Financial risk prediction method, system and device based on artificial intelligence and medium

Abstract

The invention relates to a financial risk prediction method, system and device based on artificial intelligence and a medium, and the method comprises the steps: constructing a unified time sequence feature through real-time synchronous processing of ETF option order book, spot quotation and market maker asset balance sheet data; aiming at the problem that the asymmetric conduction of the option implicit volatility and the spot volatility cannot be quantified in the prior art, solving an average option volatility curved surface gradient and verifying the distortion effectiveness in combination with spot volatility correlation; aiming at a mobility illusion risk caused by capital constraint of the market operator, modeling a market operator order withdrawal behavior probability through reinforcement learning; the fluctuation gradient, the behavior probability and the distortion mark are further fused, the conduction suppression effect is analyzed, and a comprehensive risk score is dynamically generated; a hierarchical early warning signal is output through nonlinear transformation and threshold mapping, collaborative prediction of asymmetric conduction risk and mobility illusion is realized, the limitation of a traditional model on separation analysis of two types of risks is broken through, and the accuracy and timeliness of risk identification in an extreme market environment are improved.

Description

Technical Field

[0001] This invention relates to the field of risk prediction technology, specifically to a method, system, device, and medium for predicting financial risks based on artificial intelligence. Background Technology

[0002] In financial trading markets, particularly in the realm of exchange-traded funds (ETFs), the core value of risk prediction methods lies in the early identification of abnormal market fluctuations and sudden liquidity shifts. Traditional risk models primarily focus on single-dimensional market indicator analysis, such as predicting price risk through historical volatility or assessing liquidity risk based on bid-ask spreads. While these methods are effective in stable market environments, they face fundamental challenges when the market experiences asymmetric transmission between implied volatility of options and volatility in the spot market (i.e., sudden shifts in sentiment in the derivatives market are not simultaneously reflected in the spot market) coupled with the liquidity illusion caused by market maker capital constraints (i.e., apparent ample liquidity masking the risk of instantaneous order cancellations).

[0003] Existing technologies suffer from three major drawbacks: First, traditional GARCH-like models cannot quantify the time-varying transmission strength of implied volatility jumps in options to spot volatility, especially during periods of sharp market volatility, where the transmission lag effect leads to a significant delay in risk signals. Second, liquidity indicators based on static quote data (such as the Amihud index) cannot identify instantaneous order cancellations by market makers due to a sudden increase in financing costs, creating false signals of ample liquidity. Third, the fragmented analysis of asymmetric transmission and liquidity illusion makes it impossible to capture the multiplier amplification effect of both types of risks. For example, when panic transmission in the options market is obstructed, coupled with order cancellations by market makers, it will trigger a chain reaction of liquidity vacuum and price flash crashes. Summary of the Invention

[0004] Based on this, the purpose of this invention is to provide an artificial intelligence-based financial risk prediction method, system, device, and medium that can simultaneously capture asymmetric transmission risk and liquidity illusion, and quantify their synergistic effect.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides a financial risk prediction method based on artificial intelligence, comprising the following steps:

[0007] S1: Perform clock synchronization processing on real-time acquired ETF option order book data, spot market data and market maker balance sheet data, and unify heterogeneous data sources to the same time base through a timestamp alignment engine to generate a synchronization feature dataset;

[0008] S2: Solve the second-order partial derivative of the at-the-money option volatility surface for the implied volatility data in the synchronous feature dataset to generate a volatility gradient vector. Combine this with the spot volatility data in the synchronous feature dataset and calculate the correlation coefficient between the volatility gradient vector and spot volatility within a preset time window to generate a distortion indicator that indicates the effectiveness of volatility distortion.

[0009] S3: Reinforcement learning modeling is performed on the volatility gradient vector and capital status data in the synchronization feature dataset. The Q-learning algorithm is used to simulate the decision-making behavior of market makers in a volatile distortion environment and generate a probability distribution of market maker behavior that includes order cancellation probability.

[0010] S4: Perform covariance analysis on the volatility gradient vector and the order cancellation probability in the market maker behavior probability distribution. Generate a transmission inhibition factor by calculating the statistical correlation between the change in volatility gradient and the change in order cancellation behavior within a preset time window.

[0011] S5: Based on the dynamic coefficient adjustment formula, a comprehensive risk score is calculated for volatility gradient vector, transmission inhibition factor and distortion indicator. The comprehensive risk score is processed by nonlinear transformation and a graded early warning signal is generated by mapping through a preset risk threshold range. The graded early warning signal is used to indicate the comprehensive risk level of asymmetric transmission risk and liquidity illusion in the ETF market.

[0012] In one embodiment, S1 of the artificial intelligence-based financial risk prediction method provided by the present invention specifically includes the following steps:

[0013] S11: Extract implied volatility from real-time acquired ETF option order book data, and generate an implied volatility matrix by analyzing option quote data with different strike prices and expiration dates;

[0014] S12: Perform volatility calculation on real-time acquired spot market data, calculate the standard deviation of returns within the rolling time window based on the high-frequency transaction price series, and generate a spot volatility vector;

[0015] S13: Extract capital status from real-time market maker balance sheet data and generate a capital status vector by quantifying the rate of change in financing costs and the trend of leverage ratio changes.

[0016] S14: Time alignment is performed on the implied volatility matrix, spot volatility vector, and capital state vector. A network delay compensation algorithm is used to eliminate the transmission time difference between different data sources and generate a synchronous feature dataset.

[0017] In one embodiment, step S2 of the artificial intelligence-based financial risk prediction method provided by the present invention specifically includes the following steps:

[0018] S21: Perform surface gradient calculation on the implied volatility data in the synchronous feature dataset, and generate a volatility gradient vector by solving the second-order mixed partial derivative of the three-dimensional surface in the at-the-money option region.

[0019] S22: Perform correlation analysis on volatility gradient vector and spot volatility data, and use sliding window covariance calculation to perform joint calculation of covariance and standard deviation on vector data within a fixed time window to generate time series correlation coefficient;

[0020] S23: Perform distortion verification processing on the absolute value of the volatility gradient vector and the correlation coefficient of the time series. Simultaneously detect the gradient magnitude and the negative correlation strength through a dual threshold comparison mechanism to generate a distortion indicator that indicates the effectiveness of volatility distortion.

[0021] In one embodiment, step S3 of the artificial intelligence-based financial risk prediction method provided by the present invention specifically includes the following steps:

[0022] S31: Perform state space construction processing on the volatility gradient vector and capital state vector, and concatenate the multi-dimensional feature vectors according to the time series to form a unified state representation to generate reinforcement learning input state;

[0023] S32: Perform behavioral value update processing on the input state of reinforcement learning, and use the priority experience replay mechanism to adjust the training weights of historical experience data of Q-learning algorithm through importance sampling to generate behavioral value function matrix;

[0024] S33: Perform probability transformation on the behavior value function matrix, and use the softmax function with temperature parameter to perform exponential scaling and normalization calculation on the behavior value to generate a three-dimensional behavior probability distribution including the order cancellation probability.

[0025] In one embodiment, step S4 of the artificial intelligence-based financial risk prediction method provided by the present invention specifically includes the following steps:

[0026] S41: Perform time window slicing on the volatility gradient vector and order cancellation probability, extract continuous time series data segments based on a preset duration, and generate a covariance analysis dataset.

[0027] S42: Perform statistical correlation calculation on the covariance analysis dataset, solve the eigenvalues ​​of the covariance matrix using joint moment analysis, and generate statistical correlation measures.

[0028] S43: Perform suppression factor transformation on the statistical correlation measure, and use covariance standardization to convert the correlation coefficient into a suppression strength index in the 0-1 interval to generate a transmission suppression factor.

[0029] In one embodiment, step S5 of the artificial intelligence-based financial risk prediction method provided by the present invention specifically includes the following steps:

[0030] S51: The volatility gradient vector, transmission inhibition factor and distortion indicator are weighted and processed, and multi-dimensional risk features are fused through dynamic coefficient formula to generate the original risk score.

[0031] S52: Perform nonlinear transformation on the original risk score, and use an S-curve function to compress the score value to the 0-1 range to eliminate distribution bias and generate a normalized risk score.

