discovery algorithm is then employed, traversing the compressed data matrix to explore causal dependencies between different variables. Based on methods such as conditional independence tests, the algorithm automatically discovers frequently occurring causal patterns in the data, reflecting potential driving relationships between variables.

[0085] Structural equation modeling (SEM) is used to test the causal relationships of frequent causal patterns discovered by multivariate causal discovery algorithms. Combining factor analysis and path analysis, SEM can examine whether the hypothesized causal relationships between variables fit the data, thereby verifying the statistical significance of frequent causal patterns and determining the true causal relationships between variables in financial transaction data. Furthermore, combinations of frequently co-occurring variables are further screened from the frequent causal patterns; these combinations are called co-occurrence rules. Co-occurrence rules reflect the association between variables rather than a necessary causal direction. The causal relationships validated by SEM and the co-occurrence rules screened from frequent causal patterns are combined to form a key rule set containing various types of associations. This key rule set forms the basis for subsequent evaluation and screening.

[0086] In practice, information gain, Pearson correlation coefficient, and Manhattan distance are used as multivariate evaluation metrics to comprehensively evaluate each causal relationship and co-occurrence rule in the key rule set. Information gain, based on information theory, measures the degree to which the uncertainty of one variable is reduced when the value of another is known. Calculating the information gain of the variables involved in the causal relationships and co-occurrence rules in the key rule set yields a correlation score reflecting the strength of the association between variables. The Pearson correlation coefficient is used for linear correlation analysis of the causal relationships and co-occurrence rules in the key rule set. It calculates the degree and direction of the linear correlation between two variables. Using the Pearson correlation coefficient to calculate the correlation coefficient of variable pairs in the key rule set provides a correlation score reflecting the strength of the linear relationship between variables. Manhattan distance is used to evaluate the dispersion of causal relationships and co-occurrence rules. It calculates the sum of the absolute differences between two variables in a multidimensional space. Using the Manhattan distance to calculate the discrete values ​​of variables in the causal relationships and co-occurrence rules yields a dispersion score reflecting the degree of difference in variable distribution. The correlation score, relevance score, and dispersion score all illustrate the importance of the rules from different perspectives.

[0087] A weighted scoring process is applied to each causal relationship and co-occurrence rule in the key rule set, combining correlation score, relevance score, and dispersion score. This weighted scoring process assigns appropriate weights to each scoring dimension. Through linear weighting or more complex nonlinear functions, the three scores are merged into a single aggregate score that comprehensively reflects the importance of the rule. This aggregate score is the primary basis for rule selection. Based on the calculated aggregate score, patterns with higher aggregate scores are selected from frequent causal patterns as candidate core risk indicators. These candidate core risk indicators represent potential important risk drivers. In practice, a Support Vector Machine (SVM) model is used to assign weights to the candidate core risk indicators. SVM is a supervised learning model that classifies or regresses data by finding a maximum margin hyperplane. In the weight assignment process, the SVM model learns the relationship between candidate core risk indicators and historical risk labels, assigning a weight to each candidate core risk indicator. The weight value represents the importance of the indicator in distinguishing risk states. Finally, based on the weights assigned by the support vector machine model, one or more indicators with the highest weights are extracted from the candidate core risk indicators as the final core risk indicators, which will be used for subsequent risk exposure assessment.

[0088] In some embodiments, the construction of a multidimensional data matrix needs to ensure data consistency and integrity, as outlier sequences may originate from different data sources and have different scales. Independent Component Analysis (ICA) algorithms assume that the source signals are statistically independent and non-Gaussian; these preconditions need to be validated before application. Multivariate causal discovery algorithms, such as PC or FCI, can handle situations with potentially confounding variables, but their computational complexity is high. The construction of structural equation modeling (SEM) requires setting an initial path graph based on certain theories or prior knowledge. Goodness-of-fit indices for SEM, such as CFI and RMSEA, are used to assess the model's fit to the data. Information gain is particularly effective for categorical variables; when used for continuous variables, discretization may be necessary. Pearson correlation coefficients can only capture linear relationships and are insensitive to nonlinear relationships. Manhattan distance is less sensitive to outliers than Euclidean distance, but it is computationally efficient in datasets with multiple features. Weight allocation in weighted scoring can be determined based on domain knowledge, or it can be determined using data-driven methods such as principal component analysis. Support vector machine (SVM) models are sensitive to the choice of kernel function. Commonly used kernel functions for SVM models include linear kernel, polynomial kernel, and radial basis kernel.

