zidisha v6
TRANSCRIPT
Identifying sustainable interest rates while helping African small businesses grow
Jack ChaiInsight Data Science Fellow2014
Den
sity
Den
sity
Minimal increase in average interest rate from 6% to 6.8%Would have cut losses in 2014 by $17K or 89%
Minimal increase in average interest rate from 6% to 6.8%Would have cut losses in 2014 by $17K or 89% Would have cut losses from 2009 onwards by $240K or 82%
Den
sity
Den
sity
Predictive model created from combination of logistic regression and machine learning (SVM)
• Basic probability theory to deal with class bias
Predictive model created from combination of logistic regression and machine learning (SVM)
• Basic probability theory to deal with class bias
𝑃 (𝑙𝑜𝑠𝑠 )=𝑃 (𝑑𝑒𝑓𝑎𝑢𝑙𝑡 )∗(1−𝑃 (𝑠𝑜𝑚𝑒𝑝𝑎𝑦𝑚𝑒𝑛𝑡|𝑑𝑒𝑓𝑎𝑢𝑙𝑡 ))
Predictive model created from combination of logistic regression and machine learning (SVM)
• Basic probability theory to deal with class bias
𝑃 (𝑙𝑜𝑠𝑠 )=𝑃 (𝑑𝑒𝑓𝑎𝑢𝑙𝑡 )∗(1−𝑃 (𝑠𝑜𝑚𝑒𝑝𝑎𝑦𝑚𝑒𝑛𝑡|𝑑𝑒𝑓𝑎𝑢𝑙𝑡 ))
Predictive model created from combination of logistic regression and machine learning (SVM)
• Basic probability theory to deal with class bias• Logistic regression identified 4 features that could predict risk• “Riskier population”• Borrower allowed maximum interest rate• Loan Category• Country of applicant
Predictive model created from combination of logistic regression and machine learning (SVM)
• Basic probability theory to deal with class bias• Logistic regression identified 4 features that could predict risk• “Riskier population”• Borrower allowed maximum interest rate• Loan Category• Country of applicant
Higher Risk Associated with Borrowers who entered between August 2012 and August 2013
Den
sity
Augu
st 2
012
Augu
st 2
013
Predictive model created from combination of logistic regression and machine learning (SVM)
• Basic probability theory to deal with class bias• Logistic regression identified 4 features that could predict risk• “Riskier population”• Borrower allowed maximum interest rate• Loan Category• Country of applicant
• Used identified features to train kernel SVM with 10 fold cross validation (cut losses by 89%)
• Impact/Significance• Project to cut losses by $48,000 over the next year• Over 5 year period, for every $1 million invested, recovers additional
$110,000 that can continue to be reinvested
• Actions already taken• Implement the model the risk model for interest rates
• Actions to be taken• Find policy change that allowed for risky population
Conclusions