Financial risk prediction system and method based on deep learning

Abstract

The invention discloses a financial risk prediction method based on deep learning, and the method comprises the following steps: S1, collecting original financial data, carrying out the preprocessing of the original data, and extracting the deep features of the preprocessed original financial data through a graph convolution network, processing the deep feature by using an adaptive weighted nonlinear wavelet decomposition technology to obtain a long-term trend component and a short-term fluctuation component; and S2, optimizing the long-term trend component and the short-term fluctuation component by using a dynamic fuzzy system model, inputting a quantum perception enhancement converter model for training to obtain a comprehensive financial risk, and generating a detailed risk assessment report according to the comprehensive financial risk. According to the method, the precision and stability of data analysis can be improved, more accurate, comprehensive and stable financial risk assessment is provided, and the effectiveness of risk management and decision making is improved. The invention further discloses a financial risk prediction system based on deep learning.

Description

Technical Field

[0001] The present invention relates to the field of financial risk prediction, and in particular to a financial risk prediction system and method based on deep learning. Background Art

[0002] Financial risk prediction has important application value in the financial market. Traditional methods usually rely on statistical models and expert systems, which have certain limitations when processing financial data, especially when dealing with large amounts of high-dimensional data and complex nonlinear relationships. In recent years, deep learning technology has developed rapidly, providing new solutions for financial risk prediction. Deep learning methods, especially neural networks, have achieved remarkable results in many fields such as image recognition and speech processing due to their powerful learning ability and nonlinear modeling capabilities, prompting researchers to introduce deep learning technology into financial risk prediction in the hope of improving the accuracy and efficiency of predictions.

[0003] However, existing deep learning technologies still face significant challenges when processing financial data. First, neural network models are prone to overfitting, especially overfitting training data during the training process, resulting in poor performance when processing new data and lack of generalization ability. Second, financial data is usually complex and changeable, and the data set contains noisy, incomplete and inconsistent data, which poses a threat to the accuracy of the model. In addition, the source of financial data is single, which makes it difficult to fully reflect market conditions, resulting in insufficient data quality and diversity. Financial markets change rapidly, and real-time and response speed are key challenges that financial risk prediction systems must address, but existing technologies often find it difficult to ensure sufficient prediction speed and real-time response capabilities when processing large-scale, high-frequency data. Summary of the invention

[0004] The embodiment of the present invention proposes a financial risk prediction method based on deep learning to solve the problems existing in the current neural network model, such as poor performance when processing new data and lack of generalization ability, the complexity and variability of financial data affecting the accuracy of the model, the difficulty of financial data in fully reflecting the market situation, resulting in insufficient data quality and diversity, and the problem of insufficient prediction speed and real-time response capability of the prior art when processing large-scale, high-frequency data.

[0005] An embodiment of the present invention provides a financial risk prediction method based on deep learning, comprising the following steps:

[0006] S1: Collect and preprocess the original financial data, use graph convolutional networks to extract the deep features of the preprocessed original financial data, and use adaptive weighted nonlinear wavelet decomposition technology to process the deep features to obtain long-term trend components and short-term volatility components;

[0007] S2: Use the dynamic fuzzy system model to optimize the long-term trend component and the short-term volatility component, input the quantum perception enhanced converter model for training, obtain the comprehensive financial risk, and generate a detailed risk assessment report based on the comprehensive financial risk.

[0008] As a specific implementation, in step S1, "using adaptive weighted nonlinear wavelet decomposition technology to process deep-level features to obtain long-term trend components and short-term fluctuation components" includes:

[0009] The adaptive weighted nonlinear wavelet decomposition technology is used to perform multi-scale decomposition of deep-level features, extract the original long-term trend component and the original short-term fluctuation component, calculate the adaptive weights of the original long-term trend component and the original short-term fluctuation component respectively, and perform weighted processing on the original long-term trend component and the original short-term fluctuation component to obtain the long-term trend component and the short-term fluctuation component.

