Banque Populaire Chair in Microfinance Burgundy School of Business (ESC Dijon) Elements of Artificial Intelligence in microfinance - the use of Credit Scoring and Poverty Scoring by the MFIs in Latin America Vitalie. BUMACOV @ ESCDijon. eu
Setting the STAGE a. If no stable (satisfactory) employment, then Self-employment; b. Self-employment depends on the Capacity to generate Added Value; c. High added value requires Certain combination of Factors of Production; d. Most factors of production can be acquired if Access to Finance; e. Global Awareness f. Microfinance receives donations, volunteering, subsidized capital and market funds
Self-SELECTION Self-selection is problematic in microfinance: - Many do not have the capacity to borrow; - Who have the capacity (and the need), do not apply for credit. Informal environment; Information asymmetry; Reliance on guaranties is cherry-picking. GOOD* BORROWERS THAT ARE POOR** COMPLEXITY * Good = Creditworthy; ** Poor = in a state of poverty, revenue below a threshold.
Loan officers INTELLIGENCE very general mental capability that involves the ability to reason, plan, learn quickly and learn from experience*. Short loan cycles & big number of active borrowers per loan officer => many (loan) iterations => large array of good and bad cases to LEARN (FROM EXPERIENCE). The mind infers patterns specific to: - bad borrowers and/or bad businesses; - good borrowers. Relational vs. Transactional lending: - Relational -> Loan officer s intelligence; - Transactional -> Artificial intelligence. Experience + Intuition + Emotions + Self-interest + Irrationality + Private info VS. Artificial Intelligence N.B.: Intuition is matching patterns in order to quickly suggest feasible courses of action * famous 1993 statement of 50+ academics
Artificial INTELLIGENCE Classification basic concept in I & AI. Pattern matching basis of classifiers. Credit Scoring -> AI oriented at classifying borrowers into GOOD & BAD. Poverty Scoring -> AI classifying borrowers into POOR & WEALTHY. S algorithms learn patterns from experience: - regressions, - decision trees, - neural networks, - support vector machines There is no one best technique. Rule: lowest type II error. Real Good Real Bad Classified G 90% 10% Classified B 15% 85%
Borrowers that are BAD In banking / MFIs: - loss-making activities not financed by law / policy / principle; - risky activities excluded by policies; - economic cycles (Recession x to Recession x+1) too long. In microfinance loan officers screen (mainly) applicants with bad character MFIs don t capture such data. Reflection of borrower s character in: - socio-demographic data - business-demographic data - loan preferences - past credit behavior. Do microfinance institutions use such classifiers for identifying good / bad credit risks and poor / not poor applicants?
TOOLS 2 on-line surveys using surveymonkey.com - Worldwide English in November 2012 113 answers (19 LAS) - Spanish LA in Spanish in May 2013 147 answers 19 + 156 = 175 answers from LAS MFIs => 150 distinct MFIs 30% answer rate / confidence interval 6.6% @ 95% level 17 cases of double, triple and quadruple answers: - 1 inconsistency in declaration of use / no-use of CS; - 2 + 1 inconsistencies in PS use / no-use. Used the more Senior / Recent answer.
STATISTICS No Country MFIs Borrowers Portfolio LAS Targets Answers 1 Argentina 13 39,915 42,029,933 Y 44 7 2 Belize 1 4,559 14,398,437 N 3 Bolivia 25 1,026,190 3,047,940,788 Y 26 14 4 Brazil 22 2,007,737 1,838,679,288 N 5 Chile 5 263,756 1,639,740,254 Y 6 2 6 Colombia 30 2,289,703 5,311,133,769 Y 35 16 7 Costa Rica 14 16,445 65,168,419 Y 17 2 8 Cuba Y 0 0 9 Dominican R. 10 407,570 644,509,950 Y 12 5 10 Ecuador 45 845,309 2,232,689,054 Y 105 20 11 El Salvador 12 147,190 367,179,889 Y 14 7 12 Guatemala 19 367,722 198,215,391 Y 23 6 13 Haiti 4 116,828 70,140,328 N 14 Honduras 21 180,114 235,289,405 Y 21 11 15 Jamaica 1 13,895 9,041,762 N 16 Mexico 59 6,067,058 1,991,632,893 Y 67 18 17 Nicaragua 22 315,248 272,883,574 Y 29 11 18 Panama 3 14,866 22,136,530 Y 4 2 19 Paraguay 6 500,660 986,204,462 Y 7 4 20 Peru 60 3,637,566 8,767,546,100 Y 60 24 21 Uruguay 1 2,113 7,324,248 Y 1 1 22 Venezuela 1 44,874 121,911,409 Y 2 0 Total 472 150
