Operation risk early warning method, electronic equipment, storage medium and product

Abstract

The invention discloses an operation risk early warning method, electronic equipment, a storage medium and a product, and relates to the technical field of risk early warning, and the method comprises the steps: collecting historical operation data of a to-be-detected platform, determining each risk index, and calculating a risk dimension weight corresponding to each risk index; calculating the current operation risk probability of the to-be-detected platform based on each risk index and the risk dimension weight corresponding to each risk index; inputting the historical operation data, each risk index and the risk dimension weight into a preset risk early warning model to obtain a future operation risk probability of the to-be-detected platform; and outputting the current operation risk probability and the future operation risk probability, judging the risk level of the to-be-detected platform based on the current operation risk probability and the future operation risk probability, and outputting a corresponding early warning suggestion based on the risk level. The technical problem that the operation risk early warning effect is poor is solved.

Description

Technical Field

[0001] This application relates to the field of risk warning technology, and in particular to an operational risk warning method, electronic device, storage medium and product. Background Technology

[0002] With the development of internet finance information intermediary platforms, traditional financial business models have been reshaped, providing efficient financing channels for micro and small enterprises and individual investors. However, the continuous expansion of platform business has also significantly exacerbated the risk pressures faced in their daily operations. Current mainstream risk detection methods typically rely on static weight allocation and single-indicator threshold early warning mechanisms, which struggle to effectively capture the dynamic correlations between risk factors, resulting in delayed early warning signals. Therefore, current risk detection methods for internet finance information intermediary platforms suffer from poor operational risk early warning effectiveness.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main purpose of this application is to provide an operational risk early warning method, system, electronic device, storage medium, and computer program product, aiming to solve the technical problem of poor operational risk early warning effect.

[0005] To achieve the above objectives, this application proposes an operational risk early warning method, which includes:

[0006] Collect historical operational data of the platform to be tested, determine each risk index, and calculate the risk dimension weight corresponding to each risk index;

[0007] Based on the risk indices and the corresponding risk dimension weights of the risk indices, the probability of the current operational risk of the platform under test is calculated.

[0008] The historical operational data, the various risk indices, and the weights of the risk dimensions are input into a preset risk warning model to obtain the probability of future operational risks of the platform to be tested.

[0009] Output the current operational risk probability and the future operational risk probability, determine the risk level of the platform to be tested based on the current operational risk probability and the future operational risk probability, and output the corresponding early warning suggestion based on the risk level.

[0010] In one embodiment, the historical operational data includes first historical data, second historical data, third historical data, and fourth historical data. The various risk indices include platform-endogenous risk, borrower credit risk, industry-related risk, and liquidity risk. The step of collecting historical operational data of the platform to be tested and determining each risk index includes:

[0011] Extract the risk reserve amount, monthly transaction volume, paid-in registered capital, total outstanding principal, platform loan interest rate and industry average interest rate of the platform to be tested to obtain the first historical data, and obtain the platform's intrinsic risk based on the first historical data and the first weight corresponding to the first historical data.

[0012] The delinquency rate, standard deviation of loan amount, mean of loan amount, and number of historical defaults of each borrower in the platform to be tested are extracted to obtain the second historical data. The credit risk of the borrower is obtained based on the second historical data and the second weight corresponding to the second historical data.

[0013] Extract the number of problematic platforms in the current month, the total number of platforms in the industry, the month-on-month growth of the industry's monthly transaction volume, and the industry's transaction volume in the previous month, as well as the current month's transaction volume and historical average transaction volume of the platform to be tested, to obtain the third historical data. Based on the third historical data and the third weight corresponding to the third historical data, the industry-related risk is obtained.

[0014] The net increase in transaction volume, transaction volume, number of top lenders, and total number of lenders in the current month are extracted from the platform to be tested to obtain the fourth historical data. The liquidity risk is obtained based on the fourth historical data and the fourth weight corresponding to the fourth historical data.

[0015] In one embodiment, the step of calculating the risk dimension weights corresponding to each risk index includes:

[0016] Set a preset duration for the sliding window;

[0017] The information entropy value of each risk index is calculated within the time sliding window, and the information entropy value is obtained through the probability distribution of each risk index within the time sliding window.

[0018] The risk dimension weights of each risk index are generated based on the information entropy value, and the risk dimension weights are standardized so that the sum of the risk dimension weights of each risk index is a preset value.

[0019] In one embodiment, the preset risk warning model includes an LSTM module and a TCN module. The step of inputting the historical operating data, the various risk indices, and the risk dimension weights into the preset risk warning model to obtain the probability of future operational risks of the platform to be detected further includes:

[0020] The historical operational data and the short-term time-series fluctuation characteristics of each risk index are extracted through the stacked gate structure in the LSTM module.

[0021] The historical operational data and the various risk indices are input into the TCN module, and the long-range risk transmission features are extracted through the dilated causal convolutional network in the TCN module.

[0022] By combining the multi-head attention mechanism module, the long-range risk transmission features and the risk dimension weights are dynamically weighted and fused to obtain weighted fused features;

[0023] By combining the short-term time-series fluctuation features and the weighted fusion features, and inputting them into a preset fully connected classifier, the probability of future operational risks of the platform to be detected is generated.

[0024] In one embodiment, the stacked gate structure includes a forget gate, an input gate, and an output gate. The step of extracting the historical operational data and the short-term time-series fluctuation characteristics of each risk index through the stacked gate structure in the LSTM module includes:

[0025] Based on the historical operating data and the risk indices, a transmission term feature is generated, and the historical operating data, the risk indices, and the transmission term feature are input into the LSTM module;

[0026] The forget gate is used to attenuate historical noise information in the historical operation data in order to discard non-critical historical information in the historical operation data.

[0027] Based on the input gate, key signals in each risk index and the characteristics of the transmission term are filtered, and the key signals are enhanced.

[0028] The cell state of the LSTM module is updated based on the attenuated historical operating data and the enhanced key signals, and short-term temporal fluctuation characteristics are output through the output gate and the cell state.

[0029] In one embodiment, the dilated causal convolutional network includes dilated causal convolutional layers, wherein the dilation coefficient of each dilated causal convolutional layer increases progressively, and the step of extracting long-range risk transmission features through the dilated causal convolutional structure in the TCN module includes:

[0030] For the first dilated causal convolutional layer, the received historical operational data and various risk indices are used as the layer input features. A one-dimensional causal convolution operation is performed on the layer input features based on the convolution kernel with the minimum dilation coefficient to obtain the convolution result. The convolution result is added to the layer input features by residual to obtain the primary feature map and input to the next layer of dilated causal convolutional layer of the first layer.

[0031] For any dilated causal convolutional layer other than the first layer, the primary feature map output by the previous dilated causal convolutional layer is received. The transmission path of the hierarchical input feature at each time scale is extracted through the receptive field corresponding to the dilated causal convolutional layer. The residual of the transmission path is added to the hierarchical input feature. The time scale is determined by the corresponding receptive field.

[0032] In the case where the dilated causal convolutional layer is the last dilated causal convolutional layer, the feature vector obtained by adding the residual of the propagation path and the hierarchical input features is used as the long-range risk propagation feature output.

[0033] If the dilated causal convolutional layer is not the last dilated causal convolutional layer, the primary feature map obtained by adding the residual of the propagation path and the hierarchical input features is input into the next dilated causal convolutional layer.

[0034] In one embodiment, the multi-head attention mechanism module includes attention units, and the step of dynamically weighting and fusing the long-range risk transmission features and the risk dimension weights using the multi-head attention mechanism module to obtain weighted fused features includes:

[0035] The long-range risk transmission feature is injected into each attention unit. Based on each attention unit, the correlation weight between each feature element in the long-range risk transmission feature is calculated. Based on the risk dimension weight, the attention score of each feature element in the long-range risk transmission feature is adjusted to adjust the correlation weight between each feature element.

[0036] The long-range risk transmission features are dynamically weighted based on the adjusted association weights to generate weighted features for each attention unit. The weighted features of each attention unit are then aggregated to obtain a weighted fusion feature.

[0037] Furthermore, to achieve the above objectives, this application also proposes an operational risk early warning system, which includes:

[0038] The risk data acquisition module is used to collect historical operational data of the platform under test, determine each risk index, and calculate the risk dimension weight corresponding to each risk index.

[0039] The current risk calculation module is used to calculate the current operational risk probability of the platform under test based on each risk index and the risk dimension weights corresponding to each risk index.

[0040] The future risk calculation module is used to input the historical operating data, the various risk indices, and the risk dimension weights into a preset risk warning model to obtain the probability of future operating risks of the platform to be tested.

[0041] The risk level assessment module is used to output the current operational risk probability and the future operational risk probability, and to determine the risk level of the platform to be tested based on the current operational risk probability and the future operational risk probability, and to output corresponding early warning suggestions based on the risk level.

[0042] In addition, to achieve the above objectives, this application also proposes an electronic device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the operational risk warning method as described above.

[0043] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and which, when executed by a processor, implements the steps of the operational risk warning method described above.

[0044] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the operational risk warning method described above.

[0045] This application provides an operational risk early warning method, which includes: collecting historical operational data of the platform under test, determining various risk indices, and calculating the risk dimension weights corresponding to each risk index; calculating the current operational risk probability of the platform under test based on each risk index and its corresponding risk dimension weights; inputting the historical operational data, each risk index, and the risk dimension weights into a preset risk early warning model to obtain the future operational risk probability of the platform under test; outputting the current operational risk probability and the future operational risk probability, and determining the risk level of the platform under test based on the current operational risk probability and the future operational risk probability, and outputting corresponding early warning suggestions based on the risk level.

