BILSTM-DBN-based enterprise financial result prediction model establishment method
By adopting the BILSTM-DBN model in enterprise financial data processing, combining the advantages of BiLSTM and DBN, the shortcomings of traditional models in financial data prediction are solved, and higher prediction accuracy and reliability are achieved.
Patent Information
- Application Number
- CN202510163129.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-14
- Publication Date
- 2025-06-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
When traditional machine learning models process corporate financial time series data, the prediction effect is poor and cannot make timely and accurately predict financial data.
The enterprise financial results prediction model based on BILSTM-DBN is adopted, feature extraction and deep feature learning are performed through BiLSTM, and multi-level feature integration is combined with DBN, model training and optimization, especially in the hyperparameter adjustment process, and refined training is carried out for the specific financial situation of the enterprise.
It improves the accuracy and reliability of the prediction model, can effectively capture the long-term dependencies and complex characteristics of financial data, enhances the generalization ability when facing new data, and reduces the risk of overfitting.
Smart Images
Figure CN120106997A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of time series data processing, and in particular relates to a method for establishing a BILSTM-DBN enterprise financial result prediction model. Background Art
[0002] With the development of big data and artificial intelligence technology, more and more companies are beginning to use data-driven decision-making methods to predict financial results. This method can not only improve the accuracy of predictions, but also provide companies with deeper business insights. However, since corporate financial data usually has the characteristics of time series, traditional machine learning models often have poor prediction results when processing such data and cannot make timely and accurate financial data predictions. In order to solve this problem, we propose a corporate financial results prediction model based on a bidirectional long short-term memory network BiLSTM and DBN. This model combines the advantages of BiLSTM in processing time series data and the deep learning capabilities of DBN in feature extraction, thereby providing companies with a more accurate and reliable financial forecasting tool. Summary of the invention
[0003] The purpose of the present invention is to provide a method for establishing a corporate financial results prediction model based on BILSTM-DBN. By using BiLSTM to extract features in the S2 stage and combining it with DBN for deep feature learning, the long-term dependencies and complex features of financial data can be effectively captured, and the accuracy of the prediction model can be improved. Through model training and optimization in the S4 stage, especially in the hyperparameter adjustment link, the model can be refined for the specific financial status of the enterprise, further enhancing the reliability of the prediction, and solving the problem that traditional machine learning models often have poor prediction effects when processing such data and cannot make timely and accurate financial data predictions.
[0004] In order to solve the above technical problems, the present invention is achieved through the following technical solutions:
[0005] The present invention provides a method for establishing a BILSTM-DBN enterprise financial result prediction model, comprising the following steps:
[0006] S1: Data collection and preprocessing; S2: BiLSTM feature extraction; S3: DBN feature learning; S4: Model training and optimization; S5: Model evaluation and verification; S6: Model prediction application;
[0007] The S2 includes the following sub-steps; S21, input preparation: arrange the pre-processed financial data in chronological order to ensure that each time point corresponds to a complete set of financial indicators, including key items in the balance sheet, income statement and cash flow statement. Since the BiLSTM model processes sequence data, it is necessary to combine the financial indicators at each time point into a vector, and connect these vectors in chronological order to form time series data; S22, BiLSTM network construction: design a bidirectional long short-term memory network, which contains multiple LSTM layers, each layer consists of a forward LSTM and a backward LSTM. The forward LSTM processes the sequence from left to right, and the backward LSTM processes the sequence from right to left. The outputs of the two are merged at each time step to capture the bidirectional information in the sequence; initialize the weights and biases of the BiLSTM network, usually using small random numbers to ensure that different neurons learn different features; S23, feature extraction: input the time series data into the BiLSTM network, and calculate the activation value of the hidden layer through the forward propagation of the network. In this process, each LSTM unit updates its internal state according to the current input and the state of the previous moment, thereby capturing the long-term dependencies in the sequence. At the same time, the backward LSTM also processes the sequence, but in the opposite direction, from right to left, so that the backward LSTM can capture future information in the sequence and combine it with the information of the forward LSTM to provide a more comprehensive feature representation. After the forward and backward LSTM processing is completed, their outputs are merged at each time step to obtain a hidden layer representation containing bidirectional information, which is then used for subsequent feature learning and prediction tasks; S24, output conversion: Since the output of BiLSTM is usually a high-dimensional feature representation, directly using it as the input of the DBN model may result in excessive computational complexity. Therefore, these features need to be reduced in dimensionality. Principal component analysis or autoencoder methods can be used to convert the reduced dimensionality features into fixed-length vectors so that the DBN model can process them. This can be achieved through pooling operations to ensure that the converted vector format matches the input requirements of the DBN model, including data type, shape, and range.
