Continuing the theme of models and bots, here’s another chapter of my unpublished book, Mortgage Market Mayhem.
A later chapter covers the 2007/2008 failure of the credit models. First came the 1990/1991 failure of prepayment models.
Just a reminder: Any model based on historical data cannot predict what happens if new activity goes beyond the boundaries of that data. That’s why I put out a faint defense of the Rating Agencies last fall.
They are NRSRO’s, or Nationally Recognized Statistical Rating Organizations. It seems ludicrous to criticize them for using statistics to establish their performance projections. We might as well have Nationally Recognized Opinion Blogs if we throw out statistics based on reality.
After the break, enjoy Chapter 14 of Mortgage Market Mayhem.
The Year the Prepayment Models Broke
The mortgage market had plunged to a bottom in 1986 and 1987 and was whipsawed from a sudden rise in mortgage rates in 1988 and early 1989. Investor concerns were centered on prepayments – how many people would pay off their mortgages before the terms of those mortgages (typically fifteen-year or thirty-year) ended? With the volatility of the market, investors shifted from worrying about early call risk from prepayments to worrying about whether their mortgage securities would extend in time as prepayments dropped.
These concerns were based on models for prepayments, which were based on observations of the behavior of tens of millions of mortgages over several economic cycles. The analytics groups I worked in, and others like them, had taken a wealth of historical data and built predictive prepayment models.
Prepayments always happen and there are many different reasons why they happen. Personal events trigger them. Regardless of what else is going on in the economy, some borrowers will sell their houses. There are also a myriad of specific life events and circumstances that can trigger prepayments, such as losing a job, getting a new job, having another child, getting divorced, suffering a major illness, or a death in the family.
Changes in the economy can also trigger prepayments. In fact, the primary reason for prepayment is that a borrower takes out a new mortgage. Borrowers usually refinance to get a loan with a lower rate, but they can also refinance out of variable rate loans into fixed rate loans to avoid the risk of having their payments go up in the future.
Homebuyers may need to draw additional equity from their homes in a high-rate environment, causing them to prepay even when rates are higher. They may be underemployed or disqualified for refinancing for other reasons, and therefore not exercising their “in the money” option of prepaying when rates are lower.
In fact, at any given time, a typical pool of thousands of borrowers will have some people prepaying no matter where rates are, and some not prepaying no matter how attractive the refinancing rate might be.
Collecting data on the behavior of a large group of mortgages forms the foundation for prepayment model-building. Looking at the data graphically can help identify “base functions” and external variables that might affect a mortgage borrower’s behavior.
Statistical tools are used to compare the behavior of a mortgage borrower with external variables to see whether the movements in a variety of outside factors tends to lead or lag correlated movements away from the mean, or average, expected prepayments.
When my team and I built a prepayment model, the first external variable we looked at was the age of the loan. We called it the “seasoning” of the loan. Once we had an expected life cycle for an average new loan, we could study how other variables impacted mortgage borrowers.
There was also a seasonality to loan prepayments. There were enough homeowners with children to make the school year a strong effect. Winter weather in northern states tended to keep people home, not spending their weekends looking for a new house. The exception to this was resort and retirement areas, that sometimes showed a “reverse seasonality” when compared to the bulk of the U.S. housing market.
As you might expect, there was a clear correlation between falling mortgage rates and people paying off their loans and refinancing. Upfront costs like appraisals, fees, and commissions typically acted like a barrier to refinancing, so borrowers only got serious about this option when rates dropped by a couple of hundred basis points.
After factoring in the universal effects of seasoning, along with the time of the year, we could compare our “expected” prepayments with what actually happened, and see that over 70% of the variability in prepayments was explained by mortgage rates.
The next most significant outside variables turned out to be 1) employment growth in the immediate area, and 2) house prices. Those two variables seemed to be interrelated, which made sense – when employment rose, there was more demand for houses as new people arrived to take new jobs. More demand raised prices. Between these two variables, employment growth appeared to be the stronger “explanatory” variable.
The location of the loan was an important factor as well. In general, the East Coast, West Coast and Sun Belt showed greater housing turnover activity, so they had faster prepayments. Perhaps most striking within a single state was the difference between upstate New York (known for its slow mortgage prepayment) and the region closer to New York City, where mortgages pay off much faster, especially when mortgage rates drop.
We found that the coastal regions seemed to be more likely to take advantage of lower rates. We guessed that geographic areas with more concentrated populations were areas where more mortgage bankers set up shop, which then led to more refinancing.
Several additional variables were found to be significant as we combed through the data. One, we called “the USA Today effect.” If mortgage rates hit a new low, the news media would run stories about it in the paper and on TV, and, like clockwork, two, three and four months later, we’d see a surge in refinancings. This made perfect sense, since at that time, it took between one and three months to process a new loan application, and then another month for the prepayment to show up in the mortgage bond.
Another surge in prepayments could be seen shortly after mortgage rates bottomed and then headed back up. We hypothesized that this effect was caused by borrowers waiting to see the bottom before committing, so they would avoid locking in a new rate only half way down.
We had literally tens of millions of mortgage “life histories” documented from the FHA’s experience of guaranteeing mortgages. Dating back to 1970, we also had detailed history for the loans that Freddie Mac and Fannie Mae guaranteed.
That gave us plenty of data to test against. By 1986, we already had models that would predict performance for mortgage “pools” to accuracies of 95% to 97%.
This research and the explosion of structuring techniques to divide up the cash flows in mortgage securities created a new kind of competition among Wall Street firms.
In the days before computers, investors and issuers of securities often chose their investment banks based on whom they trusted to give them good advice. It was all about the people, their relationships, their industry knowledge and their integrity. Now it was largely about which firm had the best models and market penetration.