[0046] Compared to current mainstream risk detection methods that typically rely on static weight allocation and single-indicator threshold early warning mechanisms, which struggle to effectively capture the dynamic correlations between risk factors and result in delayed early warning signals, this application addresses this issue. By collecting historical operational data from the platform under test, it provides a rich information foundation for subsequent risk index calculation and model training, facilitating a more comprehensive assessment of platform risk. Calculating the risk dimension weights corresponding to each risk index adapts to dynamic changes in the risk indices, improving the sensitivity and accuracy of risk warnings. Calculating the current operational risk probability of the platform under test using each risk index and its corresponding risk dimension weights, and comprehensively considering the dynamic changes of multiple risk dimensions, makes the calculated current operational risk probability more accurate, helping to promptly identify potential risks. Calculating the future operational risk probability using a pre-set risk warning model significantly improves the accuracy and timeliness of future risk predictions. By determining the platform's risk level based on current and future operational risk probabilities, combined with pre-set threshold rules, and outputting corresponding early warning suggestions, this application achieves real-time monitoring and early warning of platform risks, providing the platform with ample time to take countermeasures. Attached Figure Description

[0047] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart illustrating an embodiment of the operational risk warning method of this application.

[0050] Figure 2 This is a flowchart illustrating Embodiment 2 of the operational risk warning method of this application;

[0051] Figure 3 This is an overall flowchart of the operational risk early warning method provided in Embodiment 2 of this application;

[0052] Figure 4 This is a schematic diagram of the module structure of the operational risk early warning system according to an embodiment of this application;

[0053] Figure 5 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the operational risk warning method in this application embodiment.

[0054] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0055] It should be understood that the first embodiment described herein is merely used to explain the technical solution of this application and is not intended to limit this application.

[0056] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0057] The main solution of the first embodiment of this application is as follows: collect historical operating data of the platform to be tested, determine each risk index, and calculate the risk dimension weight corresponding to each risk index; calculate the current operating risk probability of the platform to be tested based on each risk index and the risk dimension weight corresponding to each risk index; input the historical operating data, each risk index, and the risk dimension weight into a preset risk warning model, and output the future operating risk probability of the platform to be tested; output the current operating risk probability and the future operating risk probability, and determine the risk level of the platform to be tested based on the current operating risk probability and the future operating risk probability, and output the corresponding warning suggestion based on the risk level.

[0058] In the first embodiment, for ease of description, the operational risk early warning system will be used as the implementing entity in the following description.

[0059] Because current mainstream risk detection methods typically rely on static weight allocation and single indicator threshold early warning mechanisms, they are difficult to effectively capture the dynamic correlation between risk factors, resulting in delayed early warning signals.

[0060] This application provides a solution that, by collecting historical operational data of the platform under test, provides a rich information foundation for subsequent risk index calculation and model training, facilitating a more comprehensive assessment of platform risks. By calculating the risk dimension weights corresponding to each risk index, it can adapt to the dynamic changes of risk indices, improving the sensitivity and accuracy of risk warnings. By calculating the current operational risk probability of the platform under test through each risk index and its corresponding risk dimension weights, and comprehensively considering the dynamic changes of multiple risk dimensions, the calculated current operational risk probability is more accurate, helping to promptly identify potential risks of the platform. By calculating the future operational risk probability through a preset risk warning model, the accuracy and timeliness of future risk prediction are significantly improved. By judging the platform's risk level based on the current and future operational risk probabilities, combined with preset threshold rules, and outputting corresponding warning suggestions, real-time monitoring and warning of platform risks are achieved, providing the platform with sufficient time to take countermeasures.

[0061] It should be noted that the executing entity in the first embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, mobile phone, or other electronic device, or a system, application, or program capable of performing the above functions. The first embodiment and the following embodiments will be described using an operational risk warning system as an example.

[0062] All actions involving the acquisition of signals, information, or data in this application are carried out in accordance with the relevant data protection laws and policies of the country where the application is located, and with the authorization of the owner of the relevant device.

[0063] Based on this, the embodiments of this application provide an operational risk early warning method, referring to Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the operational risk warning method of this application.

[0064] In this embodiment, the operational risk early warning method includes steps S01 to S04:

[0065] Step S01: Collect historical operational data of the platform to be tested, determine each risk index, and calculate the risk dimension weight corresponding to each risk index.

[0066] It should be noted that the platform under testing refers to a platform requiring operational risk warning. This can be an internet finance information intermediary platform, a lending platform, or an online lending company. Historical operational data covers various historical operational indicators of the platform under testing, such as transaction volume, number of borrowers, delinquency rate, and industry environment data. The risk index is a comprehensive indicator constructed from four dimensions: platform inherent defects, borrower credit heterogeneity, industry environment correlation, and liquidity supply-demand imbalance. It includes platform inherent risk, borrower credit risk, industry-related risk, and liquidity risk. The risk dimension weights are dynamically calculated using the sliding window entropy weight method to reflect the importance of each risk dimension at different points in time.

[0067] Additionally, it should be noted that the system comprehensively collects historical operational data from the platform under test and calculates various risk indices for that platform based on this data. A sliding window entropy weighting method is used to dynamically calculate the weight of each risk index based on its volatility and importance within the historical operational data. For example, during periods of significant industry volatility, the weight of industry-related risks is increased to reflect their impact on the platform's operation.

[0068] Understandably, current mainstream risk detection methods typically rely on static weight allocation, which cannot adapt to the dynamic changes in risk factors during the operation of the platform under test (such as an internet financial information intermediary platform). Furthermore, single-indicator threshold early warning mechanisms are also difficult to fully reflect the multi-dimensional risks faced by the platform. Therefore, step S01 is performed by collecting multi-source historical operational data and dynamically calculating the risk dimension weights of each risk index. This can more accurately reflect the real-time impact of different risk factors on the platform's operational risks, solve the problem that static weight allocation methods cannot adapt to dynamic changes in risks, and improve the accuracy and timeliness of risk weights.

[0069] Step S02: Calculate the current operational risk probability of the platform under test based on each risk index and the corresponding risk dimension weights of each risk index.

[0070] It should be noted that the current operational risk probability is the probability of the platform's current operational risk occurring, calculated by weighted summation or other statistical methods, combining various risk indices and their corresponding dynamic weights. It can be used to assess the operational risk of the platform under test in the current situation.

[0071] Additionally, it should be noted that the current operational risk probability of the platform is calculated using a weighted summation method based on the risk indices and their corresponding risk dimension weights calculated in step S01. The specific formula is:

[0072]

[0073] Among them, P grade R represents the probability of current operational risk. j For the j-th risk index, there are a total of 4 dimensions, ω j (t) represents the risk dimension weight corresponding to the j-th risk index, and σ is the sigmoid activation function.

[0074] Understandably, current mainstream risk detection methods often only consider a single risk indicator, ignoring the complex correlations and interactions between risk factors, resulting in incomplete risk assessments. Therefore, step S02, by comprehensively considering multiple risk indices and their dynamic weights, can calculate a more comprehensive and accurate probability of current operational risks, which helps to promptly identify potential risks currently faced by the platform and provides strong support for risk management.

[0075] Step S03: Input historical operating data, various risk indices, and risk dimension weights into the preset risk warning model to obtain the probability of future operational risks of the platform to be tested;

[0076] It should be noted that the preset risk warning model is a pre-set model used to determine the probability of future operational risks. The probability of future operational risks is the probability of the platform's operational risks occurring within a future period of time, predicted by the preset risk warning model based on input historical operational data, risk index, and dynamic weights. It can be used to assess the operational risks of the platform under test within a preset future period (e.g., 60 days).

[0077] Additionally, it should be noted that the preset risk warning model can be an LSTM-TCN hybrid neural network model. This model combines the short-term temporal feature extraction capability of LSTM with the long-range dependency capture capability of TCN. Compared with a single LSTM model or a random forest model, the LSTM-TCN hybrid neural network model has better performance, as shown in Table 1.

[0078] Table 1

[0079]

[0080] Table 1 compares the performance of the LSTM model, the Random Forest model, and the LSTM-TCN hybrid neural network model obtained through experiments. The AUC-ROC values ​​of the LSTM-TCN training and test sets are 0.96 and 0.94, respectively, representing improvements of 5.6% and 10.6% compared to the single LSTM (0.92 / 0.89) and Random Forest (0.88 / 0.85). This indicates that it has a better ability to distinguish between risky and non-risky samples. Furthermore, the AUC-ROC of the LSTM-TCN on the test set only decreased by 0.02 compared to the training set (0.96→0.94), significantly lower than that of the single LSTM (0.92→0.89, a decrease of 0.03) and Random Forest (0.88→0.85, a decrease of 0.03). This indicates that it has stronger adaptability to unobserved data and better model stability. The recall rates of LSTM-TCN training set and test set are 0.95 and 0.93, respectively, which are 5.7% and 10.7% higher than those of single LSTM (0.90 / 0.88) and random forest (0.87 / 0.84), indicating that it has a stronger ability to capture high-risk samples (positive class) and effectively reduces the risk of missed reports. The false positive rate of LSTM-TCN training set is only 0.04 and test set is only 0.05, which are 54.5% and 66.7% lower than those of single LSTM (0.08 / 0.11) and random forest (0.12 / 0.15), respectively. This verifies its advantage in controlling false positives of low-risk samples and reduces unnecessary risk intervention costs.

[0081] Additionally, it should be noted that the historical operational data collected in step S01, the determined risk indices and their corresponding risk dimension weights are used as inputs into the preset risk warning model. The preset risk warning model, through training and learning, outputs the operational risk probability of the platform to be detected in the future preset period (e.g., 60 days).

[0082] Understandably, traditional risk detection methods struggle to effectively capture the temporal nonlinear characteristics and long-term dependencies between risk factors, resulting in insufficient predictive power for future risks. Therefore, step S03 involves using an LSTM-TCN hybrid neural network model, combining historical data and dynamic weights, to output more accurate probabilities of future operational risks. The LSTM module captures short-term temporal characteristics, while the TCN module enhances the ability to capture long-term risk transmission paths, thereby improving the accuracy and timeliness of future risk predictions.

[0083] Step S04: Output the current operational risk probability and the future operational risk probability, and determine the risk level of the platform to be tested based on the current operational risk probability and the future operational risk probability, and output the corresponding early warning suggestions based on the risk level.