[0032] S53: Perform threshold mapping processing on the normalized risk score, determine the corresponding early warning level range by comparing the preset risk level boundary values, and generate a graded early warning signal.

[0033] Secondly, the present invention provides a financial risk prediction system based on artificial intelligence, which is configured with the following modules:

[0034] The data clock synchronization module is used to perform clock synchronization processing on real-time acquired ETF option order book data, spot market data and market maker balance sheet data. It uses a timestamp alignment engine to unify heterogeneous data sources to the same time base and generate a synchronization feature dataset.

[0035] The volatility distortion determination module is used to solve the second-order partial derivative of the at-the-money option volatility surface from the implied volatility data in the synchronous feature dataset to generate a volatility gradient vector. Combined with the spot volatility data in the synchronous feature dataset, the module calculates the correlation coefficient between the volatility gradient vector and the spot volatility within a preset time window to generate a distortion flag indicating the effectiveness of volatility distortion.

[0036] The market maker behavior modeling module is used to perform reinforcement learning modeling on the volatility gradient vector and capital status data in the synchronization feature dataset. It simulates the decision-making behavior of market makers in a volatile and distorted environment through the Q-learning algorithm and generates a probability distribution of market maker behavior that includes the probability of order cancellation.

[0037] The suppression factor calculation module is used to perform covariance analysis on the volatility gradient vector and the order cancellation probability in the probability distribution of market maker behavior. By calculating the statistical correlation between the change in volatility gradient and the change in order cancellation behavior within a preset time window, a transmission suppression factor is generated.

[0038] The comprehensive risk warning module is used to calculate a comprehensive risk score based on the volatility gradient vector, transmission inhibition factor, and distortion indicator using a dynamic coefficient adjustment formula. The comprehensive risk score is then processed by nonlinear transformation, and a graded warning signal is generated by mapping through a preset risk threshold range. The graded warning signal is used to indicate the comprehensive risk level of asymmetric transmission risk and liquidity illusion in the ETF market.

[0039] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned artificial intelligence-based financial risk prediction methods.

[0040] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the aforementioned artificial intelligence-based financial risk prediction methods.

[0041] In summary, the AI-based financial risk prediction method provided in this application achieves a synergistic prediction effect on asymmetric transmission risk and liquidity illusion, which is difficult to achieve with traditional methods, through multi-source data synchronization and cross-dimensional risk coupling analysis. Addressing the transmission lag problem between implied volatility and spot volatility, a gradient vector is generated by solving the second-order partial derivative of the at-the-money option volatility surface. This, combined with the correlation with spot volatility to verify distortion indicators, achieves a precise quantification of transmission strength, overcoming the limitation of traditional GARCH models in failing to capture time-varying transmission strength. Regarding the liquidity illusion risk caused by market maker capital constraints, reinforcement learning is used to model the probability distribution of behavior and analyze the statistical correlation between order cancellation behavior and volatility gradients, enabling dynamic identification of instantaneous order cancellation risk and overcoming the misjudgment defects of static liquidity indicators. Furthermore, dynamic coefficient weighting is achieved by integrating volatility gradients, transmission inhibition factors, and distortion indicators to quantify the multiplier effect of the two types of risks. By generating tiered early warning signals through nonlinear transformation and threshold mapping, an early warning mechanism can be formed under extreme market conditions to detect blocked risk transmission paths and liquidity vacuums, significantly improving the foresight and reliability of risk prediction.

[0042] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0043] Figure 1 A flowchart illustrating an artificial intelligence-based financial risk prediction method provided in this application embodiment;

[0044] Figure 2 This is a schematic diagram of the structure of an artificial intelligence-based financial risk prediction system provided in another embodiment of this application. Detailed Implementation

[0045] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0047] In one embodiment, such as Figure 1 As shown, an artificial intelligence-based financial risk prediction method is provided. This embodiment illustrates the method's application to a terminal, but it is understood that the method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0048] S1: Perform clock synchronization processing on real-time acquired ETF option order book data, spot market data, and market maker balance sheet data. Use a timestamp alignment engine to unify heterogeneous data sources to the same time base and generate a synchronized feature dataset.

[0049] Specifically, the system obtains ETF option order book data through the exchange's designated interface. The information obtained includes multiple price levels in the buy and sell orders, order quantities, order numbers, original timestamps, option contract codes, strike prices, expiration dates, and other related information. It also obtains ETF spot market data through the securities market's real-time market data distribution system, including the spot market's opening price, closing price, highest price, lowest price, trading volume, turnover, and tick-level timestamps. Furthermore, it obtains market maker balance sheet data through the market maker compliance data reporting channel, involving information such as core capital adequacy ratio, leverage ratio, financing cost ratio, proportion of liquid assets, and data generation timestamps. The data is updated at fixed intervals.

[0050] Preferably, the system can use Coordinated Universal Time (UTC) as the unified time base, establish a clock synchronization connection with the national time service center server via the Network Time Protocol (NAT), and calibrate the local server clock to ensure accurate time synchronization. The system standardizes the format of the raw timestamps from the three types of data sources, converting timestamps of different formats into unified UTC timestamps to ensure consistency in time representation. For data sources with different time granularities, the system uses interpolation methods for time alignment and completion, generating corresponding capital status interpolation data at each target time node and controlling the error range during the interpolation process.

[0051] Furthermore, the system performs duplicate data removal based on timestamps, data types, and unique identifiers. Statistical methods are used to identify and remove data with abnormal timestamps, including timestamps outside normal trading hours and logically contradictory time-series data, ensuring data validity. The system constructs a synchronous feature dataset indexed by Coordinated Universal Time (UTC) timestamps. This dataset contains four main categories of feature fields: option implied volatility, spot volatility, market maker capital status, and basic market data. The system uses a columnar storage format to store the synchronous feature dataset, which supports efficient time-series queries and feature extraction while also offering a high data compression ratio. The system performs integrity checks on the synchronous feature dataset to ensure that the effective data coverage meets subsequent analysis requirements and that the time-series continuity meets preset standards, i.e., the number of missing data points within any specified time period does not exceed a limited limit.

[0052] S2: Solve the second-order partial derivative of the at-the-money option volatility surface from the implied volatility data in the synchronous feature dataset to generate a volatility gradient vector. Combine this with the spot volatility data in the synchronous feature dataset, and generate a distortion indicator that indicates the effectiveness of volatility distortion by calculating the correlation coefficient between the volatility gradient vector and spot volatility within a preset time window.

[0053] Specifically, the system uses the option order book data in the synchronous feature dataset and employs the Black-Scholes pricing model to inversely deduce the implied volatility of at-the-money options. At-the-money options are defined based on the degree of deviation between the strike price and the real-time price of the ETF spot market. The system selects the overnight rate from the Shanghai Interbank Offered Rate (SHIBOR) as the risk-free rate and uses the historical average dividend yield of the ETF's underlying index as the dividend yield value to ensure the rationality of parameter selection during the calculation process. The system constructs an at-the-money option volatility surface, which uses the strike price and expiration date as variables, and selects at-the-money option contracts with different expiration dates to form a surface sample set.

[0054] Preferably, the system can use a difference method to calculate the second-order partial derivatives of the volatility surface, including the second-order partial derivatives in the strike price direction, the second-order partial derivatives in the expiration date direction, and the cross-second-order partial derivatives. These partial derivatives are used to construct a volatility gradient vector, which is used to quantify the degree of distortion and dimensionality of the volatility surface. Based on spot market data in the synchronous feature dataset, the system uses a rolling window method to calculate the annualized spot volatility. By setting the rolling window length and sliding step size, the calculation is performed according to a predetermined formula to ensure the accuracy and timeliness of the calculation results.

[0055] Furthermore, the system sets a time window, calculates the correlation coefficients between each component of the volatility gradient vector and spot volatility, calculates the variance contribution of each second-order partial derivative through principal component analysis, and calculates the comprehensive correlation coefficient using the variance contribution as the weight. The system sets a correlation coefficient threshold, and generates a distortion flag based on the comparison between the comprehensive correlation coefficient and the threshold. When the absolute value of the comprehensive correlation coefficient is less than the threshold, it indicates that the distortion of implied volatility in the options has not been synchronously transmitted to the spot market, and a distortion flag indicating that the volatility distortion is valid is generated; when the absolute value of the comprehensive correlation coefficient is greater than or equal to the threshold, a distortion flag indicating that the volatility distortion is invalid is generated.

