Financial risk dynamic prediction method based on deep learning

Abstract

The invention provides a financial risk dynamic prediction method based on deep learning, and relates to the field of financial risk dynamic prediction. The financial risk dynamic prediction process provided by the invention comprises the steps of constructing a financial risk prediction data set, calculating a deviation absolute value sequence of a factor data time sequence, and fusing and mapping the time sequence and the deviation absolute value sequence into a high-dimensional embedded vector sequence through a weighted nonlinear function; constructing a finite span directed graph by taking each time step as a graph node; introducing a structure resonance factor; carrying out weighted coding and weighted fusion on a path to form an aggregation context vector; calculating a risk score; overall feature representation of factor data is extracted in combination with a bidirectional long-short-term memory network, and finally a financial risk prediction value is calculated by using a full-connection neural network. According to the method, local deviation and global time sequence information are fused, and the accuracy of financial risk prediction is improved.

Description

Technical Field

[0001] The present invention belongs to the field of dynamic prediction of financial risks, and particularly relates to a dynamic prediction method of financial risks based on deep learning. Background Art

[0002] Financial risk refers to a probabilistic event that an enterprise may face adverse consequences such as asset impairment, debt default or capital chain breakage during the production and operation, investment and financing, and capital flow processes, due to internal management out of control, capital structure imbalance or external environmental fluctuation factors. Such risks are characterized by suddenness and conductivity. If timely warning and intervention are not carried out, they often lead to a decline in the enterprise's profitability, a decrease in credit rating, and even fall into an operating crisis or bankruptcy liquidation. Therefore, accurately predicting financial risks is of great significance for improving the enterprise's risk management ability and strengthening the intelligent level of financial supervision.

[0003] The prediction of financial risks mainly relies on means such as financial statement analysis, ratio index evaluation and expert manual judgment, which have problems such as strong data lag, much human subjective intervention, and untimely prediction response, and it is difficult to meet the dynamic warning requirements of enterprise financial risks. Although the prediction models using traditional machine learning methods have improved the automation level of information processing to a certain extent, the existing methods have not effectively characterized the complex non-linear coupling relationship between different financial dimensions.

[0004] The formation of financial risks is often jointly driven by multi-source heterogeneous data, which not only shows significant non-linear change characteristics in the time series, but also presents dynamic evolution laws of multi-scale and multi-stage. Therefore, if the model for financial risk prediction can simultaneously have the sensitive response ability to local abnormal fluctuations and the modeling ability for global trend evolution, it will be more helpful to comprehensively and accurately capture potential risk signals and improve the reliability of dynamic prediction of financial risks. Summary of the Invention

[0005] The present invention provides a dynamic prediction method of financial risks based on deep learning, which constructs an absolute deviation value sequence by using the monitoring value of each time step in the time series of each factor data and its historical mean, fuses and maps the original time series and the absolute deviation value sequence into a high-dimensional embedding vector sequence through a weighted non-linear mapping function, constructs a time series graph structure with a limited maximum span based on the high-dimensional embedding vector sequence, introduces a structural resonance factor, performs weighted coding on multiple historical paths to obtain path coding vectors, forms an aggregated context vector after weighted fusion through normalized path weights, calculates a risk score by using a linear mapping and an activation function, obtains a steepness value of the change trend through weighted difference to modulate the aggregated context vector, extracts the overall feature representation of factor data by means of a bidirectional long short-term memory network, and finally splices and inputs the overall feature representations of each factor data into a fully connected neural network to obtain accurate and stable financial risk prediction values.