[0089] In some embodiments, the discovery of causal patterns may be affected by the amount and quality of data; sufficient historical data is a prerequisite for obtaining reliable causal patterns. Co-occurrence rules can be efficiently mined using association rule mining algorithms. The dimensions of association score, relevance score, and dispersion score may differ; standardization is required before comprehensive scoring to avoid any single dimension dominating the final result. A threshold for aggregated scores can be set for the election of candidate core risk indicators; patterns exceeding the threshold are selected as candidates. When performing weight allocation in the support vector machine model, sufficient historical samples are required for training; these samples should include both scenarios of risk events occurring and not occurring. The extracted core risk indicators should be business interpretable for risk management personnel to understand and use. The number of core risk indicators needs to balance comprehensiveness and simplicity; too many indicators increase model complexity and the risk of overfitting.

[0090] Optionally, Independent Component Analysis (ICA) can be combined with other dimensionality reduction methods such as Principal Component Analysis (PCA), using PCA for dimensionality reduction first and then ICA for blind source separation. Optionally, for nonlinear causal relationships, kernel-based causal discovery algorithms or methods based on regenerating kernel Hilbert spaces can be considered. In practice, Structural Equation Modeling (SEM) can handle both manifest and latent variables simultaneously, allowing multiple observed variables to be aggregated into a few latent variables. The calculation of information gain depends on the probability distribution estimation of the variables, which can be obtained through histograms or kernel density estimation methods. The significance test of the Pearson correlation coefficient helps determine whether the correlation is statistically significant. The calculation of Manhattan distance requires features to be on the same scale; otherwise, feature standardization is necessary. Weighted scoring can be performed using objective weighting methods such as entropy weighting to determine the weights of each scoring dimension. The parameters of the Support Vector Machine (SVM) model, such as the penalty coefficient C and kernel parameters, need to be tuned through cross-validation.

[0091] It is understandable that anomaly sequences are a direct result of risk identification and the starting point for causal analysis. It is also understandable that extracting core risk indicators is a crucial step in transforming data patterns into actionable knowledge. In practice, the quality of the multidimensional data matrix directly affects the effectiveness of subsequent causal discovery and indicator extraction. Independent component analysis can help reveal the independent driving factors behind the data. Multivariate causal discovery algorithms can automatically learn the causal structure between variables from the data, reducing reliance on prior knowledge. Structural equation modeling provides a rigorous statistical testing framework for causal hypotheses. Information gain can capture any type of relationship between variables, including nonlinear and non-monotonic relationships. The Pearson correlation coefficient is simple, intuitive, and easy to calculate and interpret. The robustness of Manhattan distance to outliers makes it superior to Euclidean distance in certain scenarios. Weighted scoring integrates information from multiple perspectives, contributing to a more comprehensive rule evaluation. Support vector machine (SVM) models seek models with strong generalization ability through the maximum margin principle, and SVM models help improve the stability of weight allocation.

[0092] In practice, the entire extraction process requires computational resources, especially when the data volume is massive or the variable dimensionality is high. The algorithm's parameter settings have a significant impact on the results and need to be carefully selected through experimentation or experience. The extracted core risk indicators need to be verified and interpreted in conjunction with actual business operations. The method described in this embodiment systematically transforms the original abnormal sequences into refined risk indicators, providing high-quality input features for building dynamic financial models. The effectiveness of the core risk indicators directly affects the accuracy of subsequent risk warnings; therefore, the implementation details of this embodiment are crucial. The mining of causal paths and the analysis of co-occurrence patterns help to deeply understand the mechanisms of risk generation, rather than merely identifying symptoms. The comprehensive evaluation of multiple indicators reduces the risk of the one-sidedness of a single indicator. The introduction of the support vector machine model makes the indicator selection process more objective and data-driven.

[0093] See Figure 3 This diagram presents the core findings of the causal path mining and co-occurrence pattern analysis stages in blockchain-based financial risk early warning. It is a key visual output of the multivariate causal discovery algorithm's process for extracting core risk indicators. The horizontal and vertical axes of the diagram encompass variables such as anomaly deviation A, anomaly deviation B, risk indicator 1, risk indicator 2, and risk indicator 3, with cell values ​​representing the strength of the causal relationship between variables. This diagram intuitively reveals the causal relationship structure between anomalies and risk indicators, providing a quantitative basis for subsequent extraction of core risk indicators. It helps identify the core driving paths of risk from complex financial data and is a key analytical tool for improving the interpretability and accuracy of risk early warning.

[0094] In one specific embodiment, the long-term trend component and the short-term noise component are derived from the results of a smart contract collecting real-time financial transaction data from the blockchain distributed ledger and performing a consistency comparison with a historical financial database in the cloud. Based on the long-term trend component and the short-term noise component, a time series decomposition algorithm is used to perform pattern separation processing on the real-time financial transaction data. The time series decomposition algorithm decomposes the non-stationary time series into several components. The time series decomposition model can be expressed as:

[0095]

[0096] in: Represents a point in time The observed raw real-time financial transaction data values, Represents a point in time The long-term trend components, Represents a point in time Seasonal ingredients, Represents a point in time The residual components.