[0056] S3: Reinforcement learning modeling is performed on the volatility gradient vector and capital status data in the synchronization feature dataset. The Q-learning algorithm is used to simulate the decision-making behavior of market makers in a volatile distortion environment, and a probability distribution of market maker behavior including order cancellation probability is generated.

[0057] Specifically, the system combines the volatility gradient vector and capital status data from the synchronization feature dataset to form the state space of the reinforcement learning model. The state vector includes components of the volatility gradient vector, core capital adequacy ratio, leverage ratio, and overnight funding rate for market makers. The core capital adequacy ratio is calculated according to a predetermined formula, the leverage ratio is determined by the ratio of total on- and off-balance sheet assets to core tier 1 capital, and the overnight funding rate for market makers is taken from the interbank market repo rate. The system normalizes each component in the state space, mapping each component to a fixed interval. The normalization parameters are determined based on historical data.

[0058] The system defines three types of decision-making actions for market makers: maintaining orders, partially canceling orders, and fully canceling orders. Each action is defined based on the cancellation ratio. The system designs a reward function, which is constructed based on the balance between market maker profits and risks. The profit component is calculated by subtracting transaction costs from the sum of the products of the bid-ask spread and the trading volume. The bid-ask spread is taken from the difference between the best bid price and the best ask price in the order book. The risk cost component is calculated using risk weights, cancellation probabilities, and the absolute value of the difference between the number of orders placed at the best bid price and the number of orders placed at the best ask price. The reward function uses weighting coefficients to balance the contribution of profits and risks, and these weighting coefficients are determined through an optimization algorithm.

[0059] Further, the system initializes the Q-table, sets the learning rate, discount factor, and exploration rate, with the exploration rate adjusted using a decay strategy. The system divides the time-series data of the synchronous feature dataset into training and validation sets in chronological order, ensuring the number of training samples meets the model training requirements. During training, each time step corresponds to a state. The system selects an action based on the exploration strategy, receives environmental feedback rewards, and transitions to the next state. The system iteratively updates the Q-value according to the update formula until the convergence condition is met. The convergence condition is determined by the change in the Q-value during continuous iterations and the action prediction accuracy on the validation set. After training, the system converts the Q-value into a behavioral probability distribution using the Softmax function. The order cancellation probability is the sum of the probabilities of partial and full cancellation, and this probability value falls within a fixed range.

[0060] S4: Perform covariance analysis on the volatility gradient vector and the order cancellation probability in the market maker behavior probability distribution. Generate a transmission inhibition factor by calculating the statistical correlation between the change in volatility gradient and the change in order cancellation behavior within a preset time window.

[0061] Specifically, the system sets a time window and sliding step size, performs first-order differencing on each component of the volatility gradient vector to obtain the volatility gradient change value, and then normalizes this change value, mapping it to a fixed interval. The system also performs first-order differencing on the order cancellation probability to obtain the order cancellation behavior change value, and similarly normalizes this change value to ensure dimensional consistency between the two types of change values. Finally, the system calculates the covariance matrix of each component of the volatility gradient change and the order cancellation behavior change within the window. The matrix elements are calculated using the expected value formula, which is solved based on the sample data within the window and reflects the average level of the product of the deviations of the two variables from their respective means.

[0062] Furthermore, the system performs eigenvalue decomposition on the covariance matrix to extract the largest eigenvalue, which reflects the strength of the linear correlation between variables. The system calculates a transmission inhibition factor by dividing the largest eigenvalue by the sum of all eigenvalues. This factor has a fixed range and is used to quantify the synergistic effect of volatility distortion transmission and market maker order cancellation behavior. The closer the factor value is to the upper limit of the range, the more significant the mutual reinforcement effect of the two types of risks. When the number of valid data points within the window is insufficient, the system uses a smoothing method to supplement the data. By setting a smoothing coefficient, the weight allocation of the supplementation process is controlled to ensure the statistical validity of the covariance calculation and the reliability of the transmission inhibition factor calculation results.

[0063] S5: Based on the dynamic coefficient adjustment formula, a comprehensive risk score is calculated for volatility gradient vector, transmission inhibition factor and distortion indicator. The comprehensive risk score is processed by nonlinear transformation and a graded early warning signal is generated by mapping through a preset risk threshold range. The graded early warning signal is used to indicate the comprehensive risk level of asymmetric transmission risk and liquidity illusion in the ETF market.

[0064] Specifically, the system constructs a comprehensive risk score calculation formula, which includes three parts: the L2 norm of the volatility gradient vector, a transmission inhibition factor, and a distortion indicator. Each part is weighted and summed using dynamic weighting coefficients. The L2 norm of the volatility gradient vector is calculated by taking the square root of the sum of the squares of each component, used to quantify the absolute strength of volatility distortion. The dynamic weighting coefficients are adaptively adjusted based on the current market state, which is categorized by spot volatility. Different market states correspond to different combinations of weighting coefficients. The system determines the market state by comparing the mean and standard deviation of spot volatility within a sliding window, automatically adjusting the dynamic weights to ensure the matching of weighting coefficients with market state. The system uses the Sigmoid function to map the comprehensive risk score to a fixed interval, generating a normalized comprehensive risk score. This function includes two parameters: an amplification factor and the historical mean of the comprehensive risk score. The amplification factor enhances the distinguishability of intermediate risk intervals, while the historical mean of the comprehensive risk score is calculated based on long-term historical data and updated periodically to ensure the rationality of the transformation process.

[0065] Furthermore, based on long-term market risk event backtesting analysis, the system employs an optimization algorithm to determine risk threshold boundaries, dividing the normalized comprehensive risk score into multiple risk levels, each corresponding to a different warning status. The system-generated warning signals include fields such as risk level identifier, trigger timestamp, core risk contribution factor, and warning duration, and are stored and transmitted using a standard data format. The system outputs warning signals through a predetermined interface, enabling connection with financial institutions' risk control systems and controlling signal transmission latency. The system periodically re-optimizes the threshold range based on the latest market data and risk events, ensuring that warning accuracy and false alarm rates are controlled within a reasonable range, guaranteeing the practicality and reliability of the warning signals.

[0066] In summary, the AI-based financial risk prediction method provided in this application achieves a synergistic prediction effect on asymmetric transmission risk and liquidity illusion, which is difficult to achieve with traditional methods, through multi-source data synchronization and cross-dimensional risk coupling analysis. Addressing the transmission lag problem between implied volatility and spot volatility, a gradient vector is generated by solving the second-order partial derivative of the at-the-money option volatility surface. This, combined with the correlation with spot volatility to verify distortion indicators, achieves a precise quantification of transmission strength, overcoming the limitation of traditional GARCH models in failing to capture time-varying transmission strength. Regarding the liquidity illusion risk caused by market maker capital constraints, reinforcement learning is used to model the probability distribution of behavior and analyze the statistical correlation between order cancellation behavior and volatility gradients, enabling dynamic identification of instantaneous order cancellation risk and overcoming the misjudgment defects of static liquidity indicators. Furthermore, dynamic coefficient weighting is achieved by integrating volatility gradients, transmission inhibition factors, and distortion indicators to quantify the multiplier effect of the two types of risks. By generating tiered early warning signals through nonlinear transformation and threshold mapping, an early warning mechanism can be formed under extreme market conditions to detect blocked risk transmission paths and liquidity vacuums, significantly improving the foresight and reliability of risk prediction.

[0067] In one embodiment, S1 of the artificial intelligence-based financial risk prediction method provided by the present invention specifically includes the following steps:

[0068] S11: Extract implied volatility from the real-time acquired ETF option order book data, and generate an implied volatility matrix by analyzing option quotes with different strike prices and expiration dates.

[0069] Specifically, the system receives real-time ETF option order book data, which includes information such as option buy and sell quotes, order volumes, and transaction records for different strike prices and expiration dates. The system parses the order book data, splitting data entries by option contract code, separating key contract attributes such as strike price and expiration date, and establishing a mapping relationship between contract attributes and quote data. For each independent option contract, the system determines the contract pricing benchmark based on the best bid and best ask prices in the quote data, and uses the Black-Scholes pricing model to inversely deduce implied volatility. During the pricing process, the system calls preset risk-free rate and dividend yield parameters. The risk-free rate is taken from publicly available interbank lending rate data, and the dividend yield is determined based on the historical dividend data of the ETF's underlying index.

