History vs. Data

Krugman’s latest on Reinhart and Rogoff, which makes the rather obvious point that the meaning of the data is illuminated by looking at it on a country-by-country basis and by acknowledging country specific circumstances, raises a fundamental question for the social sciences: Whether the emphasis on data analysis is undermining the quality of social science research.

The biggest problem with data analysis is that it limits the discourse to things about which we have data series. Important questions are not addressed, because like the drunk looking for his keys, researchers are disincentivized from spending their time on them in the absence of good data.

The Reinhart Rogoff kerfluffle focuses attention on a secondary problem with data analysis: that it’s all about the underlying model, and typically represents a very narrow test built on very strong assumptions.  Of course, the best data analysts understand this and some work genuinely addresses these concerns.  However, the way data analysis is consumed and the approaches taken by average data analysts often fail to reflect how very little we learn from data analysis — in no small part because an argument is always more convincing if you deflect attention from its weaknesses rather than exposing those weaknesses to open debate.

For example, there’s nothing wrong with treating each high debt episode as a single event whether it comprises 19 years or 1, after all data analysis is all about heroic assumptions. There is, however, arguably an issue with doing this and not making it crystal clear — at least in a footnote or appendix — what your heroic assumptions are.  (Reinhart and Rogoff’s critics allege that they failed to do this.) Afterall, understanding the heroic assumptions is essential to interpreting the research.

If it is the norm — and I don’t read enough of this literature carefully to know — to hide heroic assumptions and force other researchers to learn about them by reviewing your data set, then the idea that any conclusions drawn from any of this literature analyzing data are worth paying attention to is thrown into doubt.

Data analysis is a useful tool, but should be considered a secondary backup to broader, more nuanced historical analysis that goes far beyond the inherently simplistic nature of data analysis. Narrative historical approaches are currently undervalued, and it is often argued that their weakness that they are not disciplined by formal models. It is time to recognize, first, that this is also the strength of historical analysis and, second, that the fact that data analysis is disciplined by formal models is also a weakness.

Collateral damage from asset price support?

Some stylized facts:

In recent years the Fed has relied on asset price bubbles to support the economy — and keep unemployment down

Over the past few decades a shrinking fraction of economic gains has gone to labor (and a greater fraction to capital)

This was the era of the subprime loan (p. 13ff in link), where lenders put increasing value on the opportunity to take back the underlying asset and less on the borrower’s income stream


Is Federal Reserve asset price support distorting the fundamental structure of economic contracts by undervaluing labor and overvaluing capital?

Which raises another question: should the objective of monetary policy focus not on low inflation and low unemployment, but on a slowly increasing median wage and low unemployment?

HAMP loans were designed to redefault [Updated]

From the most recent SIGTARP report:

SIGTARP is concerned that the number of homeowners who have redefaulted on a HAMP permanent mortgage modification is increasing at an alarming rate. Treasury’s data shows that the longer a homeowner remains in HAMP, the more likely he or she is to redefault out of the program. As of March 31, 2013, the oldest HAMP permanent modifications, from the third and fourth quarter of 2009, are redefaulting at a rate of 46.1% and 39.1%. HAMP permanent modifications from 2010 also had high redefault rates, ranging from 28.9% to 37.6%.

But, of course, redefaults (at least in the absence of a flourishing economic recovery that raises incomes by 3% a year) are built into the HAMP modification program, because payments, that were deliberately set at the outer limit of affordability, increase in each of the sixth, seventh, and often eighth years by 8-9%. While the first HAMP payment increases will not kick in until 2014, borrowers who know their payments are going to go up unsustainably in a matter of months (and remember every single one of them has been through this before) are incentivized to default today.

SIGTARP failed to do its homework on HAMP if it’s only realizing now that large numbers of redefaults will be a result of HAMP modifications. Why were redefaults built in to the program? Presumably because any other policy would have had too harsh an effect on investors and the banks for them to be willing to participate voluntarily in the program.  Like a subprime mortgage the HAMP program has payment increases that induce default and give the lenders the option (p. 13 ff. in link) of taking the house or refinancing the mortgage depending on the relative performance of the economy and the housing market.