[0008] Furthermore, the S1 includes the following steps: S11, data collection: collect the company's historical financial data from multiple sources such as internal databases, public financial statements, industry reports, and news releases to ensure the diversity and comprehensiveness of the data in order to obtain more accurate analysis results, with a focus on collecting key financial reports such as balance sheets, income statements, and cash flow statements; S12, data cleaning: check the missing values in the data and select appropriate processing methods based on the situation. For a small number of missing values, the mean, median, or mode can be used to fill in. For a large number of missing values, it may be necessary to delete related records or use more complex interpolation methods; identify and process outliers, which may be caused by entry errors, measurement errors, or extreme events. Commonly used methods include box plot analysis and Z-score standardization and screening; and it is also necessary to ensure data consistency.
[0009] Furthermore, the S3 includes the following sub-steps; S31, DBN network construction: DBN is a generative model, which is stacked by multiple layers of restricted Boltzmann machines (RBM). Each layer of RBM is an undirected graph model, including a visible layer and a hidden layer, which is used to learn the feature representation of data. DBN is usually composed of multiple RBM layers, and each RBM layer can learn the high-level feature representation of the input data. This multi-layer structure enables DBN to extract complex features in the data layer by layer; S32, feature mapping: The features extracted by BiLSTM are used as the input of DBN. These features already contain the long-term trend and short-term fluctuation information in the time series, which is the basis for further learning. For each layer of RBM, unsupervised pre-training is performed. This process includes contrastive divergence (CD) algorithm or other optimization methods to adjust the weights of RBM so that the model can capture the high-level features in the data. After completing the training of one layer of RBM, its output is used as the input of the next layer of RBM, and training continues. In this way, layer by layer is stacked, and each layer learns deeper features based on the previous layer.
[0010] Further, the S4 includes the following sub-steps; S41, model training: select a suitable loss function according to the task type. For regression problems, the commonly used loss function is the mean square error, which measures the average square difference between the predicted value and the actual value; for classification problems, cross entropy loss is a more commonly used choice, which measures the difference between the predicted probability distribution and the actual label distribution. Use gradient descent or other optimization algorithms to minimize the loss function. Gradient descent calculates the gradient of the loss function with respect to the model parameters and updates the parameters in the opposite direction of the gradient to gradually approach the optimal solution. Common optimization algorithms include stochastic gradient descent, momentum method and Adam. Input known corporate financial data into the BiLSTM and DBN models, calculate the prediction results through forward propagation, and then calculate the loss value according to the loss function. Then, calculate the gradient of the loss function with respect to the model parameters through back propagation, and use the optimization algorithm to update the parameters. Repeat this process multiple times until the loss function converges or reaches a preset number of training rounds; S42. Cross-validation: Divide the dataset into a training set and a validation set. The commonly used division ratio is 80% for training and 20% for validation. You can also use k-fold cross-validation to divide the dataset into k subsets, use k-1 subsets as training sets each time, and the remaining subset as validation set, and repeat K times; in each training and validation process, record the performance indicators of the model on the validation set. These indicators can help us understand the generalization ability of the model, that is, its performance on new data. Through multiple training and validation, we can obtain the mean and standard deviation of the model performance; S43. Hyperparameter adjustment: Use grid search to find the optimal configuration by traversing different hyperparameter combinations. You can try different learning rates, batch sizes, and numbers of iterations. For each combination, perform a complete training and validation process and record performance indicators; Random search is to randomly select several combinations from a predefined range for trial. This method is usually faster than grid search, but may miss some local optimal solutions.
[0011] Furthermore, the S5 includes the following sub-steps; S51, performance evaluation: use an independent test set to evaluate the performance of the model, and this test set has not been used in the training and verification process, so as to ensure the authenticity and reliability of the evaluation results and avoid the risk of overfitting. Select appropriate evaluation indicators according to the task type. For classification problems, commonly used evaluation indicators include accuracy, recall, precision and F1 score; for regression problems, commonly used evaluation indicators include mean square error, mean absolute error and determination coefficient. These indicators can reflect the predictive ability of the model from different angles; S52, confusion matrix analysis: the confusion matrix can be used to analyze the performance of the model in different categories. The confusion matrix is a table in which rows represent actual categories and columns represent predicted categories. By looking at the confusion matrix, we can understand the true number of examples, false positive examples, false negative examples and true negative examples of the model in each category. By analyzing the confusion matrix, we can identify possible misclassification patterns of the model. In order to better understand the information in the confusion matrix, we can visualize it as a heat map or bar chart.
