1. Part of wind control system
1.Blaze
Blaze is a product of FICO, used for rule management, and is the predecessor of model ABC card development. When credit companies start lending, the amount of data is small, and there are few applicants, so it is difficult to build a model. Therefore, in the early stage, expert experience is generally used to judge good or bad customers, and then efficient operation is carried out through the risk control decision-making management system. Among them, blaze is a risk control decision-making management system with high efficiency that has been used for many years. However, blaze is a commercial product, and is generally used in large consumer finance companies such as big banks and Home Credit. The fee can be higher than 1 million RMB per year. If more customized services are required, the fee will be higher.
1.1 A card
definition: Application scorecard application scorecard, the rules for assigning values to the materials submitted in the credit granting stage.
For example: "Incoming" is a term used by traditional banks, which refers to the application form. A scorecard is a comprehensive judgment on a series of user information. With the increase of user information that can be collected, credit decision makers are no longer satisfied with simple if and else logic, but hope to assign weights and scores to each information, judge risks based on the final comprehensive score of users, and adjust by delineating the score line Risk tolerance, the scorecard came into being. The scorecard is a derivative of the logistic regression algorithm.
1.2 B card
definition: Behavior scorecard Behavior scorecard, the rules for scoring user information that can be collected after lending.
Example: Similar to the A card, the B card is also a set of scoring rules. After the loan is issued, by collecting the behavior data of the user after receiving the money, it is speculated whether the user will be overdue and whether the user can continue to borrow money. For example, after a user obtains a loan from a certain bank, and then applies for loans from several other banks, it can be considered that this person is short of funds and may not be able to repay the loan. If he applies for a bank loan again, he must lend money carefully. In the B-card model, there are many sub-models for stock management, including activating the silent customer model, finding out high-value customers, increasing the loan amount model, and so on.
1.3 Definition of C card
: Collection Scorecard Collection Scorecard, a scoring rule for judging the ability of users who have expired in the future to send reminders.
Example: Collection Scorecard is a derivative application of Behavior Scorecard, and its function is to predict the collection intensity of overdue users. For users with good reputation, the money can be recovered without collection or light collection. For users who tend to be overdue for a long time, it is necessary to focus on collection from the overdue. The more days overdue, the more difficult it is to collect.
Collection is generally divided into multiple agents, and the experience and business capabilities of different agents such as M1, M2, and M3 vary greatly. AI artificial intelligence is often used in the early stage of automatic collection.
The application score card, behavior score card and collection score card are often combined and called "ABC card", which is used in pre-loan, loan and post-loan management. . . . . . . .
2. Risk control indicators
- 1 Aging Analysis
Interpretation: aging analysis. Displays the lag rate from each period to the observation point, which is characterized by the same settlement end point, and combines the loans scattered in each month into one observation time point to calculate the overdue ratio.
- 2 Vintage Analysis
Interpretation: Statistics of the overdue situation of new loans each month in the following months are also aging analysis. Different from aging analysis, vintage is based on the aging of the loan, observing the overdue ratio of N months after the loan. It can also be used to analyze the follow-up quality of lending in each period, and observe the impact of the adjustment of entry rules on the quality of creditor's rights. Example: Deliqueency Vintage 30+: Remaining principal of 30+ overdue month / corresponding bill generation monthly loan amount. Risk Control Chinese and English Terminology Manual (Bank_Consumer Finance and Credit Business)_v4_Terminology Manual
- 3 C 、M
Interpretation: C and M are proper nouns describing overdue period buckets. M0 is a normal asset, Mx is overdue for x periods, and Mx+ is overdue for x periods (inclusive). The bucket with no overdue normal repayment is M0, which is C, and M1 is over 1 period (1-29 days). M2+ means more than 2 periods and above (30+). M2 and M4 are two important observation nodes. It is generally considered that M1 is the early stage, M2-M3 is the mid-term, and above M4 is the late stage. Transfers of bad debts larger than M6 are considered.