[0006] The technical method adopted by the present invention to achieve the above object specifically includes the following steps: S1. Collect factor data affecting financial risks and construct a financial risk prediction data set; S2. For the time series data of each type of factor data, calculate the absolute value of the deviation between the monitoring value at each time step and the mean value of historical time steps to obtain an absolute deviation value sequence, and map the time series data and the absolute deviation value sequence into a high-dimensional embedded vector sequence through a weighted non-linear function; S3. Use each time step in the high-dimensional embedded vector sequence as a graph node to construct a directed graph with a limited maximum span, calculate the global average representation of the nodes, combine the direction consistency and position deviation measurement between nodes, introduce a structural resonance factor, construct a path aggregation coding module, perform weighted coding on multiple paths, and obtain a path coding vector; S4. Calculate the normalized weights of the path coding vectors, and fuse multiple path coding vectors by weighting according to the weights to form an aggregated context vector for each time step; S5. Based on the aggregated context vector, calculate the risk score using a linear mapping and a Sigmoid activation function, further obtain the steepness value of the change trend using weighted differences, weight the corresponding aggregated context vector using the steepness value of the change trend, and combine a bidirectional long short-term memory network to construct a risk evolution recognition module to obtain the overall feature representation of the factor data; S6. Concatenate the overall feature representations of all factor data, input them into a fully connected neural network, and calculate the final financial risk prediction value.

[0007] Preferably, in S1, collect factor data affecting the dynamic prediction of financial risks, including current ratio, asset-liability ratio, interest coverage ratio, net profit ratio, inventory turnover rate, operating cycle, total asset turnover rate. For missing data, use linear interpolation technology for dynamic completion to ensure the continuity and integrity of the time series input, and construct a financial risk prediction data set.

[0008] Preferably, in S2, input the time series data of the th type of factor data affecting financial risks , where is the monitoring value of the th type of factor data of financial risks at the th time step, is the length of the time series. Use the absolute value of the deviation to obtain a high-dimensional embedded vector sequence to enhance the expression ability of the model for the mutation characteristics of financial factor data. Specifically, it includes the following steps: S21. Calculate the monitoring value of the th type of factor data and its previous The absolute value of the deviation of the mean value of each time step, which measures the abnormal deviation degree of the current value. The specific mathematical model is as follows: ; In the formula, is the absolute value of the deviation of the th factor data at the th time step. is the length of the sliding window. The absolute values of the deviations of all time steps of the th factor data are combined into an absolute value of deviation sequence , , is the absolute value operation; S22. Weighted non - linear fusion is performed on the time - series data and the absolute value of deviation sequence to obtain a high - dimensional embedding vector sequence. The specific mathematical model is as follows: ; In the formula, is the high - dimensional embedding vector sequence obtained through weighted non - linear fusion. , where is the high - dimensional embedding vector of the th factor data at the th time step. , are learnable weight matrices, and is the bias term.

[0009] Preferably, introducing the absolute value of deviation can help the model effectively capture the abnormal fluctuation behavior of factor data affecting financial risks in the short term. At the same time, performing non - linear weighted fusion on the absolute value of deviation sequence and time - series data can enhance the sensitivity of the model to local perturbations and improve the ability to identify potential risk signals. Compared with traditional linear embedding or static coding methods, the method has stronger dynamic expression ability and is suitable for modeling the key feature process of the evolution of financial risks from a stable state to a mutation state.

[0010] Preferably, in S3, a path aggregation encoding module is constructed, which specifically includes the following steps: S31. Each time step in the high - dimensional embedding vector sequence is used as a graph node , where is the node of the th factor data at the th time step. The edge set , where is the edge between the th factor data at the th time step and the th time step node. is the maximum span threshold, which limits the connection of only a certain number of historical nodes in adjacent time steps at each time step to obtain a graph ; S32. Obtain the global average representation of the data nodes of the -th factor, and construct a structural resonance factor by combining the direction consistency and position deviation measurement between nodes to promote the model to capture and express the implicit correlation patterns in the dynamic evolution of long-term financial risks. The specific mathematical model is: ; ; In the formula, is the global average representation of the data of the -th factor, is the structural resonance factor between the data of the -th factor at the -th time step and the -th time step, , are resonance modulation hyperparameters, is the L2 norm, , are the nodes of the data of the -th factor at the -th and -th time steps respectively, is a minimum value to prevent the denominator from being zero; S33. Denote the -th path with the node at the -th time step of the data of the -th factor as . The path contains a series of node indices from the -th time step to the end time step . , perform structural weighted coding on . The specific mathematical model is: ; In the formula, is the path coding vector of the -th path of the data of the -th factor, is the structural resonance factor between the data of the -th factor at the -th time step and the -th time step, is the structural resonance factor between the data of the -th factor at the -th time step and the -th time step, is the node of the data of the -th factor at the Time step node is the type of factor data at the time step node is the Hadamard element multiplication