[0012] Furthermore, the S6 includes the following sub-steps: S61, prediction application: input the newly collected corporate financial data into the trained model, which includes the financial statements, market conditions, and industry trends of the most recent quarter or year, and ensure that the format and characteristics of the data are consistent with the training data so that the model can process it correctly, and use the model to predict the newly input data. Depending on the type of model, this involves forward propagation and back propagation. For time series prediction models, it is necessary to perform rolling predictions on historical data to obtain prediction results for a period of time in the future; output the prediction results of the model in an appropriate form. For classification problems, the output may be a probability distribution of each category; for regression problems, the output may be a specific numerical prediction; S62, result interpretation: output the prediction results of the model in an appropriate form. For classification problems, the output may be a probability distribution of each category; for regression problems, the output may be a specific numerical prediction, and some auxiliary information may be provided to analyze the potential risks in the prediction results. If the model predicts that the market will fluctuate significantly, then the company may need to formulate a risk management plan in advance. In addition, it is necessary to consider the uncertainty of the model itself and provide decision support to management based on the prediction results, which may involve formulating new strategic directions, adjusting existing policies, and optimizing resource allocation.
[0013] The present invention has the following beneficial effects:
[0014] 1. The present invention improves prediction accuracy and reliability. By using BiLSTM for feature extraction in the S2 stage and combining DBN for deep feature learning, it can effectively capture the long-term dependencies and complex features of financial data, thereby improving the accuracy of the prediction model. Through model training and optimization in the S4 stage, especially in the hyperparameter adjustment link, the model can be refined for the specific financial status of the enterprise, further enhancing the reliability of the prediction.
[0015] 2. Through the multi-level feature integration of DBN in the S3 stage, the model of the present invention can learn more abstract and high-level features, which not only enhances the generalization ability of the model when facing new data, but also enables it to better adapt to the changes in financial data of enterprises in different industries or sizes. The performance evaluation and fine-tuning in the S5 stage ensure that the model has a robust performance on the test set, thereby reducing the risk of overfitting.
[0016] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.
[0018] Figure 1 A flowchart of a method for establishing a BILSTM-DBN enterprise financial result prediction model according to the present invention;
[0019] Figure 2 It is a schematic diagram of the process of BILSTM-DBN feature extraction of the present invention. DETAILED DESCRIPTION
[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0021] See also Figure 1 As shown, the present invention is a method for establishing a BILSTM-DBN enterprise financial result prediction model, comprising the following steps:
[0022] S1: Data collection and preprocessing; S2: BiLSTM feature extraction; S3: DBN feature learning; S4: Model training and optimization; S5: Model evaluation and verification; S6: Model prediction application;
[0023] S2 includes the following sub-steps; S21, input preparation: arrange the pre-processed financial data in chronological order to ensure that each time point corresponds to a complete set of financial indicators, including key items in the balance sheet, income statement and cash flow statement. Since the BiLSTM model processes sequence data, it is necessary to combine the financial indicators at each time point into a vector and connect these vectors in chronological order to form time series data; S22, BiLSTM network construction: design a bidirectional long short-term memory network, which contains multiple LSTM layers, each layer consists of a forward LSTM and a backward LSTM. The forward LSTM processes the sequence from left to right, and the backward LSTM processes the sequence from right to left. The outputs of the two are merged at each time step to capture the bidirectional information in the sequence; initialize the weights and biases of the BiLSTM network, usually using small random numbers to ensure that different neurons learn different features; S23, feature extraction: input the time series data into the BiLSTM network, and calculate the hidden through the forward propagation of the network The activation value of the layer. In this process, each LSTM unit will update its internal state according to the current input and the state of the previous moment, so as to capture the long-term dependencies in the sequence. At the same time, the backward LSTM will also process the sequence, but in the opposite direction, from right to left, so that the backward LSTM can capture the future information in the sequence and combine it with the information of the forward LSTM to provide a more comprehensive feature representation. After the forward and backward LSTM processing is completed, their outputs are merged at each time step to obtain a hidden layer representation containing bidirectional information. These representations are then used for subsequent feature learning and prediction tasks; S24, output conversion: Since the output of BiLSTM is usually a high-dimensional feature representation, directly using it as the input of the DBN model may result in excessive computational complexity. Therefore, these features need to be reduced in dimensionality. The principal component analysis or autoencoder method can be used to convert the reduced dimensionality features into fixed-length vectors so that the DBN model can process them. This can be achieved through pooling operations to ensure that the converted vector format matches the input requirements of the DBN model.