For those who don’t understand how HAMP is structured, this is the (Tier 1) program:

Calculate 31% of the borrower’s monthly income.  This is the target monthly mortgage payment or TMMP and under almost no circumstances should the monthly payment in a HAMP modification fall below this amount.

A step 1 modification:  (Applies mostly to borrowers who fell behind due to temporary misfortune.)  Capitalize any delinquent amounts and fees into the principal of the loan (=the new balance) and apply the existing interest rate and loan terms to the new balance.  If this requires a payment in excess of TMMP go to step 2.  (Less than 4% of permanent mods. See MHA report p. 6)

A step 2 modification:  Determine the interest rate at which the new balance can be amortized given payments of TMMP over the existing term of the loan.  If the interest rate is 2% or higher, offer a HAMP loan with an initial interest rate at that level to the borrower.  If not go to step three.  (Approximately 35% of permanent mods.)

A step 3 modification:  Determine the number of months over which payments of TMMP will amortize the new balance at 2% interest.  If this is less than 480 months, extend the maturity of the loan to that length and offer a HAMP increasing payment loan with an initial interest rate of 2%.  If not, go to step 4. (Approximately 29% of permanent mods.)

A step 4 modification:  Calculate the principal amount of a 2% fixed rate loan amortizing over 480 months that the borrower can afford given TMMP.  Subtract this amount from the new balance.  Do a net present value analysis to determine whether adding the remainder as a balloon payment due at the end of the 40 year loan is more valuable to the lender than foreclosing.  (32.6% of permanent mods.)

The terms of the HAMP loan for the borrower:

HAMP step 3 and step 4 (and to a lesser degree step 2) loans have interest rates that are fixed only for 5 years.  (According to SIGTARP: “Treasury designed the program so that homeowners could keep their modifications for up to five years.” p. 179.) The interest rate then climbs by 1% a year until it reaches the market rate for a fixed rate loan at the date the modification was made.  As a result 2009 vintage HAMP loans have interest rates that increase to 5%.  More recent HAMP mods have interest rates that increase to as little as 3.5%.

Thus in the sixth year for the majority of HAMP mods the interest rate at which the principal balance (excluding the step 4 balloon) is amortized increases by 1%, so the payment increases.  Most HAMP mods have another increase in the seventh year and the earliest HAMP mods have an increase in the eighth year.  In the sixth year the payment increases by more than 9%, in the seventh year most payments increase by almost 9% and there are many loans that have payments that increase by 8% in the eighth year.  The total payment increase from the end of the fifth year to the start of the eighth year for early vintage HAMP loans is 29%.

Given that most early vintage HAMP loans have a built in payment increase of 29%, it is far from clear how anyone who understands the nature of the HAMP program can fail to predict increasing redefaults as the program progresses.

Update 4-26-13: In case anybody thinks it’s a mistake to analyze HAMP in nominal terms and implicitly argues that a 3% nominal growth rate for income was a reasonable assumption for Treasury to make in 2009, I have to say that it’s precisely this assumption that I find so outrageous. HAMP appears to have been built on the view that a recovery was just around the corner and a rejection of the endogeneity of government policy to economic performance.

Using census data we find that nominal median income increased from 2001 to 2008 by 2.4% per annum.  In 2009, it was completely unreasonable to imagine that this performance was likely to continue over the next few years.  A financial crisis had taken place, aggressively addressing it by resolving our largest banks was likely to cause greater economic distress in the short run and failing to address it was all but certain to cause economic stagnation for years. Furthermore, in order to set HAMP up to succeed it was incumbent on the decision makers to make a conservative prediction about future economic performance.  In fact, nominal median income increased by 0.3% from 2009 to 2011 and I believe that the likelihood of such weak performance should have been built into the HAMP program.