[0011] Preferably, a structural resonance factor is introduced. By using the direction consistency and position deviation measurement between nodes, the complex dependence between different time steps in the time series data of factor data affecting financial risks can be effectively captured. The nodes in the path are weighted and fused to obtain a path encoding vector, which can further enhance the ability to identify key financial risk evolution paths

[0012] Preferably, in S4, in the dynamic prediction of financial risks, the influence degrees of different historical paths on the current financial risk state are different. Dynamic weight assignment and fusion are performed on multiple path encoding vectors to form an aggregated context vector for each time step, including the following steps S41. Normalize each path encoding vector to obtain a path weight to significantly highlight more important paths. The specific mathematical model is ; where is the normalized path weight of the th path, ensuring that the sum of all path weights is 1 is the path encoding vector of the type of factor data for the th path is the number of paths ending at the time step node S42. According to the normalized path weights, perform a weighted sum on the multiple path encoding vectors selected for the time step to form an aggregated context vector. The specific mathematical model is ; where is the aggregated context vector of the type of factor data for the time step

[0013] Preferably, by using a weight calculation mechanism, the model can dynamically identify and amplify the influence of key information paths, significantly improving the sensitivity to the latent and evolution of financial risks. The normalization process of the weights ensures the stability and rationality of information fusion. The encoding vectors of multiple historical paths are weighted and fused according to the weights to effectively aggregate the long-term dependencies of multiple sources and multiple paths in the time series, effectively avoiding over-focusing on a single path

[0014] Preferably, in step S5, a risk evolution recognition module is constructed, including the following steps: S51. Calculate the risk score through single-layer linear mapping and the Sigmoid activation function. The specific mathematical model is: ; In the formula, is the risk score of the th factor data at the th time step, is the linear mapping weight matrix, is the bias vector; S52. Calculate the weighted difference between the risk scores at the current and adjacent time steps, and comprehensively consider the second-order difference and the first-order change amplitude to obtain the steepness value of the change trend, which is used to identify the inflection point in the evolution of financial risks. The specific mathematical model is: ; In the formula, is the steepness value of the change trend of the th factor data at the th time step, , are respectively the risk scores of the th factor data at the and th time steps, is the absolute value operation; S53. Weight the corresponding aggregated context vectors according to the steepness value of the change trend to obtain a weighted vector sequence, and use the bidirectional long short-term memory network to obtain the overall feature representation of the th factor data. The specific mathematical model is: ; In the formula, is the overall feature representation of the th factor data at the last time step obtained through the bidirectional long short-term memory network, is the bidirectional long short-term memory network.

[0015] Preferably, by jointly inputting the risk score and its steepness value of the change trend into the bidirectional long short-term memory network, not only the current information of the risk is fused, but also the speed and fluctuation characteristics of the risk change are comprehensively considered, so as to more comprehensively reflect the evolution process of financial risks. Using the bidirectional mechanism of the bidirectional long short-term memory network, the model can effectively capture the long-term dependencies and short-term mutations in the risk sequence, and significantly enhance the ability to identify the latent time points of risks.

