Enterprise loan willingness assessment method based on multi-dimensional public data
Through the application of multi-channel public data fusion and machine learning models, the problems of limited data sources and single evaluation dimensions in the existing technology are solved, and a more accurate and objective assessment of corporate loan intentions is achieved.
Patent Information
- Application Number
- CN202411580614.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-07
- Publication Date
- 2025-02-14
AI Technical Summary
The existing technology has a large impact on the assessment of corporate loan intentions, limited data sources, single evaluation dimensions, and human factors that rely on expert experience, resulting in the inability to evaluate the evaluation results to be objective and comprehensive enough.
By collecting multi-channel public data, performing data fusion processing and multi-dimensional feature screening, and using machine learning models (such as XGBoost) to conduct multi-dimensional corporate loan intention assessments to reduce dependence on expert rules.
It improves the accuracy and objectivity of the evaluation results, can understand the actual situation of the company more comprehensively, reduce the influence of human factors, and improve the robustness and accuracy of the system.
Smart Images
Figure CN119444401A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of big data processing technology, and in particular to a method for evaluating corporate loan willingness based on multi-dimensional public data. Background Art
[0002] In the financial sector, facing a vast number of businesses, identifying high-potential customers with funding needs and loan willingness has always been a key requirement for banks and other financial institutions. Accurately estimating a company's loan willingness can help financial institutions conduct targeted marketing, improving marketing efficiency and ultimately conversion rates. Existing technologies primarily rely on traditional expert rule evaluation methods, which primarily assess loan willingness based on information such as financial statements and historical loan records provided by the company, combined with the expert rules within the financial institution. While these methods can provide some reference value, they have the following significant drawbacks:
[0003] 1. Limited data sources and questionable data reliability: Existing technologies rely on data proactively disclosed by companies, such as financial statements and business reports. This data may be filtered and processed by the companies, resulting in incomplete or distorted information. Due to the single data source, the assessment results may be influenced by subjective factors of the companies, and may not objectively and comprehensively reflect the true situation of the companies.
[0004] 2. Single data dimension, incomplete assessment: Existing technologies primarily assess companies from a financial and historical loan perspective, ignoring multiple dimensions such as a company's social creditworthiness, operating conditions, and market competitiveness. This single assessment dimension can lead to misjudgments of loan willingness and an inability to accurately estimate a company's willingness to lend.
[0005] 3. Reliance on expert experience, subject to significant human influence: Expert systems are limited by their own historical business experience, making it difficult to fully and accurately reflect the funding needs of different companies. The resulting rules are often granular and only applicable to certain cycles and customer groups.
[0006] In summary, existing technologies for assessing corporate loan willingness have shortcomings such as limited data sources, single evaluation dimensions, and poor generalization. These problems restrict the accuracy and efficiency of loan marketing by financial institutions. Summary of the Invention
[0007] The purpose of the present invention is to overcome the defects of the above-mentioned existing technologies and provide a method for evaluating corporate loan willingness based on multi-dimensional public data. By integrating public data and using machine learning models to perform multi-dimensional corporate loan willingness evaluation, the accuracy of the evaluation results can be improved.
[0008] The purpose of the present invention can be achieved by the following technical solution: a method for evaluating corporate loan willingness based on multi-dimensional public data, comprising the following steps:
[0009] S1. Collect public data from multiple channels;
[0010] S2. Fusion processing is performed on public data from multiple channels to obtain sample data;
[0011] S3, filtering out multidimensional feature data from the sample data;
[0012] S4. Use multi-dimensional feature data to train the machine learning model and perform optimization to establish a loan willingness assessment model;
[0013] S5. Input the multi-dimensional characteristics of the enterprise to be evaluated into the loan willingness evaluation model, and output the loan willingness score result.
[0014] Furthermore, the multi-channel public data in step S1 includes corporate finance, taxation, social security, and illegal information.
[0015] Furthermore, the fusion processing in step S2 includes data cleaning and business indicator processing, wherein the data cleaning is used to remove duplicate, erroneous or invalid information;
[0016] The business indicator processing is used to extract key business indicators that can reflect the business operating conditions and social credit status of the enterprise from the cleaned data.