[0070] Furthermore, the system categorizes and arranges the implied volatility of each contract in ascending order of strike price and order of proximity to expiration date, constructing a two-dimensional data structure. The row dimension corresponds to different strike prices, and the column dimension corresponds to different expiration dates. Data units store the implied volatility values ​​for the corresponding strike price and expiration date combination. The system performs integrity checks on the arranged two-dimensional data. For missing implied volatility data for contracts, it supplements the data by interpolating the implied volatility of adjacent strike prices and expiration dates, ensuring the integrity of the data structure. The system defines the checked and supplemented two-dimensional data as an implied volatility matrix. The matrix dimensions are jointly determined by the number of ETF option strike prices and expiration dates currently in the market. The matrix data is updated synchronously with the real-time updates of the order book, ensuring data timeliness.

[0071] S12: Perform volatility calculation on the real-time acquired spot market data, calculate the standard deviation of returns within the rolling time window based on the high-frequency transaction price series, and generate a spot volatility vector.

[0072] Specifically, the system acquires real-time high-frequency ETF spot market data, which includes continuous transaction price records, transaction timestamps, and transaction volume. The system preprocesses the high-frequency transaction price series, removing data entries with abnormal transaction prices. Abnormal data is identified through price deviation analysis; price data that deviates from the average of adjacent transaction prices within a certain range is judged as abnormal and excluded. The system sets a rolling time window, the length of which is determined based on market trading frequency and data update cycle. The window slides along the time axis in fixed steps.

[0073] For each rolling time window's price sequence, the system calculates the logarithmic return of adjacent prices, derived from the logarithmic difference between them. The system then calculates the standard deviation of the logarithmic return within each rolling time window, using this deviation as the base value for the corresponding window's spot volatility. The system annualizes this base value by multiplying it by the square root of the number of trading days in a year, yielding the annualized spot volatility. The system arranges the annualized spot volatility corresponding to each window in chronological order, forming a one-dimensional data sequence. This one-dimensional data sequence is defined as a spot volatility vector, where each element corresponds to the annualized spot volatility of a rolling time window. The vector length matches the number of rolling time windows, and new elements are added to the vector in real-time as high-frequency market data is updated.

[0074] S13: Extract capital status data from real-time market maker balance sheet data, and generate a capital status vector by quantifying the rate of change in financing costs and the trend of leverage ratio changes.

[0075] Specifically, the system receives real-time reports of market makers' balance sheet data, which includes indicators such as total assets, total liabilities, core tier 1 capital, risk-weighted assets, financing scale, and financing costs. The system extracts financing cost-related data from the balance sheet data, including current financing interest rates, total financing amount, and financing term. It calculates the rate of change in financing costs by comparing current financing data with historical data from the same period. During the calculation, the system matches historical data by financing term to ensure comparability. The system extracts core tier 1 capital and total on- and off-balance sheet assets, and calculates the leverage ratio using the ratio of total on- and off-balance sheet assets to core tier 1 capital. The leverage ratio calculation strictly follows the formula required by capital regulatory agreements.

[0076] Preferably, the system continuously tracks leverage ratio data across different periods. Trend analysis can be used to quantify leverage ratio changes, reflecting the magnitude of change by comparing the difference between leverage ratios in adjacent periods with the previous period's leverage ratio, and determining the direction of change by fitting the multi-period leverage ratio data. The system uses the quantitative results of the financing cost change rate and leverage ratio change trend as core indicators, combined with auxiliary indicators such as the core capital adequacy ratio and the proportion of liquid assets, to construct a multi-dimensional capital status parameter set. The system arranges the values ​​of each indicator in the parameter set sequentially according to a preset dimensional order, forming a one-dimensional data vector, defined as the capital status vector. Each dimension of the vector corresponds to a capital status indicator, and the vector data is updated synchronously with the balance sheet data to ensure the real-time reflection of capital status.

[0077] S14: Time alignment is performed on the implied volatility matrix, spot volatility vector, and capital state vector. A network delay compensation algorithm is used to eliminate the transmission time difference between different data sources and generate a synchronous feature dataset.

[0078] Specifically, the system acquires the original data timestamps corresponding to the implied volatility matrix, spot volatility vector, and capital state vector. These timestamps contain information such as data generation time, data transmission start time, and data reception time. The system compares and analyzes the timestamps of the three types of vectors to detect transmission time differences between different data sources. By calculating the difference between the data reception time and the data generation time, the system determines the transmission delay duration of each data source and establishes a statistical model for the delay duration. The system employs a network delay compensation algorithm to eliminate transmission time differences. The algorithm calibrates the data timestamps based on historical transmission delay patterns and real-time delay monitoring results for each data source.

[0079] During calibration, the system uses a preset unified time benchmark as a reference to adjust the time identifiers corresponding to data elements in various vectors, ensuring a one-to-one correspondence between implied volatility data, spot volatility data, and capital status data at the same time point. The system performs time consistency checks on the calibrated three types of data, verifying for any data elements with mismatched time points. Mismatched data is interpolated or removed based on the principle of temporal proximity. The system then fuses the validated implied volatility matrix, spot volatility vector, and capital status vector, using the unified time benchmark as an index to integrate feature fields from various data types, forming a comprehensive dataset containing option implied volatility features, spot volatility features, and market maker capital status features. This comprehensive dataset is defined as a synchronous feature dataset, stored in columnar format, supporting fast time-dimension queries and feature extraction. The dataset is updated in real-time as the three types of original data are updated, ensuring data synchronization and effectiveness in subsequent analysis.

[0080] In one embodiment, step S2 of the artificial intelligence-based financial risk prediction method provided by the present invention specifically includes the following steps:

[0081] S21: Perform surface gradient calculation on the implied volatility data in the synchronous feature dataset. Generate the volatility gradient vector by solving the second-order mixed partial derivative of the three-dimensional surface in the at-the-money option region.

[0082] Specifically, the system accesses the implied volatility matrix data from the synchronous feature dataset, which contains implied volatility values ​​corresponding to different strike prices and expiration dates. The system filters data from the implied volatility matrix for at-the-money option regions. These regions are defined by the deviation range between the strike price and the real-time ETF spot price. During the filtering process, a comparison between the spot price and the option strike price is established to determine the strike price range and corresponding expiration dates that meet the at-the-money criteria. Based on the filtered at-the-money option region data, the system constructs a three-dimensional volatility surface. The three dimensions of the surface correspond to the strike price, expiration date, and implied volatility, respectively. Data gridding transforms the discrete implied volatility data into a continuous surface model. The system uses numerical differentiation to solve for the second-order mixed partial derivatives of the three-dimensional surface. These second-order mixed partial derivatives include the second-order mixed partial derivative of the strike price with respect to the expiration date, the second-order mixed partial derivative of the expiration date with respect to the strike price, and the combination of the second-order partial derivatives of the strike price and expiration date with respect to the implied volatility.

[0083] During the solution process, the system divides the 3D surface into local regions. For each local region, it approximates the partial derivative value using the difference calculation of adjacent data points, and reduces numerical calculation errors through multiple iterations. The system organizes and classifies the various second-order mixed partial derivatives obtained from the solution, arranging the partial derivative values ​​according to a preset dimensional order to form a multi-dimensional dataset. The system verifies the validity of this dataset, removing abnormal partial derivative values ​​caused by data dispersion, and supplements missing partial derivative data by analyzing the trend consistency of partial derivatives in adjacent regions. The system defines the verified and supplemented multi-dimensional partial derivative data as a volatility gradient vector, where each dimension of the vector corresponds to a second-order mixed partial derivative result. The vector data is updated in real time with the update of the synchronous feature dataset, ensuring that the gradient vector can reflect the dynamic changes of the implicit volatility surface.

[0084] S22: Perform correlation analysis on volatility gradient vector and spot volatility data. Use sliding window covariance calculation to perform joint calculation on covariance and standard deviation of vector data within a fixed time window to generate time series correlation coefficient.

