A financial index volatility forecasting method, system, device and medium

Abstract

The application relates to a financial index volatility prediction method, system, device and medium. The method comprises the following steps: acquiring multi-source factor data of a financial asset index, adjusting the dynamic weight of candidate factors in the multi-source factor data based on market state representation variables to obtain weighted factor data; inputting the weighted factor data into a measurement time sequence sub-model to extract a linear time sequence characteristic vector, combining the vector and the weighted factor data to input an attention mechanism sequence processing network sub-model to obtain a nonlinear feature representation; fusing the two to generate an original volatility prediction value, correcting the original volatility prediction value through a historical residual sequence to calculate a volatility sequence change rate; comparing the volatility sequence change rate with a preset threshold value, if the volatility sequence change rate exceeds the threshold value, loading a matching hyperparameter update sub-model, and generating a financial market risk exposure evaluation result and a transaction signal based on the updated model. The method effectively solves the technical problems of a traditional volatility prediction model, such as fixed factor weight, single feature capture and lack of market state dynamic adaptation capability.

Description

Technical Field

[0001] This invention belongs to the field of risk prediction technology, and in particular relates to a method, system, device and medium for predicting the volatility of financial indices. Background Technology

[0002] Financial index volatility forecasting is a core research area in financial engineering. Its forecasting accuracy directly determines the effectiveness of financial activities such as stock index option pricing, financial market risk measurement, and quantitative trading strategy construction. The volatility forecasting results of core broad-based financial indices such as the CSI 300 are also an important basis for financial institutions and investors to make relevant financial decisions. However, existing volatility prediction schemes for financial indices still suffer from several technical shortcomings that urgently need to be addressed: traditional econometric models such as HAR-RV and GARCH family can only effectively capture the linear time-series changes in volatility, making it difficult to accurately characterize the nonlinear fluctuations and abrupt changes in volatility under extreme market conditions; deep learning models such as LSTM and Transformer, while capable of capturing nonlinear features, suffer from poor model interpretability, making them difficult to implement in financial institutions; and existing multi-factor fusion volatility prediction schemes often employ fixed factor weight designs, failing to adapt to the dynamic changes in financial market volatility. Furthermore, most prediction models lack real-time correction mechanisms for prediction results and the ability to adaptively adjust model hyperparameters after market conditions change, resulting in a significant decrease in the accuracy of volatility predictions as market conditions change. Significant prediction lag exists in different market conditions, including oscillations, trends, and extreme volatility, failing to meet the practical application requirements of accuracy, real-time performance, and dynamic adaptability in financial scenarios for predicting the volatility of financial indices. Summary of the Invention

[0003] Therefore, it is necessary to provide a financial index volatility prediction method, system, device, and medium that can improve the accuracy of volatility prediction results and effectively solve the technical problems of fixed factor weights, single feature capture, and lack of dynamic adaptation to market conditions in traditional volatility prediction models.

[0004] Firstly, this application provides a method for predicting the volatility of a financial index, including:

[0005] Obtain multi-source factor data for financial asset indicators; multi-source factor data includes daily volatility data, high-frequency volatility data, macroeconomic factor data, market sentiment factor data, and trading factor data.

[0006] Based on market state characterization variables, the candidate factors in the multi-source factor data are dynamically weighted to obtain weighted factor data; the market state characterization variables include volatility level and volatility change rate.

[0007] The weighted factor data is input into the econometric time series sub-model to obtain a linear time series feature vector. The linear time series feature vector and the weighted factor data are then input into the sequence processing network sub-model with an attention mechanism to obtain a nonlinear feature representation.

[0008] The original volatility prediction value is generated by fitting linear time series feature vectors and nonlinear feature representations. The original volatility prediction value is then corrected based on the prediction residual sequence of historical data to obtain the volatility sequence change rate.

[0009] If the rate of change in volatility exceeds the preset switching threshold, a combination of hyperparameters matching the market state after the switch is loaded into the sequence processing network sub-model with attention mechanism to generate risk exposure assessment results and trading signals for the financial market.

[0010] In one embodiment, candidate factors in multi-source factor data are dynamically weighted based on market state characterization variables to obtain weighted factor data, including:

[0011] Candidate factor sequences and market state characterization variables are extracted from multi-source factor data; market state characterization variables include volatility level and volatility change rate.

[0012] The current market is divided into state intervals based on volatility levels, resulting in a state label sequence.

[0013] The market state switching speed is calculated based on the volatility change rate, resulting in a switching intensity index.

[0014] By integrating the state label sequence and the switching intensity index, a comprehensive representation vector of market state is constructed.

[0015] For each factor in the candidate factor sequence, calculate the conditional correlation coefficient matrix of each factor under the comprehensive representation vector of market state.

[0016] The dynamic sensitivity of each factor to market conditions is determined based on the conditional correlation coefficient matrix.

[0017] By comparing the dynamic sensitivity with the preset threshold, the weight ratio of the corresponding factor is increased when the dynamic sensitivity is higher than the preset threshold, and the weight ratio of the corresponding factor is decreased when the dynamic sensitivity is lower than the preset threshold, thus obtaining the preliminary weighting coefficient of each factor.

[0018] The initial weighting coefficients are smoothed and corrected based on the comprehensive representation vector of market state to obtain the final dynamic weights of each factor.

[0019] The candidate factor sequences are weighted and fused using the final dynamic weights to obtain weighted factor data.

[0020] In one embodiment, the conditional correlation coefficient matrix is ​​calculated using the following formula:

[0021]

[0022]

[0023] in, Indicates the first A comprehensive representation vector of each factor in market state The conditional correlation coefficient matrix below, Indicates the first Time series data for each factor, This represents the comprehensive representation vector of market state. This represents the covariance calculation operator. This represents the variance calculation operator. express right The coefficient of determination of the fit, express and Effective sample size of time series data. Indicates the first The factor in the th The original value at time, Represents a vector representing the overall market state. For explanatory variables, The first result obtained by performing linear regression Fitted values ​​at time points, , Represents the regression coefficient matrix. Indicates the first The comprehensive representation vector of the market state at any given time. This represents the regression residual term. Indicates the first The mean of the time series of each factor. , This represents the state adaptation adjustment coefficient. , Indicates the first The time-series smoothing coefficients of each factor, It is determined by the time series autocorrelation coefficient of the factor.

[0024] In one embodiment, the weighted factor data is input into the econometric time series sub-model to obtain a linear time series feature vector. The linear time series feature vector and the weighted factor data are then input into a sequence processing network sub-model with an attention mechanism to obtain a nonlinear feature representation, including:

[0025] The weighted factor data is input into the econometric time series sub-model for linear time series feature extraction, resulting in a linear time series feature vector.

[0026] The econometric time series sub-model is a linear feature extraction model built on the HAR-RV model.

[0027] Linear time series eigenvectors are used to characterize the linear time series variation patterns of weighted factor data.

[0028] The linear time series feature vectors and weighted factor data are aligned according to the feature dimensions and then concatenated according to the feature dimensions to obtain concatenated feature data.

