## Appendix A. MYbank ## Source details **Canonical URL:** [Appendix A. MYbank](https://www.imf.org/-/media/files/publications/wp/2020/english/wpiea2020193-print-pdf.pdf) ## Other formats - [Markdown version](/-/media/files/publications/wp/2020/english/wpiea2020193-print-pdf.pdf.md) - [Structured JSON version](/-/media/files/publications/wp/2020/english/wpiea2020193-print-pdf.pdf.json) --- ### I. INTRODUCTION — Context and motivation - Policy challenge: - Promoting financial inclusion for vulnerable households and SMEs; access to credit for SMEs remains quite limited. - Main barriers: high cost, physical distance, lack of proper documentation. - BigTech/fintech role: - Fintech and BigTech lenders (examples cited in source) have extended loans to millions of small borrowers. - In China, three leading virtual banks—MYbank (affiliated to Alibaba), WeBank (affiliated to Tencent), and XW Bank (affiliated to Xiaomi)—provide loans to millions of small firms annually, more than 80 percent of which have no credit history. - BigTech loans: smaller, shorter duration, mainly used for operational purposes; complementary to traditional banking by reaching underserved customers. - Key enabling factors for BigTech scale: - Low marginal customer acquisition costs via digital platforms. - Big data (traditional and proprietary) enables risk assessment absent financial history/collateral. - Digital technologies (cloud computing, artificial intelligence) enable rapid processing, dynamic risk assessment, and large-volume loan restructuring. - Contact-free feature increases robustness during shocks (e.g., COVID-19). ### Central research questions - Is the fintech approach (big data + machine learning) more reliable in predicting loan defaults than the traditional method (financial data + scorecard models)? - Can traditional models fully utilize the information value of big data? - Can AI algorithms extract more information from traditional data? - For borrowers without bank credit history, is BigTech proprietary information sufficient to substitute credit registry information? - Can the fintech approach withstand business cycles and economic shocks? - What is the contribution of data versus methodology in credit risk assessment? ### Data, setting, and empirical strategy - Dataset: - Unique sample of 1.8 million MYbank SME loans (March–August 2017); repayment information through August 2018. - All loans have maturity of one year. - Borrowers: online vendors on Alibaba’s e-commerce platforms and Ant’s Alipay users. - 99.8 percent of borrowers are micro firms (annual sales less than RMB 1 million). - Geographic distribution: firms widely distributed across China; about half located in Tier-3 and Tier-4 cities. - Subset with bank credit history: 145,109 loan transactions (about 8 percent of the entire sample). - Variables: - Total variables: 76 firm-characteristic variables. - Traditional variables: 32 (asset-related, credit history with MYbank, vendor-specific, local economy). - Proprietary variables: 44 (transaction volume, network effect score, digital footprints, shop/customer ratings). - Network effect score range: 0 (lowest) to 100 (highest). - Empirical objectives: 1. Horse race: fintech vs traditional for credit risk assessment. 2. Role of big data vs bank credit history. 3. Robustness to exogenous shocks. 4. Inclusiveness across borrower sizes and city tiers. - Modeling approach: - Scorecard model as industry benchmark. - Machine learning baseline: random forest; other ML (e.g., Gradient Boosting Decision Tree) for robustness. - Sample split: training (March–May 2017) size 771,596; testing (July–August 2017) size 1,053,748. - Data confidential and not publicly available. ### Main empirical findings (high level) - Fintech approach (big data + machine learning) significantly improves accuracy of loan default prediction relative to the traditional approach. - Improvement components: - Information advantage: BigTech proprietary information reveals capacity and willingness to repay. - Modeling advantage: AI algorithms capture complex interactions. - Traditional scorecard models applied to big data yield reasonably good assessment, but some characteristics are not fully captured. - For borrowers without bank credit history, BigTech proprietary information can effectively substitute credit registry information. - Fintech model is robust to exogenous policy shocks faced by BigTech companies. - Advantages more pronounced for smaller firms and firms in smaller cities, implying complementarity with traditional banks in broadening credit access. ### Stylized facts about MYbank (selected) - Business model and reach: - Mainly online loans to micro firms; loans are small, short-duration, used for working capital. - Since mid-2015, provided loans to more than 20 million SMEs; about 80 percent micro firms with annual sales < RMB 1 million. - In the 1.8 million loan sample, 99.8 percent are micro firms; only 7.5 percent have borrowed from traditional banks. - Contact-free operations: - 3-1-0 model: user registration and application within 3 minutes; money transferred to an Alipay account within 1 second; 0 human intervention. - Loan size and duration: - MYbank average loan size reported as RMB 2,600 ($367) in text; Figure 2 caption notes RMB 2,600 equivalent to $380. - Traditional banks’ average SME loan size in sample: RMB 1 million ($150,000), more than 30 times larger. - Duration: over 60 percent of MYbank loans are less than one month; nearly 57 percent of traditional bank loans for SMEs are longer than one year (inferred). - Sensitivity to cash flows and assets: - A 1 percent increase in cash flow associated with a 0.45 percent increase in loan amount for MYbank, compared with 0.40 percent for traditional bank loans. - About 30 percent of loan quota variation explained by cash flows for MYbank vs around 10 percent for traditional banks. - Traditional banks rely more on collateral (housing ownership increases loan quota from traditional banks, with no significant impact on MYbank loans). - NPL performance: - Chinese banks’ SME loan NPLs: 3.2 percent in 2018; SME loans < RMB 5 million: 5.5 percent; overall NPL ratio: 1.9 percent. - MYbank historical average NPL ratio around 1 percent; increased since COVID-19 onset but remained below 2 percent. - MYbank uses ecosystem measures (e.g., downgrading business ratings) to manage moral hazard. - Pricing and inclusion: - MYbank annualized lending rate: between 10 and 17 percent. - Average bank lending rate: 4.35 percent (comparison). - Higher MYbank rates attributable to higher funding cost, big data infrastructure fixed costs, and serving smaller/higher-risk borrowers. - Financial inclusion primarily reflected in access rather than price. ### Descriptive statistics (selected Table 1 entries; values preserved) - House property (0/1): N = 1,822,423; Mean = 0.621; St. dev. = 0.485; Pctl (25) = 0; Median = 1; Pctl (75) = 1. Type: Traditional. - Car property (0/1): N = 1,822,423; Mean = 0.662; St. dev. = 0.473; Pctl (25) = 0; Median = 1; Pctl (75) = 1. Type: Traditional. - Number of credit cards: N = 1,825,342; Mean = 2.968; St. dev. = 3.838; Pctl (25) = 0; Median = 2; Pctl (75) = 4. Type: Traditional. - Owner’s age (years): N = 1,825,342; Mean = 29; St. dev. = 6; Pctl (25) = 25; Median = 28; Pctl (75) = 31. Type: Traditional. - Gender (male = 1; female = 0): N = 1,825,342; Mean = 0.65; St. dev. = 0.48; Pctl (25) = 0; Median = 1; Pctl (75) = 1. Type: Traditional. - Firm’s age (years): N = 1,825,342; Mean = 4.48; St. dev. = 2.24; Pctl (25) = 2.70; Median = 4.07; Pctl (75) = 5.99. Type: Traditional. - City tiers: N = 1,817,380; Mean = 2.468; St. dev. = 1.027; Pctl (25) = 2; Median = 2; Pctl (75) = 3. Type: Traditional. - Total inflow of funds in Alipay (yuan): N = 1,825,342; Mean = 516,708.00; St. dev. = 1,149,718.00; Pctl (25) = 138,534.10; Median = 264,210.00; Pctl (75) = 534,783.60. Type: Proprietary. - Shop rating: N = 1,825,342; Mean = 4.19; St. dev. = 2.96; Pctl (25) = 2; Median = 4; Pctl (75) = 6. Type: Proprietary. - Transaction volume (last six months, yuan): N = 1,825,342; Mean = 9,196.80; St. dev. = 58,973.99; Pctl (25) = 1,964.35; Median = 4,754.43; Pctl (75) = 9,954.34. Type: Proprietary. - Log-ins (number, last six months): N = 1,825,342; Mean = 68.06; St. dev. = 42.54; Pctl (25) = 33; Median = 59; Pctl (75) = 99. Type: Proprietary. - Network effect score: N = 1,825,342; Mean = 64.48; St. dev. = 27.66; Pctl (25) = 48.85; Median = 59.95; Pctl (75) = 74.09. Type: Proprietary. - Daily average Yu’eBao balance (last three months, yuan): N = 1,825,342; Mean = 974.84; St. dev. = 6,656.89; Pctl (25) = 0; Median = 106; Pctl (75) = 532.60. Type: Proprietary. - Daily payment activity (index, last year): N = 1,822,407; Mean = 1,511.85; St. dev. = 3,682.70; Pctl (25) = 647; Median = 1,007.00; Pctl (75) = 1,587.00. Type: Proprietary. - Total amount of e-commerce purchases (last six months, yuan): N = 1,821,221; Mean = 33,514.44; St. dev. = 128,465.10; Pctl (25) = 9,132.47; Median = 18,080.32; Pctl (75) = 34,788.37. Type: Proprietary. - Duration in one location (index): N = 1,815,137; Mean = 1,815.63; St. dev. = 762.28; Pctl (25) = 1,261.00; Median = 1,714.00; Pctl (75) = 2,318.00. Type: Proprietary. ### Models, sample split, and definitions (Appendix B summary) - Four models evaluated: - Model I: traditional approach (scorecard + traditional information). - Model II: scorecard + all information (traditional + proprietary). - Model III: machine learning (random forest) + traditional information. - Model IV: fintech approach (random forest + all information). - Pairwise comparisons: - Fintech vs traditional: Model IV vs Model I. - Marginal contribution of big data: Model IV vs Model III and Model II vs Model I. - Marginal contribution of machine learning: Model IV vs Model II and Model III vs Model I. - Sample split: training (March–May 2017) 771,596; testing (July–August 2017) 1,053,748. - Random forest details: - Number of trees: 500 (M=500). - Software: RandomForestClassifier in Python. - Scorecard details: - WoE transformation, Information Value (IV) for feature selection, logistic regression on WoE-transformed variables; 32 variables selected for logit model. - Scorecard training: scorecard package in R. ### Baseline results — ROC performance and AUCs (preserve reported values) - ROC ranking (best to worst): Model IV (random forest + all information) — best; Model III (random forest + traditional information) — second; Model II (scorecard + all information) — third; Model I (scorecard + traditional information) — worst. - AUC background: AUC ranges from 50 percent (random) to 100 percent (perfect); model reasonably reliable if AUC > 60 percent; performs strongly if AUC > 70 percent. - Table 3 (AUCs by information set and model; values preserved as in source): - Scorecard: 0.70 0.58 0.72 0.72 0.69 0.76 - Random forest: 0.67 0.54 0.76 0.80 0.76 0.84 - Note: columns correspond to information groups (a) asset/financial; (b) credit history (MYbank); (c) vendor-specific; (d) local economy; (e) Ant proprietary information. - Key AUC findings (Table 4 95% CIs for Models I–IV): - Model I (scorecard + traditional information): AUC 0.7246; 95% conf. interval 0.7222 0.7269. - Model II (scorecard + all information): AUC 0.7636; 95% conf. interval 0.7615 0.7658. - Model III (random forest + traditional information): AUC 0.8022; 95% conf. interval 0.8001 0.8043. - Model IV (random forest + all information): AUC 0.8414; 95% conf. interval 0.8396 0.8433. - Differences among models are statistically significant. - Incremental contributions (reported percent changes): - Adding Ant’s proprietary information to traditional approach increases AUC by 5.6 percent. - Applying machine learning techniques adds an additional 11.1 percent to AUC. - In this sample, model advantage appears greater than information advantage. ### Reduction in NPL ratio from traditional to fintech approach (Table 5; values preserved) - Method: compare NPL ratios for accepted loans under expected default rate thresholds. - Reduction in NPL ratio (percentage points) by threshold (%): - Threshold 10% → Reduction in NPL ratio 1.05 - Threshold 15% → Reduction in NPL ratio 0.76 - Threshold 20% → Reduction in NPL ratio 0.48 - Threshold 25% → Reduction in NPL ratio 0.27 - Threshold 30% → Reduction in NPL ratio 0.15 - Note: Improvement diminishes as threshold relaxes from 10 to 30 percent. ### Variable importance and model differences - Scorecard model top drivers: - Primarily driven by five variables: four on credit history and one on daily payment activity. - Top variables include: Credit history (last 1 year); Credit history (last 6 months); Credit history (last 3 months); Credit history (last 1 month); Daily payment activity (last 1 year); Daily average Yu'e Bao balance (last 3 months); Total amount of e-commerce purchases (last 6 months); Security fund in Taobao; Daily average Alipay wallet balance (last 3 months); Number of credit cards; Payment activity (last 6 months); E-commerce purchase activity (last 6 months); Time in one location; Gender; Stability of contact information (last 1 year); Transaction volume (last 6 months); Firm's age; Link to credit card (0/1); Shop rating; Owner's age. - Random forest model top drivers: - Importance more evenly distributed across a wider range of variables. - Transaction and payment data play a larger role. - Ant’s proprietary information (customers’ ratings, network effect scores) important in ML but not in scorecard top 20, suggesting nonlinear interactions with other variables. - Appendix B elaborates information value and Gini impurity measures. ### Subsample: borrowers with bank credit history (about 7.5% of sample) - ROC and AUC outcomes (Table 6 and Table 7; values preserved): - Scorecard model AUCs: - Bank credit history information: 0.74 - Ant information: 0.72 - Bank credit history information + Ant information: 0.78 - Random forest AUCs: - Bank credit history information: 0.84 - Ant information: 0.83 - Bank credit history information + Ant information: 0.87 - 95% confidence intervals (selected from Table 7): - Random forest + bank credit history + Ant information: AUC 0.8681; 95% conf. interval 0.8636 0.8725. - Random forest + bank credit history information: AUC 0.8372; 95% conf. interval 0.8318 0.8427. - Scorecard + bank credit history + Ant information: AUC 0.7767; 95% conf. interval 0.7713 0.7820. - Scorecard + bank credit history: AUC 0.7397; 95% conf. interval 0.7337 0.7456. - Key implications: - Bank credit history has AUC 0.74 in traditional approach; applying machine learning raises AUC to 0.84. - Ant information alone yields similar predictive power as bank credit history. - Best performance achieved by combining complete information set with machine learning (AUC 0.87). ### Impact of a policy shock (People’s Bank of China regulatory action) - Policy event: - November 17, 2017: People’s Bank of China issued draft guidelines to tighten asset management activity