[0010] As a specific implementation, the adaptive weight calculation formula of the original long-term trend component is as follows:

[0011]

[0012] Among them, α l (t) is the adaptive weight of the original long-term trend component at the lth decomposition layer at time t; l is the decomposition layer index of the original long-term trend component; θ l (t) is the characteristic importance parameter of the original long-term trend component at time t in the lth decomposition layer, and the initial value is set using the mean initialization method; is the original long-term trend component of the l-th decomposition layer at time t; t is the time index; i is the time index used for the summation operation; is the long-term trend component of all decomposition layers at time i; L is the total number of decomposition layers of the original long-term trend component, which is set according to the specific implementation scenario; k is the decomposition layer index of the original long-term trend component used for the summation operation; It is the original long-term trend component of the kth decomposition layer at time t.

[0013] As a specific implementation, the calculation formula of the adaptive weight of the original short-term volatility component is as follows:

[0014]

[0015] Among them, β m (t) is the adaptive weight of the original short-term fluctuation component at the mth decomposition layer at time t; m is the decomposition layer index of the original short-term fluctuation component; γ m (t) is the characteristic importance parameter of the original short-term volatility component at time t in the mth decomposition layer, and the initial value is set using the mean initialization method; is the original short-term volatility component of the mth decomposition layer at time t; ∈ is a constant, M is the total number of decomposition layers of the original short-term volatility component, which is set according to the specific implementation scenario; n is the decomposition layer index of the original short-term volatility component used for the summation operation; It is the original short-term fluctuation component of the nth decomposition layer at time t.

[0016] As a specific implementation method, the original long-term trend component is weighted to obtain the long-term trend component of adaptive weighted nonlinear wavelet decomposition, and the formula is as follows:

[0017]

[0018] in, is the long-term trend component of the original long-term trend component at time t in the l-th decomposition layer; exp is the exponential operation; l is the decomposition layer index of the original long-term trend component used for the summation operation; α l (t) is the adaptive weight of the original long-term trend component at time t in the lth decomposition layer; is the original long-term trend component of the lth decomposition layer at time t; L is the total number of decomposition layers of the original long-term trend component; k is the decomposition layer index of the original long-term trend component used for the summation operation; It is the original long-term trend component of the kth decomposition layer at time t.

[0019] As a specific implementation method, the original short-term fluctuation component is weighted to obtain the short-term fluctuation component of adaptive weighted nonlinear wavelet decomposition, and the formula is as follows:

[0020]

[0021] in, is the short-term fluctuation component of the original short-term fluctuation component at time t at the mth decomposition layer; m is the decomposition layer index of the original short-term fluctuation component used for the summation operation; β m (t) is the adaptive weight of the original short-term fluctuation component at time t in the mth decomposition layer; is the original short-term fluctuation component of the mth decomposition layer at time t; ∈ is a constant.

[0022] As a specific implementation, step S2 includes:

[0023] An adaptive dynamic fuzzy system model is established for the long-term trend component and the short-term fluctuation component respectively for training, and the optimized long-term trend component and the optimized short-term fluctuation component are obtained;

[0024] The optimized long-term trend component and the optimized short-term volatility component are predicted using the quantum-aware enhanced converter model to generate comprehensive financial risk.

[0025] As a specific implementation, in step S2, "using the quantum perception enhanced converter model to predict the optimized long-term trend component and the optimized short-term volatility component to generate comprehensive financial risks" specifically includes:

[0026] The quantum-aware enhanced converter model introduces a quantum-aware layer and a quantum-attention mechanism based on the traditional converter model.

[0027] The optimized long-term trend component and the optimized short-term fluctuation component are converted into the quantum state of the long-term trend component and the quantum state of the short-term fluctuation component through the quantum perception layer;

[0028] The quantum attention mechanism first uses the quantum state of the optimized long-term trend component and the quantum state of the optimized short-term fluctuation component to construct the original query matrix, original key matrix, and original value matrix of the optimized long-term trend component and the optimized short-term fluctuation component respectively;

[0029] Define quantum state operation functions using the superposition and entanglement properties of quantum states;

[0030] Calculate the query matrix, key matrix and value matrix of the optimized long-term trend component and the optimized short-term volatility component respectively;

[0031] Based on the query matrix, key matrix and value matrix of the optimized long-term trend component and the optimized short-term fluctuation component, the attention output of the optimized long-term trend component and the attention output of the optimized short-term fluctuation component are calculated respectively.