[0017] Furthermore, the step S3 specifically filters out multi-dimensional feature data from the sample data according to preset filtering rules.
[0018] Furthermore, the preset screening rules include:
[0019] The data coverage is greater than or equal to a preset first threshold;
[0020] The correlation between the data and the label is greater than or equal to a preset second threshold.
[0021] Furthermore, step S3 includes the following process:
[0022] S31, counting the coverage of each data, and removing feature data with a coverage below a preset first threshold;
[0023] S32: Calculate the correlation between each retained data and the corresponding label, and remove feature data with a correlation lower than a preset second threshold.
[0024] Furthermore, step S4 includes the following process:
[0025] S41. Construct a data set using multi-dimensional feature data, perform iterative training on the machine learning model, and obtain an initial model;
[0026] S42. Perform optimization work on the initial model, including model parameter adjustment, feature selection, and model structure adjustment, to obtain an optimized loan willingness assessment model.
[0027] Furthermore, in step S42, the model parameter adjustment includes learning rate adjustment and iteration number adjustment;
[0028] The features obtained after feature selection include monetary funds, current assets, owner's equity, long-term loans, short-term loans, operating income, financial costs, profit information and total liabilities.
[0029] Furthermore, the step S42 specifically uses a grid search method to adjust the model structure to determine the optimal model structure:
[0030] First, a preset large-scale search space is defined, a coarse-grained search is performed, and then the search is gradually narrowed down through bisection until the optimal model structure is obtained.
[0031] Furthermore, the machine learning model in step S4 adopts the XGBoost model.
[0032] Compared with the prior art, the present invention has the following advantages:
[0033] The present invention collects and fuses public data from multiple channels to obtain sample data. Multidimensional feature data is then filtered from the sample data for use in training a machine learning model to construct a loan willingness assessment model. Based on this fused public data, the system expands data sources, improves the integrity and authenticity of information, and provides a more comprehensive and objective understanding of a company's actual situation. It also enriches the assessment dimensions, enabling assessments from multiple perspectives, such as finance, taxation, social security, and illegal information, increasing the dimensionality and depth of the assessment. Furthermore, the integration of machine learning models reduces reliance on expert rules, avoids the influence of human factors, and improves the robustness and accuracy of the system, enabling a more comprehensive and objective assessment and accurate judgment of a company's loan willingness.
[0034] This invention integrates and processes public data from multiple channels. Data cleaning is first performed. Raw data contains duplicate, erroneous, or invalid information, so cleaning is essential to ensure data accuracy. This stage is crucial for improving data quality and the reliability of subsequent analysis. Next, business indicator processing is performed. Through in-depth analysis of the cleaned data, key business indicators that reflect the business's operating conditions and social credit status are extracted. These indicators are then further processed and manipulated to better support the business's decision-making process. This integrated processing provides a comprehensive, accurate, and consistent portrait of a business's operating data, helping to better understand the business's operating conditions and social credit status.
[0035] This method uses preset screening rules to filter multidimensional feature data from sample data. First, the fill rate of candidate features is calculated, and features with too low coverage are discarded. Then, the correlation between features and labels is calculated, and features with too low correlation are discarded to reduce model complexity and improve prediction accuracy. This method selects features from the raw data that are meaningful to the target estimate. These features cover multiple dimensions, such as the company's financial status, operating history, social credit record, industry trends, and market competition. Through feature engineering, the raw data is converted into structured data, laying the foundation for subsequent analysis and prediction.
