Data binning processing-based peasant household credit evaluation pre-model and application thereof

Abstract

The invention discloses a peasant household credit evaluation pre-model based on data binning processing and an application thereof, data is collected and verified in essential key links (land contract / circulation, irrigation water and electricity and the like) in agricultural planting, and python data programming is adopted to carry out peasant household planting data self-defined WOE binning and multi-dimensional data fusion processing to form an IV (Information Value), so that the peasant household credit evaluation pre-model can be used for evaluating the credit of the peasant household. And obtaining a real and credible peasant household credit pre-evaluation grading model based on IV statistical grading. The method comprises the following steps: taking a grain planting cycle (year) as a node, obtaining an evaluation result of batch peasant household data according to an evaluation grading model, comprehensively matching the evaluation result with a credit and credit conservation result of a peasant household, dynamically adjusting a weighting coefficient of a WOE result according to a matching result so as to approach an optimal score and conservation matching effect, and obtaining updating of a pre-model. The application of the model can bring more honest common peasant households into the credit range of banks, and supports the development of the general financial business of banks in rural areas.

Description

Technical Field

[0001] The present invention relates to the field of credit evaluation methods for rural farmers who specialize in staple food planting (hereinafter referred to as farmers). Banks need to do credit evaluation when issuing inclusive financial loans to farmers. The present invention creatively uses a customized WOE (weight of evidence) binning technology, takes multidimensional data fusion processing in the process of farmers' planting (food planting) as one of the basic conditions for farmers' credit evaluation, forms a credit evaluation pre-model based on planting and management data processing, and applies it to the bank's agricultural credit products, realizes the calculation of multidimensional farmers' planting data, and then uses the comprehensive calculation results to form a credit evaluation pre-model and apply it to agricultural credit, so that more honest and trustworthy ordinary farmers are included in the bank's credit scope, and supports the development of the bank's inclusive financial business in rural areas. Background Art

[0002] Rural areas are an important foundation for my country's economic development. The staple food planting industry in agriculture plays a vital and fundamental role in my country's economic and social development. Bank credit is one of the most important financial activities. However, the current situation in my country's rural areas is that there is a general lack of credit assets recognized by banks, which in turn hinders the development of credit, especially inclusive finance in rural areas.

[0003] The existing credit evaluation mechanism mostly uses credit assets, cash flow and other data as the evaluation basis, lacking comprehensive credit evaluation of agricultural planting multi-dimensional production and operation data, resulting in low credit coverage of ordinary farmers by banks, slow development of inclusive finance in rural areas, and lack of a healthy development mechanism. Grain planting is one of the basic industries of my country's agriculture, with the policy support of national storage. The grain planting industry also has the characteristics of low risk of industry losses and rigid demand. Inclusive finance has a good development prospect in rural areas.

[0004] Introduce multi-dimensional data on rural staple food production and management, and use Python data programming and analysis technologies to calculate land contracting, irrigation water and electricity consumption data in boxes, remove the impact of false data and outliers, form a pre-screening model for farmers' credit evaluation, and adjust it dynamically. Combined with the bank's existing general post-model, it can support the development of agricultural planting credit business, maintain the credit environment of rural planting industry, and then support the healthy development of planting industry credit business, which is in line with the long-term interests of the country and rural areas. Summary of the invention

[0005] The purpose of this invention is to provide a farmer credit evaluation pre-model based on customized binning processing technology of planting data, so as to solve the lack of credit evaluation mechanism for planting management of professional planting households in agricultural credit, and to iteratively upgrade it in actual credit applications to support the long-term development of bank credit benefiting farmers.

[0006] In order to achieve the above-mentioned purpose and ensure the development of agricultural planting credit, the technical innovation method adopted by the present invention is:

[0007] Data is collected and verified at the essential key links in agricultural planting (land contracting / transfer, irrigation water and electricity, etc.), and python data programming is used to customize WOE binning of farmers' planting data and multi-dimensional data fusion processing to form IV (Information Value). Based on IV statistical grading, a true and credible farmer credit pre-evaluation grading model is obtained. Taking the grain planting cycle (year) as the node, the data of batches of farmers are evaluated according to the evaluation grading model to obtain the evaluation results, and then the evaluation results are comprehensively matched with the farmers' credit compliance results. The weighting coefficient of the WOE result is dynamically adjusted according to the matching results to approach the best matching effect of scoring and compliance, and obtain the update of the pre-model. This model processes a large number of farmer data samples to generate rating results. Its application can bring more honest and trustworthy ordinary farmers into the scope of bank credit and support the expansion of bank inclusive financial services in rural areas. Its main steps include:

[0008] (1) Sampling, verification and abnormal processing of farmers' contracted land data, and calculation of W1 by binning;

[0009] (2) The data of farmers’ irrigation water and electricity consumption are retrieved, verified and comprehensively processed, and W2 is calculated by binning;

[0010] (3) Determine the weight coefficients a1 and a2 (set the initial value and step value) of the data obtained in step (1) and step (2), respectively. If this step is entered from step (2), it is the initial value, otherwise it is the initial value + step value * (number of times - 1), and calculate IV weightedly to obtain a comprehensive scoring result;

[0011] (4) The IV obtained in step (3) is divided into intervals, and a rating result is formed according to the IV. The rating result is matched with the compliance result in the bank's credit according to the farmer's ID. If the matching success rate is ≥ the threshold value α or the number of matches exceeds the set upper limit n, the rating result is used as the final output; if the matching success rate is less than the threshold value α, return to step (3).