[0084] It should be noted that the risk level is determined by classifying the probability of current and future operational risks according to preset early warning grading rules, such as Critical, Severe, Watch, and Normal. Early warning recommendations are management suggestions issued for different risk levels, such as limiting loan amounts, increasing the frequency of post-loan reviews, and increasing the risk reserve provision rate.

[0085] Additionally, it should be noted that the system has preset warning classification rules:

[0086]

[0087] Where P can be the current operational risk probability P grade or probability P of future operational risks model R j Let R1 be the j-th risk index. Specifically, when j=1, R1 represents the platform's endogenous risk R. pr When j=2, R2 represents the borrower's credit risk R. cr When j=3, R3 represents the industry-related risk R. ir When j=4, R4 represents the liquidity risk R. lr . Specifically, the thresholds corresponding to the risk indices of each dimension are... KDE (Kernel Density Estimation) estimation based on historical crisis events is used to analyze the distribution density of crisis events in geospatial space. By converting discrete crisis event data into a continuous density surface, it can help identify hotspots of crisis events.

[0088] Additionally, it should be noted that when assessing the risk level of a platform under test, the risk level can be determined by separately judging the current operational risk probability and the future operational risk probability. The risk level corresponding to the current operational risk probability is taken as the first risk level, and the risk level corresponding to the future operational risk probability is taken as the second risk level. If the first and second risk levels are the same, either the first or second risk level is taken as the risk level of the platform under test. If the first and second risk levels are different, the higher-level target level between the first and second risk levels is determined, and this target level is taken as the risk level of the platform under test. For example, when P... grade ∈[0.9,1.0) and At that time, the risk level (first risk level) corresponding to the current operational risk probability of the platform under test is Level 4 (Critical). model When the probability of future operational risks of the platform to be tested is ∈[0.4,0.7), the risk level (second risk level) is Level2 (Watch). Among them, Level4 (Critical) corresponds to a higher level. Therefore, the risk level (first risk level) corresponding to the current operational risk probability is taken as the risk level of the platform to be tested.

[0089] Additionally, it should be noted that when the risk level is Level 1 (Normal), the warning suggestion is to monitor the risk index through a real-time dashboard without active intervention; when the risk level is Level 2 (Watch), the warning suggestion is to increase the post-loan review frequency to twice a week and increase the loan interest rate for high-risk borrowers by 100 basis points above the benchmark interest rate; when the risk level is Level 3 (Severe), the warning suggestion is to limit the single loan amount (≤1 million yuan) and shorten the loan term (≤6 months); when the risk level is Level 4 (Critical), the warning suggestion is to trigger cross-platform liquidity assistance (freeze high-risk transactions and increase the risk reserve provision rate to 25%).

[0090] Understandably, traditional methods often lack detailed risk level classification and corresponding early warning suggestions, resulting in limited practical value for management decisions. Therefore, step S04, by combining current and future operational risk probabilities, can determine the platform's risk level and output specific early warning suggestions. This helps platform managers take timely measures to prevent risks, improves the practicality and effectiveness of risk management, and provides the platform with targeted risk management strategies to help reduce operational risks.

[0091] In one feasible implementation, in step S01, the historical operating data includes first historical data, second historical data, third historical data, and fourth historical data. Each risk index includes platform-endogenous risk, borrower credit risk, industry-related risk, and liquidity risk. The steps of collecting historical operating data of the platform to be tested and determining each risk index include steps A01 to A04.

[0092] Step A01: Extract the risk reserve amount, monthly transaction volume, paid-in registered capital, total outstanding principal, platform loan interest rate, and industry average interest rate of the platform to be tested to obtain the first historical data. Based on the first historical data and the first weight corresponding to the first historical data, the platform's intrinsic risk is obtained.

[0093] It should be noted that the first historical data includes risk reserve amount, monthly transaction volume, paid-in registered capital, total outstanding principal, platform lending rate, and industry average interest rate. This reflects the platform's capital adequacy (reflecting the platform's fundamental strength to withstand risks), business scale, capital strength, debt pressure, and interest rate pricing ability (reflecting the platform's ability to adjust risk and return through market mechanisms). Specifically, risk reserve amount refers to the total amount of the platform's special risk reserve fund (e.g., 50 million yuan); monthly transaction volume refers to the total amount of loan transactions facilitated this month (e.g., 800 million yuan); paid-in registered capital refers to the registered capital actually paid by shareholders (e.g., 100 million yuan); total outstanding principal refers to the total outstanding principal of all loans on the platform (e.g., 5 billion yuan); platform lending rate refers to the platform's current average lending rate (e.g., 12% annualized); and industry average interest rate refers to the industry average lending rate for the same period (e.g., 10% annualized). The platform's intrinsic risk is a comprehensive risk index calculated based on the first historical data and its weighting, reflecting the platform's operational stability.

[0094] Additionally, it should be noted that the formula for calculating the platform's inherent risks is as follows:

[0095]

[0096] Among them, R prThis represents an inherent risk within the platform, with a value ranging from [0,1]. A higher value indicates a higher risk. R res The risk reserve amount reflects the platform's ability to buffer against potential default losses. vol C represents the monthly transaction volume, indicating the sensitivity of business scale to capital consumption. reg To demonstrate capital strength, P... due The total outstanding principal amount measures the crowding-out effect of debt pressure on the liquidity of the testing platform. plat For the platform's loan interest rate, r ind This is the industry average interest rate. This represents the degree of interest rate deviation, used to assess abnormal interest rate pricing, i.e., the rationality of the platform's interest rate pricing strategy. Interest rates deviating excessively from the industry average may imply adverse selection risk (such as attracting high-risk borrowers) or liquidity compensation needs. α, β, and γ are the first weights, determined using the entropy weight method, satisfying α + β + γ = 1. The core of the entropy weight method is to use information entropy to measure the dispersion of each indicator, thereby determining the indicator's weight. The smaller the information entropy, the greater the dispersion of the indicator, and the more information it provides; therefore, it should be assigned a higher weight. When the threshold is exceeded, a risk warning threshold adjustment mechanism is triggered, increasing the weight of γ to mitigate the platform's inherent risk R. pr The platform is more sensitive to the risk signal of "interest rate deviation" because when the platform's interest rate is significantly higher than the industry average, it may attract high-risk borrowers (adverse selection risk), requiring close monitoring; conversely, when the platform's interest rate is significantly lower than the industry average, it may imply insufficient liquidity compensation, requiring early warning. By adjusting the weight of γ, the interest rate deviation... Platform-inherent risks R pr The increased contribution of this factor ensures that such abnormal signals are more significantly reflected in the risk composite value, avoiding delays in early warning due to the underestimation of their impact by static weights.

[0097] Step A02: Extract the delinquency rate, standard deviation of loan amount, mean of loan amount and number of historical defaults of each borrower in the platform to be tested to obtain the second historical data. Based on the second historical data and the second weight corresponding to the second historical data, the credit risk of the borrower is obtained.

[0098] It should be noted that the second set of historical data includes each borrower's delinquency rate, standard deviation of loan amount, average loan amount, and number of historical defaults, reflecting the credit characteristics and repayment stability of the borrower group. Specifically, the delinquency rate represents the percentage of borrowers who have missed payments on time (e.g., 5.2%), the standard deviation of loan amount represents the volatility of individual loan amounts (e.g., standard deviation ± 38,000 yuan), the average loan amount represents the average amount of a single loan (e.g., 85,000 yuan), and the number of historical defaults represents the cumulative number of defaults by a borrower (e.g., 2 times). Borrower credit risk is a comprehensive risk index calculated based on the second set of historical data and the second weighting, reflecting the probability of default and the potential scale of losses for the borrower group.

[0099] Additionally, it should be noted that the formula for calculating borrower credit risk is as follows:

[0100]

[0101] Among them, R cr This represents the borrower's credit risk, with a value ranging from [0,1]. A higher value indicates a higher risk. (D) ov The borrower's delinquency rate (total delinquent periods / total repayment periods) reflects repayment stability. σ loan μ represents the standard deviation of borrowers' borrowing amounts, measuring the volatility of borrowing behavior. loan N represents the average loan amount for each borrower. def The historical number of defaults represents the degree of credit deterioration of the borrower. λ, μ, and v are secondary weights, determined through principal component analysis (PCA). The aim is to extract the latent factors that have the strongest explanatory power for the borrower's default probability, where λ + μ + v = 1. PCA is a commonly used dimensionality reduction technique widely applied in data preprocessing, feature extraction, data analysis, and pattern recognition. It transforms the original data into a new coordinate system through linear transformation, maximizing the variance of the data along the coordinate axes in this new system.

[0102] Step A03: Extract the number of problematic platforms in the current month, the total number of platforms in the industry, the month-on-month growth of the industry's monthly transaction volume, and the industry's transaction volume in the previous month, as well as the current month's transaction volume and historical average transaction volume of the platform to be tested, to obtain the third historical data. Based on the third historical data and the third weight corresponding to the third historical data, the industry-related risk is obtained.

[0103] It should be noted that the third set of historical data includes the number of problematic platforms in the current month, the total number of platforms in the industry, the month-on-month growth of the industry's monthly transaction volume, the industry's transaction volume in the previous month, and the current month's transaction volume and historical average transaction volume of the platform under testing. This reflects the impact of fluctuations in the overall industry environment on individual platforms. Specifically, the number of problematic platforms in the current month refers to the number of platforms experiencing a repayment crisis in the current month (e.g., 15 platforms); the total number of platforms in the industry refers to the total number of normally operating platforms in the entire industry (e.g., 300 platforms); the month-on-month growth of the industry's monthly transaction volume refers to the change rate of the industry's transaction volume this month compared to the previous month (e.g., -2.3%); the industry's transaction volume in the previous month refers to the total transaction volume of the industry in the previous month (e.g., 80 billion yuan); the current month's transaction volume refers to the transaction volume of the platform under testing this month (e.g., 800 million yuan); and the historical average transaction volume refers to the average transaction volume of the platform under testing over a preset historical period (e.g., the past 6 months) (e.g., 750 million yuan). Industry-related risk is used to quantify the impact of fluctuations in the overall environment of the internet finance industry on the operation of individual platforms. Its core lies in identifying the transmission path and asymmetric impact of systemic risks in the industry.