[0085] Specifically, the system extracts a spot volatility vector from the synchronous feature dataset. This vector contains annualized spot volatility values ​​corresponding to each rolling time window arranged in chronological order. The system pairs the volatility gradient vector with the spot volatility vector, based on their corresponding time identifiers, ensuring a correspondence between the gradient vector data and the spot volatility data at the same time point. The system sets a sliding time window, the time span of which is determined according to the data update frequency and the time granularity requirements of the correlation analysis. The window slides continuously along the time axis at fixed time intervals. For the paired data within each sliding time window, the system first calculates the covariance between each dimension of the volatility gradient vector and the spot volatility data. The covariance is calculated by averaging the products of the deviations of the vector data from their respective means.

[0086] Simultaneously, the system calculates the standard deviation of each dimension of the volatility gradient vector within the window, as well as the standard deviation of the spot volatility data. The system calculates the single-window correlation coefficient by the ratio of the covariance to the product of the standard deviations of the two data types; this coefficient reflects the degree of linear correlation between the two types of data within the window. The system arranges the single-window correlation coefficients corresponding to each window sequentially according to the order of the sliding time windows, forming one-dimensional time series data. The system smooths this time series data to eliminate abnormal fluctuations in the coefficients caused by short-term volatility; the smoothing is achieved through a weighted average of adjacent coefficients. The system defines the smoothed one-dimensional time series data as the time series correlation coefficient. The length of the coefficient sequence is consistent with the number of sliding time windows. New coefficient values ​​are added to the sequence data in real time as the synchronous feature dataset is updated, ensuring the timeliness of the correlation analysis.

[0087] S23: Perform distortion verification processing on the absolute value of the volatility gradient vector and the correlation coefficient of the time series. Simultaneously detect the gradient magnitude and the negative correlation strength through a dual threshold comparison mechanism to generate a distortion indicator that indicates the effectiveness of volatility distortion.

[0088] Specifically, the system performs absolute value transformation on the volatility gradient vector, converting the partial derivatives of each dimension of the vector into non-negative values ​​to form the absolute value of the volatility gradient vector. This absolute value is used to quantify the gradient magnitude of the implied volatility surface. The system extracts the coefficient values ​​corresponding to the absolute values ​​of the volatility gradient vector from the time series correlation coefficients, establishing a one-to-one correspondence between gradient magnitude and correlation coefficient. The system sets a dual threshold system, including a gradient magnitude threshold and a correlation coefficient threshold. The gradient magnitude threshold is determined based on the statistical distribution of historical volatility gradient vector absolute values, while the correlation coefficient threshold is determined based on the distribution characteristics of historical time series correlation coefficients. The system initiates a dual threshold comparison mechanism, first comparing the absolute value of the volatility gradient vector with the gradient magnitude threshold to determine whether the gradient magnitude meets the preset standard; simultaneously, it compares the corresponding time series correlation coefficient with the correlation coefficient threshold to determine whether the negative correlation strength meets the preset requirements.

[0089] During the comparison process, the system simultaneously verifies two comparison results. When the absolute value of the volatility gradient vector is greater than the gradient magnitude threshold and the time series correlation coefficient is less than the correlation coefficient threshold, it indicates that the implied volatility surface distortion of the option is significant and that this distortion is negatively correlated with spot volatility, meaning that the volatility distortion has not been synchronously transmitted to the spot market. Based on this verification result, the system generates a distortion flag, which indicates the validity of the volatility distortion. A first-type distortion flag is generated when both comparison conditions are met simultaneously; a second-type distortion flag is generated when either comparison condition is not met. The distortion flag is updated synchronously with the update of the volatility gradient vector and the time series correlation coefficient, ensuring that the flag can reflect the validity status of the volatility distortion in real time.

[0090] In one embodiment, step S3 of the artificial intelligence-based financial risk prediction method provided by the present invention specifically includes the following steps:

[0091] S31: Perform state space construction processing on the volatility gradient vector and capital state vector, and concatenate the multi-dimensional feature vectors according to the time series to form a unified state representation to generate reinforcement learning input state.

[0092] Specifically, the system extracts the previously generated volatility gradient vector and capital state vector. The volatility gradient vector contains data on various second-order mixed partial derivatives of the three-dimensional surface, while the capital state vector contains data on the rate of change in financing costs, the trend of leverage ratio changes, and other capital-related indicators. The system comprehensively compares the time markers corresponding to the two vectors, establishing a time node mapping relationship to ensure accurate correspondence between volatility gradient data and capital state data at the same time node. The system standardizes the feature data of each dimension of the two types of vectors using the Z-score method, calculated as follows:

[0093]

[0094] in, These are the original feature values ​​for a single dimension. This represents the mean of historical feature data for this dimension. The standard deviation of the historical feature data for this dimension is determined by the statistical results of the feature data in historical periods, ensuring that feature data of different dimensions are mapped to the same numerical range.

[0095] Furthermore, the system horizontally concatenates the volatility gradient vector and capital state vector at the same time point in chronological order. If the volatility gradient vector has dimension m and the capital state vector has dimension n, the comprehensive state vector has dimension m+n, encompassing both market volatility characteristics and market maker capital state characteristics. The system arranges the comprehensive state vectors of each time point sequentially to form a time-series-based state set. The system validates this state set by analyzing the distribution range of feature values ​​to eliminate abnormal state vectors. For missing time point state data, linear interpolation is used to supplement it, with the interpolation calculated based on the state vectors of two adjacent valid time points. The system defines the validated and supplemented time-series-based state set as the reinforcement learning input state. This input state includes a time dimension and a state feature dimension, and is updated synchronously with the real-time updates of the original vectors to ensure a comprehensive reflection of the dynamic correlation between the market and market maker states.

[0096] S32: The behavior value is updated for the input state of reinforcement learning. The priority experience replay mechanism is used to adjust the training weights of historical experience data for Q-learning algorithm through importance sampling, and the behavior value function matrix is ​​generated.

[0097] Specifically, the system constructs an experience pool based on the reinforcement learning input states, storing historical interaction data including reinforcement learning input states, action selection results, reward values, and subsequent state data. The system calculates the temporal difference error for each historical interaction data point in the experience pool. The calculation formula is:

[0098]

[0099] in, The reward value is for environmental feedback. This is a discount factor used to weigh immediate rewards against future rewards. Given the current input state for reinforcement learning, Select a result for the action. The subsequent state after the action is executed. The estimated value of the current state's action combination. This is used to estimate the value of subsequent state-action combinations. The system sets sampling priority based on temporal difference error and adjusts training weights using an importance sampling method. The weight coefficients are used to correct sampling bias. Preferably, the system calls the Q-learning algorithm to update the value parameters based on the adjusted training weights. The update formula is:

[0100]

[0101] in, The learning rate controls the magnitude of each value update, and w is the importance sampling weight. The system arranges the value parameters corresponding to all state-action combinations according to the state and action dimensions, constructing a two-dimensional behavior-value function matrix. The row dimension corresponds to the number of states, and the column dimension corresponds to the number of action types. The system performs convergence checks on the matrix, monitoring the magnitude of changes in matrix elements during continuous iterations, and stops iteration when convergence is achieved.

[0102] S33: Perform probability transformation on the behavior value function matrix, and use the softmax function with temperature parameter to perform exponential scaling and normalization calculation on the behavior value to generate a three-dimensional behavior probability distribution including the order cancellation probability.

[0103] Specifically, the system acquires the converged behavior value function matrix and extracts all action value data corresponding to each state. The system uses a softmax function with a temperature parameter for probability transformation, and the calculation formula is as follows:

[0104]

[0105] in, For a specific state Next action The probability, The state in the behavior value function matrix With action Corresponding expected value This is a temperature parameter used to adjust the smoothness of the probability distribution. For all possible action types, the system performs exponential scaling and normalization calculations on the action value data for each state, ensuring that the sum of the probability values ​​of all actions in each state is 1.

[0106] Furthermore, the system constructs a three-dimensional behavior probability distribution along the time, state, and action dimensions. The time dimension corresponds to the time sequence of the reinforcement learning input states, the state dimension corresponds to the row dimension of the behavior value function matrix, and the action dimension corresponds to three preset actions. The system calculates the order cancellation probability using the following formula:

[0107]

[0108] in, For order cancellation probability, This is a partial cancellation action. This is a full cancellation action. The system performs an integrity check on the three-dimensional behavior probability distribution, verifying whether the probability of all actions in each state meets the requirements, removing abnormal data, and supplementing missing information by interpolating data from adjacent states.