[0012] The verification and processing of the farmer's contracted land data is characterized by using Python data analysis technology to comprehensively verify the farmer's contracted land area with land rights confirmation data, transfer and operation contracts and other data, and eliminate farmers whose difference ratio is above a specific threshold k1. The verified and corrected results are comprehensively converted into standard land area according to the land quality grade to form WOE1.

[0013] The verification and comprehensive processing of farmers' planting irrigation water and electricity consumption data is characterized by using python data analysis technology to comprehensively compare the farmers' planting irrigation electricity consumption data (crops are the main food) statistically collected within the region with the local average irrigation planting area (crops are the main food), convert it into a standard per-mu electricity consumption × planting area (unit: mu), and compare it with the individual farmers' irrigation electricity consumption data, excluding farmers with a difference ratio of more than k2, and the verification and corrected results are comprehensively converted into a standard planting land area according to the per-mu irrigation electricity consumption value to form WOE2.

[0014] The WOE calculation of farmer planting data is characterized by customizing the farmer contracted planting land area data and planting irrigation data, and independently performing the bin calculation, and finally performing weighted calculation to obtain IV. The differences in planting data of different dimensions are first isolated and then merged in order to achieve the best matching result.

[0015] The present invention first establishes a land confirmation and land transfer data space based on the authoritative data source according to the area where farmers plant (such as: county), as the standard verification data of farmers' planting area, and then collects the actual cultivated land area of ​​farmers (data sources such as contracting contracts) and crops planted as the data of the first dimension of the front model. According to the experience of tracking and studying the planting industry, python data analysis programming is used to realize data binning processing, and its low-credibility data is eliminated and then its evidence weight (ie WOE1) in the front model is calculated. In this step, on the one hand, the binning discretization processing technology is used to remove the influence of outliers whose deviation values ​​exceed the threshold k1, and the authenticity problem of the planting area data in the farmer planting sample is solved by verifying the authoritative data source and removing the outliers; on the other hand, the regional industry research foundation is used to determine the weight WOE1 of the result formed by the planting area processing in the front model, which not only reflects the basic role of the planting area data in the model, but also avoids measuring the total scale of planting operations by the planting area alone, improves the evaluation effectiveness of the model, and reduces the evaluation risk of the model.

[0016] Secondly, according to the actual situation of staple food planting in North China (Hebei, Henan, Shandong and other places), it is determined that irrigation water and electricity consumption is one of the key links in staple food planting, and the authenticity of water and electricity consumption data can be verified based on authoritative data sources. By collecting farmers' planting water and electricity consumption data (on a monthly basis), it is used as the data of the second dimension to form the pre-evaluation model. According to the follow-up study of the irrigation water and electricity consumption of staple food planting, the customized data WOE binning is made based on the statistical mean of regional irrigation water and electricity consumption, and its evidence weight in the pre-evaluation model (ie WOE2) is calculated. In this step, the difference discretization processing technology in the binning calculation is used to remove the influence of outliers whose variance deviation from the statistical mean of monthly water and electricity consumption exceeds the threshold k2, thereby increasing the evaluation ratio of farmers' planting input.

[0017] The present invention comprehensively utilizes the pandas and numpy libraries in python data analysis programming technology, comprehensively verifies the collected agricultural planting contract data samples and irrigation water and electricity data samples with the authoritative database, eliminates the influence of large deviations, and calculates WOE in relatively independent bins, then weights the WOE results to form IV, discretizes IV to form a graded evaluation result, correlates and tracks the farmer's evaluation level with the farmer's compliance limit obtained by the bank's post-model, adjusts the weighting coefficient annually according to the matching result, and finally forms an effective and usable pre-model for farmer credit evaluation.

[0018] The present invention is based on Python programming and modularization, which is convenient and easy to use, with clear applicable scenarios, and the coefficients are adjusted according to the characteristics of different regions. Through bin calculation and dynamic coefficient adjustment and other technologies, the authenticity and timeliness of farmers' planting data are enhanced, and the specific problems of farmers' actual planting data acceptance and availability in the credit granting of farmers by banks are solved. To a certain extent, the information gap between farmers' actual planting conditions and bank credit is eliminated, and it is committed to bringing ordinary farmers into the scope of bank credit and supporting the development of inclusive finance in rural areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The block diagram of the planting data custom binning processing system in the present invention

[0020] Figure 2 The principle block diagram of similarity matching and parameter adjustment in the present invention is DETAILED DESCRIPTION

[0021] like Figure 1 As shown. The present invention uses Python data analysis technology to pre-process the original collected data of farmers, including cleaning, verification and other processes, to obtain verifiable data. In this process, the pandas library is used to perform data comparison and calculation, and the abnormal values ​​of the collected data are eliminated; the binning calculation technology is used to obtain woe, and the weighted calculation is performed according to the coefficient to obtain IV.