[0032] As a specific implementation, in step S2, "using the quantum perception enhanced converter model to predict the optimized long-term trend component and the optimized short-term volatility component to generate comprehensive financial risks" also includes:

[0033] Based on the optimized attention output of the long-term trend component and the optimized attention output of the short-term volatility component, the converter model is used to fuse the optimized attention output of the long-term trend component and the optimized attention output of the short-term volatility component, and the obtained value is used as the comprehensive financial risk.

[0034] Another embodiment of the present application further discloses a financial risk prediction system based on deep learning, including:

[0035] The data preprocessing module is used to clean, denoise and normalize the raw financial data collected from multiple data sources to obtain the preprocessed raw financial data;

[0036] The graph network feature learning module is used to extract features from the original financial data preprocessed by the data preprocessing module using a graph convolutional network, and generate deep features through multi-layer graph convolution operations;

[0037] The multi-scale time series decomposition module is used to decompose the deep features of the graph network feature learning module using adaptive weighted nonlinear wavelet decomposition technology, and extract long-term trend components and short-term fluctuation components from the deep features;

[0038] The adaptive dynamic fuzzy module is used to establish adaptive dynamic fuzzy system models for the long-term trend component and short-term fluctuation component of the multi-scale time series decomposition module, respectively, and train them to obtain the optimized long-term trend component and the optimized short-term fluctuation component;

[0039] A prediction model module is used to predict the optimized long-term trend component and the optimized short-term volatility component of the adaptive dynamic fuzzy module respectively by using the quantum perception enhanced converter model to obtain the comprehensive financial risk;

[0040] The risk assessment module is used to classify risk levels according to the comprehensive financial risks of the prediction model module, analyze the changing trends of comprehensive financial risks to foresee future risks, set thresholds to trigger the alarm system, monitor market dynamics in real time, and update comprehensive financial risks in a timely manner.

[0041] The technical solution of the embodiment of the present invention has at least the following beneficial effects:

[0042] 1. By leveraging the powerful feature extraction and pattern recognition capabilities of deep learning models, the system can more accurately capture the complex nonlinear relationships and dynamic changes in the financial market, thereby significantly improving the accuracy and reliability of risk prediction.

[0043] 2. Adaptive weighted nonlinear wavelet decomposition technology can effectively separate and extract long-term trend components and short-term volatility components in financial data, improve the accuracy and stability of data analysis, and thus provide a solid foundation for more accurate financial risk prediction and more effective risk management.

[0044] 3. The use of the quantum perception enhanced transformer model (qPET model) can improve prediction accuracy and data representation capabilities, enhance real-time risk monitoring, support decision making, and improve the stability and robustness of the model by better handling noise and abnormal situations, thereby providing a more accurate, comprehensive and stable financial risk assessment and improving the effectiveness of risk management and decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A structural diagram of a financial risk prediction system based on deep learning disclosed in one embodiment of the present invention;

[0046] Figure 2 A flowchart of a financial risk prediction method based on deep learning disclosed in another embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0048] Unless defined otherwise, 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 belongs.

[0049] The following is a detailed description of a specific scheme of a financial risk prediction system and method based on deep learning provided by the present invention in conjunction with the accompanying drawings.

[0050] An embodiment of the present invention provides a financial risk prediction system based on deep learning. Figure 1 The system includes the following parts:

[0051] Data preprocessing module, graph network feature learning module, multi-scale time series decomposition module, adaptive dynamic fuzzy module, prediction model module, and risk assessment module.

[0052] The data preprocessing module is used to clean, denoise and normalize the raw financial data collected from multiple data sources to obtain the preprocessed raw financial data. The data preprocessing module is connected to the graph network feature learning module through data transmission.

