Financial lease risk scoring method based on human feedback learning

Abstract

The invention discloses a financial lease risk scoring method based on human feedback learning, and the method carries out the automatic analysis of the multi-dimensional features of a lessee enterprise, generates a risk score from the perspective of the financial lease risk, and provides an efficient and precise risk assessment decision support for a lessor. The method is characterized by comprising the following steps: constructing multi-category features for comprehensively evaluating a financing lease risk; and training the neural network model for risk scoring, and iteratively training and optimizing the neural network model according to user feedback. The interpretability of the model is enhanced by calculating the influence intensity of the features on the score; a scoring result is visually displayed in a chart form, and lease decision is assisted. The method specifically comprises an enterprise data acquisition module, a risk feature construction module, a model training module, a score prediction module, a human feedback collection module, a model interpretability module and the like. According to the method, the risk assessment requirement of the financing lease service is closely met, and the data completeness and analysis efficiency of risk assessment are remarkably improved.

Description

Technical Field

[0001] The present invention relates to a financial leasing risk scoring method based on human feedback learning. By constructing financial leasing risk characteristics, neural network model training, and human feedback learning and other steps, it realizes the automated and intelligent evaluation of the financial leasing risk of lessee enterprises. Background Art

[0002] Financial leasing is an important quasi-financial business, and its core risk lies in the default risk of lessee enterprises. Especially small and medium-sized lessee enterprises are vulnerable to various external factors due to their weak anti-risk ability, such as industry cycle fluctuations, market environment changes, adjustment of business qualification levels, fluctuations in bank credit ratings, major changes in equity and capital structures, loss of important customers, and vicious competition within the industry. These factors may lead to the deterioration of the enterprise's operating performance and the decline of its performance fulfillment ability, ultimately having an adverse impact on the cash flow of financial leasing enterprises.

[0003] In the initial business review and access stage, it is crucial to accurately assess the risks of lessee enterprises. However, traditional risk assessment methods mainly rely on the manual review of due diligence personnel. However, due to insufficient professional capabilities or lack of practical experience, due diligence personnel may miss or misjudge key data during the information collection and analysis process, thus affecting the accuracy of the assessment of lessee enterprises and increasing the risk exposure of financial leasing enterprises.

[0004] Therefore, by making full use of emerging technologies such as big data and artificial intelligence, integrating enterprise financial data and publicly available Internet data, constructing a comprehensive portrait of lessee enterprises, and developing an intelligent risk assessment model, not only can the efficiency of risk assessment be significantly improved, but also the moral risk and professional ethics risk that may exist during the due diligence process can be effectively alleviated, further enhancing the safety and reliability of financial leasing business. Summary of the Invention

[0005] Object of the Invention: The present invention discloses a financial leasing risk scoring method based on human feedback learning, constructs a feature portrait of lessee enterprises. The method automatically analyzes the multi-dimensional features of lessee enterprises, generates a risk access score from the perspective of financial leasing risks, provides scientific decision-making support for lessors, and assists them in evaluating whether to carry out financial leasing business.

[0006] Technical Solution:

[0007] A financial leasing risk scoring method based on human feedback learning includes the following steps:

[0008] Step 1 Enterprise data acquisition: Establish a data warehouse to store publicly available data from the Internet and the financial data of the enterprise. The publicly available data from the Internet includes industry data, administrative penalty data, credit data, geographical data, and legal litigation data. The financial data includes the balance sheet, cash flow statement, and income statement;

[0009] Step 2 Construction of financial leasing risk characteristics: Based on the data in the data warehouse, and aiming at the unique requirements and potential risk points of financial leasing business, design 5 core characteristic categories of the lessee enterprise: enterprise qualification, operating conditions, financial status, geographical qualification, and legal litigation. Among them, the enterprise qualification category includes: enterprise type, industry trend, and whether it is a key industrial cluster; the operating conditions category includes: number of fines, amount of fines (in ten thousand yuan), tax payment grade, and third-party rating; the financial status category includes: net profit rate (%), return on total assets (%), operating profit rate (%), asset-liability ratio (%), interest coverage ratio, quick ratio (%), operating growth rate (%), net profit growth rate (%), and total asset growth rate (%); the geographical qualification category includes: geographical score; the legal litigation category includes: number of contract disputes, number of false statements, and number of invention patent disputes;

