Method and system for predicting corporate financial distress

Abstract

The present invention relates to the technical field of financial distress, and specifically relates to a method and system for predicting corporate financial distress. The method comprises: on the basis of data features, generating a training set and a test set; on the basis of the data features in the generated training set, introducing Pearson correlation analysis to analyze the correlation between the features; aggregating the data features selected by means of Pearson correlation analysis, and on the basis of the aggregated data features, establishing a training model for predicting corporate financial distress; and introducing penalized logistic regression to perform training. In the present invention, by means of screening and extracting financial factor data and non-financial factor data of an enterprise over the years, the diversity of data used for predicting corporate financial distress is increased, thereby improving the accuracy of predicting corporate financial distress.

Description

A method and system for predicting corporate financial distress Technical Field

[0001] This invention relates to the field of financial distress technology, and in particular to a method and system for predicting corporate financial distress. Background Technology

[0002] Financial distress refers to a company facing serious financial problems, which may prevent it from repaying debts on time or bearing operating costs, and could lead to bankruptcy, liquidation, or other forms of financial crisis. Financial distress typically manifests as declining profitability, cash flow problems, and an unbalanced asset-liability structure, among other signs. It not only affects a company's operational capabilities but also increases its credit risk and can even impact its stock price and market image. Technical issues

[0003] Currently, the assessment of a company's financial distress often requires a year-end summary of its financial data. As a result, companies often experience a lag in understanding their financial difficulties, failing to reflect their true financial situation and operational risks in a timely manner. Furthermore, in the context of a rapidly changing market environment, relying solely on traditional classifiers for data analysis often fails to cope with the complexity and diversity of data in practical applications. Moreover, this method is not accurate or effective enough when dealing with multi-dimensional, non-linear, and heterogeneous corporate data. Technical solutions

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method and system for predicting corporate financial distress.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for predicting corporate financial distress, the method comprising the following steps:

[0006] (i) Collect the necessary historical financial and non-financial data of enterprises through various means, and compile the collected data into a dataset for predicting corporate financial distress;

[0007] (ii) Based on the content gathered in the dataset, the dataset content is processed according to the required data features. The generated data features are divided into financial distress features, financial factor features and non-financial factor features. The generated financial distress features, financial factor features and non-financial factor features are then gathered into training set and test set according to their intended use.

[0008] (iii) Based on the characteristics of the generated training set data, Pearson correlation analysis is introduced to analyze the correlation between features in order to identify whether there are highly correlated features, so as to selectively select features. The data features selected by Pearson correlation analysis are collected and used as a basis to build a training model for predicting corporate financial distress. Penalized logistic regression is introduced for training. The regularization process will further help the model handle the complexity of these features.

[0009] (iv) Based on the trained model for predicting corporate financial distress, test set data features are introduced for testing. The test results are compared with the collected results, and the prediction accuracy is evaluated using a prediction accuracy evaluation method to determine the predictive effect of the model.

[0010] As a further aspect of the present invention, the methods for collecting the enterprise's historical financial and non-financial data include the CSMAR database, color analysis of listed companies' annual reports, and the CNRDS database.

[0011] As a further aspect of the present invention, the data features include financial distress features, financial factor features, and non-financial factor features;

[0012] The financial distress characteristics mentioned above are used to determine the financial distress situation of an enterprise. The determination formula is ST / ST*stock = 1, otherwise 0.

[0013] As a further aspect of the present invention, the financial factors include solvency, liquidity, profitability, and market valuation.

[0014] The solvency ratio includes the debt-to-equity ratio, cash ratio, equity ratio, current ratio, and capital expenditure ratio.

[0015] The liquidity metrics include operating index, equity free cash flow, corporate free cash flow, total cash recovery rate, and net cash content of operating income.

[0016] The profitability includes return on assets, return on equity, return on invested capital, net operating profit margin, and effective tax rate.

[0017] The market valuations include Tobin's Q, price-to-earnings ratio, price-to-sales ratio, price-to-cash flow ratio, and price-to-earnings ratio.

[0018] As a further aspect of the present invention, the non-financial factors include management sentiments and attitudes, regulator sentiments and attitudes, and investor sentiments and attitudes.