[0036] This method uses multidimensional feature data to iteratively train a machine learning model to obtain an initial model. This initial model is then fine-tuned, including adjusting model parameters, feature selection, and model structure, to produce an optimized loan willingness assessment model. The model training process is optimized by adjusting parameters such as the learning rate and number of iterations. Feature selection reduces model complexity and improves prediction accuracy. Improvements to the model structure enhance its fit and generalization capabilities. Ultimately, through fine-tuning, the model demonstrates enhanced generalization capabilities and predictive accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 Schematic diagram of the method flow of the present invention;
[0038] Figure 2 Schematic diagram of the application framework of the embodiment. DETAILED DESCRIPTION
[0039] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0040] Example
[0041] like Figure 1 As shown, a method for evaluating corporate loan willingness based on multi-dimensional public data includes the following steps:
[0042] S1. Collect public data from multiple channels;
[0043] S2. Fusion processing is performed on public data from multiple channels to obtain sample data;
[0044] S3, filtering out multidimensional feature data from the sample data;
[0045] S4. Use multi-dimensional feature data to train the machine learning model and perform optimization to establish a loan willingness assessment model;
[0046] S5. Input the multi-dimensional characteristics of the enterprise to be evaluated into the loan willingness evaluation model, and output the loan willingness score result.
[0047] This embodiment applies the above solution to build Figure 2 The application framework shown here consists of three parts: a data fusion module, a machine learning model, and a user application. The data fusion module collects and integrates public data to form an enterprise database. The machine learning model analyzes and mines the data in the database, outputting a loan willingness score. The user application leverages the model output for precision marketing, profiling, product strategy development, and adaptation.
[0048] Specifically:
[0049] 1. Data Fusion Module: The data fusion module is a core component for building a comprehensive enterprise management information database. It is primarily responsible for collecting and integrating public data from various channels, such as corporate finance, taxation, social security, and illegal information. Together, these data form a complete picture of an enterprise's operating status, comprehensively reflecting its operational status and social credit standing. Data collection is the primary step in the data fusion module. This phase determines the required data type, source, accuracy, and reliability based on requirements, laying a solid foundation for subsequent data processing and analysis. Data cleaning follows. Raw data contains duplicate, erroneous, or invalid information, so cleaning is essential to ensure data accuracy. This phase is crucial for improving data quality and the reliability of subsequent analysis. Business indicator extraction is another key step in the data fusion module. Through in-depth analysis of the cleaned data, key business indicators that reflect the enterprise's operating status and social credit status are extracted and processed to better support the enterprise's decision-making process. Through data collection, data cleaning, and business indicator processing, the data fusion module provides a comprehensive, accurate, and consistent picture of enterprise operations, helping to better understand the enterprise's operating status and social credit standing.
[0050] In actual applications, in the data fusion module, financial institutions can provide their historical marketing results data, and the platform can provide public data such as assets and liabilities, revenue and profits, violations of law and trust, etc., which are spliced into samples. At the same time, some abnormal data (fields with too high missing rates and outliers that deviate greatly from the mean) are filtered out.
[0051] 2. Machine Learning Model Development: Machine learning models are the core of building a loan willingness assessment system based on enterprise data. This involves feature engineering, model selection, and model training and tuning.
[0052] In the machine learning model module, feature engineering processes such as filling in empty values, encoding non-numerical features, and discretizing continuous features are performed. After processing, the model is trained and relevant parameters are optimized. Finally, the target enterprise to be evaluated is scored to obtain the enterprise's willingness to lend.
[0053] (1) Feature engineering: Feature engineering is the first step in building a machine learning model. Feature selection is an important part of feature engineering. It selects the features that have the greatest impact on loan willingness prediction from a large number of features (feature screening mainly has the following criteria: 1. The variance of the feature itself should be large and the coverage should be high; 2. The correlation between the feature and the label should be high, such as the Pearson correlation coefficient; the overall initial screening logic is to first calculate the fill rate of the candidate features, discard the features with too low coverage, and then calculate the correlation between the features and the label, discarding the features with too low correlation) to reduce the complexity of the model and improve the accuracy of the prediction. Select features that are of practical significance to the prediction target from the original data. These features cover multiple dimensions such as the company's financial status, operating history, social credit record, industry trends, and market competition situation. Through feature engineering, the original data is converted into structured data, laying the foundation for subsequent analysis and prediction. In the feature engineering stage, data preprocessing is required to clean the original data, remove duplicate, erroneous or invalid information, and ensure the accuracy and completeness of the data. At the same time, data transformations such as normalization and standardization are performed to enable data of different magnitudes or distributions to play the same role in the model.