[0029] The concatenated feature data is input into a sequence processing network sub-model with an attention mechanism. The sub-model assigns attention weights to capture the nonlinear dependencies between different time steps and various influencing factors in the time series corresponding to the weighted factor data, thus obtaining a nonlinear feature representation.

[0030] In one embodiment, the original volatility prediction value is generated by fitting a linear time-series feature vector and a nonlinear feature representation. The original volatility prediction value is then corrected based on the prediction residual sequence of historical data to obtain the volatility sequence change rate, including:

[0031] By fusing linear time-series feature vectors and nonlinear feature representations, volatility prediction features are obtained.

[0032] Volatility prediction features include information on linear time-series changes and information on nonlinear time-series dependencies.

[0033] Numerical mapping and regression fitting are performed on the volatility prediction characteristics to generate the original volatility prediction values.

[0034] Obtain the prediction residual sequence corresponding to the historical data.

[0035] The original volatility forecast values ​​are rolled over using the prediction residual sequence to obtain the corrected volatility forecast values.

[0036] The difference between the corrected volatility forecast and the corrected volatility forecast at the previous time point is calculated to obtain the volatility series change rate.

[0037] If the rate of change of the volatility sequence exceeds the preset volatility threshold, then higher-order nonlinear components are extracted from the volatility prediction features.

[0038] The higher-order nonlinear components are weighted and fused with the predicted residual sequence to generate the adjusted volatility change rate.

[0039] In one embodiment, if the rate of change in volatility exceeds a preset switching threshold, a hyperparameter combination matching the market state after the switch is loaded into a sequence processing network sub-model with an attention mechanism to generate risk exposure assessment results and trading signals for the financial market, including:

[0040] The volatility change rate is monitored in real time and compared with a preset switching threshold. If the volatility change rate exceeds the switching threshold, it is determined that the current market state has switched between a low volatility state and a high volatility state.

[0041] Retrieve a combination of hyperparameters that matches the market state after the switch from a pre-configured library of multiple hyperparameters.

[0042] Hyperparameter combinations include core parameters such as the number of attention heads, hidden layer dimension, and learning rate.

[0043] The hyperparameter combination is loaded into the sequence processing network sub-model with attention mechanism, and the attention weight matrix and feedforward network parameters of the sub-model are updated.

[0044] The sequence processing network sub-model with attention mechanism after parameter update is used to extract features from the current financial market sequence data corresponding to the volatility change rate, so as to obtain a time series representation vector adapted to the new market state.

[0045] Based on the time series representation vector, the volatility prediction output for the next time step is calculated through regression fitting, generating risk exposure assessment results and trading signals for the financial market.

[0046] Secondly, this application also provides a financial index volatility prediction system, the system comprising:

[0047] The data acquisition module is used to acquire multi-source factor data of financial asset indicators; the multi-source factor data includes daily volatility data, high-frequency volatility data, macroeconomic factor data, market sentiment factor data, and trading factor data.

[0048] The weighting adjustment module is used to dynamically adjust the weights of candidate factors in multi-source factor data based on market state characterization variables to obtain weighted factor data; market state characterization variables include volatility level and volatility change rate.

[0049] The feature extraction module is used to input the weighted factor data into the econometric time series sub-model to obtain a linear time series feature vector, and then input the linear time series feature vector and the weighted factor data into the sequence processing network sub-model with attention mechanism to obtain a nonlinear feature representation.

[0050] The volatility prediction module is used to fuse linear time series feature vectors and nonlinear feature representations to generate original volatility prediction values. The original volatility prediction values ​​are then corrected based on the prediction residual sequence of historical data to obtain the volatility sequence change rate.

[0051] The signal generation module is used to load a combination of hyperparameters that matches the market state after the switch into a sequence processing network sub-model with an attention mechanism if the volatility change rate exceeds a preset switching threshold, thereby generating risk exposure assessment results and trading signals for the financial market.

[0052] Thirdly, 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 method described above.

[0053] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method.

[0054] The aforementioned financial index volatility prediction method, system, computer equipment, and storage medium first acquire multi-source factor data of financial asset indicators, specifically including daily volatility data, high-frequency volatility data, macroeconomic factor data, market sentiment factor data, and trading factor data. Then, based on market state representation variables composed of volatility level and volatility change rate, dynamic weight adjustment processing is performed on candidate factors in the multi-source factor data to obtain weighted factor data. Subsequently, the weighted factor data is input into an econometric time-series sub-model to extract a linear time-series feature vector. Simultaneously, this linear time-series feature vector and the weighted factor data are jointly input into a sequence processing unit with an attention mechanism. The system first extracts a sub-model of the sequence processing network to capture nonlinear feature representations. Then, it fuses linear time-series feature vectors with nonlinear feature representations and fits them to generate the original volatility prediction value. The original volatility prediction value is then corrected by combining the prediction residual sequence of historical data to calculate the volatility sequence change rate. Finally, the volatility sequence change rate is compared with a preset switching threshold. If the volatility sequence change rate exceeds the switching threshold, a hyperparameter combination that matches the market state after the switch is retrieved from the preset hyperparameter library and loaded into the sequence processing network sub-model with attention mechanism. Based on the sub-model with updated parameters, the system generates risk exposure assessment results and trading signals for the financial market. This method achieves dynamic adaptation of factor weights to market conditions through a multi-factor dynamic weight adjustment mechanism. Combining a hybrid modeling approach of an econometric time-series sub-model and a sequence processing network sub-model with an attention mechanism, it can simultaneously capture both linear and nonlinear characteristics of financial index volatility, improving the accuracy of volatility prediction results. By correcting the original predicted values ​​using historical residual sequences, the calculation accuracy of the volatility sequence change rate is further improved. Based on this, the model's hyperparameters are adaptively adjusted, allowing the model to accurately adapt to changes in market conditions. Finally, the risk exposure assessment results and trading signals generated by the updated model can directly provide data support and technical basis for financial activities such as risk management and quantitative trading decisions in the financial market. This effectively solves the technical problems of traditional volatility prediction models, such as fixed factor weights, single feature capture, and lack of dynamic adaptation to market conditions. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 A flowchart of a financial index volatility prediction method provided in an embodiment of the present invention;

[0057] Figure 2 This is a structural block diagram of a financial index volatility prediction system provided in an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] In one embodiment, such as Figure 1 As shown, this application provides a method for predicting the volatility of a financial index, which may include the following steps:

[0060] Step S101: Obtain multi-source factor data for financial asset indicators; multi-source factor data includes daily volatility data, high-frequency volatility data, macroeconomic factor data, market sentiment factor data, and trading factor data.

[0061] Specifically, the system collects raw data of multiple factors corresponding to financial asset indicators through financial data interfaces. This includes daily volatility data of financial indices, high-frequency volatility data in high-frequency intraday dimensions, macroeconomic factor data, market sentiment factor data reflecting market trading sentiment, and trading factor data characterizing market trading behavior. The collected raw data of multiple factors undergoes preprocessing operations such as standardization, spatiotemporal alignment, and outlier removal to eliminate the differences in the dimensions of data from different dimensions and data noise, resulting in multi-source factor data of financial asset indicators with a unified format that can be directly used for subsequent model calculations.