[0019] The management's emotions and attitudes include MD&A tone, MD&A positive tone ratio, MD&A negative tone ratio, and readability;

[0020] The regulators’ sentiments and attitudes include the number of online financial news articles, the proportion of positive online financial news articles, and the proportion of negative online financial news articles.

[0021] The investor sentiment and attitudes mentioned include the investor sentiment index, the comprehensive sentiment index of Chinese investors, the number of comments in stock forums, the proportion of positive comments, and the proportion of negative comments.

[0022] As a further aspect of the present invention, the data required for financial distress, solvency, liquidity, profitability, and market valuation are all sourced from the CSMAR database; the data required for management sentiment and attitude are all sourced from the color analysis of listed companies' annual reports; and the data on regulators' sentiment and attitude, as well as investors' sentiment and attitude, are all sourced from the CNRS database.

[0023] As a further aspect of the present invention, the Pearson correlation analysis is a statistical method for measuring the strength of the linear relationship between two variables. The result is a value that represents the degree of linear correlation between the variables, ranging from -1 to 1, representing a perfect negative correlation to a perfect positive correlation. If the Pearson correlation coefficient is close to 0, it indicates that there is no linear relationship between the two variables.

[0024] As a further aspect of the present invention, in the process of introducing the punitive logistic regression, cross-validation (CV) is used to select the data-driven tuning parameter λ, and the coordinate descendant algorithm is used to iteratively solve for the approximate solution.

[0025] As a further aspect of the present invention, the prediction accuracy evaluation method is to evaluate the prediction performance of the model through a binary confusion matrix, accuracy, sensitivity, specificity, and AUC value.

[0026] A system for predicting corporate financial distress, and a method for predicting corporate financial distress according to any one of claims 1-9, the system comprising:

[0027] The data collection module, based on the CSMAR database, color analysis of listed companies' annual reports, and the CNRDS database, collects the required historical financial and non-financial data of enterprises and integrates the collected data into the dataset.

[0028] The data processing module, based on the data collection module, processes the collected data to generate the required data features, and summarizes the generated data features into financial distress features, financial factor features, and non-financial factor features.

[0029] The data training module, based on the data processing module, extracts some financial distress features, financial factor features, and non-financial factor features as training sets, establishes a training model, and introduces machine learning algorithms to train the model.

[0030] The data testing module, based on the data processing and data training modules, extracts some financial distress features, financial factor features, and non-financial factor features as a test set. The required financial and non-financial factor features are then input into the model. The predicted financial distress features obtained by the model are compared with the extracted financial distress features, and a prediction accuracy evaluation method is introduced to determine the model's predictive performance. Beneficial effects

[0031] In this invention, by filtering and extracting financial and non-financial factor data of enterprises over the years, and categorizing them into financial distress characteristics, financial factor characteristics, and non-financial factor characteristics, the data is input into a training model for training and prediction. By increasing the diversity of data used in predicting corporate financial distress, the accuracy of predicting corporate financial distress is improved.

[0032] In this invention, by introducing Pearson correlation analysis into data processing, it is possible to identify the existence of highly correlated features, thereby selectively selecting features. The data features selected through Pearson correlation analysis are aggregated and used to build a training model. Penalized logistic regression is then introduced for training. The regularization process further helps the model handle the complexity of these features, thus improving the efficiency of model training. Attached Figure Description

[0033] Figure 1 is a flowchart of the method steps of the present invention;

[0034] Figure 2 is a system flowchart of the present invention. Embodiments of the present invention

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Example

[0036] Please refer to Figure 1. The present invention provides a technical solution: a method for predicting corporate financial distress, the method comprising the following steps:

[0037] (i) Collect the required historical financial and non-financial data of enterprises through various means, and compile the collected data into a dataset for predicting corporate financial distress;

[0038] (ii) Based on the content gathered in the dataset, the dataset content is processed according to the required data features. The generated data features are divided into financial distress features, financial factor features and non-financial factor features. The generated financial distress features, financial factor features and non-financial factor features are then gathered into training set and test set according to their intended use.