[0104] Additionally, it should be noted that the formula for calculating industry-related risk is as follows:

[0105]

[0106] Among them, R ir This represents industry-related risk, with values ​​ranging from [0,1]. Higher values ​​indicate higher risk. N ip N represents the number of problematic platforms in the current month, reflecting the overall risk level of the industry. tp The total number of platforms in the industry, used to define the percentage of platforms with standardized issues, ΔV ind The month-on-month growth rate of industry transaction volume measures industry volatility. V represents the industry's transaction volume last month. vol The transaction volume of the platform to be tested in the current month. The historical average transaction value of the platform under test represents the stability of its business scale. θ, φ, It belongs to the third weight and can be determined through Granger causality analysis. To capture the time-series driving relationship of industry risk factors. When ΔV ind When φ is less than 0, it automatically increases to ψ. Granger causality test is a statistical method used to test causal relationships between variables in time series data. It uses the predictive power of time series data to determine whether one variable has a causal effect on another.

[0107] Step A04: Extract the net increase in transaction volume, transaction volume, number of top lenders, and total number of lenders for the platform under test in the current month to obtain the fourth historical data. Based on the fourth historical data and the fourth weight corresponding to the fourth historical data, the liquidity risk is obtained.

[0108] It should be noted that the fourth set of historical data includes the net increase in monthly transaction volume, monthly transaction volume, number of top lenders, and total number of lenders for the platform under testing. This reflects the stability of the platform's capital flow and the concentration of lenders. Specifically, the net increase in monthly transaction volume refers to the net inflow of funds into the platform during the month (e.g., +0.5 billion yuan); the monthly transaction volume refers to the total transaction volume facilitated by the platform during the month (e.g., 800 million yuan); the number of top lenders refers to the number of lenders in the top 10% of the platform (e.g., 200); and the total number of lenders refers to the number of active lenders in the platform during the month (e.g., 5,000). Liquidity risk measures the platform's ability to cope with short-term liquidity pressures, which essentially stems from a repayment crisis caused by an imbalance between supply and demand for funds.

[0109] Additionally, it should be noted that the formula for calculating liquidity risk is:

[0110]

[0111] Among them, R lr For liquidity risk, the value ranges from [0,1], with higher values ​​indicating higher risk. ΔF represents the net increase in monthly transaction volume, reflecting the sustainability of the platform's cash flow. V vol N represents the transaction amount for the current month. top10 This refers to the number of top lenders, with the default percentage being the top 10%, N. inv This represents the total number of lenders for the month. Used to measure lender concentration, when When the weight of τ increases to q, it is because the concentration of lenders has an "amplifier" effect during a liquidity crisis. When net capital inflows deteriorate, a concentrated lender structure may lead to simultaneous withdrawals, exacerbating repayment pressure—a "domino effect" (such as other lenders following suit and withdrawing funds), causing a liquidity crisis to erupt instantly. By increasing the weight of τ, the model can more accurately capture this nonlinear shock and avoid delaying early warnings due to underestimating concentration risk by static weights. ρ and τ belong to the fourth weight and can be obtained based on regression analysis of historical crisis events, where ρ + τ = 1. Regression analysis of historical crisis events is an important tool for studying the impact of historical crisis events on the economy, finance, or other fields. Through regression analysis, the relationship between crisis events and other variables can be quantified, helping to better understand the causes, transmission mechanisms, and impacts of crises on the economy and society.

[0112] In this implementation, the platform's intrinsic risk is calculated using first historical data and corresponding first weights. This comprehensively considers the impact of multiple important factors within the platform on risk, more accurately assessing the platform's stability and risk resistance capabilities. It provides reliable intrinsic risk indicators for subsequent risk warnings, effectively enhancing the ability to identify and warn of platform-internal risks. Borrower credit risk is calculated using second historical data and second weights. This comprehensively considers multiple credit-related factors of borrowers, more comprehensively and accurately assessing their credit risk levels. This helps the platform promptly identify potential credit risks, take preventative measures in advance, and improve the effectiveness of risk warnings. Industry-related risk is calculated using third historical data and third weights. This comprehensively considers the impact of the overall industry development trend, competitive situation, and the platform's relative position within the industry on risk, more accurately assessing industry-related risks. This allows the platform to understand the impact of industry dynamics on itself in a timely manner, prepare for risk responses in advance, and improve the comprehensiveness and accuracy of risk warnings. Liquidity risk is calculated using fourth historical data and fourth weights. This comprehensively considers multiple key factors of the platform's cash flow, more accurately assessing the platform's liquidity risk level. This helps the platform promptly identify potential liquidity problems, rationally allocate funds, ensure the platform's normal operation, and improve its ability to warn of liquidity risks.

[0113] In one feasible implementation, step S01, calculating the risk dimension weights corresponding to each risk index, includes steps A11 to A13:

[0114] Step A11: Set a preset duration for the sliding window;

[0115] It should be noted that the time sliding window is a dynamic time interval, formed by looking back from the current moment according to a preset duration. As time goes by, the window continuously slides forward while always maintaining the preset duration. This provides a dynamic, localized time perspective for analyzing risk indices, enabling timely capture of short-term changes in risk. The preset duration is the fixed length of the time sliding window, which can be six months.

[0116] For example, if the preset duration is set to six months, then this time sliding window is based on the current moment and traces back a time interval of six months. As time goes by, this window will continue to slide forward, always maintaining a duration of six months.

[0117] Step A12: Calculate the information entropy value of each risk index within the time sliding window. The information entropy value is obtained through the probability distribution of each risk index within the time sliding window.

[0118] It should be noted that information entropy is an indicator that measures information uncertainty. It is used to quantify the uncertainty of the data distribution of each risk index within a time sliding window. The larger the information entropy value, the more dispersed the data distribution of the risk index and the higher the uncertainty. Conversely, the smaller the information entropy value, the more concentrated the data distribution and the lower the uncertainty. By calculating the information entropy value, we can gain a deeper understanding of the characteristics of each risk index and provide more comprehensive information for risk assessment.

[0119] Additionally, it should be noted that the formula for calculating information entropy is:

[0120]

[0121] Among them, E j (t) represents the information entropy value of the risk index of the j-th dimension within the time sliding window. When j=1, it corresponds to the platform's endogenous risk R. pr When j=2, the corresponding borrower credit risk R cr When j=3, the corresponding industry-related risk R ir When j=4, the corresponding liquidity risk R lr n is the total number of data samples within the time sliding window. For example, in a 6-month time sliding window, n = 6. ij Let be the original value of the risk index of the j-th dimension at time i. Let P be the sum of the values ​​of the risk index of the j-th dimension in the n-th time sliding window. ij Let be the probability distribution of the risk index of the j-th dimension within the n-th time sliding window. The probability distribution describes the probability of a random variable for each possible value.

[0122] Step A13: Generate the risk dimension weights of each risk index based on the information entropy value, and standardize the risk dimension weights so that the sum of the risk dimension weights of each risk index is a preset value.

[0123] It should be noted that the risk dimension weights represent the proportion of each risk index in the overall risk assessment. Different risk indices have varying degrees of impact on the overall risk. By reasonably allocating the risk dimension weights, the importance of each risk index can be accurately reflected. This allows for a more accurate risk assessment result when calculating the weights of each risk index during the overall risk assessment. The preset value is the expected sum of weights after standardization of the risk dimension weights, typically set to 1. This ensures that the sum of the standardized risk dimension weights is 1, facilitating subsequent overall risk assessment and comparison.

[0124] Additionally, it should be noted that the risk dimension weights for each risk index are calculated based on the inverse relationship between information entropy values ​​and risk dimension weights (the smaller the information entropy value, the larger the weight). The formula for calculating the risk dimension weights is as follows:

[0125]

[0126] Where, ω j (t) represents the risk dimension weight corresponding to the risk index of the j-th dimension. There are a total of 4 dimensions, E j (t) represents the information entropy value of the risk index of the j-th dimension within the time sliding window.

[0127] In this embodiment, by setting a preset time sliding window, the analysis time range can be continuously updated over time, reflecting the characteristics and trends of each risk index at different stages in real time. This allows the risk detection method to dynamically track risks, promptly identify newly emerging risk factors or changes in risk levels, and thus more accurately grasp the risk situation during platform operation. This provides a more timely and accurate data foundation for effective operational risk early warning. By calculating the information entropy value of each risk index within the time sliding window, and based on the probability distribution of each risk index within the window, the usefulness of the information contained in each risk index can be objectively quantified. The larger the information entropy value, the higher the uncertainty of the risk index and the weaker its ability to distinguish overall risks; the smaller the information entropy value, the more effectively the risk index reflects the risk situation and the greater its contribution to overall risks. This calculation method provides an objective basis for reasonably determining the weight of each risk index, avoiding the bias caused by subjective weight setting, improving the accuracy and scientific nature of risk assessment, and helping to improve the effectiveness of operational risk early warning. Generating the risk dimension weight of each risk index based on the information entropy value can objectively convert the importance of the risk index reflected by the information entropy into weight values, making the weight setting more scientific and reasonable.

[0128] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 In step S03, the preset risk warning model includes an LSTM module and a TCN module. The steps of inputting historical operational data, various risk indices, and risk dimension weights into the preset risk warning model to obtain the probability of future operational risks for the platform to be detected include steps S11 to S14:

[0129] Step S11: Extract historical operational data and short-term time-series fluctuation characteristics of various risk indices through the stacked gate structure in the LSTM module;

[0130] It should be noted that LSTM (Long Short-Term Memory) is a special type of recurrent neural network (RNN) structure. In this embodiment, it is used to process short-term dependencies in time series data, controlling the flow of information through a stacked gate structure. The stacked gate structure includes input gates, forget gates, and output gates, used to control the flow of information within the LSTM module and extract short-term time series fluctuation features. Short-term time series fluctuation features refer to the fluctuation characteristics of time series data in the short term due to various factors (such as promotional activities, holidays, etc.).