[0054] (2) Model selection: Common algorithms include logistic regression, random forest, neural network, support vector machine, etc. Among them, logistic regression is suitable for processing data with strong linear relationships, while random forest is good at processing nonlinear relationships and complex data structures, and neural network is suitable for processing large-scale, high-dimensional data. Taking into account factors such as the complexity of the problem and the interpretability of the model, this solution adopts the XGBoost model, which is an ensemble learning algorithm based on gradient boosting. It builds a strong learner by integrating multiple weak learners. It has strong fitting ability, can fit complex nonlinear relationships, and can accurately estimate the loan willingness of enterprises. At the same time, because it uses regularization techniques such as feature sampling, it can prevent the model from overfitting. When facing unknown data, it can also maintain good prediction performance and has excellent generalization ability. Even when facing large-scale data sets, it can complete training in a relatively short time. Compared with some black box models (such as neural networks), the output of xgb has certain interpretability. By looking at the feature importance and tree structure in the model, we can understand how the model makes predictions and understand the logic and decision basis behind the model.
[0055] (3) Model training and parameter adjustment: In this stage, historical data is used for training so that the model can learn patterns in the data. Through iteration and training, a model that can accurately predict corporate loan willingness is obtained. To improve the generalization ability of the model, cross-validation and regularization methods are used to prevent overfitting. After obtaining the initial model, tuning is carried out to further improve the accuracy and stability of the model. Tuning involves adjusting algorithm parameters, optimizing feature selection, improving model structure, and other aspects. The model training process was optimized by adjusting parameters such as the learning rate and number of iterations. Feature selection (primarily focusing on the company's assets and liabilities, profit information, and tax violations, specifically including cash, current assets, owner's equity, long-term loans, short-term loans, operating income, financial costs, profit information, and total liabilities) was used to reduce model complexity and improve prediction accuracy. The model structure was refined (the XGB model consists of multiple decision trees, and the number and depth of the trees significantly influence the model structure and performance. This approach uses a grid search method to search for the optimal model structure. Due to the high computational cost of grid search, this paper first defines a large search space, performs a coarse-grained search, and then gradually narrows the search using a binary search method until the optimal model structure is obtained) to improve model fit and generalization. Through fine-tuning, the model demonstrated stronger generalization and prediction accuracy. After model training and fine-tuning, the model performance was evaluated using metrics such as precision, recall, and F1 score.
[0056] 3. Regular update mechanism: As the economic environment and loan products of financial institutions change, the mapping relationship between features and loan willingness will also change over time. In order to obtain new mapping relationships in a timely manner, this embodiment also sets a regular update mechanism to ensure that the evaluation system can promptly reflect the latest business changes and social credit status of the enterprise. It regularly obtains the latest business data of the enterprise, conducts training and estimation based on the latest features, and can timely evaluate the latest status of the enterprise.
[0057] Through a regular update mechanism, changes in a company's operations and social credit status can be reflected promptly, improving the timeliness and accuracy of assessment results. This helps financial institutions adjust marketing decisions in a timely manner and adapt to changes in the market environment and customer needs.
[0058] 4. In the user application module, financial institutions will obtain the target enterprise's loan willingness score, which can be used for:
[0059] 1. Precision marketing: Helps financial institutions accurately locate high-potential customers with a strong willingness to lend among a large number of customers, thereby improving their marketing efficiency and conversion rate (for example, there are 1,000 companies in total, and 100 are randomly selected as the experimental group using the model, and 100 are randomly selected as the control group. The experimental group ultimately converts 2 companies, and the control group converts 1 company, and the conversion rate increases from 1% to 2%).
[0060] 2. Portrait analysis: Financial institutions often need to collect multi-dimensional data of enterprises, portray enterprise portraits, conduct subsequent user data analysis, attribution research, etc. Willingness to lend is an important dimension that can help financial institutions understand the capital needs of enterprises.
[0061] 3. Product adaptation: Based on the different funding needs of enterprises, financial institutions can adapt different loan products and provide different preferential plans to achieve more refined operations.