[0109] In one embodiment, step S4 of the artificial intelligence-based financial risk prediction method provided by the present invention specifically includes the following steps:

[0110] S41: Perform time window slicing on the volatility gradient vector and order cancellation probability, extract continuous time series data segments based on a preset duration, and generate a covariance analysis dataset.

[0111] Specifically, the system extracts the previously generated volatility gradient vector and order cancellation probability data. The volatility gradient vector contains multi-dimensional second-order mixed partial derivative data, while the order cancellation probability is derived data from the three-dimensional behavioral probability distribution, reflecting the likelihood of market makers canceling orders. The system comprehensively compares the original time stamps of both types of data, establishing a precise mapping relationship between time nodes to ensure a one-to-one correspondence between volatility gradient data and order cancellation probability data at the same time node, eliminating misalignment issues in the time dimension. The system sets time windows, defining the time range of each window through a window index formula:

[0112]

[0113] in, This represents the k-th time window. This represents the start time of the k-th window. This indicates the preset window duration. The window slides continuously along the time axis in a fixed manner. After each slide, a continuous time-series data segment within the corresponding time range is extracted. The data segment contains volatility gradient vector data and order cancellation probability data for all time nodes within the window. The system performs an integrity check on each extracted data segment, verifying whether the time nodes within the data segment are continuous and whether any data points are missing. Invalid data segments with missing data or broken time nodes are removed, and valid data segments with continuous and complete data are retained. The system integrates all valid data segments in chronological order to construct a dataset containing multiple time-series data segments. Each data segment contains a pair of volatility gradient vector time-series subsequences and order cancellation probability time-series subsequences. This dataset is defined as a covariance analysis dataset.

[0114] S42: Perform statistical correlation calculation on the covariance analysis dataset, solve for the eigenvalues ​​of the covariance matrix using joint moment analysis, and generate statistical correlation measures.

[0115] Specifically, the system extracts each valid data segment from the covariance analysis dataset, separating the volatility gradient vector time-series subseries and the order cancellation probability time-series subseries corresponding to each data segment. The system employs joint moment analysis to first calculate the mean of the volatility gradient vector time-series subseries and the mean of the order cancellation probability time-series subseries within each data segment. The mean is obtained by performing an arithmetic mean calculation on all data points within the sequence. Preferably, the system calculates the covariance of the two types of subseries based on the mean results. The covariance is calculated using the formula:

[0116]

[0117] in, This represents a time series subsequence of the volatility gradient vector. The time series subsequence representing the order cancellation probability. This represents the mathematical expectation operation used to calculate the mean level of a sequence. The system constructs a covariance matrix based on the covariance results. The row dimensions of the covariance matrix correspond to the number of dimensions in the volatility gradient vector, and the column dimensions correspond to the number of dimensions in the order cancellation probability. Each element in the matrix stores the covariance value of the two classes of data in the corresponding dimension. The system performs eigenvalue decomposition on the constructed covariance matrix, extracting all eigenvalues ​​through linear algebraic operations. The magnitude of the eigenvalues ​​reflects the correlation strength of data changes in the corresponding direction of the covariance matrix.

[0118] Furthermore, the system selects the largest eigenvalue from all eigenvalues ​​and defines it as a statistical correlation metric. This metric comprehensively quantifies the overall linear correlation between changes in volatility gradient and changes in order cancellation probability within each data segment. The system arranges the statistical correlation metrics in chronological order of the data segments, forming a statistical correlation metric sequence. New metrics are added synchronously as the covariance analysis dataset is updated, ensuring that the metrics can promptly reflect the dynamic changes in the correlation between the two types of data.

[0119] S43: Perform suppression factor transformation on the statistical correlation measure, and use covariance standardization to convert the correlation coefficient into a suppression strength index in the 0-1 interval to generate a transmission suppression factor.

[0120] Specifically, the system acquires a sequence of statistical correlation measures and extracts the covariance matrix corresponding to each data segment. It then calculates the trace of the covariance matrix, which is the sum of the diagonal elements and reflects the overall variance level of the covariance matrix. Preferably, the system can employ a covariance standardization method to transform the statistical correlation measures into a suppression strength index using a formula:

[0121]

[0122] in, Indicates a transmission inhibition factor. This represents the statistical correlation measure, i.e., the largest eigenvalue of the covariance matrix. Represents the covariance matrix The standardization process involves comparing the statistical correlation measure with the trace of the covariance matrix to eliminate the impact of variance differences across data segments, thus mapping the transformed results uniformly to the 0-1 interval. The system validates the standardized results, checking whether each result falls within the 0-1 interval. For outliers outside this interval, interpolation correction is performed using the suppression strength index of adjacent data segments to ensure the reasonableness of the results.

[0123] The system defines the verified and corrected suppression strength index as the transmission suppression factor. This factor quantifies the synergistic suppression effect of volatility distortion transmission and market maker order cancellation behavior. The closer the value is to 1, the more significant the mutual reinforcement effect of the two types of risks. The transmission suppression factor forms a sequence according to the time order of the corresponding data segments and is updated synchronously with the update of the statistical correlation measure, providing key parameter support for subsequent comprehensive risk score calculation.

[0124] In one embodiment, step S5 of the artificial intelligence-based financial risk prediction method provided by the present invention specifically includes the following steps:

[0125] S51: The volatility gradient vector, transmission inhibition factor and distortion indicator are weighted and processed, and multi-dimensional risk features are fused through dynamic coefficient formula to generate the original risk score.

[0126] Specifically, the system extracts the previously generated volatility gradient vector, transmission suppression factor, and distortion flag. The volatility gradient vector quantifies the implied volatility surface distortion intensity using multidimensional second-order mixed partial derivatives. The transmission suppression factor quantifies the synergistic effect of volatility distortion transmission and market maker order cancellation behavior through covariance standardization. The distortion flag indicates the effectiveness of volatility distortion through a dual-threshold comparison mechanism. The system calculates the L2 norm of the volatility gradient vector using the following formula:

[0127]

[0128] in, This represents the value of the i-th dimension of the volatility gradient vector, where n represents the total number of dimensions of the volatility gradient vector. This norm integrates multi-dimensional gradient information into a single quantitative indicator, comprehensively reflecting the overall strength of the gradient. The system constructs a dynamic coefficient fusion formula:

[0129]

[0130] in, For the original risk score, Let L2 norm be the volatility gradient vector. As a conduction inhibitor, As a marker of distortion, , , These are dynamic weighting coefficients. The determination of dynamic weighting coefficients is based on market conditions, which are categorized by analyzing the time series characteristics of the spot volatility vector. The system performs sliding window statistics on the spot volatility vector and determines the market condition category through the mean and standard deviation of the series.

[0131] Different market conditions correspond to preset coefficient combinations. These coefficient combinations are calibrated through statistical analysis of historical risk event data. The system selects risk event samples from different historical market conditions, calculates the correlation between each feature and the probability of risk event occurrence, assigns initial values ​​to each coefficient based on the correlation, and then iteratively optimizes the coefficient combinations to match the evolution of risk. The system multiplies each feature's quantified value by its corresponding dynamic weight coefficient, sums the products, and obtains the original risk score for each time point. The system verifies the rationality of the original risk score by constructing a reasonable range interval by retrieving score data from the same historical period and market conditions. The current score is compared with the interval, and for abnormal scores exceeding the interval, a weighted average of the scores from three adjacent time points is used for smoothing correction, with weights decreasing sequentially from closest to furthest in time. The original risk score is updated synchronously with the real-time updates of the volatility gradient vector, transmission inhibition factor, and distortion indicator, ensuring timely reflection of the dynamic comprehensive impact of multidimensional risk characteristics.

[0132] S52: The original risk score is subjected to nonlinear transformation. An S-curve function is used to compress the score value to the 0-1 range to eliminate distribution bias and generate a normalized risk score.