[0131] For example, taking the platform's daily transaction amount data as an example, the transaction amount may fluctuate in the short term due to factors such as promotional activities and holidays. The LSTM module can quickly capture such short-term transaction amount change characteristics, such as a sudden and significant increase or decrease in transaction amount within a few days.

[0132] Step S12: Input historical operating data and various risk indices into the TCN module, and extract long-range risk transmission features through the dilated causal convolutional network in the TCN module.

[0133] It should be noted that TCN is a temporal convolutional network that employs dilated causal convolutional networks (DCCs), which can expand the receptive field and capture long-range dependencies in time series data. The DCC is the core structure of the TCN module; it expands the receptive field through dilation operations to capture long-range risk transmission characteristics. Long-range risk transmission characteristics are the risk transmission features in time series data caused by long-term factors (such as macroeconomic conditions and industry policies).

[0134] For example, the platform's risks may be affected by long-term factors such as the macroeconomic situation and industry policies. For instance, a macroeconomic downturn may lead to a gradual decrease in the platform's transaction volume, and this impact may persist for a long period of time. The TCN module can capture this long-term risk transmission characteristic through dilated causal convolutional networks.

[0135] Step S13: Combine the multi-head attention mechanism module to dynamically weight and fuse the long-range risk transmission features and risk dimension weights to obtain weighted fused features;

[0136] It should be noted that the multi-head attention mechanism module is a mechanism that allows the model to focus on the importance of different parts, and can dynamically weight them according to the importance of different risk dimensions at different time points. The weighted fusion feature is obtained by dynamically weighting the long-term risk transmission features according to the risk dimension weights through the multi-head attention mechanism module.

[0137] For example, under different market environments, the impact of different risk dimensions on platform operational risks may vary. For instance, when market liquidity is tight, the importance of the liquidity risk dimension may increase, while the importance of the market risk dimension may decrease relatively. The multi-head attention mechanism module can dynamically adjust the degree of attention to long-term risk transmission characteristics based on the risk dimension weights to obtain more reasonable weighted fusion characteristics.

[0138] Step S14: Combine short-term time-series fluctuation features and weighted fusion features, and input them into a preset fully connected classifier to generate the probability of future operational risks of the platform to be detected.

[0139] It should be noted that the preset fully connected classifier is a common classification model that maps input features to classification results through linear transformation and nonlinear activation, and is used to generate the probability of future operational risks of the platform to be detected.

[0140] Additionally, it should be noted that the structure of the preset risk warning model can be:

[0141] P model =σ(FC(Concatenate(LSTM(X)) t ); TCN(X) t ))))

[0142] Among them, P model This represents the probability of future operational risks for the platform under test, with a value ranging from [0,1]. A higher value indicates a higher risk. This is a 6×11 dimensional time-series risk data matrix, which is formed by the input 6-month window (other lengths can be set, such as 12 months, 18 months, etc.) × 11 dimensional risk feature sequences (including four-dimensional risk indices and 7 transmission term features), using LSTM(X). t ) represents the short-term temporal fluctuation characteristics (32 dimensions) output by the LSTM module, TCN(X) t ) represents the long-range risk transmission feature (dimension 32) output by the TCN module. Concatenate(·) represents the concatenation of short-term time series fluctuation features and weighted fusion features along the feature dimension. FC(·) is a fully connected layer classifier, and σ is the sigmoid activation function used to compress the data to the [0,1] interval.

[0143] In this implementation, by extracting short-term time-series fluctuation features, the model can more keenly perceive short-term risk changes in platform operation, promptly identify potential risks, and improve the timeliness and accuracy of risk warnings, thus compensating for the shortcomings of traditional methods in short-term risk identification. By extracting long-term risk transmission features, the model can examine platform risks from a more macro and long-term perspective, taking into account the accumulation and transmission effects of risks, avoiding the limitation of focusing only on short-term risks while ignoring long-term potential risks, and improving the comprehensiveness and accuracy of risk warnings. By dynamically weighted fusion features, the importance of different features in different contexts is fully considered, enabling the model to assess platform risks more flexibly and accurately, improving the adaptability and accuracy of risk warnings, and overcoming the limitations of traditional fixed fusion methods. By splicing short-term time-series fluctuation features and weighted fusion features, two different types of complementary features are integrated together to form a more comprehensive and richer feature representation. Then, a preset fully connected classifier is input, and by utilizing the classifier's powerful nonlinear mapping capability, these features are mapped to the probability of future operational risks, thereby achieving a quantitative assessment of platform risks. This fully utilizes information from different features to assess platform risks from multiple dimensions, greatly improving the accuracy and reliability of risk warnings and effectively solving the problem of poor operational risk warning effects in current risk detection methods for internet financial information intermediary platforms.

[0144] In one feasible implementation, in step S11, the stacked gate structure includes a forget gate, an input gate, and an output gate. The step of extracting historical operational data and short-term time-series fluctuation characteristics of various risk indices through the stacked gate structure in the LSTM module includes steps B01 to B04:

[0145] Step B01: Generate transmission term features based on historical operating data and various risk indices, and input the historical operating data, various risk indices, and transmission term features into the LSTM module;

[0146] It should be noted that the transmission term features refer to the potential mutual influence and transmission relationships between different data points during the operation of the platform under test. Specifically, these transmission term features include: industry-credit transmission term (representing the strength of the lagged transmission of industry risk to borrowers' credit), endogenous-liquidity lagged transmission term (representing the quarterly transmission of platform-endogenous defects to liquidity risk), multi-dimensional volatility synergy (representing the multi-risk resonance amplification effect), credit-liquidity cross term (representing the cross-month immediate transmission of credit risk to liquidity), industry volatility amplification effect (representing the non-linear impact of industry upheavals on platform-endogenous risks), lagged third-order endogenous risk (representing the quarterly residual impact of platform-endogenous risks), and risk acceleration term (representing the sudden intensity of borrowers' credit deterioration). Each transmission term can be determined based on historical operating data and various risk indices.

[0147] For example, the formula for calculating the industry-credit transmission term is R. ir (t-1)·R cr (t-2), the formula for calculating endogenous-liquidity lag transmission is R. pr (t-4)·R ir (t), the calculation formula for multi-dimensional fluctuation coordination is (ΔR) ir +ΔR cr )ΔR lr The formula for calculating the credit-liquidity cross term is R. cr (t)· lr (t-1), the formula for calculating the industry volatility amplification effect is: The formula for calculating the third-order lag endogenous risk is R. pr (t-3), the formula for calculating the risk acceleration term is ΔR. cr (t)-ΔR cr (t-1).

[0148] Step B02: Attenuate historical noise information in historical operational data based on the forget gate to discard non-critical historical information in historical operational data;

[0149] It's important to note that the forget gate is a gate structure within the LSTM module used to determine which information in the cell state should be retained or forgotten. The cell state is the state in the LSTM module used to store long-term memory information; it can be viewed as an information container that is continuously updated when processing sequence data to retain important historical information. Historical noise information refers to random fluctuations, outliers, and other information in historical operational data that is not closely related to the platform's actual operational risks. This information can interfere with the accurate extraction of short-term risk fluctuation characteristics. Non-critical historical information refers to historical data with low correlation to the current operational risks of the platform under test, such as transaction records from a long time ago (which have little impact on risk prediction in the short term).

[0150] Additionally, it's important to note that historical operational data may contain significant amounts of noise, which can negatively impact model training and prediction performance. The forgetting gate in the LSTM module automatically analyzes the historical operational data stream, identifying and attenuating low-frequency noise and non-critical historical information (such as random fluctuations in monthly interest rates), suppressing invalid signals by adjusting cell state weights. Specifically, the system assigns higher weights to highly correlated events within a preset timeframe (e.g., two years), while selectively discarding redundant information outside this timeframe (e.g., transaction volumes from two years ago). For instance, when processing fund flow sequences, the forgetting gate reduces the weight of atypical fluctuation data from 2015 to zero, focusing on key trend changes in the second quarter of 2016, thereby extracting a pure historical risk trajectory.

[0151] Step B03: Based on the input gate, filter the key signals in each risk index and transmission term feature, and enhance the key signals;

[0152] It should be noted that the input gate is a gate structure within the LSTM module, responsible for controlling how much new input information (various risk indices and conduction term features) can enter the cell state. Key signals refer to information within the various risk indices and conduction term features that significantly indicate the platform's short-term operational risks, such as sudden, large fluctuations in the risk indices or obvious changes in correlations within the conduction term features.

[0153] Additionally, it's important to note that the input gate synchronously scans the real-time input of various risk indices and transmission term features, filtering out core signals with early warning value and strengthening them. Specifically, the system increases the decision weight of high-weight risk dimensions (such as the credit risk index during a crisis) and assigns gain coefficients to the interaction effects in the transmission term features (such as the amplifying effect of industry fluctuations on liquidity). For example, for a sudden increase in credit risk on a certain platform, the input gate increases the signal strength of the delinquency rate jump by 40%, while simultaneously strengthening the data influence of multi-dimensional fluctuation synergy terms, ensuring that key risk sources dominate the cell state update.

[0154] Step B04: Update the cell state of the LSTM module based on the attenuated historical operating data and the enhanced key signals, and output short-term temporal fluctuation characteristics through the output gate and cell state.

[0155] It should be noted that the system integrates historical data attenuated by the forgetting gate with key signals enhanced by the input gate, updates the LSTM cell state to fuse long-term and short-term risk evolution information, and finally controls the intensity of information release through the output gate to generate short-term time-series features characterizing recent risk fluctuations. The output gate is a gate structure in the LSTM module used to determine how much information in the cell state needs to be output to the next time step or as the output of the current time step.