- 4 Delinquency
Interpretation: overdue rate/delay rate. The indicators for evaluating asset quality can be divided into Coincident and Lagged observation methods.
- 5 Coincident
Interpretation: spot indicators. It is used to analyze the quality of all accounts receivable in the current period and calculate the delay rate. The calculation method is to divide the deferred amount of each bucket in the current period by the total accounts receivable (AR) in the current period. Coincident is an overview of the whole at the current observation point, so it is vulnerable to fluctuations caused by the level of accounts receivable in the current period, which is suitable for observing asset quality when the total business volume does not fluctuate greatly. Example: Coincident DPD 30+, a commonly seen indicator
- 6 Lagged
Interpretation: Deferred indicators. The same as coincident, it is also an indicator for calculating the delay rate. The difference is that the denominator of lagged is the accounts receivable of the period in which the overdue amount was generated. Lagged observes the overdue ratio generated in the current period of lending, so it is not affected by the fluctuation of accounts receivable in the current period. Example: Lagged DPD 30+$(%)= Lagged M2+Lagged M3+Lagged M4+Lagged M5+Lagged M6 Asset balance M1 at the end of the month (1-29 days): Assets at the end of the statistical month satisfying 1≤current overdue days≤29 The sum of the remaining principal of the order, the current overdue days is the current maximum overdue days of the order, excluding bad debt orders. Lagged M1 = loan balance of M1 at the end of the month/loan balance at the end of the previous month (M0~M6) Risk Control Chinese and English Terminology Manual (Bank_Consumer Finance and Credit Business)_v4_Risk Control_02
- 7.0 PD（Past Due）
For example, the letters in front of FPD1, SPD7, TPD30..., F: first, means that the first period is overdue. Similarly, S, T, and Q represent two, three, and four respectively, and will be represented by numbers later. Such as 5PD30. The following number refers to the number of days overdue. If a customer has a mark of FPD30, there must be a mark less than 30, such as FPD1 and FPD7. dpd (days past due) is the number of days past due. For loan products, it is calculated from the day after the payment deadline (usually the next account closing date). Among the 4 installments, if any one installment is overdue for more than 30 days, it will be regarded as a bad customer. One thing to note is that the PD indicators are usually mutually exclusive, that is to say, if a person has the FPD mark, he will not have the SPD mark. Customers whose payment is overdue in the second installment.
- 7 DPD
Interpretation: Days Past Due is the number of days overdue, the number of days from the day after the repayment date to the actual repayment date. Example: DPD7+/30+, past due history of more than 7 days and 30 days. The relatively strict overdue rate calculation formula in the industry is: at a given point in time, divide the total outstanding remaining principal of loan accounts that are currently overdue for more than 90 days by the total amount of accumulated contracts that may result in 90+ overdue. The concept of its numerator is that as long as it has been overdue for more than 90 days, the total remaining principal of the contract that has not been repaid is considered to be overdue, while the denominator considers that some loans with a very short aging time are absolutely impossible to generate 90+ overdue The contract amount is excluded (for example, borrowing only 2 days ago, it is impossible to be more than 90 days overdue anyway).