regulation. - Short-run effect: tightening financial conditions; credit growth dropped by 4 percentage points the following year. - Banking system NPL ratio rose from 1.74 percent before announcement to 1.87 percent one year later. - MYbank’s NPL ratio edged up (specific value not provided in supplied extract); other text references NPL ratio in 2019 as 1.3 percent and historical around 1 percent. - Discriminatory power before vs after shock (Table 8; AUC differences preserved): - Model IV – Model I (comprehensive evaluation): 0.11 (before shock); 0.14 (after shock). - Model IV – Model II (model advantage): 0.07 (before shock); 0.09 (after shock). - Model II – Model I (information advantage): 0.04 (before shock); 0.05 (after shock). - Interpretation: - Fintech outperformance increases from 0.11 to 0.14 in AUC after the shock. - Both information and model advantages more pronounced post-shock, suggesting fintech robustness to the 2017 deleveraging policy shock. ### Inclusiveness of fintech lending (SME heterogeneity) - Method: Cumulative distribution of increase in estimated default probability (fintech minus scorecard); negative differences indicate “winners” under fintech. - By firm size (annual sales turnover, RMB thousands) — AUCs (Table 9; values preserved): - Turnover groups: < 20 | 20–100 | 100–500 | 500–1,000 | > 1,000 - Traditional approach AUCs: 0.72 | 0.72 | 0.71 | 0.8 | 0.76 - Fintech approach AUCs: 0.84 | 0.84 | 0.82 | 0.89 | 0.85 - Fintech – traditional: 0.12 | 0.12 | 0.11 | 0.09 | 0.09 - Interpretation: fintech advantage larger for smaller firms. - By city tier — AUCs (Table 9; values preserved): - City tiers: Tier-4 | Tier-3 | Tier-2 | Tier-1 - Traditional approach: 0.71 | 0.73 | 0.73 | 0.73 - Fintech approach: 0.84 | 0.85 | 0.84 | 0.84 - Fintech – traditional: 0.13 | 0.12 | 0.11 | 0.11 - Interpretation: fintech yields consistent improvements across city tiers, slightly larger gains in lower-tier cities. - Descriptive subgroup outcomes: - Smallest borrowers: ~57 percent have lower estimated default probabilities under fintech. - Largest borrowers: ~27 percent have lower estimated default probabilities under fintech. - Tier-4 city borrowers: about 62 percent have lower predicted default probabilities using fintech. - Tier-1 and Tier-2 city borrowers: about 52 percent have lower predicted default probabilities using fintech. ### Main empirical conclusions (summary bullets) - Fintech approach (big data + machine learning) more accurately predicts defaults than traditional scorecard models. - Information advantage from real-time financial and behavior data; model advantage from ML capturing nonlinear/interactive effects. - Subsample with bank history: scorecards perform reasonably; ML + full data increases AUC from 0.74 to 0.87. - ML on all information without bank credit history yields AUC 0.83, indicating proprietary information can substitute credit registry data for many SMEs. - Fintech advantage is more pronounced after the 2017 deleveraging policy shock; robustness under other shock types remains to be tested. - Fintech yields largest improvements for the smallest firms and SMEs in lower-tier cities (Tier-3 and Tier-4), demonstrating inclusive features. ### Policy implications and recommendations (preserve phrasing) - Promote fintech lending while strengthening regulation: - Encourage fintech to promote financial inclusion, complemented by strict licensing and regulatory checkups. - Fintech requires access to big data and machine learning capacity; cautionary lessons from failing peer-to-peer lending industry. - Regulators should be vigilant for new risks from fintech business models in downturns; balance innovation with financial stability. - Use regulatory sandboxes to incubate new business models and practices. - Encourage collaboration between fintech lenders and traditional banks: - Fintech: superior information and analytics but limited funding capacity. - Banks: substantial financial resources and soft information but limited big-data analytics. - Example practice in China: as of end-2019, RMB 2 trillion in business loans (equivalent to 2 percent of gross domestic product) jointly issued by banks and fintech firms, where fintech provided initial screening and banks provided most funding. - Joint lending can promote financial inclusion; modalities to be designed case-by-case. - Enhance data protection and access policy: - Set clear data policy standards to maximize big data benefits while protecting privacy; emphasize data authenticity, legality, and security. - Consider fostering portability of individual data across platforms to increase competition. - Balance data access (including proprietary fintech data) with privacy protection; lack of data sharing may exacerbate concentration and network effects. - Sequence public investments: - Development of fintech requires digital infrastructure, an established online ecosystem, and big data analytical skills. - Suggested sequencing: invest in digital infrastructure, support e-commerce and digitalization, then expand fintech lending. ### Appendix A — MYbank key facts (selected figures preserved) - Founded: 2015 (headquartered in Hangzhou, Zhejiang province). - Ownership: founded by Alibaba’s affiliate firm, Ant Group, through a 30 percent stake in a joint venture comprising a group of private firms. - Borrower base and performance: - About 20 million SME borrowers; about 80 percent never borrowed from banks in the past. - NPL ratio at about 1 percent (text also reports NPL ratio was 1.3 percent in 2019). - 2019 summary figures: - Total assets: RMB 139.6 billion ($20 billion). - Average loan size: RMB 31,000 ($4,500). - Accumulated SMEs served: 20.9 million. - Capital adequacy ratio: 16.4 percent. - NPL ratio: 1.3 percent. - Operational model: data/cloud-based model with no physical branches; operates 24 hours a day, seven days a week; uses big data and machine learning for credit profiling. *Source: Appendix A, wpiea2020193-print-pdf (calculations and figures based on data from MYbank).* ### Appendix A. MYbank ..................................................................................................... ### Appendix A. MYbank ### I. INTRODUCTION — Context and motivation - Promoting financial inclusion for vulnerable households and smaller firms is a persistent policy challenge; access to credit for small and medium-size enterprises (SMEs) remains quite limited, especially in developing countries. - Main barriers to SME credit: high cost, physical distance, lack of proper documentation. - Traditional banks rely on personal guarantees or relationship lending and face high fixed costs to reach numerous, scattered SMEs. - Fintech and BigTech lenders have recently extended loans to millions of small borrowers (examples cited: Alibaba, Tencent, Mercado Credito, Paytm, Amazon Lending). - In China, three leading virtual banks—MYbank (affiliated to Alibaba), WeBank (affiliated to Tencent), and XW Bank (affiliated to Xiaomi)—provide loans to millions of small firms annually, more than 80 percent of which have no credit history. - BigTech loans are characterized as much smaller, shorter in duration, and mainly used for operational purposes; fintech lending so far plays a complementary role to traditional banking by reaching underserved customers. ### Why BigTech can reach SMEs at scale — key enabling factors - Digital platforms connect to millions of customers at very low marginal costs, enabling large-scale customer acquisition