[0024] S1 includes the following steps: S11. Data collection: Collect the company's historical financial data from multiple sources such as internal databases, public financial statements, industry reports, and news releases to ensure the diversity and comprehensiveness of the data in order to obtain more accurate analysis results, with a focus on collecting key financial reports such as balance sheets, income statements, and cash flow statements; S12. Data cleaning: Check the missing values in the data and select appropriate processing methods based on the situation. For a small number of missing values, the mean, median, or mode can be used to fill in. For a large number of missing values, it may be necessary to delete related records or use more complex interpolation methods; and it is also necessary to ensure the consistency of the data.
[0025] S3 includes the following sub-steps; S31, DBN network construction: DBN is a generative model, which is composed of multiple layers of restricted Boltzmann machines (RBMs) stacked together. Each layer of RBM is an undirected graph model, which contains a visible layer and a hidden layer, and is used to learn the feature representation of data. DBN is usually composed of multiple RBM layers, and each RBM layer can learn the high-level feature representation of the input data. This multi-layer structure enables DBN to extract complex features in the data layer by layer; S32, feature mapping: The features extracted by BiLSTM are used as the input of DBN. These features already contain the long-term trend and short-term fluctuation information in the time series, which is the basis for further learning. For each layer of RBM, unsupervised pre-training is performed. This process includes the contrastive divergence (CD) algorithm or other optimization methods to adjust the weights of the RBM so that the model can capture the high-level features in the data. After completing the training of one layer of RBM, its output is used as the input of the next layer of RBM, and training continues. In this way, layer by layer is stacked, and each layer learns deeper features based on the previous layer.
[0026] S4 includes the following sub-steps; S41, model training: select an appropriate loss function according to the task type. For regression problems, the commonly used loss function is the mean square error, which measures the average square difference between the predicted value and the actual value; for classification problems, cross entropy loss is a more commonly used choice, which measures the difference between the predicted probability distribution and the actual label distribution. Use gradient descent or other optimization algorithms to minimize the loss function. Gradient descent calculates the gradient of the loss function with respect to the model parameters and updates the parameters in the opposite direction of the gradient to gradually approach the optimal solution. Common optimization algorithms include stochastic gradient descent, momentum method and Adam. Input known corporate financial data into the BiLSTM and DBN models, calculate the prediction results through forward propagation, and then calculate the loss value according to the loss function. Then, calculate the gradient of the loss function with respect to the model parameters through back propagation, and use the optimization algorithm to update the parameters. Repeat this process many times until the loss function converges or reaches the preset number of training rounds; S42 , Cross-validation: Divide the dataset into a training set and a validation set. The commonly used division ratio is 80% for training and 20% for validation. You can also use k-fold cross-validation to divide the dataset into k subsets, use k-1 subsets as training sets each time, and the remaining subset as validation set, and repeat K times; in each training and validation process, record the performance indicators of the model on the validation set. These indicators can help us understand the generalization ability of the model, that is, its performance on new data. Through multiple training and validation, we can obtain the mean and standard deviation of the model performance; S43. Hyperparameter adjustment: By traversing different hyperparameter combinations, use grid search to find the optimal configuration. You can try different learning rates, batch sizes, and number of iterations. For each combination, perform a complete training and validation process, and record performance indicators; Random search is to randomly select several combinations from a predefined range for trial. This method is usually faster than grid search, but may miss some local optimal solutions.