[0062] Step S102: Dynamically adjust the weights of candidate factors in the multi-source factor data based on market state characterization variables to obtain weighted factor data; market state characterization variables include volatility level and volatility change rate.

[0063] Furthermore, candidate factors related to the volatility of the financial index are first extracted from the preprocessed multi-source factor data. At the same time, the volatility level and volatility change rate of the current financial market are calculated and used as market state characterization variables to quantitatively represent the current market volatility state. Based on the preset factor weight adjustment algorithm, the weights of each candidate factor are quantitatively calculated and dynamically allocated with the market state characterization variables as the core adjustment basis. The weight ratio of candidate factors with high correlation to volatility under the current market state is increased, and the weight ratio of candidate factors with low correlation is decreased. This completes the dynamic weight adjustment of candidate factors in the multi-source factor data, and finally outputs the weighted factor data with the weight allocation completed.

[0064] Step S103: Input the weighted factor data into the econometric time series sub-model to obtain the linear time series feature vector, and input the linear time series feature vector and the weighted factor data into the sequence processing network sub-model with attention mechanism to obtain the nonlinear feature representation.

[0065] The weighted factor data is input into a pre-trained econometric time series sub-model. Based on the linear time series feature extraction logic of the econometric time series sub-model, the linear patterns in the weighted factor data are mined and quantified, outputting a linear time series feature vector that characterizes the linear time series changes in financial index volatility. This linear time series feature vector is then aligned with the original weighted factor data in terms of feature dimensions to eliminate mismatches. The dimension-aligned linear time series feature vector and the weighted factor data are then used as joint inputs into a sequence processing network sub-model with an attention mechanism. Through the model's attention weight allocation and sequence feature learning, the nonlinear volatility characteristics and correlation patterns of the financial index volatility are captured, ultimately outputting the corresponding nonlinear feature representation.

[0066] Step S104: The original volatility prediction value is generated by fitting the linear time series feature vector and the nonlinear feature representation. The original volatility prediction value is corrected according to the prediction residual sequence of historical data to obtain the volatility sequence change rate.

[0067] First, the extracted linear time-series feature vectors and nonlinear feature representations are fused to form a comprehensive feature set that integrates both linear and nonlinear volatility features. This comprehensive feature set is then input into a regression fitting module, where the model's numerical mapping and regression calculations are used to fit and generate the original volatility prediction values ​​for the financial index. Simultaneously, historical data of the financial asset index is retrieved, and the corresponding volatility prediction residual sequence is calculated. This sequence represents the set of differences between the historical original volatility prediction values ​​and the actual volatility values. A rolling correction algorithm is then applied to the original volatility prediction values ​​using the historical prediction residual sequence, thus correcting the original volatility prediction values ​​and obtaining the corrected volatility prediction values. Finally, by calculating the difference between the corrected volatility prediction values ​​at adjacent time steps, the volatility sequence change rate, which characterizes the trend of volatility changes, is obtained.

[0068] Step S105: If the volatility change rate exceeds the preset switching threshold, load the hyperparameter combination that matches the market state after the switch into the sequence processing network sub-model with attention mechanism to generate risk exposure assessment results and trading signals for the financial market.

[0069] The calculated volatility sequence change rate is compared with a pre-set market state switching threshold. Based on the comparison result, it is determined whether the current financial market state has switched between low volatility and high volatility. If a market state switch is determined, a hyperparameter combination matching the switched market state is retrieved from a pre-configured and stored hyperparameter library. This hyperparameter combination includes core parameters of the attention-based sequence processing network sub-model, such as the number of attention heads, hidden layer dimension, and learning rate. The retrieved hyperparameter combination is loaded into the attention-based sequence processing network sub-model to update and reconstruct the internal attention weight matrix and feedforward network parameters. The updated sub-model is then used to extract features and accurately predict volatility from the current multi-source factor data of financial asset indicators. Based on the prediction results and pre-set financial risk measurement rules, a risk exposure assessment result for the financial market is generated. Simultaneously, based on volatility characteristics and quantitative trading rules, corresponding financial market trading signals are generated.

[0070] The aforementioned method for predicting financial index volatility first acquires multi-source factor data of financial asset indicators, including daily volatility, high-frequency volatility, macroeconomic data, market sentiment data, and trading factor data. Then, based on market state representation variables composed of volatility levels and rates of change, the candidate factors in the multi-source factor data are dynamically weighted to obtain weighted factor data. This weighted factor data is then input into a time-series econometric sub-model to extract linear time-series feature vectors. These vectors, along with the weighted factor data, are then input into a sequence processing network sub-model with an attention mechanism to obtain nonlinear feature representations. The two are then fused to generate the original volatility prediction value. This value is corrected by combining the prediction residual sequence from historical data, and the volatility sequence rate of change is calculated. The volatility sequence rate of change is compared with a preset switching threshold. If the threshold is exceeded, a matching hyperparameter combination is retrieved from the hyperparameter library and loaded into the aforementioned sub-model. Based on the parameter-updated sub-model, risk exposure assessment results and trading signals for the financial market are generated. This method achieves dynamic adaptation of factor weights to market conditions through multi-factor dynamic weight adjustment. It captures both linear and nonlinear characteristics of volatility using a hybrid modeling approach combining econometrics and deep learning, thereby improving prediction accuracy. Further refinement using historical residual sequences enhances the accuracy of volatility sequence change rate calculations, enabling adaptive adjustment of model hyperparameters. This allows the model to accurately adapt to market state transitions. The resulting evaluation results and trading signals provide data and technical support for financial risk management and quantitative trading decisions, effectively addressing the technical challenges of traditional volatility prediction models, such as fixed factor weights, limited feature capture, and lack of dynamic adaptation to market conditions.

[0071] In one embodiment, dynamically adjusting the weights of candidate factors in multi-source factor data based on market state characterization variables to obtain weighted factor data may include the following steps:

[0072] Step S201: Extract candidate factor sequences and market state characterization variables from multi-source factor data; market state characterization variables include volatility level and volatility change rate.

[0073] Step S202: Divide the current market into state intervals based on volatility levels to obtain a state label sequence.

[0074] Based on the 20th percentile (denoted as σ_low) and 80th percentile (denoted as σ_high) of volatility levels, the market is divided into low volatility ranges (σ<σ_low), medium volatility ranges (σ_low≤σ≤σ_high), and high volatility ranges (σ>σ_high). The trend direction is determined based on the sign of the volatility change rate Δσ (Δσ>0 indicates an upward trend, Δσ=0 indicates stability, and Δσ<0 indicates a downward trend). The combination of these two factors forms four market states: low volatility trending market (σ<σ_low and Δσ>0), medium volatility oscillating market (σ_low≤σ≤σ_high and Δσ≈0), high volatility extreme market (σ>σ_high and Δσ>0), and high volatility declining market (σ>σ_high and Δσ<0).

[0075] Step S203: Calculate the market state switching speed based on the volatility change rate to obtain the switching intensity index.

[0076] Step S204: Integrate the state label sequence and the switching intensity index to construct a comprehensive market state representation vector.