and monitoring of borrower activities occurring on BigTech platforms. - Big data (traditional and proprietary information) can help assess credit risk in absence of financial history and collateral: - Traditional data: basic individual/firm information (gender, age, location, profession, business) — mostly offline and costly to collect without BigTech support. - Proprietary information: digital footprints (messages exchanged, payments made, websites browsed on BigTech platforms) usable to assess financial condition and behavioral characteristics. - Digital technologies (cloud computing, artificial intelligence) allow rapid processing of massive loan applications and dynamic risk assessment using real-time data; big data and machine learning enable large-volume loan restructuring and lower operating costs. - The “contact-free feature” from customer acquisition to underwriting and restructuring increases robustness during shocks such as the COVID-19 pandemic. ### Central research questions posed - Is the fintech approach (big data + machine learning) more reliable in predicting loan defaults than the traditional method (financial data + scorecard models)? - Can traditional models fully utilize the information value of big data? - Can AI algorithms extract more information from traditional data? - For borrowers without bank credit history, is BigTech proprietary information sufficient to substitute credit registry information in risk assessment? - Can the fintech approach withstand business cycles and various economic shocks? - What is the contribution of data versus methodology in credit risk assessment? ### Data, setting, and empirical strategy - Data: unique data set of 1.8 million MYbank SME loans. - MYbank: established in May 2015; BigTech lending business inherited from Ant Group (affiliated company of Alibaba Group). Alibaba and Ant began the BigTech lending business model in 2010. - Borrowers: online vendors on Alibaba’s e-commerce platforms and Ant’s Alipay users, mostly small retailers and self-employed. - Dataset features: large volume of proprietary information (business transactions, payments, customer ratings, consumption patterns, ecosystem importance) plus traditional data and information on borrowers’ credit histories with traditional banks (if any). ### Main empirical findings - The fintech approach (big data + machine learning) significantly improves the accuracy of loan default prediction relative to the traditional approach. - Improvement reflects a combination of: - Information advantages: BigTech proprietary information reveals financial condition (capacity to repay) and behavioral characteristics (willingness to repay). - Modeling advantages: AI algorithms capture more complex interactions among variables. - Applying traditional scorecard models to big data can yield reasonably good risk assessment, although some borrower characteristics are not fully captured. - For borrowers without a bank credit history, BigTech proprietary information can effectively substitute credit registry information in risk assessment, enhancing financial inclusion. - The fintech risk assessment model is robust to exogenous policy shocks faced by BigTech companies. - Advantages of the fintech model are more pronounced for firms that are smaller and in smaller cities, suggesting a natural complementarity with traditional banks in broadening credit access. ### II. LITERATURE REVIEW — framing of credit-risk methods - Fundamental challenge: information asymmetry leading to adverse selection and moral hazard. - Commercial banks typically use three methods to analyze and mitigate credit risk: - Financial history: detailed analyses of financial data (balance sheets, income statements, cash flow) — most useful for large corporations. - Collateral or guarantees. - Soft information / relationship lending. *Source: Appendix A. MYbank, from the provided IMF working paper content.* ### Appendix A provides more information on MYbank. ### wpiea2020193-print-pdf - Appendix A provides more information on MYbank ### Background: SME lending, collateral, and fintech approaches - Collateral can help resolve adverse selection and moral hazard (Besanko and Thakor 1987a, 1987b; Stiglitz and Weiss 1981; Cerqueiro, Ongena, and Roszbach 2016; Berger, Frame, and Ioannidou 2016), but many SMEs lack collateral assets, constraining lending to micro firms and self-employed businesses. - Relationship banking uses soft information (owner, local community) to predict default (Berger and Udell 2002, 2006; Agarwal and Hauswald 2010) but requires significant investment in physical capital and human resources and broad networks. - Big data and machine learning advances: - Digital footprints can complement credit bureau information and reduce default rates (Berg et al. 2019); simple indicators such as mobile OS can reveal creditworthiness. - Mobile and social footprints outperform traditional credit scores for loan approvals and defaults in a large fintech data set from India (Agarwal et al. 2019). - BigTech lenders show information advantage over traditional credit bureaus in case studies (Frost et al. 2019); alternative information can reclassify some subprime borrowers into better loan grades (Jagtiani and Lemieux 2019). - Machine learning methods (decision trees, random forests) often outperform logistic regression in out-of-sample and out-of-time forecasts (Butaru et al. 2016; Khandani, Kim, and Lo 2010; Gambacorta et al. 2019; Fuster et al. 2020). - BigTech lenders can process mortgage applications 20 percent faster than other lenders without higher default cost (Frost et al. 2019). ### Stylized facts about MYbank - Business model and reach: - MYbank mainly offers online loans to micro firms with no access to credit; loans are small, short duration, and used primarily for operational working capital. - Since establishment in mid-2015, MYbank has provided loans to more than 20 million SMEs, about 80 percent of which are micro firms with annual sales of less than RMB 1 million and no access to bank lending. - In the sample of 1.8 million loans, 99.8 percent of the borrowers are micro firms; only 7.5 percent have borrowed from traditional banks. - Contact-free and digitalized operations: - Operates on the 3-1-0 model: user registration and application within 3 minutes, money transferred to an Alipay account within 1 second, and 0 human intervention. - Contactless feature reduces operating costs and ensures business continuity during the pandemic. - Loan size and duration: - MYbank’s average loan size is RMB 2,600 ($367) (Figure 2). - Note in Figure 2 caption: MYbank’s average loan size (RMB 2,600, equivalent to $380). - Traditional banks’ average SME loan size in the sample is RMB 1 million ($150,000), more than 30 times larger than MYbank’s average. - Duration: over 60 percent of MYbank’s loans are less than one month; nearly 57 percent of traditional bank loans for SMEs are longer than one year (inferred from MYbank borrowers who also borrowed from traditional banks). - Relationship to cash flows and assets: - MYbank loans are more sensitive to firms’ cash flows: a 1 percent increase in cash flow is associated with a 0.45 percent increase in the loan amount for MYbank, compared with 0.40 percent for traditional bank loans. - About 30 percent of the loan quota variation can be explained by cash flows for MYbank, compared with around 10 percent for traditional bank loans. - Traditional banks rely more on collateral assets (e.g., housing