In 1990, the bond market was coming off an unusual period. The shortest rates, for one month through six months, were at or above the longest rates, the rates for ten years to thirty years. This “inverted yield curve” was a rare occurrence, and nearly always led to a recession and dramatically lower interest rates.
The size and national scope of the Savings and Loan debacle had become clear. The overhang of so many assets needing to be sold, especially commercial buildings, had led the Federal Reserve to aggressively lower rates in an attempt to head off a recession. Mortgage rates came down, and people began to refinance their mortgages.
The government sponsored a new company to manage the liquidation of the assets of the failed thrifts, and called it the Resolution Trust Company, or RTC. I was working as Head of Analytics in the Mortgage Department at Daiwa Securities America and our group, along with our competitors, was very active buying mortgage loans of all types from the RTC, as well as structured bonds.
The most difficult to analyze bonds were the “residuals” from CMO deals. Even conceptually, residuals can be difficult to understand, as they are more than just the “left over” cash flow after paying the rest of the CMO classes. This unique tranche of CMO deals came into being as a consequence of the 1986 Tax Act. Legally, it is equity rather than debt, and serves as a way of providing the IRS with the taxes it requires which aren’t being paid by the buyers of the bonds in a CMO deal.
If, for example, mortgages that yielded 8.5% were carved into tranches in a CMO deal, the government wanted to be paid taxes on that full 8.5% yield, regardless of which bonds were paying off when. The bonds in the deal might be yielding anywhere from 6% to 10%, providing taxable income either lower or higher than that the 8.5% that IRS wished to tax, and some would pay off before others.
Enter the residual, a nifty tranche to fix the problem.
Since the residual tranche was required to pay enough taxes to maintain the tax revenue from a steady 8.5% income, this created “phantom income” – the investor in the residual didn’t actually receive this income but was required to pay taxes on it nonetheless. Over the life of the deal, this phantom income would be balanced out by “phantom losses” once the lower-yielding bonds had paid off and only the higher yielding bonds remained.
Needless to say, the fact that a residual bond might bring its owner five million dollars of cash flow in a year but generate six or seven million dollars of taxable income, made it both complex and attractive only to a select few investors. With its tax implications and its position as the remainder after everything else was paid, you can almost imagine the Greek chorus chanting “Toxic Waste! Toxic Waste!” next to the trading desk each time we bought one of these unusual illiquid bonds.
Our group at Daiwa started buying these residual bonds from the RTC at yields of 25% or more, even though the tax bite cut into our income. We did this partly because we thought the prepayment models based on history weren’t predicting what we felt was about to happen.
Mortgage rates were coming down, and the process of refinancing mortgages had also changed. Automated processing and standardization of documents had reduced the costs and hassle of refinancing. You couldn’t go to a party in Manhattan or Chicago (or lots of other places) without hearing people talk about how they had already refinanced their mortgages twice.
Most importantly, securitization had given banks a way to generate income from mortgage origination through the capital markets rather than through the fees paid by borrowers. If mortgage rates dropped to 7%, someone with an 8% mortgage could refinance into a 7.5% mortgage without paying any “points,” because banks and mortgage brokers were getting their 2% commissions when the capital markets paid 102% to buy those 7.5% loans to securitize.
At Daiwa, we thought it was likely that prepayments would come in twice as fast as the models were predicting. That made the residual tranches of CMO deals very attractive. The amount of time we would pay taxes on phantom income would be drastically cut, and the day we got the benefit of phantom losses would come much sooner than the rest of the market realized.
Historically, we had never seen whole classes of GNMA mortgages pay off any faster than 35% per year. Since the prepayment models were based on historical data, 35% was the fastest prepayment rate the models could predict.
In the new mortgage environment, we thought mortgage pools could actually pay off at 40%, 50% or even 60% per year. Our CMO residuals were likely to get to the point where they would be tax-advantaged much quicker than the market predicted. The only real risk to our investment was if the mortgages paid off so rapidly that we only got to enjoy our tax-advantaged cash flow for a short time.
Our solution to this was to approach the big Wall Street firms and offer to buy PO (Principal Only) bonds from them. Our competitors still trusted in the prepayment models which we felt were now obsolete. Knowing we were hungry buyers, they pushed the price up, and thought they had taken advantage of us.
I had designed the “super PO” bond four years earlier to help investors exposed to prepayment risk hedge out that exposure. This time, we needed the hedge ourselves. Once we owned lots of PO’s bought from other dealers, we did a series of CMO deals to create super PO bonds we could retain. We sold bonds that had fixed schedules amounting to two-thirds or three-fourths of the total PO bond principal, and kept a bond from each deal that would have a huge payoff if prepayments came in very fast.
The super PO bonds we kept helped us cover the potential loss of future cash flows in the CMO residuals. If the mortgages backing the CMO residuals only paid off a little faster than expected, we still got our 20% or higher return, and we still got a partially tax-sheltered cash flow from those residuals a year or two after buying them.
Taken together in a portfolio, these two basic types of investments performed very well if rates dropped, regardless of whether prepayments increased a little or a lot. That single trade idea was a big part of our $400 million in profit over the next two years, which came from a capital base of only $50 million.
When we tie ourselves to statistical models based on past performance, we can take unexpected losses or make unexpected gains when conditions that drive performance go outside the boundaries that existed in the past.
In 1990, automated underwriting, standardized documentation and very active capital markets made it much less expensive to process and fund mortgage refinancing. These changes rendered our historically based prepayment models obsolete, virtually overnight.
Predicting the shift in behavior due to radically changed market conditions or market operations can be a very rewarding (albeit risky) speculation. Given that the majority of the market is discouraged or even prohibited from speculation, episodes like these can be extraordinarily profitable for those few who can “throw away” history and bet against the crowd.