- 8 FPD
Interpretation: First Payment Deliquency, the first repayment overdue. After the user's credit approval, the first bill that needs to be repaid, the proportion of customers who have not repaid within 7 days after the last repayment date and have not applied for extension is FPD 7, and the numerator is the order placed during the observation period and has been overdue for more than 7 days The number of users, the denominator is the number of users in the observation period who placed all the first orders in the current period and met the repayment date 7 days later. Commonly used FPD indicators are FPD 30. Example: Assume that the user approved the credit on 10.1, and generated the first loan in three installments on 10.5, and set the 8th of each month as the repayment date. Then 11.08 is the repayment date of the first bill, and the repayment after the bill date and before the end of the repayment date is not considered overdue. If the loan has not been repaid on 11.16, it will be included in the numerator of the risk control Chinese and English terminology manual (bank_consumer finance and credit business)_v4_bank_03 FPD7 for the period of 10.1-10.30. Usually, users who are overdue for a few days may have forgotten to repay the loan or are short of money for a while, but the FPD 7 indicator can be used by users to evaluate the credit risk of credit-granting groups and predict the health of future assets. Similar to FPD 7, FPD 30 is also an indicator for observing the overdue situation of the user's first outstanding bill. For users who are overdue within 30 days, some losses can be recovered by increasing collection efforts. For users who are overdue for more than 30 days, the probability of collection will be greatly reduced, and outsourced collection may be carried out. If the users’ FPD 7 is relatively high for a period of time, and most of the less repayments fall into the FPD 30, it proves that the proportion of non-starters in this group of users is high, and they never thought of repaying the loan at all, otherwise it means that the users The credit risk of the group is more serious.
- 9 Cpd30mob4
cpd is used in the collection model and is a collection indicator. The repayment performance is whether the overdue time of the fourth month is more than 30 days, excluding history
- 0 maxdpd30_mob4
Within the four observation periods (months), whether overdue exceeds 30 days, including history
- 1 MOB in account month
Example of the month after disbursement: MOB0, the disbursement date to the end of the month MOB1, the second complete month after disbursement MOB2, the third complete month after disbursement mob3-3 months is the short observation period, mob6-6 months is the long observation period
- 2 Flow Rate
Interpretation: Migration rate. Observe the probability that the overdue amount in the previous period will continue to fall into the next period after collection. Example: M0-M1=Asset balance M1 at the end of month M / loan balance of M0 at the end of last month August M0-M1: loan balance entering M1 in August / loan balance of M0 at the end of July at the beginning of August Supplementary information: macro Medium and short-term economic risks can be measured using FDP, SPD, TPD; medium-term risks can be measured using 30+@MOB4; long-term risks can be measured using 90+@MOB6 etc. To measure the short-term risk, FPD, SPD, TPD could be used; To measure the middle-term risk, 30+@MOB4 could be used; To measure the long-term risk, 90+@MOB6 could be used; different products apply different indicators Fpd30 (cash loan products) maxdpd30_mob4 (existing customers) Cpd30mob4 (collection customers) ) Definition of Bad Customers for Auto Loans (for reference only) Risk Control Chinese and English Terminology Manual (Bank_Consumer Finance and Credit Business)_v4_Bank_04 Note: Due to the subdivision of scenarios, different scenarios are quite different, and the above indicators are for reference only .
3. Risk control model part
3.1
Interpretation of Benchmark: Benchmark. Each version of the new model is compared with an online benchmark model or rule set.
3.2
Interpretation of IV: information value Information value, also known as VOI, value of information, value range (0,1). This value is used to represent the predictive ability of a variable, the bigger the better. The threshold of financial risk control screening variables is 0.02. If the variable's iv is lower than 0.02, then the variable will be kicked out. As a model expert, I remind everyone that the iv value cannot be memorized by rote, and the threshold needs to be customized according to the distribution characteristics of the scene data. The distribution of variable iv values in different scenarios may vary greatly, such as loan lending, car loan and cash loan.
3.3 Interpretation of KS value
: KS refers to klmogrov-smirnov, which is an index of differentiation. The so-called differentiation power refers to the ability of the model to distinguish good and bad customers. The KS value ranges from 0-1, the bigger the better, the smaller the worse. In real scenarios, the model ks in the field of risk control can rarely exceed 0.4.
3.4 PSI
interpretation: population stability index, stability index, the lower the more stable. It is used to compare the degree of difference between the current customer group and the model development sample customer group, and evaluate whether the effect of the model meets expectations. The closer the PSI is to 0, the better the model stability. When the PSI is less than 0.1, the model is relatively stable. When the psi is between 0.1 and 0.25, the model stability fluctuates, and the model needs to be checked and redeveloped if necessary.