[0097] In practice, the core task of time series decomposition algorithms is to estimate the long-term trend component. Seasonal ingredients and residual components Through this decomposition, the time series decomposition algorithm can identify relatively smooth and persistent normal fluctuation patterns in real-time financial transaction data (as well as sudden, abnormal fluctuation patterns that are difficult to explain by trends and seasonality). The result of pattern separation processing is a detailed decomposition containing long-term trend components, seasonal components, and residual components. The decomposition result provides a structured understanding of real-time financial transaction data. Anomalies can be detected by matching the decomposition result with a preset normal fluctuation template. The preset normal fluctuation template is a statistical description learned from a large amount of historical financial transaction data during normal operating periods. The preset normal fluctuation template defines the long-term trend component under healthy conditions. Reasonable range of variation, seasonal components Typical form and amplitude and residual components The distribution characteristics. The matching process typically involves statistical hypothesis testing or control chart methods. Similarly, the long-term trend component... If it suddenly deviates from its smooth evolutionary path, or if seasonal components... Significant distortions in the shape of data may also be considered abnormal. Detecting and labeling these financial transaction data points or continuous data point sequences that do not conform to the normal fluctuation template constitutes the final abnormal deviation sequence, which is the direct input for subsequent causal analysis and risk indicator extraction.

[0098] In some embodiments, the choice of time series decomposition algorithm depends on the characteristics of the financial transaction data. For data with obvious and fixed-period seasonal fluctuations, classic seasonal decomposition algorithms can be used. For data with non-fixed seasonal patterns or no strong seasonality, moving average-based methods or methods incorporating seasonal components can be used. Consider it as zero. Long-term trend component. Extraction can be achieved using Hodrick-Prescott filtering or polynomial fitting. Residual components The calculation is as follows: The construction of a normal fluctuation template requires a sufficiently long and clean historical data period to ensure that the template can reliably represent normal behavior. The choice of threshold needs to strike a balance between sensitivity and specificity; a threshold that is too low will lead to too many false alarms, while a threshold that is too high will miss real anomalies.

[0099] In some embodiments, anomalous deviation sequences may manifest in various forms, including outlier anomalies, contextual anomalies, and collective anomalies. Detection methods need to be tailored to different types of anomalies. For residual components... For anomaly detection, besides using Z-score, boxplot methods or machine learning-based methods can also be used. For long-term trend components... Anomalies may necessitate detecting trend reversals or acceleration / deceleration changes. Seasonal components. Anomalies may manifest as abnormal increases or decreases in seasonal amplitude, or shifts in the location of seasonal peaks and troughs. Anomalies in different components of the decomposition results may indicate different types of risks; for example, anomalies in residual components... Sharp fluctuations may indicate operational risks or the impact of sudden market events, while long-term trend components... The deterioration of these factors may foreshadow a decline in fundamentals.

[0100] In some embodiments, time series decomposition algorithms are sensitive to parameter settings. For example, in STL decomposition, the window size for LOESS smoothing needs to be carefully chosen; a window size that is too small can lead to the loss of trend components. Excessive volatility and excessively large window sizes can smooth out important trend changes. Seasonal components. The period length needs to be specified precisely. The robustness of the decomposition results may be affected by outliers in the data; extreme outliers may distort the trend components. and seasonal ingredients The estimation can be robust, decomposition methods that are insensitive to outliers. Real-time financial transaction data may be high-frequency, while cloud-based historical financial databases may contain low-frequency data; therefore, data frequency alignment may be necessary before consistency comparison and decomposition. Annotated information on outlier sequences can serve as valuable supervisory information for training subsequent anomaly detection models or evaluating model performance.

[0101] Optionally, for cases with multiple related financial variables, multivariate time series decomposition can be performed. Multivariate time series decomposition considers the correlation between multiple variables simultaneously, potentially detecting collaborative anomaly patterns that are not apparent from a univariate perspective. In practice, the preset normal fluctuation template can be dynamically updated using a sliding time window, updating the template parameters only with historical data from the most recent period. This allows the template to adapt to the slow evolution of business models. The matching process not only identifies outliers but also calculates an anomaly severity score for each outlier. This score, based on the degree to which it deviates from the normal range, helps prioritize anomalies in subsequent analysis.

[0102] It is understandable that time series decomposition is the foundation for breaking down complex data signals into interpretable components. It is also understandable that the accurate detection of abnormal deviation sequences is a crucial link in the entire risk warning chain. In practical implementation, long-term trend components... This reflects the long-term direction of financial fundamentals. Seasonal component. It captured the periodic fluctuation pattern. Residual component. This includes random fluctuations and potential anomaly signals that cannot be explained by trends and seasonality. The accuracy of the normal fluctuation template directly determines the reliability of anomaly detection. The quality of the anomaly deviation sequence directly affects the effectiveness of subsequent causal path mining and core risk indicator extraction. The method described in this embodiment systematically transforms raw financial transaction data into anomaly signal sequences with clear risk indications. The robustness of the decomposition algorithm is crucial for handling noise and anomalies in real-world data. Multi-component joint analysis helps to more precisely locate the source of anomalies. The dynamic template update mechanism enables the anomaly detection system to have a certain degree of adaptability. In specific implementation, the entire anomaly detection process needs to handle the problem of missing data. Financial transaction data may have missing values ​​for various reasons, requiring appropriate imputation or processing before decomposition. The computational efficiency of the algorithm is important for real-time or near-real-time early warning applications, requiring the selection of decomposition and detection algorithms with moderate computational complexity. The detected anomaly deviation sequence needs to be verified by business operations to confirm that it truly represents a potential risk event rather than a data quality problem or a special but normal business activity.