[0010] Step 3 Training of risk assessment model: Collect the historical data of enterprises in past financial leasing projects, construct a training set D, and train and optimize the neural network model;

[0011] Step 4 Prediction of risk access score: Calculate the risk access score according to the characteristics of the lessee enterprise, evaluate its financial leasing risk, and the higher the access score, the lower the risk;

[0012] Step 5 Collection of human feedback: In order to improve the model performance, users are required to give feedback on the access score of the financial leasing risk assessment model, that is, to give human feedback on the results predicted by the neural network risk assessment model, and use the human feedback labels to strengthen the neural network model and improve the accuracy of the neural network prediction model;

[0013] Step 6 Interpretability of risk assessment model: Calculate the influence intensity of each characteristic category on the model score, so as to quantify the contribution degree of each characteristic category;

[0014] Step 7 Financial leasing decision-making: Present the results in the form of a chart report to assist financial leasing decision-making.

[0015] Furthermore, in Step 2, when constructing the financial leasing risk characteristics, determine the data sources and calculation methods of the characteristics. The specific steps are as follows:

[0016] Step 201. Construction of characteristic categories: According to the unique requirements and potential risk points of financial leasing business, design 5 characteristic categories of the lessee enterprise: enterprise qualification F 1 and operating conditions F2 , Financial status F 3 , Regional qualification F 4 and legal proceedings F 5 , that is Let x i represent the feature vector of the i-th enterprise. Then, for any enterprise, we can comprehensively describe its risk characteristics through F 1 , F 2 , F 3 , F 4 and F 5 , that is

[0017] x i = [F 1 , F 2 , F 3, F 4 , F 5 , 1 ≤ i ≤ N, where N is the number of enterprises;

[0018] Step 202. Determine the data sources and calculation methods of the features: According to the feature categories, determine the financial leasing risk features, and summarize and sort out the data sources and calculation methods of each feature. For some feature categories, such as "enterprise type" in the enterprise qualification category, it is directly obtained through data extraction; while for some other feature categories, such as "net profit margin (%)" in the financial status category, it needs to be obtained through complex data calculations. The specific calculation formula is: (net profit / operating income) * 100%; The financial leasing features are shown in Table 1.

[0019] Table 1 Financial leasing features

[0020]

[0021] Furthermore, in Step 3, construct a training set, train and optimize the risk assessment model for subsequent enterprise scoring.

[0022] The specific steps are as follows:

[0023] Step 301. Construct a training set: where M < N is the number of training samples, and y i represents the financial leasing risk access score of this enterprise. In the initial training stage of the scoring model, the training set is mainly constructed manually based on the samples of enterprises that have carried out financial leasing business. In the subsequent iterative optimization process of the scoring model, human feedback is gradually introduced into the training set to improve the model performance and prediction accuracy;

[0024] Step 302. Model training: Construct a risk assessment model based on a deep neural network, and learn the model parameters through the training set data. Take the enterprise feature vector x iAs the input of the neural network model, the ReLU is used as the activation function in the hidden layer to perform element-wise non-linear transformation on the input, and finally the risk admission score result f(x i ) is obtained in the output layer. The risk assessment and decision-making process of the model for the enterprise sample x i can be expressed as y i = f(x i ).

[0025] The loss function is defined as:

[0026]

[0027] where p is the number of enterprise pairs; λ is the weight coefficient of the pairwise loss, used to control the proportion of the pairwise loss in the total loss; m is the margin of the pairwise loss, used to ensure that the admission scores of the enterprise pairs differ by at least m. By calculating the gradient of the loss function with respect to the model parameters and using the gradient descent algorithm to update the model parameters, the training model is iteratively optimized so that the model can more accurately predict the financial leasing risk scores of enterprises.