[0133] Specifically, the system acquires the original risk score sequence, performs distribution characteristic analysis on the sequence, and calculates the skewness and kurtosis of the sequence to identify the type of distribution deviation. The formulas for calculating skewness and kurtosis are as follows:

[0134]

[0135]

[0136] in, Let n be the i-th score in the original risk scoring sequence, and n be the sequence length. The mean of the sequence. Let be the standard deviation of the sequence. The system uses a sigmoid curve function for nonlinear transformation, and the transformation formula is:

[0137]

[0138] in, For the normalized risk score, x is the original risk score, and k is the scaling factor. The scaling factor is the historical mean of the original risk scores. The scaling factor adjusts the steepness of the transformation curve, while the historical mean corrects the center position of the score distribution. Both are determined using historical statistical data of the original risk scores. The system filters long-term historical score data and calculates different k values ​​with... The transformation effect under the value combination is to select the parameter combination that makes the distribution of the transformed sequence closest to a uniform distribution.

[0139] Furthermore, the system substitutes each original risk score into the formula for calculation, maps the score to the 0-1 interval through exponential operations, and then eliminates the influence of scores of different magnitudes through normalization. The system performs distribution verification on the transformed normalized risk scores, calculates the skewness and kurtosis of the score sequence on the verification set, and compares the calculation results with the preset reasonable interval to confirm whether the distribution deviation has been eliminated. For score sequences that do not meet the distribution requirements, the system adjusts the scaling factor and historical mean, resubmits them into the formula for transformation, and repeats this process until the skewness and kurtosis indicators meet the preset standards. The system defines the verified scores as normalized risk scores, with one normalized risk score corresponding to each time point. The score sequence is updated synchronously with the updates of the original risk scores to ensure the timeliness and accuracy of the scores, providing standardized and comparable data support for risk level determination.

[0140] S53: Perform threshold mapping processing on the normalized risk score, determine the corresponding early warning level range by comparing the preset risk level boundary values, and generate a graded early warning signal.

[0141] Specifically, the system sets risk level boundary values, which are determined based on the normalized score distribution of historical risk events. The system collects long-term historical risk event data, extracts the normalized risk score corresponding to each event, statistically sorts these scores, and divides the system into multiple continuous risk level intervals based on the sorting results. The interval division expression is as follows: ,in Here, m represents the preset risk level boundary values, and m is the number of risk levels, with each interval corresponding to a specific warning level. The system acquires the normalized risk score sequence, extracts the normalized risk score for each time node, and compares the score with each boundary value in ascending order of boundary value. The comparison is first made with... The size, if the score is less than Then it is determined to belong to the first interval, if it is greater than or equal to Then continue with The comparison is repeated until the risk level range to which the current score belongs is determined. Based on the risk level range, the system generates a tiered early warning signal. The early warning signal includes core information such as a time node identifier, a normalized risk score value, the corresponding risk level, and a risk level description. The time node identifier is consistent with the timestamp of the score, and the risk level description corresponds to the early warning level of the range.

[0142] The system performs integrity verification on the generated tiered early warning signals, checking whether the signals contain all core fields, whether the field content is consistent with the scoring and risk level, whether the time node identifier format is standardized, whether the normalized risk score value is within the 0-1 range, and whether the risk level matches the description. Invalid signals with missing information, abnormal format, or mismatched content are removed. The system organizes the valid tiered early warning signals into a signal sequence in chronological order, and the signal sequence is updated in real time with new early warning signals as the normalized risk score is updated. The system outputs the tiered early warning signals to the financial institution's risk control system through a preset interface. The output process follows a preset transmission protocol to ensure the continuity and timeliness of signal transmission, enabling the risk control system to receive early warning signals in real time and take corresponding risk disposal measures based on the signal content, thereby achieving effective early warning of the combined risks of asymmetric transmission and liquidity illusion in the ETF market.

[0143] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0144] Based on the same inventive concept, this application also provides an AI-based financial risk prediction system for implementing the AI-based financial risk prediction method described above. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations in one or more AI-based financial risk prediction system embodiments provided below can be found in the limitations of the AI-based financial risk prediction method described above, and will not be repeated here.

[0145] Preferably, such as Figure 2 As shown, the present invention provides a financial risk prediction system 600 based on artificial intelligence, which is configured with the following modules:

[0146] The data clock synchronization module 610 is used to perform clock synchronization processing on real-time acquired ETF option order book data, spot market data and market maker balance sheet data. It unifies heterogeneous data sources to the same time base through a timestamp alignment engine and generates a synchronization feature dataset.

[0147] The volatility distortion determination module 620 is used to solve the second-order partial derivative of the at-the-money option volatility surface from the implied volatility data in the synchronous feature dataset to generate a volatility gradient vector. Combined with the spot volatility data in the synchronous feature dataset, the module calculates the correlation coefficient between the volatility gradient vector and the spot volatility within a preset time window to generate a distortion flag indicating the effectiveness of volatility distortion.

[0148] The Market Maker Behavior Modeling Module 630 is used to perform reinforcement learning modeling on the volatility gradient vector and capital status data in the synchronization feature dataset. It simulates the decision-making behavior of market makers in a volatile distortion environment through the Q-learning algorithm and generates a probability distribution of market maker behavior that includes order cancellation probability.

[0149] The inhibition factor calculation module 640 is used to perform covariance analysis on the volatility gradient vector and the order cancellation probability in the probability distribution of market maker behavior. By calculating the statistical correlation between the change in volatility gradient and the change in order cancellation behavior within a preset time window, a transmission inhibition factor is generated.

[0150] The comprehensive risk warning module 650 is used to calculate a comprehensive risk score based on the volatility gradient vector, transmission inhibition factor and distortion indicator based on the dynamic coefficient adjustment formula. The comprehensive risk score is processed by nonlinear transformation and a graded warning signal is generated by mapping through a preset risk threshold range. The graded warning signal is used to indicate the comprehensive risk level of asymmetric transmission risk and liquidity illusion in the ETF market.

[0151] Preferably, the data clock synchronization module 610 provided in this application is configured with the following units:

[0152] The implied volatility extraction unit is used to extract implied volatility from real-time acquired ETF option order book data. It generates an implied volatility matrix by analyzing option quote data with different strike prices and expiration dates.

[0153] The spot volatility calculation unit is used to perform volatility calculation on real-time acquired spot market data, calculate the standard deviation of returns within a rolling time window based on high-frequency trading price series, and generate a spot volatility vector.

[0154] The capital status extraction unit is used to extract capital status from real-time acquired market maker balance sheet data and generate a capital status vector by quantifying the rate of change in financing costs and the trend of leverage ratio changes.

[0155] The multi-source data alignment unit is used to perform time alignment processing on the implied volatility matrix, spot volatility vector, and capital state vector. It uses a network delay compensation algorithm to eliminate the transmission time difference between different data sources and generate a synchronous feature dataset.

[0156] Preferably, the fluctuation distortion determination module 620 provided in this application is configured with the following units:

[0157] The volatility gradient calculation unit is used to perform surface gradient calculation on the implied volatility data in the synchronous feature dataset. It generates a volatility gradient vector by solving the three-dimensional surface second-order mixed partial derivative of the at-the-money option region.

[0158] The gradient spot correlation analysis unit is used to perform correlation analysis on volatility gradient vector and spot volatility data. It uses sliding window covariance calculation to perform joint calculation of covariance and standard deviation on vector data within a fixed time window to generate time series correlation coefficient.

[0159] The volatility distortion verification unit is used to perform distortion verification processing on the absolute value of the volatility gradient vector and the time series correlation coefficient. It simultaneously detects the gradient magnitude and the negative correlation strength through a dual threshold comparison mechanism, and generates a distortion flag indicating the effectiveness of volatility distortion.

[0160] Preferably, the market maker behavior modeling module 630 provided in this application is configured with the following units:

[0161] The state space construction unit is used to construct the state space of the volatility gradient vector and the capital state vector, and to concatenate the multi-dimensional feature vectors according to the time series to form a unified state representation, thereby generating the reinforcement learning input state.

[0162] The behavior value update unit is used to update the behavior value of the reinforcement learning input state. It adopts a priority experience replay mechanism to adjust the training weights of historical experience data of the Q-learning algorithm through importance sampling, and generates a behavior value function matrix.

[0163] The behavior probability transformation unit is used to perform probability transformation processing on the behavior value function matrix. It uses a softmax function with a temperature parameter to perform exponential scaling and normalization calculation on the behavior value, generating a three-dimensional behavior probability distribution that includes the order cancellation probability.