[0016] Preferably, in step S6, the The overall feature representations of various factor data are concatenated to form the final feature vector, and the final feature vector is input into a fully connected neural network to calculate the final financial risk prediction value. The specific mathematical model is as follows: ; In the formula, is the final financial risk prediction value, is the fully connected neural network, 、 、 are the overall feature representations of the 、 、 -th factor data respectively, is the concatenation operation.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention calculates the absolute value of the deviation between the monitored value and the historical mean in the time series of factor data, generates a high-dimensional embedded vector sequence by using a weighted non-linear function mapping, constructs a directed graph with a limited span and introduces a structural resonance factor based on the high-dimensional embedded vector sequence to obtain a path-encoded vector, enhances the ability to identify the dynamic evolution path of financial risks, uses the normalized weighted fusion of the path-encoded vectors to dynamically aggregate multi-source long-term dependence information, improves the expression accuracy of time series features, combines the weighted difference of risk scores with a bidirectional long short-term memory network to achieve sensitivity to potential risk inflection points, obtains the overall feature representation of factor data, and finally concatenates and inputs the overall feature representations of each factor data into a fully connected neural network to obtain a more accurate and stable financial risk prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flowchart of the dynamic prediction steps of financial risks based on deep learning.

[0019] Figure 2 is a diagram for generating high-dimensional embedded vectors.

[0020] Figure 3 is a diagram of the path aggregation and encoding module.

[0021] Figure 4 is a diagram of the risk evolution identification module.

[0022] Figure 5 is a diagram of the effect of dynamic prediction of financial risks. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] The present invention proposes a dynamic financial risk prediction method based on deep learning. This method constructs an absolute deviation sequence by calculating the absolute value of the deviation between the monitored value at each time step and the historical time step mean in the time series of each factor data, generates a high-dimensional embedded vector sequence through weighted non-linear function mapping, constructs a directed graph with a limited maximum span based on the high-dimensional embedded vector sequence, introduces a structural resonance factor, performs weighted encoding on multiple paths to obtain path encoding vectors, weighted fuses the path encoding vectors of multiple paths to form an aggregated context vector for each time step, calculates the risk score, further obtains the steepness value of the change trend, weights the corresponding aggregated context vector, and combines a bidirectional long short-term memory network to extract the overall feature representation. Finally, the overall feature representations of each factor data are concatenated and input into a fully connected neural network to obtain the financial risk prediction result, improving the accuracy and stability of financial risk prediction. The technical solutions in the embodiments of the present invention will be described in detail and completely below, specifically including the following steps, as Figure 1 shown.

[0024] S1. Collect the factor data affecting financial risk and construct a financial risk prediction data set.

[0025] Further, in S1, collect the factor data affecting the dynamic prediction of financial risk for 92 consecutive days, including the current ratio, asset-liability ratio, interest coverage ratio, net profit margin, inventory turnover ratio, operating cycle, and total asset turnover ratio. For missing data, use linear interpolation technology for dynamic completion to ensure the continuity and integrity of the time series input, construct a financial risk prediction data set, divide the data set into a training set and a validation set in a ratio of 7:3, and use the data of the previous 7 days at each time step as the model input.

[0026] S2. For the time series data of each factor data, calculate the absolute value of the deviation between the monitored value at each time step and the historical time step mean to obtain an absolute deviation sequence, and map the time series data and the absolute deviation sequence into a high-dimensional embedded vector sequence through a weighted non-linear function.

[0027] Further, in S2, input the time series data of the th factor data affecting financial risk , , where is the monitored value of the th factor data of financial risk at the th time step, is the length of the time series, , and use the absolute deviation to obtain a high-dimensional embedded vector sequence, as shown in Figure 2 below, specifically including the following steps: S21. Calculate the monitored value of the th factor data and its previous The absolute value of the deviation of the mean value of each time step. During the implementation process , the specific mathematical model is as follows: ; In the formula, is the absolute value of the deviation of the th factor data at the th time step. The absolute values of the deviations of all time steps of the th factor data are combined into an absolute value of deviation sequence , , where || represents the absolute value operation; S22. Perform weighted non - linear fusion on the time - series data and the absolute value of deviation sequence to obtain a high - dimensional embedding vector sequence. The specific mathematical model is as follows: ; In the formula, is the high - dimensional embedding vector sequence obtained through weighted non - linear fusion, , where is the high - dimensional embedding vector of the th factor data at the th time step, , are learnable weight matrices, is the bias term.