[0028] Furthermore, in step five, the neural network model is enhanced with human feedback. The specific steps are as follows:

[0029] Step 501. Score feedback: Based on their professional knowledge and practical business experience, users review the comprehensive situation of the lessee enterprise, adjust the score results predicted by the model, and feedback the corrected results to the model training process. The method supports two feedback mechanisms: (1) Direct feedback: Users correct the predicted scores of individual enterprises; (2) Pairwise feedback: Users provide feedback on the comparison relationship for two enterprises (i.e., enterprise pairs).

[0030] Step 502. Model training based on human feedback learning: By introducing an adaptive layer, the loss function of the deep neural network is optimized, and the direct feedback and pairwise feedback information are incorporated into it, thereby guiding the iterative optimization process of the risk assessment model.

[0031] Furthermore, in step six, to improve the interpretability of the model, the influence intensity of each feature category on the model score is calculated. The steps are as follows:

[0032] Step 601. Calculate the influence intensity: k ∈ [1, 5], w k ∈ [-1, 1], f -k (x i ) represents the risk score after removing the F k feature.

[0033] Step 602. Interpret the influence intensity: By calculating w kThe value quantifies the contribution degree of each feature category to the final score. When 0 < w k < 1, it means that F k The feature has a positive impact on the risk access score, and the larger the w k value, the more significant the positive impact of the F k feature on the model prediction score, thus increasing the access score and reducing the risk of financial leasing; when -1 < w k < 0, it means that F k The feature has a negative impact on the risk access score, and the smaller the w k value, the more significant the negative impact of the F k feature on the model prediction score, which may lead to a decrease in the enterprise's access score and an increase in the risk of financial leasing.

[0034] Beneficial effects: The present invention provides an intelligent solution for the risk assessment of financial leasing financial services, significantly improving the comprehensive performance of risk assessment. Specifically, the method comprehensively improves the data completeness, assessment accuracy, and calculation efficiency of the risk assessment process by constructing a multi-dimensional enterprise feature portrait and a scoring model based on a deep neural network. From a macro perspective, the present invention provides important technical support for strengthening the local financial governance ability, and helps to promote the standardization and sustainable development of the financial leasing industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a schematic flowchart of the method provided in Embodiment 1 of the present invention

[0036] Figure 2 is an example schematic diagram of a financial leasing decision provided in Embodiment 1 of the present invention

[0037] Figure 3 is an example schematic diagram of human feedback provided in Embodiment 1 of the present invention DETAILED DESCRIPTION OF THE INVENTION

[0038] The present invention will be further explained below with reference to the accompanying drawings. Figure 1 is a schematic flowchart of a financial leasing risk scoring method based on human feedback learning provided in Embodiment 1 of the present invention. As Figure 1 shown, this embodiment includes the following steps:

[0039] Step 1: First, start the data collection task, and crawl the public data of the enterprises on the Internet platform according to the list of enterprises that have provided financial data. Subsequently, use the comprehensively collected data to build an underlying data warehouse, laying a foundation for subsequent feature construction and model prediction work.

[0040] Step 2: Then, according to the unique requirements and potential risk points of the financial leasing business, five categories of financial leasing characteristics are constructed, namely enterprise qualifications, business conditions, financial conditions, regional qualifications, and legal proceedings. For each characteristic category, determine the data sources and values of specific characteristics. When selecting enterprise prediction access scores later, the method will extract corresponding data from the data warehouse according to the set characteristics for cleaning and transformation.

[0041] For example, enterprise x 1 and x 2 The characteristics after cleaning and transformation are shown in Table 2.

[0042] Step 3: First, based on the historical data of the financial leasing business, after cleaning and transformation, the corresponding characteristics are extracted, and the risk access score is used as a label to construct a training data set. Subsequently, we start to construct a risk assessment model based on a deep neural network, input the training data set into the model for training, and thus obtain the initial form of the risk assessment model.

[0043] Then we select some enterprises from all enterprises to form a test data set, and input this test data set into the model to obtain its risk access score. According to the feedback of the test results, we perform

[0044] Table 2 Enterprise Characteristics

[0045]

[0046]

[0047] tuning processing on the prediction accuracy of the model. The specific methods include adjusting model parameters, increasing the number of training samples, etc., to achieve iterative training of the model.

