Oded Netzer
Columbia Business School

When Words Sweat: Identifying Signals for Loan Default in the Text of Loan
Applications

In this research we propose that borrowers' writing, when requesting an online
crowdfunded loan, may leave some traces, similar to body language detected by
loan officers, that can improve predictions of future repayment behavior.
Specifically, we investigate whether borrowers leave traces of their intentions,
circumstances, emotional states, and personality in the text they write that are
predictive of whether they will default on their loan up to three years after
the text was written. To answer these questions we apply text-mining and
machine-learning tools to a large dataset of loans from the crowdfunding
platform Prosper. We began by creating an ensemble of predictive models
consisting of decision trees and regularized logit models. We find that the
predictive ability of the textual information alone is of similar magnitude to
that of the financial and demographic information. Moreover, supplementing the
financial and demographical information with the textual information improves
predictions of default by as much as 4.03%. Next we used a multi-method approach
to uncover the words and writing styles that are most predictive of default.
Using a naÌøve Bayes and an L1 regularization binary logistic model we find that
loan requests written by defaulting borrowers are more likely to include words
related to the borrower's family, financial and general hardship, mentions of
God and the near future, as well as pleading lenders for help. We use a latent
Dirichlet analysis (LDA) to identify the loan purpose, life circumstances, and
writing styles that are most associated with loan default. We find that loans
whose purpose is to help with a business or medical circumstances are riskier
than other types loans in terms of their default likelihood. Consistent with the
naÌøve Bayes results, we find that pleading lenders for help and providing
explanations are also associated with higher risk of default,. We further
explore the writing styles and personality traces embedded in the loan request
text using the Linguistic Inquiry and Word Count dictionary (LIWC; Tausczik and
Pennebaker 2010). We find that defaulting loan requests are written in a manner
consistent with the writing style of extroverts. This is not surprising as
extroverts aspire to live exciting lives, more so than others, and hence spend
their money on experiences and worry less about the financial consequences
(Brown and Taylor 2014). We further find that defaulting loan requests are
written in a manner consistent with the writing style of liars. While we are
unable to claim that defaulting borrowers were intentionally deceptive when they
wrote the loan request, we believe their writing style may have reflected their
doubts in their ability to repay the loan. Our research shows that in an
environment characterized by high uncertainty, verifiable and unverifiable data
have similar predictive ability. While borrowers can truly write whatever they
wish in the textbox of the loan application -- supposedly "cheap talk"
(Farrell and Rabin 1996) -- their word usage reveals something about them that we
could not have inferred from their financial dossier alone, namely their
personality and intentions.