[0103] See Figure 4 This involves a complete process of automatically collecting real-time financial transaction data from a blockchain distributed ledger using smart contracts, and then comparing this data with a historical financial database in the cloud to separate long-term trend components from short-term noise components. Dedicated smart contracts for data collection are deployed on each participating node of the blockchain network. These smart contracts are pre-written and automatically executed code deployed on the blockchain, periodically triggering data collection tasks. By accessing the blockchain's application programming interface (API), the smart contracts scan and capture newly added data records related to financial transactions in the blockchain distributed ledger; these records constitute the real-time financial transaction data. Simultaneously with the capture of real-time financial transaction data, the smart contracts' built-in data verification logic is activated. This logic includes verifying the digital signatures of transactions to confirm the legitimacy of the transaction initiator, and verifying the compliance of the transaction content according to preset business rules, thereby ensuring the authenticity of the collected real-time financial transaction data and the rationality of the business logic.

[0104] In practice, real-time financial transaction data, initially verified by smart contracts, is synchronously transmitted to a historical financial database in the cloud. This cloud-based historical financial database is a centralized data warehouse for long-term storage of multi-dimensional historical financial data. On the cloud-based historical financial database side, a time-series alignment operation needs to be performed. This operation aims to precisely map newly synchronized real-time financial transaction data with different timestamps to the historical financial data stored in the cloud-based historical financial database along the timeline. The time-series alignment operation may involve resampling irregular time-series data and aligning it to a common time point index, laying the foundation for subsequent consistency comparisons and data analysis. A moving average algorithm is then applied to smooth the aligned time-series data. The moving average algorithm smooths out short-term random fluctuations in the data by calculating the average value of the data within a specified time window. The window size of the moving average algorithm is a key parameter, determining the degree of smoothing. A window that is too small will result in poor smoothing, while a window that is too large may mask the true trend changes. By using the moving average algorithm, the long-term trend component reflecting the long-term development direction can be separated from the raw, noisy real-time financial transaction data. The difference between the raw data and the long-term trend component mainly contains short-term noise components and possible seasonal fluctuations.

[0105] Referring to Table 1, cryptographic hash functions are used to verify the integrity of the long-term trend component and short-term noise component after the data separation process. Before or after synchronizing real-time financial transaction data to the cloud-based historical financial database, the original hash value of the real-time financial transaction data can be calculated. After successfully separating the long-term trend component and the short-term noise component, the hash values ​​of the long-term trend component and the short-term noise component are calculated separately. By comparing the hash values ​​of the separated components with the pre-calculated or stored hash values ​​of the original data, it is possible to rigorously verify whether any tampering or damage has occurred to the long-term trend component and the short-term noise component during transmission and processing, ensuring the integrity and reliability of the data from acquisition to separation.

[0106] Table 1: Key Aspects of Data Acquisition and Processing

[0107]

[0108] In some embodiments, the triggering frequency of smart contracts can be configured according to business needs. The triggering frequency can be based on block height, fixed time intervals, or driven by on-chain events. The data verification logic built into smart contracts may need to be updated and maintained according to specific financial rules and regulations. Cloud-based historical financial databases need to be highly available and scalable to handle the continuous inflow of real-time data. Time series alignment operations may need to handle complex situations such as time zone conversions and holidays. Moving average algorithms have various variations, including simple moving averages, weighted moving averages, or exponential moving averages. The choice of moving average algorithm depends on the characteristics of the data and the analysis objectives. The choice of cryptographic hash function needs to consider its collision resistance and computational efficiency; SHA-256 is a commonly used choice. Integrity verification can be performed at multiple stages; it is not limited to after the separation stage but can also be initially verified immediately after data synchronization.

[0109] Optionally, in addition to collecting data, smart contracts can also record relevant metadata about data collection events on the blockchain, such as collection time and block height, enhancing audit trail capabilities. Optionally, for boundary effects generated by moving average algorithms, methods such as symmetric padding, forward padding, or backward padding can be used to handle boundary data. In practical implementation, the separation effect between long-term trend components and short-term noise components directly affects the accuracy of subsequent time series decomposition and anomaly detection. An overly smooth long-term trend component may mask some important intermediate-term changes, while a short-term noise component containing too much residual trend can interfere with anomaly detection. Therefore, the selection of the moving average algorithm window size often requires repeated testing and adjustment based on historical data. The separated long-term trend components and short-term noise components need to be stored in a structured format and marked with their correspondence with the original data for subsequent use. The results of integrity verification can also be recorded on the blockchain as proof of the credibility of the data processing process.