[0156] In this implementation, by generating transmission term features, the potential relationship between historical operational data and various risk indices can be quantitatively expressed. The forget gate attenuates historical noise information in the historical operational data, selectively discarding non-critical historical information, allowing the model to focus more on data related to current and future risks. This helps improve the model's accuracy in identifying risk features and reduces interference from irrelevant information, thereby enhancing the accuracy of operational risk warnings. The input gate acts like a "filter," selecting key signals from various risk indices and transmission term features, and enhancing these key signals. This allows the model to highlight information that has a significant impact on risk prediction, improving the model's sensitivity to key risk factors. Updating the cell state of the LSTM module based on the attenuated historical operational data and enhanced key signals organically combines historical information and current key signals to form a dynamic internal state that reflects short-term risk changes. The output gate and cell state output short-term time-series fluctuation features, which accurately characterize the fluctuation of platform operational risks in the short term, providing strong support for subsequent risk warnings.

[0157] In one feasible implementation, in step S12, the dilated causal convolutional network includes dilated causal convolutional layers, and the dilation coefficient of each dilated causal convolutional layer increases progressively. The step of extracting long-range risk transmission features through the dilated causal convolutional structure in the TCN module includes steps B11 to B14:

[0158] Step B11: For the first dilated causal convolutional layer, the received historical operational data and various risk indices are used as layer input features. A one-dimensional causal convolution operation is performed on the layer input features based on the convolution kernel with the minimum dilation coefficient to obtain the convolution result. The convolution result is added to the layer input features by residual to obtain the primary feature map and input into the next layer of dilated causal convolutional layer of the first layer.

[0159] It's important to note that the dilated causal convolutional layer is the basic building block of the dilated causal convolutional network. By introducing a dilation coefficient, the receptive field of the convolutional kernel is expanded, enabling the capture of features over a longer time span. The dilation coefficient controls the spacing between elements in the convolutional kernel; a larger coefficient results in a larger receptive field and a wider range of information captured. In the first dilated causal convolutional layer, the dilation coefficient is minimized to facilitate the initial extraction of local features. The hierarchical input features refer to the historical operational data and various risk indices input to the current dilated causal convolutional layer. This data encompasses the platform's operational status and risk situation at different points in time. The convolutional kernel is a set of weight parameters used for the convolution operation. By performing convolution operations with the input features, specific patterns are extracted from the input features. In one-dimensional causal convolution, the kernel slides along the time axis. One-dimensional causal convolution is a special type of convolution operation that ensures that the convolution calculation only depends on the input features from the current and previous times, without utilizing information from future times. This aligns with the causality of time series data. For example, when analyzing daily transaction data of a platform, the analysis of today's transaction situation does not depend on data from the next few days. The convolution result is the output obtained by performing a convolution operation between the convolution kernel and the hierarchical input features, reflecting the feature patterns of the input features within a local range. Residual addition is an operation that adds the convolution result to the hierarchical input features element-wise. Its purpose is to preserve information from the original input features while introducing new features extracted by the convolution operation, helping to alleviate the vanishing gradient problem and improve the network's training performance. The primary feature map is the feature representation obtained after the residual addition operation. It contains local information from the hierarchical input features and new features extracted by the convolution operation, serving as the input to the next dilated causal convolutional layer.

[0160] Additionally, it should be noted that the system uses historical operational data and various risk indices as hierarchical input features to input the first dilated causal convolutional layer. This layer uses a convolutional kernel with the minimum dilation coefficient to perform a one-dimensional causal convolution operation, scanning the input sequence of adjacent time steps to generate convolutional results that quantify local risk correlation patterns (such as capturing the intra-month co-fluctuation of borrower credit risk and cash flow). Subsequently, the system adds the convolutional results to the original hierarchical input features by residual addition, preserving the right skewness and peak characteristics of the risk distribution, and outputs the fused primary feature map to the next dilated causal convolutional layer. The minimum dilation coefficient constraint ensures that the operation only relies on historical data, which can prevent future information leakage.

[0161] For example, assuming historical operational data includes the platform's daily transaction volume and various risk indices such as credit risk index and liquidity risk index, the first dilated causal convolutional layer uses a convolutional kernel with the minimum dilation coefficient (e.g., a dilation coefficient of 1) to perform one-dimensional causal convolution operations on the daily transaction volume and various risk indices. For example, with a kernel size of 3, it performs convolution calculations on data from three consecutive days, obtaining a convolution result. Then, this convolution result is added to the original three-day data (hierarchical input features) as residuals to obtain a primary feature map. This primary feature map contains richer local feature information than the original data.

[0162] Step B12: For any dilated causal convolutional layer other than the first layer, receive the primary feature map output by the previous dilated causal convolutional layer, extract the transmission path of the layer input features at each time scale through the receptive field corresponding to the dilated causal convolutional layer, and add the residual of the transmission path to the layer input features. The time scale is determined by the corresponding receptive field.

[0163] It's important to note that the receptive field refers to the area that the convolutional kernel can cover on the input features. As the dilation coefficient increases, the receptive field gradually expands, capturing features over a longer time span. For example, when the dilation coefficient in the first layer is 1, the receptive field may only cover data from a few time points, while in subsequent layers with increased dilation coefficients, it may cover data from dozens or even hundreds of time points. The time scale is determined by the receptive field and represents the range of time the convolutional kernel can capture. Features at different time scales reflect changes in platform operations and risks over different time spans. For example, a short-term time scale may reflect daily transaction fluctuations on the platform, while a long-term time scale may reflect the impact of changes in the platform's macroeconomic environment on its operations. The transmission path refers to the correlation and transmission pattern between hierarchical input features at different time scales. By calculating the dilated causal convolutional layer, these transmission paths can be extracted, thus revealing long-term risk transmission relationships. For example, when analyzing credit risk indices, the transmission path may reflect the transmission and accumulation process of credit risk at different time points.

[0164] Additionally, it should be noted that for the intermediate dilated causal convolutional layer, the system receives the primary feature map output from the previous layer as the layer's input features. By progressively increasing the dilation coefficient, the receptive field of the convolution is expanded (e.g., a time span from quarterly to annual), scanning the intertemporal dependencies in the feature map and extracting the transmission paths at each time scale (e.g., identifying the lagged impact of a surge in problematic platforms on credit risk three months later). The system then adds the residuals of the transmission paths to the layer's input features, maintaining the integrity of the original risk distribution. The formula for adding the residuals is:

[0165] y = Conv(u) + u

[0166] Where y is the superimposed output of the transmission path and the hierarchical input features, preserving the risk distribution characteristics, u is the original input data of the current dilated causal convolutional layer, and Conv(u) is the intertemporal transmission pattern output after performing dilated convolution on u.

[0167] For example, suppose the dilation coefficient of the second dilated causal convolutional layer is 2, and its receptive field is larger than that of the first layer. It receives the primary feature map output by the first layer and extracts features over a longer time range through convolution operations. For instance, when considering platform trading volume and risk index, the first layer may capture local fluctuation features over a few days, while the second layer can capture trend change features over several weeks. By adding the residuals of the propagation path extracted by the second layer to the primary feature map of the first layer, a new feature map that integrates features from different time scales is obtained.

[0168] Step B13: When the dilated causal convolutional layer is the last dilated causal convolutional layer, the feature vector obtained by adding the residual of the propagation path and the hierarchical input features is used as the output of the long-range risk propagation feature.

[0169] It should be noted that when processing the last dilated causal convolutional layer, the system expands the receptive field to a complete time window (e.g., 6 months) through the convolutional kernel with the maximum dilation coefficient, extracts the systemic risk transmission path (e.g., the whole chain evolution of "decline in capital adequacy → redemption crisis six months later"), adds the residuals of the high-level transmission path and the hierarchical input features, and generates a long-range risk transmission feature that integrates global dependencies. This feature vector carries the cross-cycle risk evolution law (e.g., the trajectory of liquidity risk accumulation from March to August) and is transmitted as the core output to the downstream multi-bullish attention mechanism, which can achieve crisis warning n days (e.g., 65.2 days) in advance.

[0170] Step B14: If the dilated causal convolutional layer is not the last dilated causal convolutional layer, input the primary feature map obtained by adding the residual of the propagation path and the hierarchical input features into the next dilated causal convolutional layer.

[0171] It should be noted that if the current dilated causal convolutional layer is not the last layer, the system will input the primary feature map generated by adding the residuals into the next layer. In each iteration, the dilation coefficient increases (e.g., doubling → quadrupling), and the time scale of the convolutional receptive field is expanded layer by layer (from the medium term to the long term evolution), so as to realize the continuous deepening of the risk signal.

[0172] In this implementation, historical operational data and risk indices are processed using a minimum dilation coefficient convolution kernel in the first dilated causal convolutional layer. The residuals are then summed to output a primary feature map for the next layer. This achieves preliminary feature extraction and information retention of the input data, solving the problem of losing key information during initial feature extraction. It preserves fundamental and crucial features for subsequent analysis, improving the accuracy of risk assessment. Through dilated causal convolutional layers other than the first layer, feature transmission paths at different time scales are extracted using corresponding receptive fields, and the residuals are summed. This enables the mining of risk transmission features across multiple time scales, solving the problem of difficulty in capturing risk transmission relationships across different time spans. This allows the model to comprehensively perceive the transmission of risk over time. To enhance the comprehensiveness of risk warning, the last layer of dilated causal convolutional layer is used to output the feature vector obtained by adding the propagation path and the residual of the layer input features as the long-range risk propagation feature. This achieves accurate extraction and output of key risk features, solves the problem of not being able to effectively extract the core features of long-range risk propagation, provides an accurate basis for subsequent risk prediction, and improves the accuracy of warning. By using a non-last layer of dilated causal convolutional layer to input the primary feature map after adding the residuals into the next layer, the effective propagation and deep mining of features in the multi-layer network are realized. This solves the problem of insufficient deep feature extraction due to easy decay of feature propagation, ensures that the model can gradually mine deeper risk features, and strengthens the ability to identify complex risks.