Use of CS Do your MFI use CS? Q LAS Int. Yes 53 35.3% 38.8% Intend, Short Term 46 30.7% 33.0% 54.7% Intend, Long Term 36 24.0% 18.4% No, don't intend 15 10.0% 9.7% Define it: 1. Algorithm developed using sample of past Good and Bad borrowers + statistics. 2. Algorithm developed by credit experts by agreeing on the factors & weights. 3. Algorithm acquired from consulting company. 4. Score from a credit bureau. Of the 53 MFIs that use CS Q LAS Int. 1. Statistics / Mathematics 19 35.8% 55.0% 2. Rating 17 32.1% 25.0% 3. Buy from CS company 8 15.1% 12.5% 4. Credit bureau 9 17.0% 7.5% Real Use of CS: 24% (29.1% Int); Confidence Interval: 5.7 @ 95%
Definition of CS Of the 97 MFIs that don t use CS Q LAS Int. 1. Statistics / Mathematics 48 49.5% 52.4% 2. Rating 21 21.6% 30.2% 3. Buy from CS company 1 1.0% 4.8% 4. Credit bureau 27 27.8% 12.7% Wrong CS definition: 21.6% (25.3% including those who say use ) Rating AI Big emphasis on Credit Bureaus! Those that don t intend to use CS (15 MFIS): Total Bad previous experience with CS 0 We trust more the decision of a loan officer / no need for CS 10 We don't know the principles of CS 0 We can not find the required budget. 3 We think that economic advantages of CS are below the costs 2
Use of PS Do your MFI use PS? LAS Int. Yes 22 14.7% 39.8% Intend, Short Term 55 36.7% 23.3% 63.4% Intend, Long Term 40 26.7% 21.4% No, don't intend 33 22.0% 15.5% Define it: 1. Algorithm developed using sample of own Poor and Wealthy clients + statistics. 2. Algorithm developed using sample from national/regional survey + statistics. 3. Algorithm developed by Poverty experts by agreeing on the factors & weights. 4. Acquired from institution, company, NGO, website (progressoutofpoverty.org). Real Use of PS: 12% (⅓ Int); Confidence Interval: 4.3 @ 95% Of the 22 MFIs that use PS Q LAS Int. 1. Statistics/Math. (own sample) 6 23.8% 29.3% 2. Statistics/Math. (notional survey) 5 23.8% 29.3% 3. Rating 4 19.0% 17.1% 4. NGO, website 7 33.3% 24.4%
Use of CS: 24% - less than ¼ (29.1% Int) Use of PS: 12% (33% Int) Joint use of CS & PS: 4.7% (18.4% Int) Joint use of CS & PS Conclusions Only 10 MFIs - 6.7% (3.5%) declared no interest in using CS & PS Averages Use CS (1) Not use CS (2) Use PS (3) Not use PS (4) PAR 30 (Risk) 6.3% (6.8) 6.1% (5.4) 4.6% (6.8) 6.4% (5.5) Poor rate 42.9% (65.6) 50.9% (59.4) 61.0% (62.4) 47.3% (60.6) Female borrow. 59.6% (69.8) 63.9% (67.6) 65.1% (75.2) 62.4% (65.2)
CORRELATIONS Extra * Poor Fem PAR Real CS Real PS CS&PS Scale Poor 1.00 0.19-0.03-0.11 0.15 0.11-0.22 Fem 0.19 1.00-0.06-0.09 0.04-0.06 0.05 PAR -0.03-0.06 1.00 0.01-0.06 0.00-0.20 Real CS -0.11-0.09 0.01 1.00 0.13 0.39 0.18 Real PS 0.15 0.04-0.06 0.13 1.00 0.60 0.01 CS&PS 0.11-0.06 0.00 0.39 0.60 1.00 0.08 Scale -0.22 0.05-0.20 0.18 0.01 0.08 1.00 Next steps: - Regressions (extra info from MIX Market); - Deeper investigation of the targeting process with PS Is it targeting or measurement and if high, reporting? * Significance check not performed
Survey QUESTIONS Extra 1. La IMF que usted representa utiliza actualmente la herramienta de Scoring de Crédito en el proceso de selección para distinguir entre solicitantes de crédito potencialmente buenos y potencialmente malos? 2. En su opinión, cuál de las siguientes afirmaciones describe mejor el Scoring de Crédito utilizado en su IMF? Por favor, si su IMF no utiliza Credit Scoring, señale cuál de las siguientes afirmaciones en su opinión describe mejor el concepto de Scoring de Crédito? 3. Si su institución no utiliza el Scoring de Crédito, pero tiene la intención de usarlo algún día, cuál es el motivo principal de no haber utilizado dicho método? 4. Si no se utiliza el Scoring de Crédito y no tiene intención de utilizarlo, es a causa de que: 5. La institución de microfinanzas que usted representa utiliza actualmente el Scoring de Pobreza en el proceso de identificación de solicitantes de crédito pobres y no pobres? 6. Cuál de las siguientes definiciones describe mejor el concepto de Scoring de Pobreza que se utiliza en la IMF que usted representa? Por favor, si su IMF no utiliza Scoring de Pobreza, señale cuál de las siguientes definiciones, en su opinión, describe mejor este concepto? 7. Cuántos clientes con microcrédito activo tiene su institución de microfinanzas actualmente? 8. Aproximadamente, cuál es el porcentaje de prestatarios pobres entre sus micro prestatarios actuales y cuál es el porcentaje de mujeres prestatarias entre sus micro prestatarios actuales? Cuál es el PAR30 de su institución)? 9. En qué pais de América Latina se encuentra la Institución de microfinanzas que usted representa?