[0077] Step S205: For each factor in the candidate factor sequence, calculate the conditional correlation coefficient matrix of each factor under the comprehensive representation vector of market state.

[0078] Step S206: Determine the dynamic sensitivity between each factor and the market state based on the conditional correlation coefficient matrix.

[0079] Step S207: Compare the dynamic sensitivity with the preset threshold. When the dynamic sensitivity is higher than the preset threshold, increase the weight ratio of the corresponding factor. When the dynamic sensitivity is lower than the preset threshold, decrease the weight ratio of the corresponding factor to obtain the preliminary weighting coefficient of each factor.

[0080] Step S208: Based on the comprehensive representation vector of market state, smooth and correct each preliminary weighting coefficient to obtain the final dynamic weight of each factor.

[0081]

[0082] in, Indicates the first The final dynamic weights of each factor Indicates the first The initial weighting coefficients of the factors, This represents the comprehensive representation vector of market state. Indicates the first The fit coefficients of the comprehensive representation vector of each factor and the market state are obtained by regression analysis of the factor time series and Z. This represents the total number of factors in the candidate factor sequence. This represents the smoothing adjustment coefficient, with a value range of [value missing]. This is used to balance the initial weighting coefficients with the weighting percentages after adapting to market conditions.

[0083] Step S209: The candidate factor sequences are weighted and fused using the final dynamic weights to obtain weighted factor data.

[0084] Specifically, candidate factor sequences and market state characterization variables are extracted from multi-source factor data of financial asset indicators. These market state characterization variables include volatility level and volatility change rate. Based on the numerical characteristics of volatility level, the current financial market's corresponding volatility state interval is divided, generating a state label sequence representing the market's basic volatility state. Simultaneously, based on the numerical value of volatility change rate, the state switching speed of the financial market is calculated to obtain a switching intensity index that quantifies the degree of market volatility state switching. Feature fusion processing is performed on the state label sequence and switching intensity index to construct a comprehensive market state characterization vector that comprehensively represents the overall volatility state of the financial market. For each factor in the candidate factor sequence, individual calculations are performed... The conditional correlation coefficient matrix of factors under the comprehensive market state representation vector is used as the quantitative basis to determine the dynamic sensitivity of each factor to the current market state. The dynamic sensitivity of each factor is compared with a preset threshold. Factors with dynamic sensitivity higher than the preset threshold are weighted, while factors with dynamic sensitivity lower than the preset threshold are weighted, thus obtaining the preliminary weighting coefficients of each factor. Based on the comprehensive market state representation vector, the preliminary weighting coefficients of each factor are smoothed to eliminate the impact of weight abrupt changes, resulting in the final dynamic weights of each factor. The final dynamic weights of each factor are used to perform weighted fusion calculations on the candidate factor sequence, outputting the weighted factor data with completed weight allocation.

[0085] This embodiment achieves accurate quantification of the overall volatility of the financial market by extracting market state characterization variables from multi-source factor data and constructing a comprehensive market state characterization vector. Based on this, factor weights are adjusted to ensure a deep fit between factor weight allocation and the current market state, solving the technical problem of fixed factor weights in traditional multi-factor models that cannot adapt to dynamic changes in market state. By calculating the conditional correlation coefficient matrix, the dynamic sensitivity of factors to market state is determined, providing a quantitative basis for factor weight adjustment and ensuring the objectivity and rationality of weight allocation. Furthermore, the initial weighting coefficients are smoothed and corrected using the comprehensive market state characterization vector, further improving the stability of factor weights. The final weighted factor data highlights the role of factors highly correlated with the current market state and weakens the interference of low-correlation factors, providing factor data that fits the actual market situation for subsequent volatility feature extraction and prediction model input, effectively improving the accuracy and adaptability of subsequent financial index volatility prediction.

[0086] In one embodiment, the conditional correlation coefficient matrix can be calculated using the following formula:

[0087]

[0088]

[0089] in, Indicates the first A comprehensive representation vector of each factor in market state The conditional correlation coefficient matrix below, Indicates the first Time series data for each factor, This represents the comprehensive representation vector of market state. This represents the covariance calculation operator. This represents the variance calculation operator. express right The coefficient of determination of the fit, express and Effective sample size of time series data. Indicates the first The factor in the th The original value at time, Represents a vector representing the overall market state. For explanatory variables, The first result obtained by performing linear regression Fitted values ​​at time points, , Represents the regression coefficient matrix. Indicates the first The comprehensive representation vector of the market state at any given time. This represents the regression residual term. Indicates the first The mean of the time series of each factor. , This represents the state adaptation adjustment coefficient. , Indicates the first The time-series smoothing coefficients of each factor, It is determined by the time series autocorrelation coefficient of the factor.

[0090] This embodiment calculates the conditional correlation coefficient matrix using the aforementioned formula, accurately quantifying the conditional correlation between individual factors and the comprehensive representation vector of market state. This provides an objective and quantitative mathematical basis for determining the dynamic sensitivity between each factor and the market state, giving the dynamic sensitivity analysis rigorous numerical support. The introduction of the fitting determination coefficient in the formula effectively eliminates the interference of linear fitting error between the factor and the comprehensive representation vector of market state on the quantification of the correlation relationship, improving the accuracy of the correlation coefficient calculation results. The synergistic integration of the state adaptation adjustment coefficient and the time series smoothing coefficient not only achieves the adaptation adjustment of the coefficient to the market state, but also smooths the calculation results by combining the time series autocorrelation characteristics of the factor itself, avoiding calculation deviations caused by extreme values. At the same time, the calculation logic of this formula fits the time series data characteristics of financial factors, ensuring the rationality and effectiveness of the conditional correlation coefficient matrix calculation results. Based on this, the factor dynamic sensitivity determination and subsequent weight adjustment can make the allocation of factor weights highly consistent with the actual volatility of the current financial market, providing accurate quantitative support for the generation of the final weighted factor data and improving the scientificity and rationality of the multi-factor dynamic weight adjustment process.

[0091] In one embodiment, the weighted factor data is input into the econometric time series sub-model to obtain a linear time series feature vector. The linear time series feature vector and the weighted factor data are then input into the sequence processing network sub-model with an attention mechanism to obtain a nonlinear feature representation. This may include the following steps:

[0092] Step S301: Input the weighted factor data into the econometric time series sub-model to extract linear time series features and obtain the linear time series feature vector.

[0093] Preferably, the econometric time series sub-model is a linear feature extraction model built based on the HAR-RV model. The linear time series feature vector is used to characterize the linear change pattern of the weighted factor data over time.

[0094] Specifically, weighted factor data is used as input data and imported into a pre-trained and validated convergent econometric time-series sub-model for linear time-series feature extraction. This econometric time-series sub-model is a linear feature extraction model built on the HAR-RV model, optimized for the feature distribution characteristics of financial time-series data, and capable of analyzing the time-series linear patterns of multi-dimensional weighted factor data. During the extraction process, the weighted factor data is first regularized according to the time-series dimension to maintain the temporal correlation and dimensional consistency of the data before being input into the econometric time-series sub-model. The model relies on the core logic of time-series aggregation and linear regression of the HAR-RV model to quantitatively mine and extract features from the linear correlation patterns and time-series change trends of the weighted factor data under different time windows. Through the model's internal computational layer, feature mapping and dimensionality reduction are performed on the input weighted factor data, ultimately outputting a linear time-series feature vector that accurately represents the overall time-series linear change pattern corresponding to the weighted factor data.