[0022] 1. Use Python programming to analyze and verify the original data

[0023] Farmers collect data and process it in a structured way to form computable structured data and process the contracted land area. The following is a code example of python processing data

[0024]

[0025]

[0026] 2. Use the phthon tool library to perform binned WOE and IV calculations. Perform binned calculations according to the values ​​of the farmer samples, and perform woe and IV calculations on the processed data. The following is a code diagram:

[0027] df = copy.copy(data)

[0028] df['pos_rate']=(df['pos']+1) / df['pos'].sum()#Calculate the response ratio

[0029] df['neg_rate']=(df['neg']+1) / df['neg'].sum()#Calculate the percentage of non-response

[0030] df['woe'] = np.log(df['pos_rate'] / df['neg_rate']) # Calculate group WOE

[0031] IV=\sum_{i=1}^{2}(P(Yes)_i-P(No)_i)\times WOE_i#Calculate IV

[0032] 3. Discretize IV according to the statistical value interval to form a seven-level evaluation level

[0033] Based on a typical data sample of farmers in a certain place in the north (obtained legally through official channels), the IV results calculated by the planting data are discretized into intervals to form a seven-level evaluation model. Level 1 is the highest, representing the highest level of farmers' planting scale and operating input in typical planting areas in the north (reference interval is planting more than 1,000 mu and credit limit of 1 million yuan), and level 7 is the lowest (reference interval is planting less than 50 mu and credit limit less than 50,000 yuan), representing the lowest level of creditable planting scale and operating input.

[0034] like Figure 2 As shown in the figure. The application of this model matches the graded evaluation results of the batch of farmers who have obtained bank credit with the credit limit of farmers obtained by the bank's post-model, and dynamically adjusts the parameters according to the matching results. In this process, pandas data matching is used, and the matching results are used as the basis for parameter adjustment; the parameter adjustment uses Δ as the step value, and the two parameters are added and subtracted respectively.

[0035] The following is a schematic code of the matching and dynamic parameter adjustment process

[0036]

Claims

1. A farmer credit evaluation pre-model based on data binning and its application, characterized by: Data is collected and verified at the essential key links in grain planting (land contracting / transfer, irrigation water and electricity, etc.), and python data programming is used to customize WOE binning of farmers' planting data and multi-dimensional data fusion processing to form IV (Information Value). Based on IV statistical grading, a true and credible farmer credit pre-evaluation grading model is obtained. Taking the grain planting cycle (year) as the node, the data of batches of farmers are evaluated according to the evaluation grading model to obtain evaluation results, and then the evaluation results are comprehensively matched with the credit compliance results of farmers. The weighting coefficient of the WOE results is dynamically adjusted according to the matching results to approach the best matching effect of scoring and compliance, and obtain an updated pre-model. The use of this model can bring more honest and trustworthy ordinary farmers into the scope of bank credit and support the development of bank inclusive financial services in rural areas. The main steps include: (1) Sampling, verification and abnormal processing of farmers' contracted land data; (2) Retrieval, verification and comprehensive processing of farmers' irrigation water and electricity consumption data; (3) Determine the weight coefficients a1 and a2 (set the initial value and step value) for the data obtained in step (1) and step (2), respectively. If the step is entered from step (2), the initial value is used; otherwise, the initial value + step value * (number of times - 1) is used. The weighted calculation (WOE weight of evidence) is performed to obtain the comprehensive scoring result. (4) The scoring results obtained in step (3) are divided into intervals to form rating results, and the rating results are matched with the compliance results in the bank's credit according to the farmer's ID. If the matching success rate is ≥ the threshold value α or the number of matches exceeds the set upper limit n, the rating result is used as the final output; if the matching success rate is less than the threshold value α, return to step (3).

2. The verification and processing of farmer contracted land data according to claim 1 is characterized in that Python data analysis technology is used to comprehensively check the contracted planting land area of ​​farmers with land rights confirmation data, transfer and operation contracts and other data, and farmers whose difference ratio is above a specific threshold k1 are eliminated. The corrected results are converted into standard land area according to the land quality grade to form WOE1.

3. The verification and comprehensive processing of water and electricity consumption data for irrigation of farmers according to claim 1 is characterized in that Python data analysis technology was used to comprehensively compare the farmers' irrigation electricity consumption data (crops are the main food) in the regional area with the local average irrigated planting area (crops are the main food), and convert it into a standard per-mu electricity consumption × planting area (in mu), and then compare it with the individual farmers' irrigation electricity consumption data, eliminating farmers with a difference ratio of more than k2, and the verified and corrected results were converted into a standard planting land area according to the per-mu irrigation electricity value to form WOE2.

4. The custom binning calculation of farmer planting data according to claim 1 is characterized in that The data of contracted planting land area and planting irrigation data of farmers are divided into custom bins, and the bin calculations are performed independently, and finally the IV is obtained by weighted calculation. The differences in planting data of different dimensions are isolated and then merged in order to achieve the best matching results.

Images