ownership significantly increases loan quota from traditional banks, with no significant impact on MYbank loans). - Nonperforming loan (NPL) performance and moral hazard management: - Historical average NPL ratio context: many SME loan portfolios typically have higher default rates. - In 2018, the share of NPLs of Chinese banks’ SME loans was 3.2 percent, and 5.5 percent for SME loans that were less than RMB 5 million, compared with the overall NPL ratio of 1.9 percent. - MYbank has kept its average NPL ratio at around 1 percent historically; since the onset of the COVID-19 pandemic, MYbank’s average NPL ratio has increased but remained contained at below 2 percent. - MYbank uses unique means (e.g., downgrading business ratings in the e-commerce ecosystem) to address moral hazard, in contrast to traditional banks’ reliance on bankruptcy procedures or disposing NPLs to asset management companies. - Pricing and financial inclusion: - MYbank’s annualized lending rate is between 10 and 17 percent, similar to prevailing private lending rates in China (e.g., Wenzhou composite lending rate), but higher than the average bank lending rate of 4.35 percent. - Reasons for higher rates include higher funding cost (disadvantage in attracting retail deposits and ineligibility for preferential policies), high fixed costs to establish big data infrastructure, and serving smaller, higher-risk borrowers excluded by banks. - Financial inclusion is reflected mainly in credit access rather than price. ### Data description (sample and variables) - Sample and coverage: - Unique data set of 1.8 million SME loans granted by MYbank between March and August in 2017. - All loans have maturity of one year. - 99.8 percent of borrowers are micro firms (annual sales less than RMB 1 million). - Borrowers are online vendors on Alibaba’s e-commerce platforms and users of Ant’s Alipay. - Data include loan repayment information through August 2018. - Geographic distribution: firms widely distributed across China; about half located in Tier-3 and Tier-4 cities. - Variable taxonomy: - Total variables used: 76 variables on firm characteristics. - Traditional data variables: 32, classified into: - (a) asset-related information (e.g., housing property) - (b) credit history with MYbank - (c) vendor-specific information (gender, age, business) - (d) local (provincial and municipal) economy information - Proprietary information variables: 44, including transaction volume, network effect score, and digital footprints (social network activities, locations, online consumption patterns, financial transaction styles, shop ratings, customer ratings/reviews). - Network effect score: measures relative importance in the BigTech network, with 0 representing the lowest impact and 100 the highest impact. - Descriptive statistics (selected entries from Table 1): - House property (0/1): N = 1,822,423; Mean = 0.621; St. dev. = 0.485; Pctl (25) = 0; Median = 1; Pctl (75) = 1. Type: Traditional. - Car property (0/1): N = 1,822,423; Mean = 0.662; St. dev. = 0.473; Pctl (25) = 0; Median = 1; Pctl (75) = 1. Type: Traditional. - Number of credit cards: N = 1,825,342; Mean = 2.968; St. dev. = 3.838; Pctl (25) = 0; Median = 2; Pctl (75) = 4. Type: Traditional. - Owner’s age (years): N = 1,825,342; Mean = 29; St. dev. = 6; Pctl (25) = 25; Median = 28; Pctl (75) = 31. Type: Traditional. - Gender (male = 1; female = 0): N = 1,825,342; Mean = 0.65; St. dev. = 0.48; Pctl (25) = 0; Median = 1; Pctl (75) = 1. Type: Traditional. - Firm’s age (years): N = 1,825,342; Mean = 4.48; St. dev. = 2.24; Pctl (25) = 2.70; Median = 4.07; Pctl (75) = 5.99. Type: Traditional. - City tiers: N = 1,817,380; Mean = 2.468; St. dev. = 1.027; Pctl (25) = 2; Median = 2; Pctl (75) = 3. Type: Traditional. - Total inflow of funds in Alipay (yuan): N = 1,825,342; Mean = 516,708.00; St. dev. = 1,149,718.00; Pctl (25) = 138,534.10; Median = 264,210.00; Pctl (75) = 534,783.60. Type: Proprietary. - Shop rating: N = 1,825,342; Mean = 4.19; St. dev. = 2.96; Pctl (25) = 2; Median = 4; Pctl (75) = 6. Type: Proprietary. - Transaction volume (last six months, yuan): N = 1,825,342; Mean = 9,196.80; St. dev. = 58,973.99; Pctl (25) = 1,964.35; Median = 4,754.43; Pctl (75) = 9,954.34. Type: Proprietary. - Log-ins (number, last six months): N = 1,825,342; Mean = 68.06; St. dev. = 42.54; Pctl (25) = 33; Median = 59; Pctl (75) = 99. Type: Proprietary. - Network effect score: N = 1,825,342; Mean = 64.48; St. dev. = 27.66; Pctl (25) = 48.85; Median = 59.95; Pctl (75) = 74.09. Type: Proprietary. - Daily average Yu’eBao balance (last three months, yuan): N = 1,825,342; Mean = 974.84; St. dev. = 6,656.89; Pctl (25) = 0; Median = 106; Pctl (75) = 532.60. Type: Proprietary. - Daily payment activity (index, last year): N = 1,822,407; Mean = 1,511.85; St. dev. = 3,682.70; Pctl (25) = 647; Median = 1,007.00; Pctl (75) = 1,587.00. Type: Proprietary. - Total amount of e-commerce purchases (last six months, yuan): N = 1,821,221; Mean = 33,514.44; St. dev. = 128,465.10; Pctl (25) = 9,132.47; Median = 18,080.32; Pctl (75) = 34,788.37. Type: Proprietary. - Duration in one location (index): N = 1,815,137; Mean = 1,815.63; St. dev. = 762.28; Pctl (25) = 1,261.00; Median = 1,714.00; Pctl (75) = 2,318.00. Type: Proprietary. - Bank credit history availability: - A subset of borrowers have bank credit history: 145,109 loan transactions (about 8 percent of the entire sample), allowing comparison with traditional bank credit information. ### Empirical strategy and models - Four empirical objectives: 1. Horse race between the fintech approach and the traditional approach for credit risk assessment to quantify information and model advantages. 2. Assess the role of big data relative to bank credit history information; test whether BigTech proprietary information alone is sufficient for reliable credit risk assessment absent bank credit history. 3. Test robustness of fintech outperformance to an exogenous shock (e.g., adoption of a new regulatory policy). 4. Explore inclusiveness of the fintech approach by comparing performance across borrower sizes and city tiers. - Horse races and modeling approach: - Compare contributions of traditional data versus MYbank proprietary information using the same models. - Compare predictive power of traditional scorecard models and machine learning models using identical information sets. - Scorecard model used as industry benchmark; machine learning baseline uses the random forest model à la Butaru et al. (2016). - Other machine learning models (such as the Gradient Boosting Decision Tree) are used for robustness checks; conclusions remain unchanged. - Data confidentiality: - Data were provided by MYbank on a confidential basis and are not publicly available. *Source: Appendix A, wpiea2020193-print-pdf (calculations and figures based on data from MYbank).