This section is only viewable on larger screen sizes. Please use a tablet or bigger. For best experience, please view on a desktop computer and use Google Chrome browser.

Large Scale Data Visualization

Total Return Distribution Returns and Defaults by Grade Risk parameters vs. risk grade Underwriting drift? Performance drift?
  • A critical step, before applying machine learning algorithms, is to dig into the data and develop intuition around it.
  • With millions of rows and hundreds of columns of data to wade through, this is best accomplished via data visualization.
  • Data visualization also informs our AI approach and is very useful for coming up with a gameplan for the machine learning phase.
  • In this section, we present of few of the charts we generated before beginning our AI task.
  • As would be expected, the total return distribution of charged off loans is vastly different from that of loans that paid off.
  • The high correlation between a loan's final status (charged off or fully paid) and the total return generated implies that if we can use a loan's attributes to predict its final status, we should be able to construct a portfolio that's largely superior to the baseline.
  • We can therefore train a binary classifier to distinguish between the charged off and paid off loan populations and expect some degree of success.
  • The risk grades do a decent job of discerning loan default performance.
  • For vintages that have fully resolved (all loans having paid off or charged off), moving a notch up in risk grade implied roughly 5 percentage points of additional defaults. (The exception being an outperforming grade A.)
  • Investors were largely compensated for taking on additional risk. The exception being grade F where the increase in APR was not enough to offset the increase in defaults.
  • The top left violin plot shows that FICO bands change significantly between grades A and B before legging down with each grade between B and D. Surprisingly, only grade A saw a meaningful difference in the distributions of fully paid vs. charged off loans.
  • Debt-to-income (DTI) ratios creep up with each grade, with defaulted populations having higher DTIs for each grade.
  • Interestingly, income levels rise with risk grades. Since income may or may not have been verified by the P2P platform, this made us investigate the 'stated income' effect. The bar chart on the bottom right shows that the platform (rightly) verified borrower income for a higher proportion of riskier loans.
  • FICO and DTI have drifted significantly over time. After the initial years, the platform's increase in bandwidth came at the cost of negative drifts before turning around in late 2015-early 2016. DTIs seem to be edging up again and would need to be monitored.
  • Offsetting the above were positive drifts in applicants' average years of employment and home ownership status (set against the backdrop of an improving economy). These trends have remained relatively stable for the last couple of years.
  • A natural question from the previous section is - what is the net effect of all these interactions?
  • Generally, performance has worsened over time and correlates fairly well with changes in underwriting.
  • Performance for the 2015 vintage seems to be saturating before and lower than preceding vintages. But it's still early days. Something to keep an eye on for the next few quarters.