Can Artificial Intelligence be used to gain an investment edge in peer-to-peer lending?

A condensed, mobile-readable version of the full case study.

The question

P2P lending platforms bucket and grade borrower applications based on the borrowers' risk profile i.e. the predicted ability to repay the loan.
More than a 100 parameters are used to adjudge a borrower's application. These range from credit score (FICO) and debt-to-income ratio (DTI) to home ownership status, monthly income, and credit card utilization data.
A borrower's risk grade determines the interest rate on their loan.

We present IRRs for all loans made by a large P2P platform between 2007 and 2012.
Investors garnered annualized returns of 5% to 10%, generally with a higher return for higher risk.
While the gap between the APR on the loans and the IRR realized by investors increases with a worsening in grade, no grade is immune to the effects of charge-offs.
This chart serves as a baseline for our analysis, as we look to narrow this gap in returns, across grades.
Our goal is to utilize machine learning and AI to construct a superior portfolio that will outperform these baseline IRRs.

Since the beginning of the decade, the P2P platform we analyze has issued more than 1.6 million loans.
Between application data from the borrower and data pulled from credit bureaus and other third parties, we have access to more than 150 fields of data per loan.
We have nearly 30 million rows of time series data detailing month-by-month payment history of every loan.

We construct a 1,000+ dimensional dataset to train our binary classifier.
We run and evaluate various algorithms including Gaussian Naive Bayes, Support Vector Classifier, Decision Tree, Random Forest, Logistic Regression and several Deep Neural Networks.
In the associated chart, the Deep Neural Network shown was the best predictor, while the Gaussian Naive Bayes was the worst.
The dashed line (an AUC of .5), represents the baseline to benchmark against. In other words, a model with this performance will make us buy all loans originated (or a representative sample), leading to returns that would mirror the overall platform experience.

The general approach was to train the model on a random subset of loans (training set) and search the hyperparameter space for the best settings using cross-validation. The results were then evaluated on unseen loan data (test set).
Learning is performed on the training data while hyperparameters are tuned using the cross-validation set. The (out of sample) test set is used for the evaluation of model efficacy.

Of the parameters we show, debt-to-income ratio (DTI) was the most powerful in explaining default performance, followed by FICO.
Besides FICO, number of credit accounts and total credit limit were positively correlated with performance. Parameters related to a recently heavier need and use of credit were negatively correlated.

Here, we present the returns generated on an out-of-sample portfolio by a binary classifier.
We use the same vintages as previously presented to compare apples to apples.
As can be seen, this portfolio outperformed the baseline returns by 8% to 20%+, with more outperformance for riskier grades.
The returns (and outperformance) are dependent on the portfolio strategy implemented. The strategies, depending on the mandate, could be as diverse as targeting a certain proportion of overall production or production by grade or equal portfolio-weighting by grade or more actively managed strategies, etc.

Backtesting results are presented on out-of-sample data from 2007-2012 vintages.
We only report results where our algos were able to select at least 500 loans (± small tolerance) for a portfolio given the constraints we overlay.
The basic investment strategy is benchmarked to baseline IRRs. In other words, it is a 'relative value' strategy.
We are not aware of the existence of any efficient mechanism to sell a baseline portfolio as a hedge against our artificial intelligence-based portfolio. This ties our fate, to a certain degree, to the overall sector's performance.
We are also unsure of the volume and efficiency of trading peer-to-peer loans on secondary exchanges. Effectively, the strategy is a 'hold to maturity' strategy.
The returns presented here are gross, unlevered returns.

The full case study, with every chart and the complete commentary, is laid out for larger screens.