[0028] S3. Take each time step in the high - dimensional embedding vector sequence as a graph node, construct a directed graph with a limited maximum span, calculate the global average representation of the nodes, combine the direction consistency and position deviation measurement between nodes, introduce a structural resonance factor, construct a path aggregation encoding module, and perform weighted encoding on multiple paths to obtain a path encoding vector.

[0029] Furthermore, in S3, to construct a path aggregation encoding module, as shown in Figure 3 , the specific steps are as follows: S31. Take each time step in the high - dimensional embedding vector sequence as a graph node , where is the th factor data at the th time step node. The edge set , where is the edge between the th factor data at the th time step and the th time step nodes. is the maximum span threshold, which is set to 3 during the implementation process, restricting each time step to only connect to several historical nodes in adjacent time steps, obtaining graph ; S32. Obtain the global average representation of the data nodes of the th factor, and construct a structural resonance factor by combining the direction consistency and position deviation measurement between nodes. The specific mathematical model is as follows: ; ; In the formula, is the global average representation of the data of the th factor, is the structural resonance factor between the th factor data at the th time step and the th time step, , are resonance modulation hyperparameters. During the implementation process, is set to 0.6 for the initial value, and is set to 0.4 for the initial value during the implementation process, is the L2 norm, , are the th factor data nodes at the th and th time steps respectively, is a minimum value to prevent the denominator from being zero. The initial value is set to 0.000001 during the implementation process; S33. Denote the th factor data node at the th time step as the end point of the th path as . The path contains a series of node indices from the th time step to the end time step . , perform structural weighted encoding on . The specific mathematical model is as follows: ; In the formula, is the path encoding vector of the th factor data for the th path, is the structural resonance factor between the th factor data at the th time step and the th time step, is the structural resonance factor between the th factor data at the th time step and the th time step, is the th factor data node at the Time step node is the type of factor data at the time step node is Hadamard element multiplication

[0030] S4. Calculate the normalized weights of the path encoding vectors, and fuse multiple path encoding vectors by weighting according to the weights to form an aggregated context vector for each time step

[0031] Furthermore, in S4, dynamically allocate and fuse multiple path encoding vectors to form an aggregated context vector for each time step, including the following steps S41. Normalize each path encoding vector to obtain path weights to significantly highlight more important paths. The specific mathematical model is ; In the formula, is the normalized path weight of the th path, ensuring that the sum of all path weights is 1 is the type of factor data at the th path encoding vector of the is the number of paths with the time step node as the end point S42. According to the normalized path weights, perform weighted summation on the multiple path encoding vectors selected at the time step to form an aggregated context vector. The specific mathematical model is ; In the formula, is the type of factor data at the time step aggregated context vector

[0032] S5. Based on the aggregated context vector, calculate the risk score using a linear mapping and a Sigmoid activation function, further obtain the steepness value of the change trend using weighted difference, weight the corresponding aggregated context vector using the steepness value of the change trend, and combine with a bidirectional long short-term memory network to construct a risk evolution recognition module to obtain the overall feature representation of the factor data

[0033] Furthermore, in S5, construct a risk evolution recognition module, as shown in Figure 4 including the following steps S51. Calculate the risk score through a single-layer linear mapping and a Sigmoid activation function. The specific mathematical model is ; In the formula, is the The time step risk score is the linear mapping weight matrix and is the bias vector; ; In the formula, is the change trend steepness value of the th factor data at the th time step, , are the risk scores of the th factor data at the th and th time steps respectively, is the absolute value operation; S53. Weight the corresponding aggregated context vectors according to the change trend steepness value to obtain a weighted vector sequence, and use a bidirectional long short-term memory network to obtain the overall feature representation of the th factor data. The specific mathematical model is: ; In the formula, is the overall feature representation of the th factor data at the last time step obtained by the bidirectional long short-term memory network, is the bidirectional long short-term memory network.