[0110] It is understandable that the automatic execution characteristic of smart contracts is fundamental to ensuring the timeliness and consistency of data collection. It is also understandable that cloud-based historical financial databases provide the necessary context and historical benchmarks for consistency comparison and trend analysis. In practical implementation, time series alignment is a prerequisite for ensuring data comparability; without precise time alignment, any comparison across time periods will be meaningless. The moving average algorithm is a classic and effective data smoothing technique, capable of resisting the interference of outliers to a certain extent. Cryptographic hash functions provide strong protection for data integrity; even minor modifications to the data will lead to significant changes in the generated hash value. The data collection, verification, synchronization, alignment, decomposition, and integrity verification steps described in this embodiment together constitute a complete front-end data processing pipeline. This pipeline ensures that the data source relied upon by the subsequent risk analysis model is authentic, complete, pre-cleaned, and consistent high-quality data. The collaborative work of smart contracts and cloud databases demonstrates the effective combination of blockchain technology and traditional centralized systems in specific application scenarios. The separation of long-term trend components and short-term noise components is the first step in decomposing overall changes in time series analysis, creating conditions for more refined component decomposition.

[0111] See Figure 5 This chart presents the core visualization results of the data collection and trend / noise separation stages in a blockchain-based financial risk early warning method. It corresponds to the process of smart contract-based data collection, comparison with a cloud-based historical database, and separation of long-term trends and short-term noise, providing a data foundation for time series decomposition. The chart uses time period as the horizontal axis and data values ​​as the vertical axis, presenting key information from three dimensions. This chart clearly demonstrates the value of blockchain technology in the data collection stage. By ensuring the authenticity of the data source through smart contracts and then separating trends and noise through algorithms, it lays a reliable data foundation for subsequent anomaly detection, causal analysis, and risk assessment. It is a crucial visual support for achieving pre-risk verification and improving the accuracy of early warnings.

[0112] In a specific embodiment, the abnormal deviation sequence is a sequence of financial transaction data that deviates from the normal pattern, detected by a time series decomposition algorithm. The core risk indicator is a key risk driver extracted from the abnormal deviation sequence using a multivariate causal discovery algorithm. The risk threshold is a risk level boundary defined by a dynamic financial model and a random forest model. The abnormal deviation sequence, core risk indicator, and risk threshold are integrated to form an input feature vector. The input feature vector is a multidimensional numerical array, with each dimension representing a feature. For example, the input feature vector may include statistics of the abnormal deviation sequence at the most recent N time points, the current value and rate of change of the core risk indicator, and the relative distance between the financial entity's actual risk score and each risk threshold. The construction of the input feature vector aims to provide the model with a comprehensive, quantitative snapshot of the current risk situation.

[0113] A deep neural network model is trained using pre-constructed input feature vectors and their corresponding historical labels. A deep neural network is an artificial neural network containing multiple hidden layers. Through multiple layers of nonlinear transformations, it learns the complex mapping relationship between input features and the output target. The goal of training a deep neural network model is to enable it to capture risk patterns embedded in the input feature vectors that are closely related to future financial risks. The training process typically involves forward propagation to calculate the loss, backpropagation to calculate the gradient, and updating the network weights using optimization algorithms.

[0114] A trained deep neural network model is used to predict the probability of financial risks in future accounting periods. The input feature vector at the current time point is fed into the trained deep neural network model, which then performs forward propagation calculations, ultimately producing a probability distribution at the output layer. The specific form of the probability distribution can vary; for example, it can be a simple scalar representing the probability of a financial risk event occurring within a specific future accounting period; or it can be a multi-dimensional vector, where each element represents the probability of a financial entity falling into different risk levels. The probability distribution output by the model quantifies the uncertainty of future risks.

[0115] Specific risk warning messages are generated based on the probability distribution output by the deep neural network model. The generation of risk warning messages must adhere to pre-defined warning trigger rules. For example, the rules could be defined as follows: if the predicted probability of a financial risk event exceeds 70%, a high-risk warning message is generated; if the probability is between 30% and 70%, a medium-risk warning message is generated; and if the probability is below 30%, a low-risk warning message or no warning message is generated. The content of risk warning messages typically needs to include key information, such as the risk entity identifier, the predicted risk level, the predicted time window for the risk occurrence, the main risk drivers, and a summary of recommended measures. The generated risk warning messages need to be securely stored and distributed.