3.6 Logloss
Interpretation: logarithmic loss function
As the predicted probability approaches 1, the log loss drops off slowly. But the log loss increases rapidly as the predicted probability decreases. The larger the log loss value, the less accurate the model is, and vice versa.
3.7 Training Sample
Interpretation: Modeling samples, a set of expressive user data used to train the model. There is also an off-time sample (verification sample) in conjunction with this sample. Both samples take the same user dimension. Usually, the model trained by the modeling sample is used for verification on the verification sample.
3.8
Interpretation of WOE: weight of ecidence, weight of evidence, value range (-1,1). The proportion of breached items is higher than normal items, and WOE is negative. The higher the absolute value, the stronger the ability of this group of factors to distinguish good and bad customers. The data of the scorecard model needs to convert the original data into woe data, so as to reduce the variance of the variables and make them smooth. The IV value is also converted from the woe value. Since woe has certain defects in evaluating variables, it generally uses the iv value to evaluate the importance of variables.
3.9 Bad Capture Rate
Interpretation: bad user capture rate. This is an indicator for evaluating the performance of the model, and the higher the ratio, the better.
Example: Top 10% Bad Capture Rate refers to the ratio of the top 10% of the worst users evaluated by the model to bad users in the sample.
3.10 Population
definition: All Population, all sample users, including modeling samples and verification samples.
3.11 Variable
Interpretation: variable name. Each model relies on many basic and derived variables as input parameters. The naming of variables needs to conform to the specification, easy to understand and expand. Variables need to be screened before modeling. In big data models, more than 90% of the variables are noise variables. Really useful variables are very few of them.
3.12
Interpretation of CORR: correlation coefficient. The closer the absolute value of Corr is to 1, the higher the degree of linear correlation, and the closer to 0, the lower the degree of correlation. The calculation of the correlation coefficient depends on the data distribution. If the data presents a normal distribution, the accuracy of the Pearson method is higher; otherwise, the Spearman method is more appropriate.
3.13 confusion matrix confusion matrix
Sensitivity: Under true positive conditions, the test is also positive
specificity: Under true negative conditions, the test is also negative
FALSE positive: The test is positive under true negative conditions
FALSE negative: The test is negative under true positive conditions
3.14 Model Algorithm
Logistic regression
Logistic regression is a generalized linear model, so it has many similarities with multiple linear regression analysis. Their model forms are basically the same, both have w'x+b, where w and b are the parameters to be sought, and the difference is that their dependent variables are different. Multiple linear regression directly uses w'x+b as the dependent variable, that is, y =w'x+b, and logistic regression uses the function L to correspond w'x+b to a hidden state p, p =L(w'x+b), and then determines the dependent variable according to the size of p and 1-p value. If L is a logistic function, it is logistic regression, and if L is a polynomial function, it is polynomial regression.
The dependent variable of logistic regression can be binary or multi-category, but the binary category is more commonly used and easier to explain, and multi-category can be processed using the softmax method. The most commonly used in practice is the logistic regression of the two classifications.
scorecard model
The scorecard model is a derivative of the logistic regression algorithm. Apply woe binning and score stretching techniques to convert logistic regression probability scores into standard scores. Standard scores are similar to FICO scores or Sesame Credit scores, ranging from 300 to 900. The figure below shows the scoring mode of the scorecard
Scorecard-related tutorials: https://edu.csdn.net/course/detail/30611
Support Vector Machine (SVM)
Support Vector Machine (SVM) is a kind of generalized linear classifier (generalized linear classifier) that performs binary classification on data according to supervised learning, and its decision boundary is the maximum margin for solving the learning samples Hyperplane (maximum-margin hyperplane). SVM was proposed in 1964, developed rapidly after the 1990s and derived a series of improved and extended algorithms, which have been applied in pattern recognition problems such as portrait recognition and text classification. The support vector machine algorithm works better on small sample data, but it takes a long time to train large data.