[0048] We repeat the above test and optimization operations. After multiple rounds of iterative training and optimization, we finally obtain a risk assessment model with stable performance and accurate prediction. This model will be applied to the subsequent risk assessment work of enterprises.

[0049] Step 4: When the user selects an enterprise, the method will automatically obtain all the data of the enterprise, convert it into corresponding characteristics and input them into the model, so as to obtain its risk access score.

[0050] If the enterprise selected by the user is x in Table 2 1 and x 2 , then after model prediction, the risk access score of enterprise x 1 is 96 points, and the risk access score of enterprise x 2 is 59 points. The model risk assessment results are as Figure 2 shown.

[0051] Step Five: If the user believes that the enterprise risk access score obtained in Step Four needs to be adjusted, the user can provide feedback on the model result on the feedback page. The feedback page is as shown in Figure 3 and the selected enterprise for feedback is x 1 .

[0052] First, the user can provide direct feedback on enterprise x 1 , that is, the user fills in a new access score, such as 94 points, at the manually corrected score section on the direct feedback interface in Figure 3 . Then the risk access score of enterprise x 1 is corrected to 94 points;

[0053] Second, the user can provide paired feedback on enterprise x 1 , that is, the user selects a similar enterprise that has completed risk assessment for comparison at the paired feedback entry in Figure 3 . After the user selects an enterprise, the user will enter the paired feedback interface in Figure 3 . In the current interface, the user checks the characteristics and access scores of the two enterprises, and judges whether the comparison of the access scores of the current two enterprises conforms to the facts. If it conforms to the facts, the user selects "Yes", otherwise selects "No". The result is returned to the method in the form of enterprise pairs.

[0054] Step Six: After the enterprise obtains the risk access score, the method will calculate the influence intensity of each characteristic category of the enterprise on the risk access score, that is, calculate the influence intensity of characteristic category F k . Taking enterprise x 1 as an example, when calculating the influence intensity of the financial status category, first remove the characteristics of the financial status category and calculate the score f -财务状况 (x 1 ) after removing the characteristics. By calculating , the influence intensity of the financial status is obtained as 0.43. Repeat the above steps until the influence intensities of all categories are calculated. The influence intensities of all categories are between -1 and 1, and the results are presented in a chart form below the risk access score, as shown in Figure 2 .

[0055] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for scoring financial leasing risks based on human feedback learning, characterized in that: The steps include: Step 1: Enterprise data acquisition: Establish a data warehouse to store public data from the Internet and the company's financial data. Public data from the Internet includes industry data, administrative penalty data, credit data, regional data, and legal litigation data. Financial data includes balance sheets, cash flow statements, and income statements. Step 2: Constructing the risk characteristics of financial leasing: Based on the data in the data warehouse, and targeting the unique needs and potential risk points of the financial leasing business, design the five core characteristic categories of the lessee: enterprise qualifications, operating conditions, financial conditions, regional qualifications and legal proceedings, and define the data source and calculation method of the characteristics; Step 3: Risk assessment model training: Collect historical data of enterprises in past financial leasing projects, construct training set D, and train and optimize the neural network model; Step 4: Risk access score prediction: Calculate the risk access score based on the characteristics of the lessee to assess its financial leasing risk. The higher the access score, the lower the risk. Step 5: Human feedback collection: In order to improve model performance, users are required to provide feedback on the access score of the financial leasing risk assessment model, that is, to provide human feedback on the results predicted by the neural network risk assessment model, and to strengthen the neural network model with the help of human feedback marking to improve the accuracy of the neural network prediction model; Step 6: Interpretability of the risk assessment model: Calculate the impact of each feature category on the model score to quantify the contribution of each feature category; Step 7: Financial leasing decision: Present the results in a chart report to assist in financial leasing decision-making.

2. The method for scoring risk of financial leasing based on human feedback learning according to claim 1, characterized in that: Enterprise qualification categories include: enterprise type, industry trends, and whether it is a key industry cluster; The business status categories include: number of fines, fine amount, tax level, and third-party rating; Financial status categories include: net profit margin, return on total assets, operating profit margin, debt-to-asset ratio, times interest earned, quick ratio, operating growth rate, net profit growth rate, and total asset growth rate; The regional qualification categories include: regional score; Legal litigation categories include: number of contract disputes, number of false statements, and number of invention patent disputes.