[0095] Step S302: Align the linear time series feature vector and weighted factor data according to the feature dimension and then perform feature dimension concatenation to obtain concatenated feature data.

[0096] Step S303: Input the spliced ​​feature data into the sequence processing network sub-model with attention mechanism, and use the sub-model to assign attention weights to capture the nonlinear dependency relationship between different time steps and various influencing factors in the time series corresponding to the weighted factor data, so as to obtain the nonlinear feature representation.

[0097] Specifically, the weighted factor data is input into a pre-constructed econometric time series sub-model to perform linear time series feature extraction, outputting a linear time series feature vector. This econometric time series sub-model is a linear feature extraction model based on the HAR-RV model, and the extracted linear time series feature vector is used to characterize the linear time series variation law corresponding to the weighted factor data. Subsequently, feature dimension alignment is performed on the linear time series feature vector and the weighted factor data to eliminate the mismatch between them in feature dimensions. Based on the completion of dimension alignment, the feature dimensions of the two types of data are concatenated to generate concatenated feature data. Finally, the concatenated feature data is input into a sequence processing network sub-model with an attention mechanism. Through the attention weight allocation mechanism of this sub-model, the nonlinear dependencies between different time steps and various influencing factors in the time series corresponding to the weighted factor data are mined and quantified, and finally the corresponding nonlinear feature representation is output.

[0098] This embodiment utilizes a dedicated econometric time-series sub-model built on the HAR-RV model to extract linear features. This accurately uncovers the linear variation patterns in the weighted factor data, ensuring that the extracted linear features match the linear time-series characteristics of financial index volatility. The feature dimension alignment effectively avoids splicing errors caused by dimensional inconsistencies between the linear time-series feature vectors and the weighted factor data, improving the completeness and effectiveness of the spliced ​​feature data. Leveraging the attention weight allocation capability of the sequence processing network sub-model with an attention mechanism, it can specifically capture the nonlinear dependencies between different time steps and various influencing factors in the time-series sequence, compensating for the technical shortcomings of econometric models in capturing nonlinear features. This achieves precise dimensional extraction of linear and nonlinear features from the weighted factor data, providing comprehensive and accurate feature data support for the subsequent fusion of volatility prediction features and the generation of original volatility prediction values. It solves the technical problem that traditional single models cannot simultaneously capture both linear and nonlinear features, thus improving the accuracy of subsequent financial index volatility predictions at the feature level.

[0099] In one embodiment, the original volatility prediction value is generated by fitting linear time-series feature vectors and nonlinear feature representations. The original volatility prediction value is then corrected based on the prediction residual sequence of historical data to obtain the volatility sequence change rate. This may include the following steps:

[0100] Step S401: Fuse the linear time series feature vector and the nonlinear feature representation to obtain the volatility prediction feature.

[0101] Preferably, the volatility prediction features include time-series linear change information and time-series nonlinear dependency information.

[0102] Step S402: Perform numerical mapping and regression fitting on the volatility prediction features to generate the original volatility prediction value.

[0103] Step S403: Obtain the prediction residual sequence corresponding to the historical data.

[0104] Step S404: The original volatility prediction value is rolled over using the prediction residual sequence to obtain the corrected volatility prediction value.

[0105]

[0106] in, Indicates the first Volatility forecast after time adjustment Indicates the first Original volatility forecast value at time point This indicates the preset rolling correction window length (take a positive integer to adapt to the characteristics of financial time series, such as 60 or 90 trading days). Indicates the first Volatility prediction residual at time point , Indicates the first Actual observed value of volatility at any given time.

[0107] Schematic, the predicted residual sequence is the set of differences between the original volatility predictions and the actual volatility observations for each historical time step of the financial index, and has already undergone time dimension alignment and standardization preprocessing with the original volatility predictions. During the correction process, a rolling correction window length adapted to the characteristics of the financial time series is preset. The predicted residual sequence is then processed by sliding the window according to the chronological order of the time series, extracting the residual subsequence that matches the original volatility prediction value at the current time step in the time dimension. Finally, a preset residual correction algorithm is used to quantify the statistical characteristics of this residual subsequence. The system applies a feature to the original volatility forecast value at the corresponding time step, thereby eliminating systematic and random biases generated during historical forecasting. During the rolling correction process, independent window matching and residual correction operations are performed on the original volatility forecast value at each time step to maintain the continuity of the volatility forecast value time series and avoid cross-time step correction interference. Finally, the corrected volatility forecast value after completing the rolling correction process for each time step is output. This value effectively reduces the bias of the original volatility forecast value, improves the accuracy of the volatility forecast result, and maintains temporal consistency with the original volatility forecast value.

[0108] Step S405: Calculate the difference between the corrected volatility forecast and the corrected volatility forecast at the previous moment to obtain the volatility series change rate.

[0109] Step S406: If the rate of change of the volatility sequence exceeds the preset volatility threshold, then extract the higher-order nonlinear component from the volatility prediction features.

[0110] Step S407: The higher-order nonlinear components and the predicted residual sequence are weighted and fused to generate the adjusted volatility change rate.

[0111] Specifically, a volatility prediction feature is generated by fusing linear time-series feature vectors and nonlinear feature representations. This volatility prediction feature integrates time-series linear change information and time-series nonlinear dependency information characterizing the volatility pattern of financial indices. Based on a pre-defined numerical mapping method, the volatility prediction feature is transformed into a feature space. Then, a regression fitting algorithm is used to calculate the original volatility prediction value of the financial index. Simultaneously, historical data of financial asset indicators are retrieved to calculate the corresponding volatility prediction residual sequence. This sequence is the set of differences between the historical original volatility prediction values ​​and the actual volatility values. A rolling correction algorithm is then used to refine the prediction residual sequence. The system is used to perform rolling corrections on the original volatility predictions and output the corrected volatility predictions. By calculating the difference between the corrected volatility predictions at the current time and the corrected volatility predictions at the previous time, the volatility sequence change rate, which can quantify the trend of financial index volatility, is obtained. The volatility sequence change rate is compared with a preset volatility threshold. If the volatility sequence change rate exceeds the preset volatility threshold, a high-order nonlinear component that can characterize extreme volatility is extracted from the volatility prediction features. Finally, a weighted fusion algorithm is used to fuse the extracted high-order nonlinear component with the prediction residual sequence to generate the adjusted volatility change rate.