Neural network
Logical thinking refers to the process of reasoning according to logical rules; it first converts information into concepts and expresses them with symbols, and then performs logical reasoning in serial mode according to symbolic operations; this process can be written as serial instructions, allowing the computer to implement. Intuitive thinking, however, is the synthesis of distributed stored information and the result is a sudden idea or solution to a problem. The fundamental point of this way of thinking lies in the following two points: 1. Information is stored on the network through the distribution of excitation patterns on neurons; 2. Information processing is completed through the dynamic process of simultaneous interaction between neurons.
Note: The operating principles of computer neural networks and human brain biological neural networks are different.
A little bit: efficient in processing big data, can handle complex and multi-dimensional data, flexible and fast
Disadvantage: data needs to be preprocessed
xgboost
XGBoost is an optimized distributed gradient boosting library designed to be efficient, flexible and portable. It implements machine learning algorithms under Gradient Boosting framework. XGBoost provides parallel tree boosting (also known as GBDT, GBM) that can solve many data science problems quickly and accurately. The same code runs on major distributed environments (Hadoop, SGE, MPI) and can solve problems beyond billions of examples. xgboost is an ensemble tree algorithm, invented by Chen Tianqi, which has won many championships in kaggle competitions
lightgbm
Light Gradient Boosted Machine, LightGBM for short, is an open source library that provides an efficient implementation of the gradient boosting algorithm. The algorithm developed by Microsoft has better overall performance than xgboost.
LightGBM extends the gradient boosting algorithm by adding a form of automatic feature selection and focusing on boosted examples with larger gradients. This can significantly speed up training and improve prediction performance.
Compared with other boosting related frameworks, it has the following advantages -
-
Train faster without compromising efficiency.
-
Memory usage is also low.
-
It provides better accuracy.
-
It supports both parallel and GPU learning methods.
-
It has the ability to handle large-scale data.
cat boost
Russian search giant Yandex announced that it will submit a gradient boosting machine learning library CatBoost to the open source community. It is capable of "teaching" machine learning with sparse data. Especially when there is no sensory data like video, text, images, CatBoost can also operate on transactional or historical data.
catboost features:
Little or no need to adjust parameters, the default parameters work very well
Support for categorical variables
GPU support
Catboost related tutorials: https://edu.csdn.net/course/detail/30742
4. Basic vocabulary of risk control
4.1 Interpretation of APR
: Annual percentage rate, annual percentage rate, the interest rate of annual compound interest calculation. nominal APR nominal interest rate, effective APR actual interest rate.
4.2
Interpretation of AR: accounts receivable, current accounts receivable.
4.3 Application fraud
interpretation: counterfeit application
4.4 Transaction fraud
interpretation: fraudulent transaction
4.5 Balance Transfer
interpretation: balance compensation, that is, credit card repayment business.
4.6 Collection
Interpretation: Collection. According to the user's collection time from short to long, it is divided into Early collection (early collection), Front end (front-end collection), Middle range (middle-stage collection), Hot core (later-stage collection) Recovery (bad debt collection/bad debt income) Several stages correspond to different collection methods and frequencies.
4.7
Interpretation of DBR: debit burden ratio, debt ratio. Generally, the overall unsecured liabilities of the debtor in each channel should not exceed 22 times of its average monthly income.
4.8
Interpretation of Installment: installment payment
4.9
Interpretation of IIP: provision for bad debts
4.10
Interpretation of PIP: asset impairment loss
4.11 NCL
interpretation: net credit loss, net loss rate. The amount of bad debts transferred in the current period minus the recovery of bad debts in the current period is the net loss amount.
4.12 Loan Amount
Interpretation: total loan amount
4.13 MOB
Interpretation: month on book aging
Example: MOB0, loan date to the end of the current month. MOB1, the second full month after disbursement
4.14 Non-starter
interpretation: Malicious overdue customers
4.15 Payday Loan
interpretation: payday loan. Unsecured credit loan, fast lending speed, low amount, short term but high interest rate. Low quota and high interest rate are necessary conditions for this model.
