For traders of residential mortgage backed securities (NASDAQ:RMBS), collateralized mortgage obligations (CMO's) and other prepayment sensitive securities, the likelihood of prepayment across a range of scenarios is the operative factor in assessing risk and opportunity.
Traditional regression models assess the likelihood of pre-payment by measuring historical pre-payments and drawing correlations to those historical prepayments with various indicators, including but not limited to: interest rates, household income, unemployment rates and average LTV at origination.
Rudimentary regression models, including linear, logistic and hedonic regression models, are generally considered to be out-dated for pre-payment prediction, and studies dispute their accuracy for this purpose (Sirignano, Sadhwani, Giesecke 2015), despite the demonstrative evidence of their usefulness in valuing collateral (i.e. homes and buildings) quickly and in an automated fashion (Monson 2009).
(Source: Geanakoplos, Axtell, Farmer, Howitt, Conlee, Goldstein, Hendrey, Palmer, Yang 2012)
For about 20 years, hedge funds and other sophisticated traders have used more effective models that incorporate agent-based modeling of various types. The most effective agent-based models, such as the ones developed by John Geanakoplos and the research team at Kidder Peabody in the early 1990's, focused on the cost to prepay. This 'cost' is broken down into easily quantifiable factors including basic closing costs, but also includes more arcane factors such as time cost, inconvenience, psychological cost, etc.
Many agent based models that are employed to predict pre-payments include a variety of cost factors, but often times, the more complex factors used to assess the costs to pre-pay are simply derivatives of the simpler ones.
The advent of advanced neural networks, commoditized computing power, cloud computing, and most importantly, the ability to collect and deploy massive of data about individuals via basic services like Facebook may make for a golden age of trading for those who can take advantage of those tools, before they are generally accepted and thus priced in.
Inside one database (Facebook) exists every piece of personal data that a predictor of RMBS pre-payments could ever want to know, including current and past employers, educational history, religious views, relationship status (and the health of those relationships), location history and about 80 others.
Would the likelihood of divorce in a household or the sentiment of a neighborhood be a useful factor for predictors of mortgage defaults and prepayments? I think so. That's why I'm looking very closely at how social media data can be used to construct effective models for trading residential mortgage backed securities and other pre-payment sensitive securities.
Disclosure: I/we have no positions in any stocks mentioned, and no plans to initiate any positions within the next 72 hours.