[0112] This embodiment obtains volatility prediction features containing dual-dimensional volatility information by fusing linear and nonlinear features, providing comprehensive feature data support for volatility prediction. This allows the generation of the original volatility prediction value to take into account both the linear and nonlinear characteristics of financial index fluctuations. Rolling corrections are performed on the original volatility prediction value using the prediction residual sequence of historical data, effectively offsetting the impact of cumulative prediction bias and improving the calculation accuracy of the corrected volatility prediction value. The volatility sequence change rate obtained through differential calculation can accurately quantify the volatility trend of the financial index. For volatility sequence change rates exceeding a preset threshold, higher-order nonlinear components are extracted and weighted and fused with the prediction residual sequence to generate an adjusted volatility change rate. This compensates for the shortcomings of conventional correction methods in handling extreme market conditions, making the calculation results of the volatility change rate adaptable to extreme volatility scenarios in the financial market. The overall process, through multi-stage feature fusion, prediction, and correction optimization, significantly improves the calculation accuracy and scenario adaptability of volatility prediction values ​​and volatility change rates, effectively solving the technical problems of prediction lag and insufficient accuracy of traditional prediction methods under extreme market conditions.

[0113] In one embodiment, if the volatility change rate exceeds a preset switching threshold, a hyperparameter combination matching the market state after the switch is loaded into a sequence processing network sub-model with an attention mechanism to generate risk exposure assessment results and trading signals for the financial market. This may include the following steps:

[0114] Step S501: Monitor the volatility change rate in real time and compare it with a preset switching threshold. If the volatility change rate exceeds the switching threshold, it is determined that the current market state has switched between a low volatility state and a high volatility state.

[0115] Preferably, the market conditions include four categories: low-volatility trending market, medium-volatility oscillating market, high-volatility extreme market, and high-volatility declining market.

[0116] The volatility change rate is monitored in real time. Based on the volatility level quantiles (20th percentile, 80th percentile), low / medium / high volatility ranges are divided. Combined with the trend direction of the volatility change rate (rising / falling / stable), the current market state is divided into four categories: low volatility trend market, medium volatility oscillating market, high volatility extreme market, and high volatility pullback market. The volatility change rate is compared with a preset switching threshold. If the volatility change rate meets the preset state switching judgment condition (if the volatility change rate exceeds the switching threshold for three consecutive time steps, and the volatility level remains within the target state range during the period, the switching threshold includes an upward switching threshold and a downward switching threshold, and the upward switching threshold is greater than the downward switching threshold), then the current market state is determined to have switched.

[0117] The switching thresholds for the four market states are as follows: Δσ1 = 0.005 for low-volatility trending market → medium-volatility oscillating market; Δσ2 = 0.01 for medium-volatility oscillating market → high-volatility extreme market; Δσ3 = -0.008 for high-volatility extreme market → high-volatility pullback market; Δσ4 = -0.005 for high-volatility pullback market → medium-volatility oscillating market; and Δσ5 = -0.003 for medium-volatility oscillating market → low-volatility trending market.

[0118] State switching determination requires meeting two conditions: ① The volatility change rate exceeds the corresponding switching threshold for three consecutive trading days (e.g., for a medium-volatility oscillating market to a high-volatility extreme market, Δσ>0.01 for three consecutive trading days); ② The volatility level remains within the target state range for three consecutive trading days (e.g., for the above-mentioned switching, σ>σ_high for three consecutive trading days). At the same time, the difference between the upward and downward switching thresholds is set to 0.002-0.005 (e.g., the upward threshold Δσ2=0.01 for a medium-volatility oscillating market to a high-volatility extreme market, and the downward threshold Δσ4=-0.005 for a high-volatility extreme market to a medium-volatility oscillating market, with a difference of 0.005), forming a buffer zone to avoid frequent switching caused by market noise.

[0119] Step S502: Retrieve a combination of hyperparameters that matches the market state after the switch from a pre-configured library of multiple hyperparameters.

[0120] Hyperparameter combinations include core parameters such as the number of attention heads, hidden layer dimension, and learning rate.

[0121] The hyperparameter library contains exclusive hyperparameter combinations corresponding to four market states: Low volatility trend market is adapted to 4 attention heads, 128 hidden layer dimensions, and a learning rate of 0.001; Medium volatility oscillating market is adapted to 6 attention heads, 256 hidden layer dimensions, and a learning rate of 0.0005; High volatility extreme market is adapted to 8 attention heads, 512 hidden layer dimensions, and a learning rate of 0.0001; High volatility pullback market is adapted to 6 attention heads, 256 hidden layer dimensions, and a learning rate of 0.0003.

[0122] Step S503: Load the hyperparameter combination into the sequence processing network sub-model with attention mechanism, and update the attention weight matrix and feedforward network parameters of the sub-model.

[0123] Preferably, the hyperparameter combination retrieved from the hyperparameter library and matching the market state after the switch, along with the original sequence processing network sub-model with attention mechanism, are used as input. First, according to the parameter hierarchy mapping relationship of the sub-model, the core parameters contained in the hyperparameter combination, such as the number of attention heads, hidden layer dimension, and learning rate, are accurately matched to the corresponding parameter configuration layer of the sub-model, thus completing the loading of the hyperparameter combination. Based on the loaded new hyperparameters, the attention weight matrix of the sub-model is reconstructed and updated. The matrix dimension is adjusted according to the new number of attention heads, and the attention weight allocation coefficient is recalculated to adapt to the feature focus under the new market state. At the same time, the feedforward network parameters of the sub-model are updated synchronously. According to the new hidden layer dimension and learning rate, the weights and bias terms of each layer of the feedforward network are adjusted to complete the iterative optimization of the feedforward network parameters, ultimately achieving the overall update of the core parameters inside the sequence processing network sub-model with attention mechanism.

[0124] Step S504: The sequence processing network sub-model with attention mechanism after parameter update is used to extract features from the current financial market sequence data corresponding to the volatility change rate, so as to obtain a time series representation vector adapted to the new market state.

[0125] Using the updated sequence processing network sub-model with attention mechanism and the current financial market sequence data on which the volatility change rate is based as core inputs, the current financial market sequence data is first preprocessed, performing time dimension alignment and feature dimension regularization operations to eliminate data dimensional bias and time series disorder, ensuring that the data dimensions match the input dimension requirements of the updated sub-model. The preprocessed current financial market sequence data is then input into the updated sub-model. The model, relying on the updated attention weight matrix, requantifies and reassigns the importance of information at different time steps and different feature dimensions in the financial market sequence data, accurately capturing the core temporal characteristics of financial index fluctuations under the new market conditions. Then, the updated feedforward network performs nonlinear transformation and feature dimension fusion on the initially extracted features, completing the in-depth mining and quantitative representation of features, and finally outputting a temporal representation vector adapted to the new market conditions. This vector can accurately characterize the overall temporal change pattern and volatility characteristics of the financial market sequence data under the new market conditions.

[0126] Step S505: Calculate the volatility prediction output for the next time step based on the time series representation vector through regression fitting, and generate risk exposure assessment results and trading signals for the financial market.