[0053] The graph network feature learning module is used to extract features from the preprocessed raw financial data using a graph convolutional network and generate deep features through multi-layer graph convolution operations. The graph network feature learning module is connected to the multi-scale time series decomposition module through data transmission.

[0054] The multi-scale time series decomposition module uses adaptive weighted nonlinear wavelet decomposition technology to decompose deep-level features and extract long-term trend components and short-term fluctuation components from the deep-level features. The multi-scale time series decomposition module is connected to the adaptive dynamic fuzzy module through data transmission.

[0055] The adaptive dynamic fuzzy module is used to establish adaptive dynamic fuzzy system models for the long-term trend component and the short-term fluctuation component respectively for training to obtain the optimized long-term trend component and the optimized short-term fluctuation component. The adaptive dynamic fuzzy module is connected to the prediction model module through data transmission.

[0056] The prediction model module, (QPET model) predicts the optimized long-term trend component and the optimized short-term volatility component respectively to obtain the comprehensive financial risk. The prediction model module is connected to the risk assessment module through data transmission.

[0057] The risk assessment module is used to classify risk levels according to comprehensive financial risks, analyze the changing trends of comprehensive financial risks to foresee future risks, set thresholds to trigger the alarm system, monitor market dynamics in real time and update comprehensive financial risks in a timely manner.

[0058] Another embodiment of the present invention provides a financial risk prediction method based on deep learning. Figure 2 , the method comprises the following steps:

[0059] S1: Collect and preprocess the original financial data, use graph convolutional networks to extract the deep features of the preprocessed original financial data, and use adaptive weighted nonlinear wavelet decomposition technology to process the deep features to obtain long-term trend components and short-term volatility components.

[0060] Specifically, the data preprocessing module of the deep learning-based financial risk prediction system collects raw financial data from multiple data sources, and preprocesses the collected raw financial data, and obtains the preprocessed raw financial data through cleaning, denoising and normalization; the graph network feature learning module uses the graph convolution network to extract features from the preprocessed raw financial data, and obtains the extracted deep features through multi-layer graph convolution operations. The deep features are complex, high-dimensional and multi-scale information extracted from the preprocessed raw financial data, reflecting the complex, high-dimensional and multi-scale information in the data, including complex nonlinear relationships in the raw financial data, integration of multiple different types of raw financial data indicators, time dependence and dynamic changes, as well as potential connections and mutual influences in the market. Through deep learning technologies such as graph convolution networks, deep features are effectively captured and extracted.

[0061] The specific preprocessing technology and graph convolutional network technology are well known to those skilled in the art and will not be elaborated here.

[0062] Adaptive weighted nonlinear wavelet decomposition technology is an improved technology of discrete wavelet decomposition technology. It uses discrete wavelet decomposition technology on the deep features extracted by graph convolutional network technology to obtain the original long-term trend component. and the original short-term volatility component Discrete wavelet decomposition technology is a well-known technical means for those skilled in the art and will not be elaborated here.

[0063] The cumulative impact of the original long-term trend component is introduced, and the adaptive weight of the original long-term trend component is calculated by the following formula:

[0064]

[0065] Among them, α l (t) is the adaptive weight of the original long-term trend component at the lth decomposition layer at time t; l is the decomposition layer index of the original long-term trend component; θ l (t) is the characteristic importance parameter of the original long-term trend component at time t in the lth decomposition layer, and the initial value is set using the mean initialization method; is the original long-term trend component of the l-th decomposition layer at time t; t is the time index; i is the time index used for the summation operation; is the long-term trend component of all decomposition layers at time i; L is the total number of decomposition layers of the original long-term trend component, which is set according to the specific implementation scenario; is between time 1 and time t The cumulative average value reflects the cumulative impact of the original long-term trend component; k is the decomposition layer index of the original long-term trend component used for the summation operation; It is the original long-term trend component of the kth decomposition layer at time t.