[0039] (iii) Based on the characteristics of the generated training set data, Pearson correlation analysis is introduced to analyze the correlation between features in order to identify whether there are highly correlated features, so as to selectively select features. The data features selected by Pearson correlation analysis are collected and used as a basis to build a training model for predicting corporate financial distress. Penalized logistic regression is introduced for training. The regularization process will further help the model handle the complexity of these features.

[0040] (iv) Based on the trained model for predicting corporate financial distress, test set data features are introduced for testing. The test results are compared with the collected results, and the prediction accuracy is evaluated using a prediction accuracy evaluation method to determine the predictive effect of the model.

[0041] The collection channels for historical financial and non-financial data of enterprises include the CSMAR database, color analysis of listed companies' annual reports, and the CNRDS database. Data required for financial distress, solvency, liquidity, profitability, and market valuation are all sourced from the CSMAR database. Data required for management sentiment and attitude are sourced from color analysis of listed companies' annual reports. Data on regulators' sentiment and attitude, as well as investors' sentiment and attitude, are sourced from the CNRDS database.

[0042] Data features include financial distress features, financial factor features, and non-financial factor features. Financial distress features are used to determine the financial distress situation of a company. The judgment formula is ST / ST* stock = 1, otherwise 0.

[0043] Financial factors include solvency, liquidity, profitability, and market valuation. Solvency includes debt-to-equity ratio, cash ratio, equity ratio, current ratio, and capital expenditure ratio. Liquidity includes operating index, free cash flow to equity, free cash flow to the enterprise, total cash recovery rate, and net cash content of operating income. Profitability includes return on assets, return on equity, return on invested capital, net operating profit margin, and effective tax rate. Market valuation includes Tobin's Q, price-to-earnings ratio, price-to-sales ratio, price-to-cash flow ratio, and price-to-earnings ratio.

[0044] Non-financial factors include management sentiment and attitude, regulator sentiment and attitude, and investor sentiment and attitude. Management sentiment and attitude include MD&A tone, the proportion of positive words in MD&A, the proportion of negative words in MD&A, and readability. The MD&A tone is calculated as (number of affirmative words in the "Management Discussion and Analysis" text - number of negative words in the "Management Discussion and Analysis" text) / (Number of affirmative words in the "Management Discussion and Analysis" text - a color tone analysis of listed company annual reports; number of negative words in the "Management Discussion and Analysis" text / number of positive words in the "Management Discussion and Analysis" text - number of negative words in the "Management Discussion and Analysis" text), MD&A positive tone ratio is the number of positive words in the "Management Discussion and Analysis" text / total vocabulary in the "Management Discussion and Analysis" text; MD&A negative tone ratio is the number of negative words in the "Management Discussion and Analysis" text / total vocabulary in the "Management Discussion and Analysis" text; readability is the total number of words in the "Management Discussion and Analysis" text / number of punctuation marks in the "Management Discussion and Analysis" text; regulators' sentiment and attitude include the number of online financial news articles, the proportion of positive online financial news articles, and the proportion of negative online financial news articles, among which online financial... The number of news items is calculated as positive news + neutral news + negative news. The proportion of positive online financial news is calculated as positive news / (positive news + neutral news + negative news), and the proportion of negative online financial news is calculated as negative news / (positive news + neutral news + negative news). Investor sentiment and attitudes include the investor sentiment index, the China Investor Composite Sentiment Index, the number of comments in stock forums, the proportion of positive comments, and the proportion of negative comments. The investor sentiment index can be judged by the VIX index. A high VIX index usually indicates tense market sentiment and pessimistic investors about future volatility, which may indicate market uncertainty or panic. A low VIX index usually indicates relatively stable market sentiment and high investor confidence. The China Investor Composite Sentiment Index is based on the analysis of a large amount of market data, combined with the results of investor behavior and sentiment surveys. The calculation method can rely on various data sources, such as market data, investor surveys, social media analysis, and macroeconomic data. The number of comments in stock forums is calculated as positive comments + neutral comments + negative comments. The proportion of positive comments is calculated as positive comments / (positive comments + neutral comments + negative comments), and the proportion of negative comments is calculated as negative comments / (positive comments + neutral comments + negative comments).