[0062] In summary, applying this solution in practice can improve the marketing success rate of financial institutions: by accurately estimating loan willingness, financial institutions can accurately locate high-potential enterprises with funding needs, reduce marketing to unwilling users, save marketing costs, and improve marketing success rate, ultimately achieving performance improvement.
[0063] As a third-party information service platform, it enables the desensitized flow of information. Financial institutions provide feedback on their marketing effectiveness, and the model accesses industry-wide data. This data is used to train and optimize the model. The trained model only contains mappings between features and labels, and does not store any specific corporate information, eliminating information leakage. The model serves as an information platform, enabling the desensitized use of information between different institutions.
[0064] Promoted the healthy development of the financial market: The application of this solution helps financial institutions more accurately assess the willingness of enterprises to lend, allowing funds to flow more efficiently to enterprises in need, improving the efficiency of fund distribution in the financial market and helping it develop more healthily and rapidly.
Claims
1. A method for evaluating corporate loan willingness based on multi-dimensional public data, characterized in that: The following steps are involved: S1. Collect public data from multiple channels; S2. Perform fusion processing on public data from multiple channels to obtain sample data; S3, filtering out multi-dimensional feature data from sample data; S4. Use multi-dimensional feature data to train the machine learning model and perform tuning to establish a loan willingness assessment model; S5. Input the multi-dimensional characteristics of the enterprise to be evaluated into the loan willingness evaluation model, and output the loan willingness score result.
2. According to claim 1, a method for evaluating corporate loan willingness based on multi-dimensional public data is characterized in that: The multi-channel public data in step S1 includes corporate finance, taxation, social security, and illegal information.
3. According to claim 1, a method for evaluating corporate loan willingness based on multi-dimensional public data is characterized in that: The fusion processing in step S2 includes data cleaning and business indicator processing, wherein the data cleaning is used to remove duplicate, erroneous or invalid information; The business indicator processing is used to extract key business indicators that can reflect the business operating conditions and social credit status of the enterprise from the cleaned data.
4. According to claim 1, a method for evaluating corporate loan willingness based on multi-dimensional public data is characterized in that: The step S3 specifically involves filtering out multi-dimensional feature data from the sample data according to preset filtering rules.
5. The method for evaluating corporate loan willingness based on multi-dimensional public data according to claim 4 is characterized in that: The preset screening rules include: The coverage rate of the data is greater than or equal to a preset first threshold; The correlation between the data and the label is greater than or equal to a preset second threshold.
6. The method for evaluating corporate loan willingness based on multi-dimensional public data according to claim 5, characterized in that: Step S3 The process includes: S31, counting the coverage of each data, and removing feature data whose coverage is lower than a preset first threshold; S32, calculating the correlation between each retained data and the corresponding label, and eliminating the feature data whose correlation is lower than a preset second threshold.
7. The method for evaluating corporate loan willingness based on multi-dimensional public data according to claim 2 is characterized in that: Step S4 The process includes: S41, constructing a data set using multi-dimensional feature data, and iteratively training the machine learning model to obtain an initial model; S42. Perform optimization work on the initial model, including model parameter adjustment, feature selection, and model structure adjustment, to obtain an optimized loan willingness assessment model.
8. The method for evaluating corporate loan willingness based on multi-dimensional public data according to claim 7 is characterized in that: In step S42, the model parameter adjustment includes learning rate adjustment and iteration number adjustment; The features obtained after feature selection include monetary funds, current assets, owner's equity, long-term loans, short-term loans, operating income, financial costs, profit information and total liabilities.
9. The method for evaluating corporate loan willingness based on multi-dimensional public data according to claim 7, characterized in that: The step S42 specifically uses a grid search method to adjust the model structure to determine the optimal model structure: First, a preset large-scale search space is defined, a coarse-grained search is performed, and then the search is gradually narrowed down through the binary search method until the optimal model structure is obtained.
10. A method for evaluating corporate loan willingness based on multi-dimensional public data according to any one of claims 1 to 9, characterized in that: The machine learning model in step S4 adopts the XGBoost model.