[0116] Distributed storage of risk warning messages is achieved through a blockchain network. The generated risk warning message is encapsulated as a special transaction, which is broadcast to all nodes in the blockchain network. Nodes in the network verify this transaction containing the risk warning message through a consensus mechanism and reach an agreement. Once verified, the transaction is added to a new block and recorded on the immutable distributed ledger. This storage method ensures the immutability, traceability, and transparency of risk warning messages to all authorized participants.

[0117] In some embodiments, the dimensionality of the input feature vector may be high, necessitating feature scaling to improve the training stability and convergence speed of the deep neural network model. The structure of the deep neural network model requires careful design; the number of layers, the number of neurons per layer, the choice of activation function, and whether to use regularization techniques such as Dropout to prevent overfitting are all key hyperparameters. Training a deep neural network model requires a large amount of labeled historical data. If the number of positive samples is scarce, oversampling techniques or adjustments to the loss function may be necessary to address class imbalance. The interpretation of probability distributions must be cautious; a probability value itself does not equate to certainty or non-occurrence, but rather represents a possibility. Decision-makers need to combine this with other information for a comprehensive judgment.

[0118] In some embodiments, the setting of warning trigger rules is subjective. The setting of these rules needs to balance the sensitivity of the warning and the false alarm rate. Overly sensitive rules will generate a large number of false alarms, potentially leading to warning fatigue, while overly lenient rules may miss real risks. The format of risk warning messages can be standardized to facilitate automatic parsing and processing. For example, JSON or XML formats can be used to structure the various fields of the risk warning message. The choice of blockchain network can be a public chain, a consortium chain, or a private chain, depending on the requirements for data privacy, performance, and cost. In consortium chain or private chain scenarios, access permissions for risk warning messages can be finely controlled. Risk warning messages recorded on the blockchain can serve as part of an audit trail, providing irrefutable evidence for post-event tracing and liability determination.

[0119] Optionally, deep neural network models can employ more complex architectures, such as Long Short-Term Memory (LSTM) networks or gated recurrent units (GRUs), which are better able to handle time-series-related dependencies in the input feature vectors. Optionally, in addition to point prediction, interval estimations of future risk probabilities can be attempted, for example, by outputting confidence intervals of predicted values ​​through quantile regression or Bayesian neural networks, providing more information for decision-making. Optionally, after risk warning messages are generated, in addition to being stored on the blockchain, they can also be automatically sent to relevant risk management personnel through an integrated push notification system for real-time alerts. In practical implementation, the integration process needs to ensure that heterogeneous data such as abnormal deviation sequences, core risk indicators, and risk thresholds are time-aligned and that numerical units are consistent. Training deep neural network models is a computationally intensive task and may require acceleration using graphics processing units (GPUs). The model needs to be retrained or incrementally learned periodically using the latest data to maintain its predictive capabilities to adapt to potentially changing business environments. Distributed storage of risk warning messages on a blockchain network may incur storage overhead and latency, requiring a trade-off between benefits and costs during the design phase.

[0120] It's understandable that integrating abnormal deviation sequences, core risk indicators, and risk thresholds is to comprehensively utilize the analytical results from previous stages. It's also understandable that the advantage of deep neural network models lies in their ability to automatically learn high-level risk patterns from complex features. In practical implementation, the quality of the input feature vector directly determines the upper limit of the deep neural network model's predictive performance. The trained deep neural network model acts as a risk pattern recognition engine. The output of the probability distribution propels risk warnings from qualitative judgment to quantitative assessment. Risk warning messages are the ultimate carrier of the entire system's value realization. The distributed storage of risk warning messages on the blockchain network endows the warning results with extremely high credibility and auditability.

[0121] Example 2

[0122] This embodiment 2 provides a non-transitory computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, they implement the blockchain-based financial risk early warning method described above. The method includes:

[0123] The system automatically collects real-time financial transaction data from the blockchain distributed ledger through smart contracts, and compares the real-time financial transaction data with the historical financial database in the cloud to separate the long-term trend components and short-term noise components in the real-time financial transaction data.

[0124] The time series decomposition algorithm is applied to process the long-term trend component and the short-term noise component in order to detect abnormal deviation sequences in the real-time financial transaction data;

[0125] A multivariate causal discovery algorithm is used to perform causal path mining and co-occurrence pattern analysis on the abnormal deviation sequence to extract core risk indicators;

[0126] The core risk indicators are input into the dynamic financial model, combined with industry benchmark data streams, to construct a standard financial trajectory. By quantifying the deviation between the standard financial trajectory and the actual financial trajectory, a random forest model is integrated to assess the risk exposure of the financial entity and classify risk thresholds.

[0127] Based on the abnormal deviation sequence, the core risk indicators, and the risk threshold, a deep neural network model is used to predict the probability of financial risks in future accounting periods and generate risk warning messages.