[0066] The adaptive weight of the original short-term volatility component is calculated by the following formula:

[0067]

[0068] Among them, β m (t) is the adaptive weight of the original short-term fluctuation component at the mth decomposition layer at time t; m is the decomposition layer index of the original short-term fluctuation component; γ m (t) is the characteristic importance parameter of the original short-term volatility component at time t in the mth decomposition layer, and the initial value is set using the mean initialization method; is the original short-term fluctuation component of the mth decomposition layer at time t; ∈ is a constant to prevent the denominator from being zero, which can be set according to the specific implementation scenario; It is used to calculate the current time t and time t-1 The volatility difference; M is the total number of decomposition layers of the original short-term volatility component, which is set according to the specific implementation scenario; n is the decomposition layer index of the original short-term volatility component used for the summation operation; It is the original short-term fluctuation component of the nth decomposition layer at time t.

[0069] The feature importance parameters of the original long-term trend component and the original short-term volatility component are dynamically adjusted using the following formula:

[0070]

[0071] Among them, θ l(t) is the characteristic importance parameter of the adjusted original long-term trend component; θ l (t-1) is the characteristic importance parameter of the original long-term trend component before adjustment; γ m (t) is the characteristic importance parameter of the adjusted original short-term volatility component; γ m (t-1) is the characteristic importance parameter of the original short-term volatility component before adjustment; η θ is the learning rate for adjusting the feature importance parameter of the original long-term trend component, η γ is the learning rate for adjusting the feature importance parameters of the original short-term volatility component; η θ and η γ Obtained through experimental tuning; tanh is the hyperbolic tangent function; sinh is the hyperbolic sine function.

[0072] The original long-term trend component and the original short-term fluctuation component are weighted to obtain the long-term trend component and short-term fluctuation component of adaptive weighted nonlinear wavelet decomposition. The formula is as follows:

[0073]

[0074] in, is the long-term trend component of the original long-term trend component at time t in the l-th decomposition layer,

[0075] It is the short-term volatility component of the original short-term volatility component at time t in the mth decomposition layer; exp is the exponential operation; l is the decomposition layer index of the original long-term trend component used for the summation operation; m is the decomposition layer index of the original short-term volatility component used for the summation operation.

[0076] In summary, the adaptive weighted nonlinear wavelet decomposition technology is used to perform multi-scale decomposition of deep-level features, extract the original long-term trend component and the original short-term fluctuation component, and calculate the adaptive weights of the original long-term trend component and the original short-term fluctuation component respectively. The original long-term trend component and the original short-term fluctuation component are weighted to obtain the long-term trend component and the short-term fluctuation component.

[0077] The long-term trend component reflects the changes in financial data over a long period of time, revealing the macro trends of the financial market, such as the overall upward or downward trend of the stock market or the economy; it reflects changes in the economic cycle, such as economic expansion or recession phases; and the impact of long-term policies, market environment and corporate fundamentals. The long-term trend component can help us understand the long-term changing patterns and trends of the financial market.

[0078] The short-term volatility component reflects the fluctuation of financial data within a short period of time, including daily price fluctuations and changes in trading volume, the amplitude and frequency of market fluctuations, and the impact of short-term market events (such as announcements and news). The short-term volatility component is used to identify and analyze the short-term volatility and risks of the financial market.

[0079] S2: Use the dynamic fuzzy system model to optimize the long-term trend component and the short-term volatility component, input the quantum perception enhanced converter model (QPET model) for training, obtain the comprehensive financial risk, and generate a detailed risk assessment report based on the comprehensive financial risk.

[0080] Specifically, adaptive dynamic fuzzy system models are established for the long-term trend component and the short-term fluctuation component for training, and the optimized long-term trend component is obtained. and optimized short-term volatility components The optimized long-term trend component is smoother, can more stably and reliably reflect the long-term trend, improves the prediction accuracy, and enables the model to more accurately predict the future long-term trend; the optimized short-term fluctuation component eliminates noise, can more accurately reflect the real situation of short-term fluctuations, and improves the prediction accuracy of short-term fluctuations. The adaptive dynamic fuzzy system model is an existing technology and will not be described in detail here.