[0164] Preferably, the inhibition factor calculation module 640 provided in this application is configured with the following units:

[0165] The time window slicing unit is used to perform time window slicing on the volatility gradient vector and order cancellation probability, extracting continuous time series data segments based on a preset duration to generate a covariance analysis dataset.

[0166] The statistical correlation calculation unit is used to perform statistical correlation calculation on the covariance analysis dataset, solve the eigenvalues ​​of the covariance matrix through joint moment analysis, and generate statistical correlation measures.

[0167] The suppression factor conversion unit is used to perform suppression factor conversion on statistical correlation measures. It uses covariance standardization to convert the correlation coefficient into a suppression strength index in the 0-1 interval, generating a transmission suppression factor.

[0168] Preferably, the comprehensive risk warning module 650 provided in this application is configured with the following units:

[0169] The risk feature weighting unit is used to perform feature weighting on volatility gradient vector, transmission inhibition factor and distortion indicator, and to fuse multi-dimensional risk features through dynamic coefficient formula to generate original risk score;

[0170] The scoring nonlinear transformation unit is used to perform nonlinear transformation on the original risk score. It uses an S-curve function to compress the score value to the 0-1 range to eliminate distribution bias and generate a normalized risk score.

[0171] The scoring threshold mapping early warning unit is used to perform threshold mapping processing on the normalized risk score, determine the corresponding early warning level range by comparing preset risk level boundary values, and generate a graded early warning signal.

[0172] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described artificial intelligence-based financial risk prediction method.

[0173] In one embodiment, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned artificial intelligence-based financial risk prediction method.

[0174] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0175] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0176] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all 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.

Claims

1. A financial risk prediction method based on artificial intelligence, characterized in that, Includes the following steps: S1: Perform clock synchronization processing on real-time acquired ETF option order book data, spot market data and market maker balance sheet data, and unify heterogeneous data sources to the same time base through a timestamp alignment engine to generate a synchronization feature dataset; S2: Solve the second-order partial derivative of the at-the-money option volatility surface for the implied volatility data in the synchronous feature dataset to generate a volatility gradient vector. Combine the spot volatility data in the synchronous feature dataset with the correlation coefficient between the volatility gradient vector and the spot volatility within a preset time window to generate a distortion indicator indicating the effectiveness of volatility distortion. S3: Perform reinforcement learning modeling on the volatility gradient vector and the capital status data in the synchronization feature dataset, simulate the decision-making behavior of market makers in a volatile distortion environment through the Q-learning algorithm, and generate a probability distribution of market maker behavior that includes the probability of order cancellation. S4: Perform covariance analysis on the volatility gradient vector and the order cancellation probability in the market maker behavior probability distribution, and generate a transmission inhibition factor by calculating the statistical correlation between the change in volatility gradient and the change in order cancellation behavior within a preset time window. S5: Calculate the comprehensive risk score of the volatility gradient vector, the transmission inhibition factor and the distortion indicator based on the dynamic coefficient adjustment formula, perform nonlinear transformation processing on the comprehensive risk score, and generate a graded early warning signal by mapping through a preset risk threshold range. The graded early warning signal is used to indicate the comprehensive risk level of asymmetric transmission risk and liquidity illusion in the ETF market.

2. The method according to claim 1, characterized in that, S1 includes: S11: Extract implied volatility from real-time acquired ETF option order book data, and generate an implied volatility matrix by analyzing option quote data with different strike prices and expiration dates; S12: Perform volatility calculation on real-time acquired spot market data, calculate the standard deviation of returns within the rolling time window based on the high-frequency transaction price series, and generate a spot volatility vector; S13: Extract capital status from real-time market maker balance sheet data and generate a capital status vector by quantifying the rate of change in financing costs and the trend of leverage ratio changes. S14: Perform time alignment processing on the implied volatility matrix, the spot volatility vector, and the capital state vector, and use a network delay compensation algorithm to eliminate the transmission time difference between different data sources to generate a synchronous feature dataset.

3. The method according to claim 1, characterized in that, S2 includes: S21: Perform surface gradient calculation on the implied volatility data in the synchronous feature dataset, and generate a volatility gradient vector by solving the second-order mixed partial derivative of the three-dimensional surface in the at-the-money option region. S22: Perform correlation analysis on the volatility gradient vector and the spot volatility data, and use sliding window covariance calculation to perform joint calculation of covariance and standard deviation on the vector data within a fixed time window to generate time series correlation coefficient; S23: Perform distortion verification processing on the absolute value of the volatility gradient vector and the correlation coefficient of the time series, and simultaneously detect the gradient magnitude and negative correlation strength through a dual threshold comparison mechanism to generate a distortion flag indicating the effectiveness of volatility distortion.

4. The method according to claim 1, characterized in that, S3 includes: S31: Perform state space construction processing on the volatility gradient vector and capital state vector, and concatenate the multi-dimensional feature vectors according to the time series to form a unified state representation to generate reinforcement learning input state; S32: Perform behavioral value update processing on the reinforcement learning input state, and use a priority experience replay mechanism to adjust the training weights of historical experience data for the Q-learning algorithm through importance sampling to generate a behavioral value function matrix; S33: Perform probability transformation processing on the behavior value function matrix, and use the softmax function with temperature parameter to perform exponential scaling and normalization calculation on the behavior value to generate a three-dimensional behavior probability distribution including the order cancellation probability.

5. The method according to claim 1, characterized in that, S4 includes: S41: Perform time window slicing on the volatility gradient vector and order cancellation probability, extract continuous time series data segments based on a preset duration, and generate a covariance analysis dataset. S42: Perform statistical correlation calculation on the covariance analysis dataset, solve the eigenvalues ​​of the covariance matrix using joint moment analysis, and generate statistical correlation measures. S43: Perform suppression factor transformation on the statistical correlation measure, and use covariance standardization to convert the correlation coefficient into a suppression strength index in the 0-1 interval to generate a transmission suppression factor.

6. The method according to any one of claims 1-5, characterized in that, S5 includes: S51: Perform feature weighting on the volatility gradient vector, the transmission inhibition factor and the distortion indicator, and fuse multi-dimensional risk features through a dynamic coefficient formula to generate an original risk score; S52: Perform nonlinear transformation on the original risk score, and use an S-curve function to compress the score value to the 0-1 range to eliminate distribution bias and generate a normalized risk score. S53: Perform threshold mapping processing on the normalized risk score, determine the corresponding early warning level range by comparing the preset risk level boundary values, and generate a graded early warning signal.

7. A financial risk prediction system based on artificial intelligence, characterized in that, The system includes: The data clock synchronization module is used to perform clock synchronization processing on real-time acquired ETF option order book data, spot market data and market maker balance sheet data. It uses a timestamp alignment engine to unify heterogeneous data sources to the same time base and generate a synchronization feature dataset. The volatility distortion determination module is used to solve the second-order partial derivative of the at-the-money option volatility surface from the implied volatility data in the synchronous feature dataset to generate a volatility gradient vector. Combined with the spot volatility data in the synchronous feature dataset, the module calculates the correlation coefficient between the volatility gradient vector and the spot volatility within a preset time window to generate a distortion flag indicating the effectiveness of volatility distortion. The market maker behavior modeling module is used to perform reinforcement learning modeling on the volatility gradient vector and the capital status data in the synchronization feature dataset. It simulates the decision-making behavior of market makers in a volatile distortion environment through the Q-learning algorithm and generates a market maker behavior probability distribution that includes order cancellation probability. The suppression factor calculation module is used to perform covariance analysis on the volatility gradient vector and the order cancellation probability in the market maker behavior probability distribution. By calculating the statistical correlation between the change in volatility gradient and the change in order cancellation behavior within a preset time window, a transmission suppression factor is generated. The comprehensive risk warning module is used to calculate the comprehensive risk score of the volatility gradient vector, the transmission inhibition factor and the distortion indicator based on the dynamic coefficient adjustment formula, perform nonlinear transformation processing on the comprehensive risk score, and generate a graded warning signal by mapping through a preset risk threshold range. The graded warning signal is used to indicate the comprehensive risk level of asymmetric transmission risk and liquidity illusion in the ETF market.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images