* ### Appendix B provides a more detailed elaboration of the formulas and algorithms. ### wpiea2020193-print-pdf - Appendix B provides a more detailed elaboration of the formulas and algorithms. ### Models, sample split, and definitions - Four models evaluated: - Model I: traditional approach (scorecard models + traditional information) - Model II: scorecard models + all information (traditional + proprietary data) - Model III: machine learning models (random forest) + traditional information - Model IV: fintech approach (random forest + all information) - Pairwise comparisons: - Fintech versus traditional approach: Model IV versus Model I - Marginal contribution of big data: Model IV versus Model III and Model II versus Model I - Marginal contribution of machine learning model: Model IV versus Model II and Model III versus Model I - Out-of-sample tests: - Sample split into two subperiods: March to May 2017 (training set) and July to August 2017 (testing set). - Training set size: 771,596 loan transactions. - Testing set size: 1,053,748 loan transactions. - Note: Results are robust to alternative subsample periods. ### Baseline results — ROC performance and AUCs - ROC curve ranking (best to worst): - Model IV (random forest + all information) — best (purple line) - Model III (random forest + traditional information) — second (blue line) - Model II (scorecard + all information) — third (green line) - Model I (scorecard + traditional information) — worst (red line) - Interpretation: - At a given level of specificity (false positive), higher sensitivity (true positive) indicates superior performance. - Conclusion from controlled sample and variable lists: - The fintech approach is more reliable than the traditional approach in predicting defaults. - Replacing traditional information with all information (adding proprietary information) improves performance (information advantage). - Replacing scorecard with machine learning improves performance (model advantage). - In this case, the model advantage appears greater than the information advantage (sample-dependent). - AUC background: - AUC ranges from 50 percent (purely random) to 100 percent (perfect). - Model reasonably reliable if AUC > 60 percent; performs strongly if AUC > 70 percent. - Table 3: AUCs by information set and model (values preserved as in source) - Scorecard: 0.70 0.58 0.72 0.72 0.69 0.76 - Random forest: 0.67 0.54 0.76 0.80 0.76 0.84 - Note: (a) asset and financial information; (b) credit history (only from MYbank); (c) vendor-specific information; (d) local economy information; (e) Ant proprietary information. - Key AUC findings: - With relatively few variables, machine learning models show no advantage (smaller AUCs than scorecard models in first two columns of Table 3) — implies machine learning more powerful for large data sets. - Ant’s proprietary information alone ((e) in Table 3) has AUC of 0.76; limited advantage over Model I (traditional variables + scorecard model, AUC of 0.72). - Adding more variables improves credit risk assessment. - Combination of large data set and machine learning greatly improves assessment: - Adding Ant’s proprietary information to the traditional approach increases AUC by 5.6 percent. - Applying machine learning techniques adds an additional 11.1 percent to the AUC. - Table 4: 95% confidence intervals of AUCs for Models I–IV - Model I (scorecard + traditional information): AUC 0.7246; 95% conf. interval 0.7222 0.7269 - Model II (scorecard + all information): AUC 0.7636; 95% conf. interval 0.7615 0.7658 - Model III (random forest + traditional information): AUC 0.8022; 95% conf. interval 0.8001 0.8043 - Model IV (random forest + all information): AUC 0.8414; 95% conf. interval 0.8396 0.8433 - Differences in AUCs among the four models are significant. ### Reduction in NPL ratio from traditional to fintech approach - Method: For alternative thresholds of expected default rate, loans with expected default rates below the threshold are accepted; compare NPL ratios of loans accepted by fintech and traditional approaches; report reduction in NPL ratio (percentage points). - Table 5: Reduction in NPL ratio (percentage points) by threshold (%) - Threshold 10% → Reduction in NPL ratio 1.05 - Threshold 15% → Reduction in NPL ratio 0.76 - Threshold 20% → Reduction in NPL ratio 0.48 - Threshold 25% → Reduction in NPL ratio 0.27 - Threshold 30% → Reduction in NPL ratio 0.15 - Note: Improvement becomes smaller as threshold relaxes from 10 to 30 percent. NPL = nonperforming loans. ### Variable importance and model differences - Top 20 contributing variables reported separately for scorecard and random forest models; importance metrics differ across models (information value vs Gini impurity) and are not numerically comparable. - Scorecard model (top drivers): - Primarily driven by five variables: four on credit history and one on daily payment activity. - Among top 20, most variables relate to transactions and payments. - Example top variables listed in Figure 8 (panel a) with information values on x-axis: Credit history (last 1 year); Credit history (last 6 months); Credit history (last 3 months); Credit history (last 1 month); Daily payment activity (last 1 year); Daily average Yu'e Bao balance (last 3 months); Total amount of e-commerce purchases (last 6 months); Security fund in Taobao; Daily average Alipay wallet balance (last 3 months); Number of credit cards; Payment activity (last 6 months); E-commerce purchase activity (last 6 months); Time in one location; Gender; Stability of contact information (last 1 year); Transaction volume (last 6 months); Firm's age; Link to credit card (0/1); Shop rating; Owner's age. - Random forest model (top drivers): - Information values more evenly distributed across a wider range of variables. - Transaction and payment data play a more important role than in the scorecard model. - Ant’s proprietary information (e.g., customers’ ratings of vendors, network effect scores) is important in machine learning models but does not appear in the scorecard model’s top 20. - Interpretation: Proprietary variables likely affect predicted defaults through nonlinear interactions with other variables (e.g., high network effect score useful only when combined with healthy cash flows). - Appendix B provides detailed elaboration of information value and Gini impurity. ### Subsample: borrowers with bank credit history (7.5% of total) - ROC results: - Machine learning + all information remains most efficient. - Scorecard + bank credit history information is least efficient. - Adding Ant’s information or replacing scorecard with machine learning significantly improves assessment. - Table 6: AUCs for traditional bank borrowers by information set and model - Scorecard model: - Bank credit history information: 0.74 - Ant information: 0.72 - Bank credit history information + Ant information: 0.78 - Random forest: - Bank credit history information: 0.84 - Ant information: 0.83 - Bank credit history information + Ant information: 0.87 - Note: Ant’s information includes all information (traditional information and Ant’s proprietary information) used in the baseline model. Sample only includes borrowers with bank credit history. - Table 7: 95% confidence intervals of AUCs for these four models - Random forest + bank credit history information + Ant information: AUC 0.8681; 95% conf. interval 0.8636 0.8725 - Random forest + bank credit history information: AUC 0.8372; 95% conf. interval 0.8318 0.8427 - Scorecard model + bank credit history information + Ant information: AUC 0.7767; 95% conf. interval 0.7713 0.7820 - Scorecard model + bank credit history information: AUC 0.7397; 95% conf. interval 0.7337 0.7456 - Key implications: - Bank credit history information has AUC 0.74 in traditional approach; effective for predicting defaults. - Applying machine learning raises AUC to 0.84. - Ant information alone generates similar results as bank credit history information in both scorecard and machine learning models. - Many SMEs lacking registration in the credit registration system could be covered by fintech lenders using Ant information. - Best outcome achieved by combining complete information set with machine learning. ### Impact of a policy shock (People’s Bank of China regulatory action) - Context: - On November 17, 2017, the People’s Bank of China issued draft guidelines to tighten regulations on asset management activities to curb shadow banking risks. - Short-run effect: tightening of financial conditions, with credit growth dropping by 4 percentage points the following year. - SMEs were heavily affected due to reliance on shadow banking. - Observed system NPL changes: - Banking system NPL ratio rose from 1.74 percent before the announcement to 1.87 percent one year later. - MYbank’s NPL ratio edged up (value not provided in the supplied extract). - Purpose: assess robustness of fintech approach under financial cycle change induced by exogenous regulatory shock. - General conclusion (from analysis summary): - The fintech approach’s robustness needs testing when financial conditions change; the study analyzes the impact of the exogenous regulatory shock to address this concern. *Source: Calculations using data from MYbank.