[0027] S5 includes the following sub-steps; S51. Performance evaluation: Use an independent test set to evaluate the performance of the model. This test set has not been used in the training and validation process. This can ensure the authenticity and reliability of the evaluation results and avoid the risk of overfitting. Select appropriate evaluation indicators according to the task type. For classification problems, commonly used evaluation indicators include accuracy, recall, precision and F1 score; for regression problems, commonly used evaluation indicators include mean square error, mean absolute error and determination coefficient. These indicators can reflect the predictive ability of the model from different angles; S52. Confusion matrix analysis: The confusion matrix can be used to analyze the performance of the model in different categories. The confusion matrix is a table. By looking at the confusion matrix, we can understand the true number of examples, false positive examples, false negative examples and true negative examples of the model in each category. By analyzing the confusion matrix, we can identify possible misclassification patterns of the model. In order to better understand the information in the confusion matrix, we can visualize it as a heat map or bar chart.
[0028] S6 includes the following sub-steps: S61. Forecast application: Input the newly collected corporate financial data into the trained model. These data include the financial statements, market conditions, and industry trends of the most recent quarter or year. Ensure that the format and characteristics of the data are consistent with the training data so that the model can process it correctly. Use the model to predict the newly input data. Depending on the type of model, this involves forward propagation and back propagation. For time series prediction models, rolling predictions of historical data are required to obtain prediction results for a period of time in the future; output the prediction results of the model in an appropriate form. For classification problems, the output may be the probability distribution of each category; for regression problems, the output may be a specific numerical prediction; S62. Result interpretation: Output the prediction results of the model in an appropriate form. For classification problems, the output may be the probability distribution of each category; for regression problems, the output may be a specific numerical prediction. Some auxiliary information can also be provided to analyze the potential risks in the prediction results. In addition, the uncertainty of the model itself needs to be considered, and decision support is provided to management based on the prediction results. This may involve formulating new strategic directions, adjusting existing policies, and optimizing resource allocation.
[0029] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.
[0030] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.
Claims
1. A method for establishing a BILSTM-DBN enterprise financial results prediction model, characterized by: The steps include: S1: data collection and preprocessing; S2: BiLSTM feature extraction; S3: DBN feature learning; S4: Model training and optimization; S5: Model evaluation and validation; S6: Model prediction application; The S2 includes the following sub-steps: S21, input preparation: Arrange the pre-processed financial data in chronological order to ensure that each time point corresponds to a complete set of financial indicators, including key items in the balance sheet, income statement and cash flow statement. Since the BiLSTM model processes sequence data, it is necessary to combine the financial indicators at each time point into a vector, and connect these vectors in chronological order to form time series data; S22. BiLSTM network construction: Design a bidirectional long short-term memory network, which contains multiple LSTM layers. Each layer consists of a forward LSTM and a backward LSTM. The forward LSTM processes the sequence from left to right, and the backward LSTM processes the sequence from right to left. The outputs of the two are merged at each time step to capture the bidirectional information in the sequence. Initialize the weights and biases of the BiLSTM network, usually using small random numbers to ensure that different neurons learn different features. S23, feature extraction: input the time series data into the BiLSTM network, and calculate the activation value of the hidden layer through the forward propagation of the network. In this process, each LSTM unit will update its internal state according to the current input and the state of the previous moment, so as to capture the long-term dependency in the sequence. At the same time, the backward LSTM will also process the sequence, but in the opposite direction, from right to left, so that the backward LSTM can capture the future information in the sequence and combine it with the information of the forward LSTM to provide a more comprehensive feature representation. After the forward and backward LSTM processing is completed, their outputs are merged at each time step to obtain the hidden layer representation containing bidirectional information; S24. Output conversion: Since the output of BiLSTM is usually a high-dimensional feature representation, directly using it as the input of the DBN model may result in excessive computational complexity. Therefore, these features need to be reduced in dimensionality. Principal component analysis or autoencoder methods can be used to convert the reduced-dimensional features into vectors of fixed length so that the DBN model can process them. This can be achieved through pooling operations to ensure that the converted vector format matches the input requirements of the DBN model.
2. According to claim 1, a method for establishing a BILSTM-DBN enterprise financial result prediction model is characterized in that: The S1 includes the following steps: S11, data collection: collect the company's historical financial data from a variety of sources including the company's internal database, public financial statements, industry reports, and news releases to ensure the diversity and comprehensiveness of the data in order to obtain more accurate analysis results, with a focus on collecting key financial reports such as the balance sheet, income statement, and cash flow statement; S12. Data cleaning: Check the missing values in the data and choose the appropriate processing method according to the situation. For a small number of missing values, you can use the mean, median or mode to fill in. For a large number of missing values, you may need to delete the relevant records; identify and handle outliers, and also ensure the consistency of the data.