[0045] Pearson correlation analysis is a statistical method used to measure the strength of the linear relationship between two variables. The result is a value representing the degree of linear correlation between the variables, ranging from -1 to 1, indicating a perfect negative correlation to a perfect positive correlation. If the Pearson correlation coefficient is close to 0, it means that there is no linear relationship between the two variables. Therefore, Pearson correlation analysis can identify highly correlated features, such as correlation coefficients close to +1 or -1, which can help to remove some redundant features, reduce model complexity, and avoid affecting the stability of the penalized logistic regression model due to multicollinearity.

[0046] In the computation of penalized logistic regression, cross-validation (CV) is used to select the data-driven tuning parameter λ, and the coordinate descendant algorithm is used to iteratively solve for an approximate solution. Penalized logistic regression typically refers to the technique of introducing a penalty term into a logistic regression model. This technique aims to prevent overfitting and improve the model's generalization ability on unseen data. In the computation of penalized logistic regression, the tuning parameter λ controls the magnitude of the penalty; the larger λ is, the stronger the penalty, and the fewer parameters are selected in the model. Therefore, in the computation process, a data-driven λ can be selected using the CV method to obtain the best estimation or prediction performance. In machine learning or statistical modeling, CV... Cross-validation, a model evaluation method, is typically used to assess the performance of machine learning algorithms, particularly when selecting hyperparameters such as tuning parameters, as it can effectively prevent overfitting. Coordinate descendant methods, on the other hand, are iterative methods for solving optimization problems, especially when dealing with large-scale and high-dimensional problems, where they can effectively find local optima. They are commonly used in machine learning and data science, particularly for solving constrained optimization problems or sparsity problems. The parameter estimation formula for penalized logistic regression is not only nonlinear but also nonconvex, therefore, the coordinate descendant algorithm is used to iteratively find approximate solutions.

[0047] The prediction accuracy is evaluated using a binary confusion matrix, accuracy, sensitivity, specificity, and AUC value. ROC curves are plotted using specificity and sensitivity metrics. The horizontal axis (1 - specificity) represents the false positive rate (FPR). A lower FPR means fewer samples are predicted as class 1 out of 2. The vertical line represents sensitivity, representing the true positive rate (TPR). A higher TPR means more samples are predicted as class 1. By plotting combinations of false positive and true positive rates at different thresholds, ROC curves can be obtained. The shaded area under the curve represents the AUC metric. Generally, a higher AUC value indicates a better model fit. An AUC value between 0.5 and 1, with values ​​closer to 1 indicating better prediction performance, is considered optimal. When the AUC is 0.5, the classifier is essentially ineffective, and the prediction result is similar to random guessing. Example

[0048] Please refer to Figure 2, which shows a further improvement based on Example 1.

[0049] A system for predicting corporate financial distress, the system comprising:

[0050] The data collection module, based on the CSMAR database, color analysis of listed companies' annual reports, and the CNRDS database, collects the required historical financial and non-financial data of enterprises and integrates the collected data into the dataset.

[0051] The data processing module, based on the data collection module, processes the collected data to generate the required data features, and summarizes the generated data features into financial distress features, financial factor features, and non-financial factor features.

[0052] The data training module, based on the data processing module, uses extracted financial distress features, financial factor features, and non-financial factor features as the training set to build a training model. Pearson correlation analysis, penalized logistic regression, cross-validation machine, and coordinate descendant algorithm are introduced to train the model.

[0053] The data testing module, based on the data processing module and the data training module, extracts some financial distress features, financial factor features, and non-financial factor features as a test set. The required financial factor features and non-financial factor features are input into the model. The predicted financial distress features obtained by the model are compared with the extracted financial distress features, and a prediction accuracy evaluation method is introduced to determine the prediction effect of the model.

[0054] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for predicting financial distress of a business, characterized by, The method comprises the following steps: (1) Collecting the required enterprise historical financial data and non-financial data through various channels, and gathering the collected data into a data set for predicting enterprise financial distress; (2) Based on the content gathered in the data set, processing the data set content according to the required data characteristics, dividing the generated data characteristics into financial distress characteristics, financial factor characteristics and non-financial factor characteristics, and gathering the generated financial distress characteristics, financial factor characteristics and non-financial factor characteristics into a training set and a test set according to the purpose; (3) Based on the generated training set data characteristics, introducing Pearson correlation analysis for analyzing the correlation between characteristics, so as to identify whether there are highly correlated characteristics, and selectively perform feature selection, gather the data characteristics selected by Pearson correlation analysis, and establish a training model for predicting enterprise financial distress based thereon, introduce a penalty logistic regression for training, and the regularization process will further help the model to process the complexity of these characteristics; (4) Based on the trained prediction model of enterprise financial distress, introducing the test set data characteristics for testing, comparing the test results with the collected results, and using a prediction accuracy evaluation method to evaluate the prediction accuracy to determine the prediction effect of the model.