[0034] S6. Concatenate the overall feature representations of all factor data, input them into a fully connected neural network, and calculate the final financial risk prediction value.

[0035] Furthermore, in the above S6, concatenate the overall feature representations of the collected factor data to form the final feature vector, input the final feature vector into a fully connected neural network, and calculate the final financial risk prediction value. The specific mathematical model is: ; In the formula, is the final financial risk prediction value, is the fully connected neural network, , , are the overall feature representations of the th, th, th factor data respectively, is the concatenation operation.

[0036] Furthermore, the model is implemented using the Python 3.8 programming language and the PyTorch framework, and runs in the CUDA 11.3 environment. When training, an NVIDIA 3090 24GB GPU is used to accelerate the calculation process. During the training process, to ensure efficiency and stability, the initial learning rate is , the batch size is , and the loss function is loss function. The optimizer is used to iteratively update the parameters.

[0037] Furthermore, the prediction effect of this method is as Figure 5 shown, where the ordinate is the financial risk value (%), and the abscissa is the time (days). The gray dashed line in the figure represents the actual evaluation value of the financial risk, and the black solid line is the financial risk prediction value generated by this method. It can be seen from Figure 5 that the difference between the prediction value and the actual evaluation value is small, indicating that there is a good fitting effect between the prediction value and the actual evaluation value, and it shows that the proposed method has high accuracy and stability in the task of dynamic prediction of financial risks.

[0038] The above is only the preferred embodiment of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the creative concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A dynamic financial risk prediction method based on deep learning, characterized in that, It includes the following steps: S1. Collect the factor data affecting financial risks and construct a financial risk prediction data set; S2. For the time series data of each factor data, calculate the absolute value of the deviation between the monitoring value at each time step and the mean value of historical time steps to obtain an absolute deviation value sequence, and map the time series data and the absolute deviation value sequence into a high-dimensional embedded vector sequence through a weighted non-linear function; S3. Take each time step in the high-dimensional embedded vector sequence as a graph node, construct a directed graph with a limited maximum span, calculate the global average representation of the nodes, combine the direction consistency and position deviation measurement between nodes, introduce a structural resonance factor, construct a path aggregation coding module, perform weighted coding on multiple paths, and obtain a path coding vector; S4. Calculate the normalized weights of the path coding vectors, and weighted-fuse multiple path coding vectors according to the weights to form an aggregated context vector for each time step; S5. Based on the aggregated context vector, calculate the risk score using a linear mapping and a Sigmoid activation function, further obtain the steepness value of the change trend using weighted difference, weight the corresponding aggregated context vector using the steepness value of the change trend, and combine with a bidirectional long short-term memory network to construct a risk evolution recognition module to obtain the overall feature representation of the factor data; S6. Concatenate the overall feature representations of all factor data, input them into a fully connected neural network, and calculate the final financial risk prediction value.

2. The financial risk dynamic prediction method based on deep learning according to claim 1, characterized in that In S1, collect the factor data affecting the dynamic prediction of financial risks, including the current ratio, asset-liability ratio, interest coverage ratio, net profit ratio, inventory turnover ratio, operating cycle, total asset turnover ratio. For missing data, use linear interpolation technology for dynamic filling to construct a financial risk prediction data set.