[0127] In a schematic manner, using a time-series representation vector adapted to the new market state as input, the vector is first imported into a pre-defined regression fitting module. Based on a regression fitting algorithm adapted to the volatility characteristics of financial indices, the module performs feature space numerical mapping and future volatility trend deduction on the time-series representation vector. The model calculates the predicted volatility output for the next time step, which includes specific volatility prediction values ​​and corresponding confidence intervals, ensuring the quantification and completeness of the prediction results. Based on this volatility prediction output, combined with pre-defined financial market risk measurement rules, quantitative indicators such as the degree of risk exposure and risk level in the financial market are obtained through quantitative calculation, generating standardized financial market risk exposure assessment results. Simultaneously, based on the volatility prediction output and pre-defined quantitative trading rules, combined with the volatility characteristics of the new market state, the market trading trend is determined and corresponding trading decision logic is triggered, generating financial market trading signals containing specific instructions such as opening positions, closing positions, and position adjustments. The generated risk exposure assessment results and trading signals can directly provide accurate quantitative data support and decision-making basis for financial market risk management, quantitative trading decisions, and other financial activities.

[0128] Specifically, the volatility change rate is monitored in real time, and the monitored volatility change rate is compared with a pre-set market state switching threshold. If the volatility change rate exceeds the switching threshold, it is determined that the current financial market state has switched between a low volatility state and a high volatility state. From a pre-configured and stored hyperparameter library, a hyperparameter combination matching the determined market state after the switch is retrieved. This hyperparameter combination includes core parameters of the attention mechanism sequence processing network sub-model, such as the number of attention heads, hidden layer dimension, and learning rate. The retrieved hyperparameter combination is loaded into the attention mechanism sequence processing network sub-model to update and reconstruct the internal attention weight matrix and feedforward network parameters of the sub-model. Using the parameter-updated attention mechanism sequence processing network sub-model, feature extraction processing is performed on the current financial market sequence data on which the volatility change rate is based to obtain a time series representation vector adapted to the new market state. This time series representation vector is input into the regression fitting module for calculation to obtain the volatility prediction output for the next time step. Based on this volatility prediction output, the corresponding financial market risk exposure assessment results and trading signals are generated.

[0129] This embodiment achieves accurate and timely determination of the transition between low and high volatility states in the financial market by real-time monitoring of the rate of change of volatility and threshold comparison, providing an objective and quantitative trigger basis for the dynamic adaptation of the model. By retrieving hyperparameter combinations that match the new market state from a preset hyperparameter library and loading them to update the model, the core parameters of the sequence processing network sub-model with attention mechanism can adaptively adjust with the market state transition, solving the technical problem of fixed hyperparameters in traditional models that cannot adapt to changes in market volatility. The time-series representation vector extracted from the current financial market sequence data by the updated model can accurately match the volatility characteristics of the new market state. Based on this, the volatility prediction output for the next time step obtained through regression fitting has higher accuracy and scenario adaptability. Finally, the financial market risk exposure assessment results and trading signals generated based on this prediction output can accurately match the current market volatility state, providing more targeted technical support for risk management, quantitative trading decisions, and other activities in the financial market, effectively improving the scientific nature of risk assessment and the effectiveness of trading signals.

[0130] In one embodiment, such as Figure 2 As shown, this application also provides a financial index volatility prediction system, which may include:

[0131] The data acquisition module 601 is used to acquire multi-source factor data of financial asset indicators; the multi-source factor data includes daily volatility data, high-frequency volatility data, macroeconomic factor data, market sentiment factor data, and trading factor data.

[0132] The weight adjustment module 602 is used to dynamically adjust the weights of candidate factors in multi-source factor data based on market state characterization variables to obtain weighted factor data; the market state characterization variables include volatility level and volatility change rate.

[0133] The feature extraction module 603 is used to input the weighted factor data into the econometric time series sub-model to obtain a linear time series feature vector, and input the linear time series feature vector and the weighted factor data into the sequence processing network sub-model with attention mechanism to obtain a nonlinear feature representation.

[0134] The volatility prediction module 604 is used to fuse linear time series feature vectors and nonlinear feature representations to generate the original volatility prediction value. The original volatility prediction value is corrected based on the prediction residual sequence of historical data to obtain the volatility sequence change rate.

[0135] The signal generation module 605 is used to load a combination of hyperparameters that matches the market state after the switch into a sequence processing network sub-model with an attention mechanism if the volatility change rate exceeds a preset switching threshold, thereby generating risk exposure assessment results and trading signals for the financial market.

[0136] The aforementioned financial index volatility prediction system includes a data acquisition module, a weight adjustment module, a feature extraction module, a volatility prediction module, and a signal generation module. Each module sequentially completes data flow and collaborative processing. Specifically: the data acquisition module collects multi-source factor data of financial asset indicators, including daily volatility data, high-frequency volatility data, macroeconomic factor data, market sentiment factor data, and trading factor data, providing basic data input for subsequent module processing; the weight adjustment module takes the multi-source factor data output by the data acquisition module as input, and performs dynamic weight adjustment processing on candidate factors in the multi-source factor data based on market state representation variables including volatility level and volatility change rate, outputting weighted factor data; the feature extraction module receives the weighted factor data output by the weight adjustment module, inputs it into the econometric time series sub-model to extract linear time series feature vectors, and then... The linear time-series feature vector and weighted factor data are input into the sequence processing network sub-model with attention mechanism to capture and output nonlinear feature representation. The volatility prediction module takes the linear time-series feature vector and nonlinear feature representation output by the feature extraction module as input, fuses them and fits them to generate the original volatility prediction value, and then combines the prediction residual sequence of historical data to correct the original volatility prediction value, calculates and outputs the volatility sequence change rate. The signal generation module receives the volatility sequence change rate output by the volatility prediction module, compares it with the preset switching threshold, and if the volatility change rate exceeds the switching threshold, it retrieves the hyperparameter combination that matches the market state after the switch from the preset hyperparameter library, loads the hyperparameter combination into the sequence processing network sub-model with attention mechanism, and generates and outputs the risk exposure assessment results and trading signals of the financial market based on the parameter-updated sub-model.

[0137] Each module in this embodiment addresses specific technical pain points in traditional volatility prediction. The dynamic weight adjustment mechanism of the weight adjustment module allows factor weights to adapt to changes in market state variables, solving the problem of fixed factor weights in traditional multi-factor models and improving the fit of factor data to market conditions. The feature extraction module extracts linear and nonlinear features through dual models, achieving accurate capture of different types of features in financial index volatility and overcoming the limitations of single-model feature extraction. The residual correction processing of the volatility prediction module effectively reduces the deviation of the original volatility prediction values, improves the calculation accuracy of the volatility sequence change rate, and provides a reliable quantitative basis for market state judgment. The signal generation module achieves adaptive loading and updating of model hyperparameters based on the threshold judgment of volatility change rate, allowing the model to dynamically adjust with market state changes, ensuring the relevance and effectiveness of subsequent risk exposure assessment results and trading signals.

[0138] 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.

[0139] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a financial index volatility prediction method as described above.

[0140] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0141] 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.

[0142] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for predicting the volatility of a financial index, characterized in that, The method includes: Obtain multi-source factor data for financial asset indicators; the multi-source factor data includes daily volatility data, high-frequency volatility data, macroeconomic factor data, market sentiment factor data, and trading factor data; Based on market state characterization variables, the candidate factors in the multi-source factor data are dynamically weighted to obtain weighted factor data; the market state characterization variables include volatility level and volatility change rate. The weighted factor data is input into the econometric time series sub-model to obtain a linear time series feature vector. The linear time series feature vector and the weighted factor data are then input into the sequence processing network sub-model with attention mechanism to obtain a nonlinear feature representation. The original volatility prediction value is generated by fitting the linear time series feature vector and the nonlinear feature representation together. The original volatility prediction value is then corrected based on the prediction residual sequence of historical data to obtain the volatility sequence change rate. If the volatility change rate exceeds a preset switching threshold, a hyperparameter combination matching the market state after the switch is loaded into the sequence processing network sub-model with attention mechanism to generate risk exposure assessment results and trading signals for the financial market.