[0081] The Quantum Perception Enhanced Transformer Model (QPET Model) innovates on the traditional transformer model by introducing a quantum perception layer and a quantum attention mechanism. The parallel processing capability of quantum computing and the quantum perception mechanism are used to enhance the prediction performance of the traditional transformer model, and the QPET model is used to predict the optimized long-term trend component and the optimized short-term volatility component to generate comprehensive financial risks.

[0082] First, the optimized long-term trend component and the optimized short-term fluctuation component are converted into the quantum state Q of the long-term trend component through the quantum perception layer. T and the quantum state Q of the short-term fluctuation component S ,Quantum state coding technology is an existing technology and will not be elaborated here.

[0083] Next, the quantum attention mechanism first uses the quantum state Q of the optimized long-term trend component T and the quantum state Q of the optimized short-term fluctuation component S The original query matrix, original key matrix, and original value matrix of the components are constructed as follows:

[0084]

[0085] in, are the original query matrix, original key matrix and original value matrix of the optimized long-term trend component, respectively. are the original query matrix, original key matrix and original value matrix of the optimized short-term volatility component respectively.

[0086] The quantum state operation function f is defined by using the superposition and entanglement characteristics of quantum states. Quantum state superposition is realized by linear combination, and quantum state entanglement is realized by introducing the entanglement operator. The superposition state and the entangled state are combined to obtain the quantum state operation function f formula:

[0087]

[0088] Among them, A and B are two input matrices for quantum state operations, which can be Any two of Any two of and is the superposition coefficient, which is used to control the weight of quantum state superposition, and is the weight parameter, which is used to control the combined weight of the superposition state and the entangled state. and The value can be set according to the specific implementation scenario and is not limited here. It is an outer product operation that combines two quantum states into a higher-dimensional quantum state, where |0) and |1> are the ground states of the quantum bit.

[0089] Compute the optimized long-term trend component of the query matrix Key Matrix Sum Matrix The formula is as follows:

[0090]

[0091] Compute the query matrix of the optimized short-term volatility component Key Matrix Sum Matrix The formula is as follows:

[0092]

[0093] Based on the query matrix, key matrix, and value matrix of the optimized long-term trend component and the optimized short-term fluctuation component, the attention scores of the optimized long-term trend component and the optimized short-term fluctuation component are calculated respectively. The formula is as follows:

[0094]

[0095] Among them, W T is the attention output of the optimized long-term trend component, W S is the attention output of the optimized short-term fluctuation component, is the attention score of the optimized long-term trend component, is the attention score of the optimized short-term fluctuation component, and They are and The transposed matrix, d k yes The dimension of yes The softmax function is used to perform normalization operations. h is a nonlinear transformation function, which can be set according to the specific implementation scenario and is not limited here.

[0096] The quantum attention mechanism uses the superposition and entanglement characteristics of quantum states to define the quantum state operation function f, increasing the nonlinearity and complexity of the QPET model, enabling the QPET model to better adapt to and process complex patterns and relationships in financial data, improve the accuracy of predictions, adapt to large-scale data processing needs, and provide a more comprehensive risk assessment.

[0097] Based on the optimized attention output of the long-term trend component and the optimized attention output of the short-term volatility component, the converter model is used to fuse the optimized attention output of the long-term trend component and the optimized attention output of the short-term volatility component, and the obtained value is used as the comprehensive financial risk. The converter model is an existing technology and will not be elaborated here.

[0098] Comprehensive financial risk is a holistic assessment of all potential risks in the data source of the deep learning-based financial risk prediction system. It provides a holistic perspective for assessing and managing financial risks by providing a single indicator to quantify the overall risk level.