* ### 1.23 to 1.3 percent. ### wpiea2020193-print-pdf - 1.23 to 1.3 percent. ### Discriminatory power of models and response to a policy shock - Comparison periods: October 2017–November 2017 (before the shock) and December 2017–January 2018 (after the shock). - Discriminatory power measured by AUC (area under the receiver operating characteristics curve). - AUC differences (Table 8): - Model IV – Model I (comprehensive evaluation): 0.11 (before shock); 0.14 (after shock). - Model IV – Model II (model advantage): 0.07 (before shock); 0.09 (after shock). - Model II – Model I (information advantage): 0.04 (before shock); 0.05 (after shock). - Interpretation: - The fintech approach (random forest + all information) outperforms the traditional approach (scorecard model + traditional information) and its outperformance increases from 0.11 to 0.14 in AUC after the shock. - The fintech approach shows both information and model advantages that are more pronounced after the policy shock. - Mechanisms: fintech uses real-time data and behavior information that are more up-to-date and more stable; machine learning models capture dynamic and interactive relationships, whereas scorecard models are linear and some traditional data can become outdated. ### Inclusiveness of fintech lending (SME heterogeneity) - Question: Does fintech credit assessment advantage vary across SMEs by size and city tier? - Method: Cumulative distribution function of increase in estimated default probability (fintech minus scorecard) by subgroup; borrowers with negative difference are “winners” under fintech. - Key subgroup findings (Figure 10; descriptive): - By firm size (annual sales turnover): - Among the smallest borrowers, ~57 percent have lower estimated default probabilities under the fintech model. - Among the largest borrowers, ~27 percent have lower estimated default probabilities under the fintech model. - Conclusion: fintech benefits smaller firms proportionately more than larger firms. - By city tier: - Borrowers in Tier-4 cities: about 62 percent have lower predicted default probabilities using the fintech model. - Borrowers in Tier-1 and Tier-2 cities: about 52 percent have lower predicted default probabilities using the fintech model. - Conclusion: SMEs in lower-tier cities benefit more from fintech, likely because they lack traditional data and rely more on proprietary information. ### Reliability of fintech predictions across borrower groups (AUCs) - Table 9: AUCs for different borrower groups under traditional and fintech approaches. - By firm size (RMB, thousands) — AUCs: - Borrowers, by annual turnover: < 20 | 20–100 | 100–500 | 500–1,000 | > 1,000 - Traditional approach: 0.72 | 0.72 | 0.71 | 0.8 | 0.76 - Fintech approach: 0.84 | 0.84 | 0.82 | 0.89 | 0.85 - Fintech – traditional: 0.12 | 0.12 | 0.11 | 0.09 | 0.09 - Interpretation: both approaches perform better for relatively larger firms (annual turnover of at least RMB 0.5 million), but the fintech advantage (difference) is larger for smaller firms. - By firm location (city tier) — AUCs: - Borrowers, by city tier: Tier-4 | Tier-3 | Tier-2 | Tier-1 - Traditional approach: 0.71 | 0.73 | 0.73 | 0.73 - Fintech approach: 0.84 | 0.85 | 0.84 | 0.84 - Fintech – traditional: 0.13 | 0.12 | 0.11 | 0.11 - Interpretation: fintech yields consistent AUC improvements across city tiers, with slightly larger gains in lower-tier cities. ### Main empirical conclusions - Overall performance: - Fintech approach (big data + machine learning) more accurately predicts loan defaults than traditional scorecard models with standard financial information. - Information advantage arises from real-time financial data and behavior information; model advantage arises from machine learning capturing nonlinear and interactive relationships. - Bank-history subsample: - Scorecard models with bank credit history predict defaults reasonably well. - Adding traditional data and proprietary information and replacing scorecards with machine learning increases AUC from 0.74 to 0.87. - Machine learning on all information without bank credit history yields AUC of 0.83, indicating proprietary information can complement or replace bank credit history. - Implication: fintech lending can potentially serve the vast number of unbanked customers (more than 90 percent of the sample). - Robustness to policy shock: - Fintech advantage is more pronounced in the after-shock period (larger AUC gains), indicating robustness to the 2017 deleveraging policy shock. Performance under other shock types remains to be tested. - Inclusiveness: - Fintech improvement in default prediction is largest for the smallest firms and for SMEs in lower-tier cities (Tier-3 and Tier-4), demonstrating fintech’s inclusive features. ### Policy implications and recommendations - Promote fintech lending while strengthening regulation: - Governments could encourage fintech lending to promote financial inclusion, but should complement encouragement with strict licensing and regulatory checkups. - Fintech requires access to big data and machine learning capacity; lessons from the failing peer-to-peer lending industry warn against pure online lending platforms without big data analysis. - Regulators should remain vigilant for new risks from fintech business models, especially in downturns, and balance innovation with financial stability. - Use of regulatory sandboxes recommended to incubate new business models and practices. - Encourage collaboration between fintech lenders and traditional banks: - Fintech: superior information and analytics but limited funding capacity. - Banks: substantial financial resources and soft information but limited big-data analytics. - Example practice in China: as of the end of 2019, RMB 2 trillion in business loans (equivalent to 2 percent of gross domestic product) were jointly issued by banks and fintech firms, where fintech provided initial screening and banks provided most of the funding. - Joint lending can promote financial inclusion; modalities can be designed case-by-case. - Enhance data protection and access policy: - Governments should set clear data policy standards to maximize big data benefits while protecting privacy; emphasize data authenticity, legality, and security. - Consider fostering portability of individual data across platforms to increase competition. - Ensure data access (including proprietary fintech