2. The method according to claim 1, characterized in that, The dynamic weighting adjustment of candidate factors in the multi-source factor data based on market state characterization variables to obtain weighted factor data includes: Candidate factor sequences and market state characterization variables are extracted from the multi-source factor data; the market state characterization variables include volatility level and volatility change rate. Based on the volatility level, the current market is divided into state intervals to obtain a state label sequence; The market state switching speed is calculated based on the volatility change rate to obtain the switching intensity index; By integrating the state label sequence and the switching intensity index, a comprehensive market state representation vector is constructed. For each factor in the candidate factor sequence, calculate the conditional correlation coefficient matrix of each factor under the comprehensive representation vector of the market state; The dynamic sensitivity between each factor and the market state is determined based on the conditional correlation coefficient matrix. By comparing the dynamic sensitivity with a preset threshold, the weight ratio of the corresponding factor is increased when the dynamic sensitivity is higher than the preset threshold, and the weight ratio of the corresponding factor is decreased when the dynamic sensitivity is lower than the preset threshold, thus obtaining the preliminary weighting coefficient of each factor. Based on the comprehensive market state representation vector, the initial weighting coefficients are smoothed and corrected to obtain the final dynamic weights of each factor. The candidate factor sequences are weighted and fused using the final dynamic weights to obtain weighted factor data.

3. The method according to claim 2, characterized in that, The conditional correlation coefficient matrix is ​​calculated using the following formula: in, Indicates the first A comprehensive representation vector of each factor in market state The conditional correlation coefficient matrix below, Indicates the first Time series data for each factor, This represents the comprehensive representation vector of market state. This represents the covariance calculation operator. This represents the variance calculation operator. express right The coefficient of determination of the fit, express and Effective sample size of time series data. Indicates the first The factor in the th The original value at time, Represents a vector representing the overall market state. For explanatory variables, The first result obtained by performing linear regression Fitted values ​​at time points, , Represents the regression coefficient matrix. Indicates the first The comprehensive representation vector of the market state at any given time. This represents the regression residual term. Indicates the first The mean of the time series of each factor. , This represents the state adaptation adjustment coefficient. , Indicates the first The time-series smoothing coefficients of each factor, It is determined by the time series autocorrelation coefficient of the factor.

4. The method according to claim 1, characterized in that, The process involves inputting the weighted factor data into a time-series sub-model to obtain a linear time-series feature vector, and then inputting the linear time-series feature vector and the weighted factor data into a sequence processing network sub-model with an attention mechanism to obtain a nonlinear feature representation, including: The weighted factor data is input into the econometric time series sub-model for linear time series feature extraction to obtain a linear time series feature vector. The time series measurement sub-model is a linear feature extraction model built based on the HAR-RV model; The linear time-series feature vector is used to characterize the time-series linear change pattern corresponding to the weighted factor data; The linear time-series feature vector and the weighted factor data are aligned according to the feature dimension and then concatenated according to the feature dimension to obtain concatenated feature data. The spliced ​​feature data is input into a sequence processing network sub-model with an attention mechanism. The sub-model assigns attention weights to capture the nonlinear dependencies between different time steps and various influencing factors in the time series corresponding to the weighted factor data, thereby obtaining a nonlinear feature representation.

5. The method according to claim 1, characterized in that, The process of fusing the linear time-series feature vector and the nonlinear feature representation to fit and generate the original volatility prediction value, and then correcting the original volatility prediction value based on the prediction residual sequence of historical data to obtain the volatility sequence change rate, includes: By fusing the linear time-series feature vector and the nonlinear feature representation, volatility prediction features are obtained. The volatility prediction features include time-series linear variation information and time-series nonlinear dependency information; Numerical mapping and regression fitting are performed on the volatility prediction features to generate the original volatility prediction values. Obtain the prediction residual sequence corresponding to historical data; The original volatility prediction value is rolled over and corrected using the predicted residual sequence to obtain the corrected volatility prediction value. The difference between the corrected volatility forecast and the corrected volatility forecast at the previous time point is calculated to obtain the volatility series change rate. If the rate of change of the volatility sequence exceeds a preset volatility threshold, then a higher-order nonlinear component is extracted from the volatility prediction feature; The higher-order nonlinear components are weighted and fused with the predicted residual sequence to generate an adjusted volatility change rate.

6. The method according to claim 1, characterized in that, If the volatility change rate exceeds a preset switching threshold, a hyperparameter combination matching the market state after the switch is loaded into the sequence processing network sub-model with attention mechanism to generate risk exposure assessment results and trading signals for the financial market, including: The volatility change rate is monitored in real time and compared with a preset switching threshold. If the volatility change rate exceeds the switching threshold, it is determined that the current market state has switched between a low volatility state and a high volatility state. Retrieve a combination of hyperparameters that matches the market state after the switch from a pre-configured library of hyperparameters; The hyperparameter combination includes core parameters such as the number of attention heads, hidden layer dimension, and learning rate. The hyperparameter combination is loaded into the sequence processing network sub-model with attention mechanism, and the attention weight matrix and feedforward network parameters of the sub-model are updated. The sequence processing network sub-model with attention mechanism after parameter update is used to extract features from the current financial market sequence data corresponding to the volatility change rate, so as to obtain a time series representation vector adapted to the new market state. Based on the time series representation vector, the volatility prediction output for the next time step is calculated through regression fitting, generating risk exposure assessment results and trading signals for the financial market.

7. A financial index volatility prediction system, characterized in that, The system includes: The data acquisition module is used to acquire multi-source factor data of financial asset indicators; the multi-source factor data includes daily volatility data, high-frequency volatility data, macroeconomic factor data, market sentiment factor data, and trading factor data; The weighting adjustment module is used to dynamically adjust the weights of candidate factors in the multi-source factor data based on market state characterization variables to obtain weighted factor data; the market state characterization variables include volatility level and volatility change rate. The feature extraction module is used to input the weighted factor data into the econometric time series sub-model to obtain a linear time series feature vector, and input the linear time series feature vector and the weighted factor data into the sequence processing network sub-model with attention mechanism to obtain a nonlinear feature representation; The volatility prediction module is used to fuse the linear time series feature vector and the nonlinear feature representation to generate an original volatility prediction value, and to correct the original volatility prediction value based on the prediction residual sequence of historical data to obtain the volatility sequence change rate. The signal generation module is used to load a combination of hyperparameters matching the market state after the switch into the sequence processing network sub-model with attention mechanism if the volatility change rate exceeds a preset switching threshold, thereby generating risk exposure assessment results and trading signals for the financial 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 steps of the method according to 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 steps of the method according to any one of claims 1 to 6.

Images