[0099] According to the comprehensive financial risk value, the risk level is divided into low, medium and high risk intervals for risk level assessment; by analyzing the changing trend of comprehensive financial risk, the potential risk increase or decrease trend is identified, so as to foresee the future risk trend; set the risk threshold, when the comprehensive financial risk value exceeds the set threshold, trigger the alarm system, prompting to take corresponding risk management measures; monitor the market dynamics in real time, and update the comprehensive financial risk in time. The above-mentioned technologies are all well-known technical means for those skilled in the art, and will not be elaborated here.

[0100] In summary, a financial risk prediction system and method based on deep learning has been completed.

[0101] The order of the embodiments of the invention is for description only and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0102] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0103] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A financial risk prediction method based on deep learning, characterized in that: The following steps are involved: S1: Collect and preprocess the original financial data, use graph convolutional networks to extract the deep features of the preprocessed original financial data, and use adaptive weighted nonlinear wavelet decomposition technology to process the deep features to obtain long-term trend components and short-term volatility components; S2: Use the dynamic fuzzy system model to optimize the long-term trend component and the short-term volatility component, input the quantum perception enhanced converter model for training, obtain the comprehensive financial risk, and generate a detailed risk assessment report based on the comprehensive financial risk.

2. The financial risk prediction method based on deep learning according to claim 1, characterized in that: In step S1, "using adaptive weighted nonlinear wavelet decomposition technology to process deep-level features to obtain long-term trend components and short-term fluctuation components" includes: The adaptive weighted nonlinear wavelet decomposition technology is used to perform multi-scale decomposition of deep-level features, extract the original long-term trend component and the original short-term fluctuation component, calculate the adaptive weights of the original long-term trend component and the original short-term fluctuation component respectively, and perform weighted processing on the original long-term trend component and the original short-term fluctuation component to obtain the long-term trend component and the short-term fluctuation component.

3. The financial risk prediction method based on deep learning as claimed in claim 2, characterized in that: The adaptive weight calculation formula of the original long-term trend component is as follows: Among them, α l (t) is the adaptive weight of the original long-term trend component at the lth decomposition layer at time t; l is the decomposition layer index of the original long-term trend component; θ l (t) is the characteristic importance parameter of the original long-term trend component at time t in the lth decomposition layer, and the initial value is set using the mean initialization method; is the original long-term trend component of the l-th decomposition layer at time t; t is the time index; i is the time index used for the summation operation; is the long-term trend component of all decomposition layers at time i; L is the total number of decomposition layers of the original long-term trend component, which is set according to the specific implementation scenario; k is the decomposition layer index of the original long-term trend component used for the summation operation; It is the original long-term trend component of the kth decomposition layer at time t.

4. The financial risk prediction method based on deep learning as claimed in claim 2, characterized in that: The adaptive weight of the original short-term volatility component is calculated as follows: Among them, β m (t) is the adaptive weight of the original short-term fluctuation component at the mth decomposition layer at time t; m is the decomposition layer index of the original short-term fluctuation component; γ m (t) is the characteristic importance parameter of the original short-term volatility component at time t in the mth decomposition layer, and the initial value is set using the mean initialization method; is the original short-term volatility component of the mth decomposition layer at time t; ∈ is a constant, M is the total number of decomposition layers of the original short-term volatility component, which is set according to the specific implementation scenario; n is the decomposition layer index of the original short-term volatility component used for the summation operation; It is the original short-term fluctuation component of the nth decomposition layer at time t.

5. The financial risk prediction method based on deep learning as claimed in claim 3, characterized in that: The original long-term trend component is weighted to obtain the long-term trend component of adaptive weighted nonlinear wavelet decomposition. The formula is as follows: in, is the long-term trend component of the original long-term trend component at time t in the l-th decomposition layer; exp is the exponential operation; l is the decomposition layer index of the original long-term trend component used for the summation operation; α l (t) is the adaptive weight of the original long-term trend component at time t in the lth decomposition layer; is the original long-term trend component of the lth decomposition layer at time t; L is the total number of decomposition layers of the original long-term trend component; k is the decomposition layer index of the original long-term trend component used for the summation operation; It is the original long-term trend component of the kth decomposition layer at time t.