4.16
Interpretation of Revolving: revolving credit. Tiqianle Credit Wallet provides users with revolving quotas, corresponding to special quotas for medical aesthetics and education.
4.17 Interpretation of WO
: Write-off, transfer of bad debts, usually overdue for more than 6 periods.
4.18 AR
AR credit approval rate = SUM (loan application approval account) / SUM (application account)
4.19 DR
DR default rate = SUM (default account) / SUM (used credit account)
4.20 EAD
EAD credit exposure=SUM(C0+M1+M2+…+M6+)
4.2 1 credit conversion rate
Credit conversion rate = SUM (used credit account) / SUM (applied account)
4.22 Delay rate/delay rate (flow through%)
The calculation can be divided into two methods: coincidental and lagged. In addition to the lag rate of each bucket, the lag rate above a specific bucket will also be observed. For indicators such as M2+lagged% and M4+lagged%, taking M2+lagged% as an example, the denominator is the accounts receivable two months ago, and the numerator is the overdue amount of M2 (including) that has not been transferred to bad debts this month. In the risk management of consumer finance, M2 and M4 are two important observation points. The reason is that the customer may be too busy or forget to cause the account to be overdue, but after M1 collection, it still falls above M2, which can be confirmed as inability to pay. or deliberate default.
4.23**** Defective rate (bad%)
In addition to the general risk analysis for the application of bad, the construction of the credit scoring model also needs to realize the definition of bad.
In addition to overdue accounts and high-risk accounts, the general definition of bad is mainly overdue accounts.
4.24**** Bad debt transfer rate (write-off%)
It is abbreviated as wo%, the amount of bad debts transferred in the current month/accounts receivable in the month of overdue start. After annualization, the monthly bad debt rate is converted into an annual loss rate.
4.2****5 Net Loss Ratio (NCL)
It is defined as: the amount of bad debts transferred in the current period - the recovery of bad debts in the current period, which is the concept of net loss. From the perspective of overall risk management performance, bad debt recovery is also an important part, so NCL% and WO% are often displayed together.
4.26 Appropriation rate
Also known as the allocation-to-loan ratio, it refers to the ratio of provisions to total loans. The higher the allocation-to-loan ratio, the stronger the bank's ability to defend against bad debt risks. The calculation formula is: provision balance/total loans=provision coverage ratio*non-performing loan ratio.
4.27 Provision coverage ratio
Also known as the provision adequacy ratio, it is actually the proportion of bad debt reserves that may occur in bank loans. The provision coverage ratio is the ratio of the actual provision for loan losses to non-performing loans. The best ratio is 100%. The calculation formula is: loan loss reserve/non-performing loan balance.
4.28 Non-performing loan ratio
Refers to the ratio of non-performing loans of financial institutions to the total loan balance. Non-performing loans mean that when estimating the quality of implied loans, loans are divided into five categories based on risk: normal, special mention, substandard, doubtful and loss, of which the latter three categories are collectively called non-performing loans. Calculation formula: non-performing loan ratio = (substandard loans + doubtful loans + loss loans) / various loans * 100% = loan provision ratio / provision coverage ratio * 100%. Loan provision ratio, non-performing loan ratio, and provision coverage ratio are the three basic indicators of the asset quality of the commercial banking industry.
4.29 Debt Ratio (DBR)
Debit burden ratio (DBR) is the main indicator that banks pay attention to. A common indicator for measuring a borrower's repayment stress is total unsecured debt outstanding (credit cards, debit cards, lines of credit)/average monthly income.
4.30 Malicious delay rate (non-starter%)
The original definition is "customers who have never paid after the loan", and the main purpose is to find cases of vicious fraud.
4.31 hit rate (hit%)
It is used in credit card halfway credit and early warning reports. The so-called hit rate refers to the probability of customer delay within a certain period of time after control. A hit rate that is too low may indicate flooding or misdirection of risk judgment.