[0128] Example 3

[0129] This embodiment 3 provides a computer device, including a memory and a processor, wherein the processor and the memory communicate with each other, and the memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute the blockchain-based financial risk early warning method described above, the method including:

[0130] The system automatically collects real-time financial transaction data from the blockchain distributed ledger through smart contracts, and compares the real-time financial transaction data with the historical financial database in the cloud to separate the long-term trend components and short-term noise components in the real-time financial transaction data.

[0131] The time series decomposition algorithm is applied to process the long-term trend component and the short-term noise component in order to detect abnormal deviation sequences in the real-time financial transaction data;

[0132] A multivariate causal discovery algorithm is used to perform causal path mining and co-occurrence pattern analysis on the abnormal deviation sequence to extract core risk indicators;

[0133] The core risk indicators are input into the dynamic financial model, combined with industry benchmark data streams, to construct a standard financial trajectory. By quantifying the deviation between the standard financial trajectory and the actual financial trajectory, a random forest model is integrated to assess the risk exposure of the financial entity and classify risk thresholds.

[0134] Based on the abnormal deviation sequence, the core risk indicators, and the risk threshold, a deep neural network model is used to predict the probability of financial risks in future accounting periods and generate risk warning messages.

[0135] Example 4

[0136] This embodiment 4 provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device performs the blockchain-based financial risk early warning method described above, including:

[0137] The system automatically collects real-time financial transaction data from the blockchain distributed ledger through smart contracts, and compares the real-time financial transaction data with the historical financial database in the cloud to separate the long-term trend components and short-term noise components in the real-time financial transaction data.

[0138] The time series decomposition algorithm is applied to process the long-term trend component and the short-term noise component in order to detect abnormal deviation sequences in the real-time financial transaction data;

[0139] A multivariate causal discovery algorithm is used to perform causal path mining and co-occurrence pattern analysis on the abnormal deviation sequence to extract core risk indicators;

[0140] The core risk indicators are input into the dynamic financial model, combined with industry benchmark data streams, to construct a standard financial trajectory. By quantifying the deviation between the standard financial trajectory and the actual financial trajectory, a random forest model is integrated to assess the risk exposure of the financial entity and classify risk thresholds.

[0141] Based on the abnormal deviation sequence, the core risk indicators, and the risk threshold, a deep neural network model is used to predict the probability of financial risks in future accounting periods and generate risk warning messages.

[0142] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A blockchain-based method for early warning of financial risks, characterized in that, include: The system automatically collects real-time financial transaction data from the blockchain distributed ledger through smart contracts, and compares the real-time financial transaction data with the historical financial database in the cloud to separate the long-term trend components and short-term noise components in the real-time financial transaction data. The time series decomposition algorithm is applied to process the long-term trend component and the short-term noise component in order to detect abnormal deviation sequences in the real-time financial transaction data; A multivariate causal discovery algorithm is used to perform causal path mining and co-occurrence pattern analysis on the abnormal deviation sequence to extract core risk indicators; The core risk indicators are input into the dynamic financial model, combined with industry benchmark data streams, to construct a standard financial trajectory. By quantifying the deviation between the standard financial trajectory and the actual financial trajectory, a random forest model is integrated to assess the risk exposure of the financial entity and classify risk thresholds. Based on the abnormal deviation sequence, the core risk indicators, and the risk threshold, a deep neural network model is used to predict the probability of financial risks in future accounting periods and generate risk warning messages.

2. The blockchain-based financial risk early warning method according to claim 1, characterized in that, The process involves inputting the core risk indicators into a dynamic financial model, combining industry benchmark data streams to construct a standard financial trajectory, and quantifying the deviation between the standard financial trajectory and the actual financial trajectory. An integrated random forest model is then used to assess the risk exposure of the financial entity and define risk thresholds. This includes: using the core risk indicators as input, fusing industry benchmark data streams, applying an exponential smoothing algorithm to denoise multi-source historical financial data to obtain purified financial data, and using a multinomial regression model to curve-fit the purified financial data to generate a standard financial trajectory. The multi-source historical financial data includes: financial reports from companies in the same industry and the target financial entity's actual transaction flows over multiple accounting periods. The system uses water, historical default records, and credit rating logs. It compares the actual financial trajectory with the standard financial trajectory, calculates the offset between the actual and standard financial trajectories, and measures the offset using the mean squared error index to obtain an offset score. Based on the offset score, it uses a random forest model to assess the financial entity's risk exposure, generating an initial risk assessment value. An evolutionary strategy algorithm is then used to optimize the initial risk assessment value, resulting in a risk score for the financial entity. Using this risk score, combined with a seasonal autoregressive model, it simulates the financial entity's risk dynamics within a predetermined time window, obtaining risk simulation results. These results are then interpreted, and risk thresholds are defined.