3. A dynamic financial risk prediction method based on deep learning according to claim 2, characterized in that In S2, input the time series data of the data of the th factor affecting financial risk , where is the monitoring value of the data of the th factor of financial risk at the th time step, is the length of the time series, and a high-dimensional embedding vector sequence is obtained by using the absolute value of the deviation, which specifically includes the following steps: S21. Calculate the absolute value of the deviation of the monitored value of the -th factor data from the mean value in the previous time steps, and combine the absolute values of the deviations of all time steps of the -th factor data into an absolute value of deviation sequence ; S22. Multiply the time series data and the absolute deviation sequence by their respective weight matrices respectively, then perform weighted combination, combine the bias term, and after being processed by the activation function, obtain the high-dimensional embedding vector sequence , where is the high-dimensional embedding vector of the th factor data at the th time step.

4. A dynamic financial risk prediction method based on deep learning according to claim 3, characterized in that In S3, to construct a path aggregation coding module, it specifically includes the following steps: S31. Take each time step in the high-dimensional embedding vector sequence as a graph node , where is the th factor data at the th time step node, and the edge set , where is the edge between the th factor data at the th time step and the th time step node. is the maximum span threshold, which restricts each time step to only connect to several historical nodes in adjacent time steps, obtaining the graph ; S32. Calculate the global average representation of the data nodes of the th factor , and construct a structural resonance factor by combining the direction consistency and position deviation measurement between nodes. The specific mathematical model is as follows: ； In the formula, is the global average representation of the th factor data, is the th factor data at the th time step and the th time step, which is the structure resonance factor between them, , are resonance modulation hyperparameters, is the L2 norm, , are respectively the th factor data at the th and th time step nodes, is a minimum value to prevent the denominator from being zero; S33. Denote the th factor data, and the th path ending at the time step node is . The path contains a series of node indices from time step to the end time step . . Calculate the weighting coefficients between adjacent nodes on the path, and perform structural weighted coding on . The specific mathematical model is: ​ ； In the formula, is the path coding vector of the th factor data for the th path, is the structural resonance factor between the th factor data at the th time step and the th time step, is the structural resonance factor between the th factor data at the th time step and the th time step, is the th factor data at the th time step node, is the th factor data at the th time step node, is the Hadamard element multiplication.

5. A dynamic financial risk prediction method based on deep learning according to claim 4, characterized in that In S4, for dynamic weight allocation and fusion of multiple path coding vectors to form an aggregated context vector for each time step, it includes the following steps: S41. Perform exponential weighting on the L2 norm of the path encoding vector with the time step node as the end point path, and perform normalization. The specific mathematical model is as follows: ； Wherein, is the normalized path weight of the th path, is the path encoding vector of the th factor data for the th path, is the number of paths with the th time step node as the end point; S42. According to the normalized path weights, perform a weighted sum on the encoded vectors of multiple paths selected at the time step to form an aggregated context vector. The specific mathematical model is as follows: ； In the formula, is the aggregated context vector of the -th factor data at the -th time step.

6. A dynamic financial risk prediction method based on deep learning according to claim 5, characterized in that In S5, to construct a risk evolution recognition module, it includes the following steps: S51. Calculate the risk score through a single-layer linear mapping and a sigmoid activation function to obtain the risk score of the th factor data at the th time step ; S52. Calculate the weighted difference between the risk scores at the current and adjacent time steps, comprehensively consider the second-order difference and the first-order change amplitude to obtain the steepness value of the change trend. The specific mathematical model is: ； In the formula, is the steepness value of the change trend of the th factor data at the th time step, , are the risk scores of the th factor data at the th and th time steps respectively, is the absolute value operation; S53. Weight the corresponding aggregated context vectors according to the steepness value of the change trend to obtain a weighted vector sequence, and use a bidirectional long short-term memory network to obtain the overall feature representation of the data of the th factor. The specific mathematical model is as follows: ； In the formula, is the overall feature representation of the th factor data at the last time step obtained by the bidirectional long short-term memory network, is the bidirectional long short-term memory network.

7. A dynamic financial risk prediction method based on deep learning according to claim 6, characterized in that In S6, the overall characteristics representations of the types of factor data are spliced to form the final feature vector, and the final feature vector is input into a fully connected neural network to calculate the final financial risk prediction value .

Images