6. The financial risk prediction method based on deep learning according to claim 4, characterized in that: The original short-term fluctuation component is weighted to obtain the short-term fluctuation component of adaptive weighted nonlinear wavelet decomposition. The formula is as follows: in, is the short-term fluctuation component of the original short-term fluctuation component at time t at the mth decomposition layer; m is the decomposition layer index of the original short-term fluctuation component used for the summation operation; β m (t) is the adaptive weight of the original short-term fluctuation component at time t in the mth decomposition layer; is the original short-term fluctuation component of the mth decomposition layer at time t; ∈ is a constant.

7. The financial risk prediction method based on deep learning according to claim 1, characterized in that: The step S2 comprises: An adaptive dynamic fuzzy system model is established for the long-term trend component and the short-term fluctuation component respectively for training, and the optimized long-term trend component and the optimized short-term fluctuation component are obtained; The optimized long-term trend component and the optimized short-term volatility component are predicted using the quantum-aware enhanced converter model to generate comprehensive financial risk.

8. The financial risk prediction method based on deep learning according to claim 7, characterized in that: In step S2, "using the quantum perception enhanced converter model to predict the optimized long-term trend component and the optimized short-term volatility component to generate comprehensive financial risks" specifically includes: The quantum-aware enhanced converter model introduces a quantum-aware layer and a quantum-attention mechanism based on the traditional converter model. The optimized long-term trend component and the optimized short-term fluctuation component are converted into the quantum state of the long-term trend component and the quantum state of the short-term fluctuation component through the quantum perception layer; The quantum attention mechanism first uses the quantum state of the optimized long-term trend component and the quantum state of the optimized short-term fluctuation component to construct the original query matrix, original key matrix, and original value matrix of the optimized long-term trend component and the optimized short-term fluctuation component respectively; Define quantum state operation functions using the superposition and entanglement properties of quantum states; Using quantum state operation functions, the query matrix, key matrix and value matrix of the optimized long-term trend component and the optimized short-term volatility component are calculated respectively; Based on the query matrix, key matrix and value matrix of the optimized long-term trend component and the optimized short-term fluctuation component, the attention output of the optimized long-term trend component and the attention output of the optimized short-term fluctuation component are calculated respectively.

9. The financial risk prediction method based on deep learning as claimed in claim 8, characterized in that: In step S2, "using the quantum perception enhanced converter model to predict the optimized long-term trend component and the optimized short-term volatility component to generate comprehensive financial risks" also includes, Based on the optimized attention output of the long-term trend component and the optimized attention output of the short-term volatility component, the converter model is used to fuse the optimized attention output of the long-term trend component and the optimized attention output of the short-term volatility component, and the obtained value is used as the comprehensive financial risk.

10. A financial risk prediction system based on deep learning, characterized in that: include: The data preprocessing module is used to clean, denoise and normalize the raw financial data collected from multiple data sources to obtain the preprocessed raw financial data; A graph network feature learning module, used to extract features from the original financial data preprocessed by the data preprocessing module using a graph convolutional network, and generate deep features through multi-layer graph convolution operations; A multi-scale time series decomposition module, used to decompose the deep features of the graph network feature learning module using an adaptive weighted nonlinear wavelet decomposition technique, and extract long-term trend components and short-term fluctuation components from the deep features; An adaptive dynamic fuzzy module is used to establish adaptive dynamic fuzzy system models for the long-term trend component and the short-term fluctuation component of the multi-scale time series decomposition module, respectively, to train them, so as to obtain an optimized long-term trend component and an optimized short-term fluctuation component; A prediction model module, used to predict the optimized long-term trend component and the optimized short-term volatility component of the adaptive dynamic fuzzy module respectively by using a quantum perception enhanced converter model to obtain a comprehensive financial risk; The risk assessment module is used to classify risk levels according to the comprehensive financial risks of the prediction model module, analyze the changing trends of the comprehensive financial risks to foresee future risks, set thresholds to trigger the alarm system, monitor market dynamics in real time, and update the comprehensive financial risks in a timely manner.

Images