3. The blockchain-based financial risk early warning method according to claim 2, characterized in that, The process of assessing the risk exposure of a financial entity based on the offset score using a random forest model to generate an initial risk assessment value, and then optimizing the initial risk assessment value using an evolutionary strategy algorithm to obtain a risk score for the financial entity, includes: using a convolutional neural network to perform deep feature extraction on the offset score to generate a basic risk assessment value; based on the basic risk assessment value, using a hybrid frog leaping algorithm and an artificial bee colony algorithm to collaboratively optimize the random forest model to obtain an optimized risk model; inputting the offset score into the optimized risk model, outputting a model assessment value, and applying a gray wolf optimization algorithm and a flame algorithm to refine the model assessment value to generate a refined assessment value; and using a Bayesian averaging method to fuse the basic risk assessment value, the optimized risk model, and the refined assessment value to generate a risk score for the financial entity.

4. The blockchain-based financial risk early warning method according to claim 3, characterized in that, The process of inputting the offset score into the optimized risk model, outputting a model evaluation value, and applying the Grey Wolf Optimization Algorithm and the Flame Algorithm to refine the model evaluation value to generate a refined evaluation value includes: calling the optimized risk model, using the offset score as an input feature, performing risk inference calculation, and outputting a preliminary risk score; using the Grey Wolf Optimization Algorithm to perform a population iterative search on the preliminary risk score to locate the optimal solution, adjusting the preliminary risk score based on the optimal solution to obtain the Grey Wolf optimization result; and based on the Grey Wolf optimization result, combining it with the Flame Algorithm to perform a local fine-grained search to further optimize the preliminary risk score and generate a refined evaluation value.

5. The blockchain-based financial risk early warning method according to claim 1, characterized in that, The method employs a multivariate causal discovery algorithm to perform causal path mining and co-occurrence pattern analysis on the abnormal deviation sequences, extracting core risk indicators, including: A data matrix is ​​constructed based on the abnormal deviation sequence. Independent component analysis is used to perform feature dimensionality reduction on the data matrix to obtain a compressed data matrix. A multivariate causal discovery algorithm is applied to traverse the compressed data matrix to discover frequent causal patterns. The frequent causal patterns are verified by structural equation model causality test, the causal relationships of the financial transaction data are determined, and co-occurrence rules are screened from the frequent causal patterns. The causal relationships and the co-occurrence rules are then combined to form a key rule set. Information gain, Pearson correlation coefficient, and Manhattan distance are used as evaluation metrics to assess the causal relationships and co-occurrence rules of the key rule set and generate aggregate scores. Based on the aggregated score, candidate core risk indicators are selected from the frequent causal patterns. The support vector machine model is used to perform weight allocation processing on the candidate core risk indicators to extract the core risk indicators.

6. The blockchain-based financial risk early warning method according to claim 5, characterized in that, The evaluation uses information gain, Pearson correlation coefficient, and Manhattan distance as assessment metrics to evaluate the causal relationships and co-occurrence rules of the key rule set, generating an aggregate score, including: Calculate the information gain of causal relationships and co-occurrence rules in the key rule set to obtain an association score; Using the Pearson correlation coefficient, a linear correlation analysis is performed on the causal relationships and co-occurrence rules in the key rule set, and the correlation coefficients of the causal relationships and co-occurrence rules in the key rule set are calculated to obtain a correlation score; The Manhattan distance is used to evaluate the dispersion of the causal relationship and the co-occurrence rule, and the discrete values ​​of the causal relationship and the co-occurrence rule are calculated to obtain the dispersion score; By combining the correlation score, the relevance score, and the dispersion score, a weighted score is applied to each causal relationship and co-occurrence rule in the key rule set to generate an aggregate score.

7. A blockchain-based financial risk early warning system, characterized in that, include: The acquisition module is used to automatically collect real-time financial transaction data from the blockchain distributed ledger through smart contracts, and compare the real-time financial transaction data with the cloud historical financial database to separate the long-term trend component and short-term noise component in the real-time financial transaction data. The extraction module is used to process the long-term trend components and short-term noise components using a time series decomposition algorithm to detect abnormal deviation sequences in the real-time financial transaction data; and to use a multivariate causal discovery algorithm to perform causal path mining and co-occurrence pattern analysis on the abnormal deviation sequences to extract core risk indicators. The segmentation module is used to input the core risk indicators into the dynamic financial model, combine industry benchmark data streams to construct a standard financial trajectory, and integrate a random forest model to assess the risk exposure of financial entities and segment risk thresholds by quantifying the deviation between the standard financial trajectory and the actual financial trajectory. The prediction module is used to predict the probability of financial risk in future accounting periods based on the abnormal deviation sequence, the core risk indicators, and the risk threshold using a deep neural network model, and generate risk warning messages.

8. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the blockchain-based financial risk early warning method as described in any one of claims 1-6.

9. A computer device, characterized in that, The system includes a memory and a processor, the processor and the memory communicating with each other, the memory storing program instructions executable by the processor, and the processor invoking the program instructions to execute the blockchain-based financial risk warning method as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions that implement the blockchain-based financial risk warning method as described in any one of claims 1-6.

Images