data) is balanced with privacy protection and proper usage; lack of data sharing may exacerbate concentration and network-effect issues. - Sequence public investments: - Development of fintech requires digital infrastructure, an established online ecosystem, and big data analytical skills. - Policy sequencing: invest in digital infrastructure, support e-commerce and broader digitalization, then expand fintech lending. ### Appendix A — MYbank key facts (selected) - Founded: 2015 (headquartered in Hangzhou, Zhejiang province). - Ownership: founded by Alibaba’s affiliate firm, Ant Group, through a 30 percent stake in a joint venture comprising a group of private firms. - Borrower base and performance: - About 20 million SME borrowers, about 80 percent of whom have never borrowed from banks in the past. - NPL ratio at about 1 percent (text also reports NPL ratio was 1.3 percent in 2019). - 2019 summary figures: - Total assets: RMB 139.6 billion ($20 billion). - Average loan size: RMB 31,000 ($4,500). - Accumulated SMEs served: 20.9 million. - Capital adequacy ratio: 16.4 percent. - NPL ratio: 1.3 percent. - Operational model: data/cloud-based model with no physical branches; operates 24 hours a day, seven days a week; uses big data and machine learning for credit profiling (e.g., e-commerce and cash flow). ### Appendix B — Random Forest (algorithmal note) - Random forest builds a large collection of uncorrelated decision trees and averages them. - Decision tree elements: - Internal node: test on an attribute. - Branch: outcome of the test. - Leaf node: class label (default or not). - Splitting criterion: - Uses Gini index to measure node impurity. - For node m with N_m observations and C classes, p_mc denotes proportion of class c in node m. - For two-class default/not-default case, Gini index Gini(D_m) = 2 p_m (1 − p_m), where p_m is proportion of defaults in node m. - A binary split on attribute A partitions D_m into D_m,1 and D_m,2; Gini_A(D_m) is the weighted sum of impurities after the split. - Reduction in impurity by split on A: ΔGini(A) = Gini(D_m) − Gini_A(D_m). - Attribute and split-point maximizing ΔGini(A) are selected as the splitting criterion. - Decision tree built by iteratively selecting optimal splits (steps summarized in the source). *Source: Calculations using data from MYbank.* ### 1. The tree starts as a single node, representing the training data set in D. In the process of building ### wpiea2020193-print-pdf - 1. The tree starts as a single node, representing the training data set in D. In the process of building ### Decision tree and random forest construction - Decision tree growth procedure: - The tree starts as a single node, representing the training data set in D. - If all observations are of the same class, node m becomes a leaf and is labeled with that class. - If there are no remaining attributes on which the observations may be further partitioned, then node m becomes a leaf and is labeled with the most common class in D (majority voting). - Otherwise, the splitting criterion automatically separates or partitions the tuples in D into individual classes, producing partitions D_m. - The algorithm recursively repeats steps 1 and 2 to form a decision tree for the observations at each resulting partition D_m, until the minimum node size n_min is reached. - Random forest implementation: - Random forest builds an ensemble of trees by bootstrapping M subsets from training data and trains M de-correlated trees on these subsets. - Each decision tree in the random forest outputs a class prediction and the class with the most votes becomes the model’s prediction. - Implementation specifics in the analysis: - Data set: 76 feature variables and a response variable indicating whether to default or not. - Number of trees: 500 decision trees (M=500). - Software: open-source package RandomForestClassifier in Python used to train random forest models. ### Feature importance - Definition and calculation: - Feature importance measures the predictive power of each variable. - It is measured by the improvement in the split criterion attributed to the splitting variable at each split in each tree and accumulated over all the trees in the forest. ### Credit scorecard methodology (three-step algorithm) - Overview: - The credit scorecard model is widely used in the financial industry to evaluate borrowers’ creditworthiness. - The algorithm contains three steps. - Step 1 — Weight of Evidence (WoE) transformation: - Purpose: Transform all independent variables using the weight of evidence (WoE) method to measure the “strength” of grouping for differentiating good and bad risk and to attempt to find a monotonic relationship between the independent variables and the target variable (0 if no default, 1 if default). - Procedure: - Split the data into several bins of each independent variable (e.g., age and income group). - Calculate WoE for bin i of independent variable j as defined: - WOE_{i,j} = ln( (B_i / B_T) / (G_i / G_T) ) - Where B_i is the number of bad borrowers in bin i and B_T is the number of total bad borrowers. - G_i and G_T are the numbers of good borrowers in bin i and the whole sample, respectively. - Replace the raw data with the calculated WoE values. - Benefits: Helps to build a strict linear relationship with log-odds, and can handle missing values and outliers. - Step 2 — Information Value (IV) for feature selection: - Purpose: Use information value to select informative independent variables (features). - Definition: - IV_i = ln( (B_i / B_T) / (G_i / G_T) ) * ( (B_i / B_T) - (G_i / G_T) ) - Rule of thumb in the analysis: - Variables with IV less than 0.02 can be dropped. - Of the 76 variables, 32 were selected and entered into the logit model later. - Step 3 — Logistic regression on WoE-transformed variables and score scaling: - Logistic regression model: - Assume N dependent variables (N=32 in this paper). - Model form given: - y = 1 / (1 − e^{−β_0 + ∑_{j=1}^N β_j * WoE_j} ) - Where y takes 0 if there is no default, and 1 in the case of defaults, and WoE_j is the transformed WoE for independent variable X_j. - Scorecard scaling: - Use logistic regression coefficients and WoE values to scale the model into a scorecard. - For each independent variable X_j, its corresponding score is: - sc_WoE_j = (β_j * WoE_j + α_G) * factor_f_Ws + W_ffs_G - (As stated in the source: "푠푠푐푐푊푊 푠푠푒푒_푗푗 = �훽훽_푗푗 ∗ 푊푊푊푊푊푊_푗푗 + 훼훼_퐺퐺� ∗ 푓푓푓푓푐푐 푓푓푊푊푠푠 + 푊푊𝑓𝑓푓𝑓𝑠𝑠𝑒𝑒𝑓𝑓_𝐺𝐺") - Total score is the sum of individual scores: - Total_score = ∑_{i=1}^n sc_WoE_i - Application in the analysis: - The analysis did not calculate the credit score itself but used the logit model results to calculate expected default probability and plot the area under the receiver operating characteristics curve. - Software: open-source package scorecard in R used to train the scorecard models. ### Implementation and data specifics highlighted - Dataset: 76 feature variables; response variable indicates default or not. - Feature selection outcome: 32 variables selected for the logit model. - Random forest configuration: M=500 trees; RandomForestClassifier in Python. - Scorecard training: scorecard package in R. - Evaluation: Expected default probability and area under the receiver operating characteristics curve plotted (scorecard not calculated). *Italic: Source: wpiea2020193-print-pdf (IMF Working Paper content as provided).* --- _Source: https://www.imf.org/-/media/files/publications/wp/2020/english/wpiea2020193-print-pdf.pdf_