3. The method for establishing a BILSTM-DBN enterprise financial result prediction model according to claim 1, characterized in that: The S3 includes the following sub-steps; S31, DBN network construction: DBN is a generative model, which is stacked by multiple layers of restricted Boltzmann machines. Each layer of RBM is an undirected graph model, including a visible layer and a hidden layer, which is used to learn the feature representation of data. DBN is usually composed of multiple RBM layers, and each RBM layer can learn the high-level feature representation of input data. This multi-layer structure enables DBN to extract complex features in the data layer by layer; S32. Feature mapping: The features extracted by BiLSTM are used as the input of DBN. These features already contain the long-term trend and short-term fluctuation information in the time series, which is the basis for further learning. For each layer of RBM, unsupervised pre-training is performed. This process includes contrastive divergence algorithm or other optimization methods to adjust the weights of RBM so that the model can capture the high-level features in the data. After completing the training of one layer of RBM, its output is used as the input of the next layer of RBM to continue training. In this way, each layer learns deeper features based on the previous layer.
4. The method for establishing a BILSTM-DBN enterprise financial result prediction model according to claim 1, characterized in that: The S4 includes the following sub-steps; S41, model training: select a suitable loss function according to the task type. For regression problems, the commonly used loss function is the mean square error, which measures the average square difference between the predicted value and the actual value; for classification problems, cross entropy loss is a more commonly used choice, which measures the difference between the predicted probability distribution and the actual label distribution. Use gradient descent or other optimization algorithms to minimize the loss function. Gradient descent calculates the gradient of the loss function with respect to the model parameters and updates the parameters in the opposite direction of the gradient to gradually approach the optimal solution. Input known corporate financial data into the BiLSTM and DBN models, calculate the prediction results through forward propagation, and then calculate the loss value according to the loss function. Then, calculate the gradient of the loss function with respect to the model parameters through back propagation, and use the optimization algorithm to update the parameters. Repeat this process multiple times until the loss function converges or reaches a preset number of training rounds; S42, cross validation: The data set is divided into a training set and a validation set. The commonly used division ratio is 80% for training and 20% for validation. You can also use k-fold cross validation to divide the data set into k subsets, each time using k-1 subsets as the training set and the remaining subset as the validation set, repeating K times; S43. Hyperparameter adjustment: Use grid search to find the optimal configuration by traversing different hyperparameter combinations. You can try different learning rates, batch sizes, and numbers of iterations. For each combination, a complete training and validation process is performed, and performance indicators are recorded. Random search randomly selects several combinations from a predefined range for trial. This method is usually faster than grid search, but may miss some local optimal solutions.
5. The method for establishing a BILSTM-DBN enterprise financial result prediction model according to claim 1, characterized in that: The S5 includes the following sub-steps: S51, performance evaluation: using an independent test set to evaluate the performance of the model, which has not been used in the training and verification process, so as to ensure the authenticity and reliability of the evaluation results and avoid the risk of overfitting. According to the task type, appropriate evaluation indicators are selected, which can reflect the prediction ability of the model from different angles; S52. Confusion matrix analysis: The confusion matrix can be used to analyze the performance of the model on different categories. The confusion matrix is a table in which rows represent actual categories and columns represent predicted categories. By looking at the confusion matrix, we can understand the number of true positive examples, false positive examples, false negative examples, and true negative examples of the model in each category. By analyzing the confusion matrix, we can identify possible misclassification patterns of the model. In order to better understand the information in the confusion matrix, we can visualize it in the form of a heat map or a bar chart.
6. The method for establishing a BILSTM-DBN enterprise financial result prediction model according to claim 1, characterized in that: The S6 includes the following sub-steps: S61, forecast application: input the newly collected enterprise financial data into the trained model, ensure that the format and characteristics of the data are consistent with the training data so that the model can process it correctly, and use the model to predict the newly input data. Depending on the type of model, this involves forward propagation and back propagation. For time series forecasting models, rolling forecasts of historical data are required to obtain forecast results in the future; S62. Result interpretation: Output the model's prediction results in an appropriate form. For classification problems, the output may be the probability distribution of each category; for regression problems, the output may be a specific numerical prediction. It can also provide some auxiliary information to analyze the potential risks in the prediction results. In addition, it is necessary to consider the uncertainty of the model itself and provide decision support to the management based on the prediction results. This may involve formulating new strategic directions, adjusting existing policies, and optimizing resource allocation.