4.32 Available Balance (OTB)
It often appears together with the hit rate indicator. The calculation method is to first find out the customers who have confirmed the control hit, and then adjust the credit card available balance of these customers when they are controlled. This figure can be regarded as the reduced loss of the bank due to the control.
4.33 Bad debt recovery rate
Recovery rate of bad debts in the current period = recovery of bad debts in the current period / amount transferred to bad debts in the current period
Total bad debt recovery rate for the current period = bad debt recovery for the current period / total balance of bad debts in the previous period
Bad debt recovery rate for this year = total bad debt recovery amount for this year / average bad debt balance for this year
Bad debt recovery rate for the last 12 periods = total amount of bad debt recovery for the last 12 periods / average bad debt balance for the last 12 periods
The 12-period recovery rate after the transfer of bad debts = the total recovery amount of the 12 periods after the transfer of bad debts / the average bad debts of the 12 periods after the transfer of bad debts
balance
5. Data dictionary
client_no: client account
apply_time: application time
gender: gender
age: age
income_range: income range
education: education level;
carreer: job;
credit_score:
credit score; credit_score_range
: credit
score range
; : Whether it is overdue;
used_time: Loan usage times;
credit_approved: Credit approved amount
5. Financial risk control modeling actual combat classic teaching case study
5.1 German credit data set (German credit)
5.2 Kaggle model competition give me some credit dataset
5.3 Jiangsu City Investment Enterprise Credit Rating
5.1-5.3 related tutorials: https://edu.csdn.net/course/detail/30611
5.4 Credit Dataset of US Fintech Company Lendingclub
5.5 Portraits of Consumers—Intelligent Credit Scoring
Organizer Fujian Provincial Digital Fujian Construction Leading Group Office & Fujian Provincial Department of Industry and Information Technology & Fuzhou Municipal People's Government & China Electronics Information Industry Development Research Institute & Digital China Research Institute & China Internet Investment Fund
5.4-5.5 related tutorials: https://edu.csdn.net/course/detail/30742
6. Websites for collecting financial information
6.1tradingeconomics****
The official website https://tradingeconomics.com/ contains hundreds of economic indicators from all over the world, including GDP, CPI, PPI, debt ratio, commodity price index and so on.
6.2 FRED economic data
Official website https://fred.stlouisfed.org/, financial data query
6.3 Bank of Japan
https://www.boj.or.jp/
6.4 wind database
Official website: https://www.wind.com.cn/Default.html, CICC Financial Industry Database
6.5 Paper Gold
Gold price and trading volume query, specific data download http://www.zhijinwang.com/etf/
6.6 Stock/bond market public opinion analysis and early warning related websites
Wind (https://www.wind.com.cn/)
East Money Network (https://www.eastmoney.com/)
Hexun Data (http://data.hexun.com/)
Bloomberg (https://www.bloomberg.net/)
6.7 Anti-Money Laundering Investigation
FATFhttp://www.fatf-gafi.org/
Financial Action Task Force on Money Laundering. An intergovernmental international organization established in Paris in 1989 by seven western countries to study the hazards of money laundering, prevent money laundering and coordinate international actions against money laundering. One of the world's largest international organizations. Currently comprising 36 member jurisdictions and 2 regional organisations, representing most major financial centers around the world. Its forty recommendations on anti-money laundering and nine special recommendations on anti-terrorist financing (referred to as FATF 40+9 recommendations) are the most authoritative documents on anti-money laundering and anti-terrorist financing in the world
6.8 Intelligent extraction of corporate financial announcement information to help bank account managers in marketing
Juchao Information Network (http://www.cninfo.com.cn/new/index)
The Banker's Almanac (https://accuity.com/)
Dow Jones (https://www.dowjones.com/)
Copyright statement: The article comes from the official account (python risk control model), without permission, no plagiarism. Following the CC 4.0 BY-SA copyright agreement, please attach the original source link and this statement for reprinting.