[0173] In one feasible implementation, in step S13, the multi-head attention mechanism module includes various attention units. The step of dynamically weighting and fusing long-range risk transmission features and risk dimension weights in combination with the multi-head attention mechanism module to obtain weighted fused features includes steps B21 to B22:

[0174] Step B21: Inject long-range risk transmission features into each attention unit, calculate the correlation weights between each feature element in the long-range risk transmission features based on each attention unit, and adjust the attention scores of each feature element in the long-range risk transmission features based on the risk dimension weights to adjust the correlation weights between each feature element.

[0175] It's important to note that the attention unit is a fundamental component of the multi-head attention mechanism module. Each attention unit has the ability to independently calculate attention weights. Different attention units can focus on information in long-term risk transmission characteristics from different perspectives, much like multiple observers examining the same thing from different viewpoints, thus capturing a more comprehensive and richer set of feature relationships. For example, in a model processing financial risk data, some attention units may focus more on the relationships between short-term risk fluctuation characteristics, while others may focus on the correlations between long-term risk trend characteristics. Long-term risk transmission characteristics can be viewed as a multi-dimensional vector, where the value in each dimension is a feature element. These feature elements represent risk information extracted from historical operational data and risk indices from different aspects. The correlation weight is used to measure the importance or correlation between the feature elements in the long-term risk transmission characteristics. It reflects the contribution of each feature element to other feature elements when calculating weighted features. For example, if the correlation weight between feature element A and feature element B is large, it means that feature element A has a greater impact on feature element B when calculating weighted features. The attention score is a numerical value obtained by the attention unit after evaluating each feature element. It reflects the importance of the feature element under the current attention unit. The risk dimension weight will adjust the attention score so that feature elements under different risk dimensions can obtain more reasonable attention scores according to their risk importance. For example, for feature elements with high credit risk dimension, their attention score may be appropriately increased to highlight the importance of credit risk in the overall risk.

[0176] Additionally, it should be noted that the system inputs long-term risk transmission characteristics into each attention unit of the multi-head attention mechanism. Each attention unit acts as an independent information processor, automatically scanning the intrinsic connections between feature elements. Based on the risk dimension weights, the system synchronously adjusts the attention scores of each feature element: increasing attention to high-weight dimensions (such as signals of a sudden increase in delinquency rates) and reducing the influence of low-weight dimensions (such as liquidity indicators during stable periods).

[0177] Step B22: Dynamically weight the long-range risk transmission features based on the adjusted association weights to generate weighted features for each attention unit, and aggregate the weighted features of each attention unit to obtain weighted fusion features.

[0178] It's important to note that weighted features are feature representations obtained by weighting long-range risk transmission features according to adjusted correlation weights. They emphasize the role of key feature elements in overall risk, enabling the model to identify risks more accurately. For example, in a long-range risk transmission feature vector containing multiple feature elements, after weighting, the values ​​of feature elements related to important risks will be relatively large, thus playing a greater role in subsequent risk prediction. Weighted fusion features, on the other hand, are features obtained by aggregating the weighted features generated by each attention unit. They integrate risk information captured from different perspectives by multiple attention units, possessing a more comprehensive and richer risk representation capability. For instance, in internet finance risk detection, weighted fusion features can simultaneously include information such as short-term risk fluctuations, long-term risk trends, and complex relationships between different risk dimensions, helping to improve the accuracy of risk prediction.

[0179] Additionally, it should be noted that the system performs dynamic weighting operations on long-range risk transmission features based on the adjusted correlation weights, giving higher weights to features of key transmission paths. Each attention unit outputs weighted features that strengthen the core signal, and the system merges the weighted features of all attention units to generate a 64-dimensional weighted fusion feature.

[0180] In this embodiment, after injecting long-range risk transmission features into each attention unit, the correlation weights between each feature element in the long-range risk transmission features can be calculated based on the internal mechanism of the attention unit. This allows the model to automatically learn the interrelationships between different feature elements. By introducing risk dimension weights, the model can pay more attention to feature elements that perform well in important risk dimensions, making risk assessment more consistent with reality. Based on the adjusted correlation weights, the long-range risk transmission features are dynamically weighted, and different weights can be assigned according to the importance of each feature element, so that important feature elements occupy a larger proportion in the final weighted features. By aggregating the weighted features of each attention unit, the information of each attention unit can be comprehensively utilized to obtain a more comprehensive and accurate risk feature representation, which helps to improve the model's ability to identify different types of risks, thereby solving the problem of poor operational risk warning effect in the current risk detection methods of Internet financial information intermediary platforms.

[0181] For example, to aid in understanding the technical concept or principles of this application, please refer to Figure 3 , Figure 3The system provides an overall flowchart of the operational risk early warning method. It collects historical operating data and extracts a four-dimensional risk index. Then, it generates dynamic weights corresponding to the four-dimensional risk index using the sliding window entropy weight method. Based on the four-dimensional risk index and the corresponding dynamic weights, it calculates the current risk probability. Simultaneously, the data is input into the LSTM-TCN hybrid model. The LSTM is used to extract short-term fluctuation features, and the TCN is used to capture long-term transmission relationships. The data is then fused using a multi-head attention mechanism. Finally, a fully connected classifier outputs the probability. Subsequently, the system performs joint analysis on the two probabilities to determine the risk level and outputs customized contingency plans to achieve risk early warning and management.

[0182] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the operational risk warning method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0183] This application also provides an operational risk early warning system; please refer to [link / reference]. Figure 4 The operational risk early warning system includes:

[0184] The risk data acquisition module 10 is used to collect historical operational data of the platform to be tested, determine each risk index, and calculate the risk dimension weights corresponding to each risk index.

[0185] The current risk calculation module 20 is used to calculate the current operational risk probability of the platform to be tested based on each risk index and the risk dimension weights corresponding to each risk index.

[0186] The future risk calculation module 30 is used to input historical operating data, various risk indices and risk dimension weights into a preset risk warning model to obtain the probability of future operating risks of the platform to be tested.

[0187] The risk level assessment module 40 is used to output the current operational risk probability and the future operational risk probability, and to determine the risk level of the platform to be tested based on the current operational risk probability and the future operational risk probability, and to output corresponding early warning suggestions based on the risk level.

[0188] Optionally, historical operational data includes first historical data, second historical data, third historical data, and fourth historical data. Each risk index includes platform-endogenous risk, borrower credit risk, industry-related risk, and liquidity risk. The risk data collection module 10 is also used for:

[0189] Extract the risk reserve amount, monthly transaction volume, paid-in registered capital, total outstanding principal, platform loan interest rate and industry average interest rate of the platform to be tested to obtain the first historical data. Based on the first historical data and the first weight corresponding to the first historical data, the platform's intrinsic risk is obtained.

[0190] The delinquency rate, standard deviation of loan amount, mean of loan amount, and number of historical defaults of each borrower in the platform to be tested are extracted to obtain the second historical data. The credit risk of the borrower is obtained based on the second historical data and the second weight corresponding to the second historical data.

[0191] Extract the number of problematic platforms in the current month, the total number of platforms in the industry, the month-on-month growth of the industry's monthly transaction volume, and the industry's transaction volume in the previous month, as well as the current month's transaction volume and historical average transaction volume of the platform to be tested, to obtain the third historical data. Based on the third historical data and the third weight corresponding to the third historical data, the industry-related risk is obtained.

[0192] Extract the net increase in transaction volume, transaction volume, number of top lenders, and total number of lenders for the platform under test in the current month to obtain the fourth historical data. Based on the fourth historical data and the fourth weight corresponding to the fourth historical data, the liquidity risk is obtained.

[0193] Optionally, the risk data acquisition module 10 is also used for:

[0194] Set a preset duration for the sliding window;

[0195] The information entropy value of each risk index is calculated within a time sliding window. The information entropy value is obtained through the probability distribution of each risk index within the time sliding window.

[0196] The risk dimension weights of each risk index are generated based on the information entropy value, and the risk dimension weights are standardized so that the sum of the risk dimension weights of each risk index is a preset value.

[0197] Optionally, the preset risk warning model includes an LSTM module and a TCN module, and the future risk calculation module 30 is also used for:

[0198] Historical operational data and short-term time-series fluctuation characteristics of various risk indices are extracted using the stacked gate structure in the LSTM module.

[0199] Historical operational data and various risk indices are input into the TCN module, and long-range risk transmission features are extracted through the dilated causal convolutional network in the TCN module.

[0200] By combining the multi-head attention mechanism module, long-range risk transmission characteristics and risk dimension weights are dynamically weighted and fused to obtain weighted fused characteristics;

[0201] By combining short-term temporal fluctuation features and weighted fusion features, and inputting them into a preset fully connected classifier, the probability of future operational risks of the platform to be tested is generated.

[0202] Optionally, the future risk calculation module 30 is also used for:

[0203] Based on historical operational data and various risk indices, transmission term features are generated, and the historical operational data, various risk indices, and transmission term features are input into the LSTM module;

[0204] The forget gate is used to attenuate historical noise information in historical operational data in order to discard non-critical historical information in historical operational data.

[0205] Key signals in various risk indices and transmission term features are filtered based on the input gate, and these key signals are then enhanced.

[0206] The cell state of the LSTM module is updated based on the decayed historical operating data and the enhanced key signals, and the short-term temporal fluctuation characteristics are output through the output gate and the cell state.

[0207] Optionally, the dilated causal convolutional network includes dilated causal convolutional layers, with the dilation coefficient of each layer increasing progressively. The future risk calculation module 30 is also used for:

[0208] For the first dilated causal convolutional layer, the received historical operational data and various risk indices are used as the layer input features. A one-dimensional causal convolution operation is performed on the layer input features based on the convolution kernel with the minimum dilation coefficient to obtain the convolution result. The convolution result is added to the layer input features by residual to obtain the primary feature map and input into the next layer of dilated causal convolutional layer after the first layer.

[0209] For any dilated causal convolutional layer other than the first layer, the primary feature map output by the previous dilated causal convolutional layer is received. The transmission path of the layer input features at each time scale is extracted through the receptive field corresponding to the dilated causal convolutional layer. The transmission path is added to the residual of the layer input features. The time scale is determined by the corresponding receptive field.