3. The method for scoring risk of financial leasing based on human feedback learning according to claim 1 is characterized in that: In step 2, construct the risk characteristics of financial leasing and determine the data source and calculation method of the characteristics. The specific steps are as follows: Step 201. Construct feature categories: Based on the unique needs and potential risk points of the financial leasing business, design five feature categories for the lessee: enterprise qualification F1, operating status F2, financial status F3, regional qualification F4 and legal proceedings F5, namely 1≤k≤5,m k is the number of features of the kth category, let x i Represents the characteristic vector of the i-th enterprise. For any enterprise, its risk characteristics are fully described by F1, F2, F3, F4 and F5, that is, x i =[F1,F2,F 3, F4, F5], 1≤i≤N, N is the number of enterprises; Step 202. Determine the data source and calculation method of the feature: Determine the risk characteristics of financial leasing based on the feature category, and summarize the data source and calculation method of each feature.

4. The method for scoring risk of financial leasing based on human feedback learning according to claim 2 is characterized in that: The categories, characteristics, data types and values, and data sources of financial leasing risk characteristics are shown in the following table:

5. The method for scoring risk of financial leasing based on human feedback learning according to claim 1 is characterized in that: In step three, a training set is constructed to train and optimize the risk assessment model for subsequent enterprise scoring. The specific steps are as follows: Step 301. Construct a training set: y i ∈ [0, 100], where M < N is the number of training samples, and y i represents the risk access score for financial leasing of this enterprise; In the initial training phase of the scoring model, the training set is manually constructed based on samples of companies that have already carried out financial leasing business; In the subsequent iterative optimization process of the scoring model, human feedback is gradually introduced into the training set to improve model performance and prediction accuracy; Step 302. Model training: Build a risk assessment model based on a deep neural network. The model is constructed using a multi-layer perceptron, which consists of an input layer, multiple hidden layers, and an output layer. The model parameters are learned through training set data. The model is used for enterprise sample x i The risk assessment and decision-making process can be expressed as i =f(x i ).

6. The method for scoring risk of financial leasing based on human feedback learning according to claim 1 is characterized in that: In step 5, the risk assessment model described in step 302 is optimized using human feedback, and the specific steps are as follows: Step 501. Rating feedback: Based on their own professional knowledge and actual business experience, the user reviews the comprehensive situation of the lessee enterprise, adjusts the scoring results predicted by the model, and feeds the revised results back to the model training process. Two feedback mechanisms are supported: (1) direct feedback: the user corrects the predicted score of a single enterprise; (2) paired feedback: the user provides feedback on the comparative relationship between two enterprises; Step 502. Model training based on human feedback learning: By introducing an adaptive layer, the loss function of the deep neural network is optimized, and direct feedback and paired feedback information are integrated into it, thereby guiding the iterative optimization process of the risk assessment model.

7. The method for scoring risk of financial leasing based on human feedback learning according to claim 1 is characterized in that: In step 6, we improve the interpretability of the model and calculate the impact of each feature category on the model prediction score. The steps are as follows: Step 601. Calculate the impact strength: k∈[1,5],w k ∈[-1,1], f -k (x i ) means to remove F k The risk score after the feature, f(x i ) represents the risk assessment model; Step 602. Explain the impact strength: By calculating w k The value of quantifies the contribution of each feature category to the final score.

8. A method for scoring risk of financial leasing based on human feedback learning according to claim 7, characterized in that When 0 <w k <1, which means F k The feature has a positive impact on the risk admission score, and w k The larger the value, the higher the F k The more significant the positive impact of the feature on the model prediction score, the higher the admission score and the lower the risk of financial leasing; <w k <0, which means F k The feature has a negative impact on the risk admission score, and w k The smaller the value, the higher the F k The more significant the negative impact of a feature on the model's prediction score, the more likely it is that the company's admission score will be lowered and the risk of financial leasing will increase.

Images