2. The method for predicting financial distress of a business according to claim 1, wherein, The collection channels of the enterprise historical financial data and non-financial data include CSMAR database, color tone analysis of listed company annual reports and CNRDS database.

3. The method for predicting financial distress of a business of claim 1, wherein, The data characteristics include financial distress characteristics, financial factor characteristics and non-financial factor characteristics; The financial distress characteristics are used to judge the situation of enterprise financial distress, and the judgment formula is ST / ST* stock = 1, otherwise 0.

4. The method for predicting financial distress of a business according to claim 3, wherein, The financial factor characteristics include solvency, liquidity, profitability and market valuation; The solvency includes asset-liability ratio, cash ratio, equity ratio, current ratio and capital expenditure ratio; The liquidity includes operation index, equity free cash flow, enterprise free cash flow, full cash recovery rate and net cash content of operating income; The profitability includes asset return rate, equity return rate, investment capital return rate, net operating profit rate and actual tax rate; The market valuation includes Tobin Q, price-earnings ratio, price-sales ratio, price-to-cash flow ratio and price-earnings ratio.

5. The method for predicting financial distress of a business according to claim 3, wherein, The non-financial factor characteristics include management emotion and attitude, regulator emotion and attitude, and investor emotion and attitude; The management emotion and attitude include MD&A tone, MD&A positive tone ratio, MD&A negative tone ratio and readability; The regulator emotion and attitude include the number of online financial news, the proportion of positive online financial news and the proportion of negative online financial news; The investor emotion and attitude include investor sentiment index, China investor comprehensive sentiment index, the number of comments in stock forums, the proportion of positive comments and the proportion of negative comments.

6. The method for predicting financial distress of a business according to claim 5, wherein, The financial distress, solvency, liquidity, profitability and market valuation required data are from the CSMAR database, the management sentiment and attitude required data are from the color tone analysis of listed company annual report, and the regulator's emotion and attitude and investor's emotion and attitude data are from the CNRDS database.

7. The method for predicting financial distress of a business according to claim 3, wherein, The Pearson correlation analysis is a statistical method for measuring the strength of linear relationship between two variables, and the result is a value representing the degree of linear correlation between variables, ranging from -1 to 1, representing complete negative correlation to complete positive correlation. If the Pearson correlation coefficient is close to 0, it means that there is no linear relationship between the two variables.

8. The method for predicting financial distress of a business of claim 1, wherein, In the introduction calculation process of the penalized logistic regression, cross-validation CV is used to select data-driven tuning parameters λ, and the coordinate descent algorithm is used to iteratively solve the approximate solution.

9. The method for predicting financial distress of a business of claim 1, wherein, The prediction accuracy evaluation method is to evaluate the prediction effect of the model by binary confusion matrix, accuracy, sensitivity, specificity and AUC value.

10. A system for predicting financial distress of an enterprise, characterized by, The system comprises: A data collection module collects the required enterprise historical financial data and non-financial data based on the CSMAR database, color tone analysis of listed company annual report and CNRDS database, and integrates the collected data into a data set; A data processing module processes the collected data to generate the required data features based on the data collection module, and induces the generated data features into financial distress features, financial factor features and non-financial factor features; A data training module extracts part of the financial distress features, financial factor features and non-financial factor features as a training set based on the data processing module, establishes a training model, and introduces a machine learning algorithm to train the model; A data testing module extracts part of the financial distress features, financial factor features and non-financial factor features as a test set based on the data processing module and the data training module, inputs the required financial factor features and non-financial factor features into the model, compares the predicted financial distress features obtained through the model with the extracted financial distress features, and introduces a prediction accuracy evaluation method to determine the prediction effect of the model.

Images