[0210] In the case where the dilated causal convolutional layer is the last dilated causal convolutional layer, the feature vector obtained by adding the residual of the propagation path and the hierarchical input features is used as the output of the long-range risk propagation feature.

[0211] In the case where the dilated causal convolutional layer is not the last dilated causal convolutional layer, the primary feature map obtained by adding the residual of the propagation path and the hierarchical input features is input into the next dilated causal convolutional layer.

[0212] Optionally, the multi-head attention mechanism module includes various attention units, and the future risk calculation module 30 is also used for:

[0213] Long-range risk transmission features are injected into each attention unit. Based on each attention unit, the correlation weights between each feature element in the long-range risk transmission features are calculated. The attention scores of each feature element in the long-range risk transmission features are adjusted based on the risk dimension weights to adjust the correlation weights between each feature element.

[0214] The long-range risk transmission features are dynamically weighted based on the adjusted association weights to generate weighted features for each attention unit. The weighted features of each attention unit are then aggregated to obtain a weighted fusion feature.

[0215] The operational risk early warning system provided in this application, employing the operational risk early warning method in the above embodiments, can solve the technical problem of poor operational risk early warning effect. Compared with the prior art, the beneficial effects of the operational risk early warning system provided in this application are the same as those of the operational risk early warning method provided in the above embodiments, and other technical features of the operational risk early warning system are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0216] This application provides an electronic device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the operational risk warning method in Embodiment 1 above.

[0217] The following is for reference. Figure 5 The diagram illustrates a structural schematic of an electronic device suitable for implementing embodiments of this application. The electronic devices in these embodiments may include, but are not limited to, mobile terminals such as mobile phones, laptops, and PADs (Portable Application Description: Tablet computers), as well as fixed terminals such as digital TVs and desktop computers. Figure 5 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0218] like Figure 5As shown, the electronic device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the electronic device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. The communication device 1009 allows the electronic device to communicate wirelessly or wiredly with other devices to exchange data. Although the diagrams show electronic devices with various systems, it should be understood that it is not required to implement or have all of the systems shown. More or fewer systems may be implemented alternatively.

[0219] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0220] The electronic device provided in this application, employing the operational risk warning method in the above embodiments, can solve the technical problem of poor operational risk warning effect. Compared with the prior art, the beneficial effects of the electronic device provided in this application are the same as those of the operational risk warning method provided in the above embodiments, and other technical features of the electronic device are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0221] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0222] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0223] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the operational risk warning method in the above embodiments.

[0224] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0225] The aforementioned computer-readable storage medium may be included in an electronic device or may exist independently without being assembled into an electronic device.

[0226] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by an electronic device, the operational risk warning device is able to collect historical operational data of the platform under test, determine each risk index, and calculate the risk dimension weights corresponding to each risk index; calculate the current operational risk probability of the platform under test based on each risk index and its corresponding risk dimension weights; input the historical operational data, each risk index, and the risk dimension weights into a preset risk warning model to obtain the future operational risk probability of the platform under test; determine the risk level of the platform under test based on the current and future operational risk probabilities, and output corresponding warning suggestions based on the risk level.

[0227] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0228] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0229] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0230] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described operational risk warning method, thereby solving the technical problem of poor operational risk warning effectiveness. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the operational risk warning method provided in the above embodiments, and will not be repeated here.

[0231] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the operational risk warning method described above.

[0232] The computer program product provided in this application can solve the technical problem of poor operational risk early warning effect. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the operational risk early warning method provided in the above embodiments, and will not be repeated here.

[0233] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. An operational risk early warning method, characterized in that, The operational risk early warning methods include: Collect historical operational data of the platform to be tested, determine each risk index, and calculate the risk dimension weight corresponding to each risk index; Based on the risk indices and the corresponding risk dimension weights of the risk indices, the probability of the current operational risk of the platform under test is calculated. The historical operational data, the various risk indices, and the weights of the risk dimensions are input into a preset risk warning model to obtain the probability of future operational risks of the platform to be tested. Output the current operational risk probability and the future operational risk probability, determine the risk level of the platform to be tested based on the current operational risk probability and the future operational risk probability, and output the corresponding early warning suggestion based on the risk level.

2. The operational risk early warning method as described in claim 1, characterized in that, The historical operational data includes first historical data, second historical data, third historical data, and fourth historical data. The various risk indices include platform-endogenous risk, borrower credit risk, industry-related risk, and liquidity risk. The steps for collecting historical operational data of the platform under test and determining each risk index include: Extract the risk reserve amount, monthly transaction volume, paid-in registered capital, total outstanding principal, platform loan interest rate and industry average interest rate of the platform to be tested to obtain the first historical data, and obtain the platform's intrinsic risk based on the first historical data and the first weight corresponding to the first historical data. The delinquency rate, standard deviation of loan amount, mean of loan amount, and number of historical defaults of each borrower in the platform to be tested are extracted to obtain the second historical data. The credit risk of the borrower is obtained based on the second historical data and the second weight corresponding to the second historical data. Extract the number of problematic platforms in the current month, the total number of platforms in the industry, the month-on-month growth of the industry's monthly transaction volume, and the industry's transaction volume in the previous month, as well as the current month's transaction volume and historical average transaction volume of the platform to be tested, to obtain the third historical data. Based on the third historical data and the third weight corresponding to the third historical data, the industry-related risk is obtained. The net increase in transaction volume, transaction volume, number of top lenders, and total number of lenders in the current month are extracted from the platform to be tested to obtain the fourth historical data. The liquidity risk is obtained based on the fourth historical data and the fourth weight corresponding to the fourth historical data.

3. The operational risk early warning method as described in claim 1, characterized in that, The steps for calculating the risk dimension weights corresponding to each risk index include: Set a preset duration for the sliding window; The information entropy value of each risk index is calculated within the time sliding window, and the information entropy value is obtained through the probability distribution of each risk index within the time sliding window. The risk dimension weights of each risk index are generated based on the information entropy value, and the risk dimension weights are standardized so that the sum of the risk dimension weights of each risk index is a preset value.

4. The operational risk early warning method as described in claim 1, characterized in that, The preset risk warning model includes an LSTM module and a TCN module. The step of inputting the historical operating data, the various risk indices, and the risk dimension weights into the preset risk warning model to obtain the probability of future operational risks of the platform to be detected further includes: The historical operational data and the short-term time-series fluctuation characteristics of each risk index are extracted through the stacked gate structure in the LSTM module. The historical operational data and the various risk indices are input into the TCN module, and the long-range risk transmission features are extracted through the dilated causal convolutional network in the TCN module. By combining the multi-head attention mechanism module, the long-range risk transmission features and the risk dimension weights are dynamically weighted and fused to obtain weighted fused features; By combining the short-term time-series fluctuation features and the weighted fusion features, and inputting them into a preset fully connected classifier, the probability of future operational risks of the platform to be detected is generated.

5. The operational risk early warning method as described in claim 4, characterized in that, The stacked gate structure includes a forget gate, an input gate, and an output gate. The step of extracting the historical operating data and the short-term time-series fluctuation characteristics of each risk index through the stacked gate structure in the LSTM module includes: Based on the historical operating data and the risk indices, a transmission term feature is generated, and the historical operating data, the risk indices, and the transmission term feature are input into the LSTM module; The forget gate is used to attenuate historical noise information in the historical operation data in order to discard non-critical historical information in the historical operation data. Based on the input gate, key signals in each risk index and the characteristics of the transmission term are filtered, and the key signals are enhanced. The cell state of the LSTM module is updated based on the attenuated historical operating data and the enhanced key signals, and short-term temporal fluctuation characteristics are output through the output gate and the cell state.

6. The operational risk early warning method as described in claim 4, characterized in that, The dilated causal convolutional network includes dilated causal convolutional layers, wherein the dilation coefficient of each dilated causal convolutional layer increases progressively. The step of extracting long-range risk transmission features through the dilated causal convolutional structure in the TCN module includes: For the first dilated causal convolutional layer, the received historical operational data and various risk indices are used as the layer input features. A one-dimensional causal convolution operation is performed on the layer input features based on the convolution kernel with the minimum dilation coefficient to obtain the convolution result. The convolution result is added to the layer input features by residual to obtain the primary feature map and input into the next layer of dilated causal convolutional layer of the first layer. For any dilated causal convolutional layer other than the first layer, the primary feature map output by the previous dilated causal convolutional layer is received. The transmission path of the hierarchical input feature at each time scale is extracted through the receptive field corresponding to the dilated causal convolutional layer. The residual of the transmission path is added to the hierarchical input feature. The time scale is determined by the corresponding receptive field. In the case where the dilated causal convolutional layer is the last dilated causal convolutional layer, the feature vector obtained by adding the residual of the propagation path and the hierarchical input features is used as the long-range risk propagation feature output. If the dilated causal convolutional layer is not the last dilated causal convolutional layer, the primary feature map obtained by adding the residual of the propagation path and the hierarchical input features is input into the next dilated causal convolutional layer.

7. The operational risk early warning method as described in claim 4, characterized in that, The multi-head attention mechanism module includes various attention units. The step of dynamically weighting and fusing the long-range risk transmission features and the risk dimension weights using the multi-head attention mechanism module to obtain the weighted fused features includes: The long-range risk transmission feature is injected into each attention unit. Based on each attention unit, the correlation weight between each feature element in the long-range risk transmission feature is calculated. Based on the risk dimension weight, the attention score of each feature element in the long-range risk transmission feature is adjusted to adjust the correlation weight between each feature element. The long-range risk transmission features are dynamically weighted based on the adjusted association weights to generate weighted features for each attention unit. The weighted features of each attention unit are then aggregated to obtain a weighted fusion feature.

8. An electronic device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the operational risk warning method as described in any one of claims 1 to 7.

9. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the operational risk warning method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the operational risk warning method as described in any one of claims 1 to 7.

Images