Comparison of Lewtan's ABSNet Loan HomeVal with FHFA and Case-Shiller 20 Methods

Lewtan's ABSNet Loan HomeVal property valuations provide quite stark valuation differences on underlying properties of RMBS transactions relative to Federal Housing Finance Agency (FHFA), Case-Shiller 20 (CS20), and combinations of the two.
FHFA and CS20 generally tend to overvalue properties thereby resulting in undervaluation of RMBS deal weighted-average loan-to-value ratios (WALTVs).
As the government moratorium on foreclosures is lifted the borrower nonperformance pipeline will bulge still more thereby further depressing home prices in the not too distant future. If potential losses, as reflected through the FHFA and CS20 HPI, are already underestimated, this will further exacerbate loss estimation error for bondholders.
Lewtan's ABSNet Loan Homeval module, however, uses market-leading automated valuation models (AVMs) and all-sales (versus matched-pair repeat-sale only) time-adjust indexes to estimate current property values in RMBS deals. AVMs target the specific characteristics of the subject property and time-adjust indexes reach as far down as the neighborhood level of granularity.
Side-by-side comparisons of current HomeVal estimates with FHFA, CS20, and combinations of FHFA and CS20 are significant and informative. First, deal WALTVs for 748 ABSNet Loan first lien, securitized deals (comprising 348,992 active loans) are compared in X-Y scatter-plots.
Next a simple linear regression is used to illustrate the relationship between Homeval AVMs and actual sales prices of properties. A final example involves a comparison of the differential impact on bond performance associated with FHFA versus HomeVal current property value estimates and expected losses on loans in default.
Figure 1 (see next page) compares July 2010 reporting period deal WALTVs for all 748 deals based on five different property valuation approaches. The figure 1-A chart compares HomeVal property value estimates with those based solely on FHFA. In this approach, when a property does not map to one of the FHFA MSAs, the appropriate non-MSA State index is applied.
The figure 1-B chart compares HomeVal with a so-called pure CS method (pure means that only those properties which geographically map to one of the CS20 indexes were used to calculate a deal WALTV).
The figure 1-C chart compares HomeVal to an FHFA-CS blend approach. This blend results from applying the CS20 HPI to the loans in a deal which map to a CS20 MSA and for the loans which do not map to a CS20, the appropriate FHFA MSA (or non-MSA) HPI is applied.
The figure 1-D chart compares HomeVal and an FHFA-CS regression model approach. A July 2010 data property value simple linear regression was estimated between loans with a CS value and an FHFA value. The vast majority of RMBS deal loans map to one of the many FHFA indexes, while only a rather small fraction map to one of the CS20. This equation is used to estimate a CS value for those loans which do not map to a CS20 index.
The figure 1-E chart compares WALTVs based on HomeVal AVMs and ZIP-Code level all-sales indexes. Actual sales prices are the dependent variable in AVMs, hence the tighter relation.
The red "1:1" line in each chart is used to illustrate a perfect association between valuation approaches. The FHFA (top left) and FHFA-CS Blend (top right) approaches exhibit the greatest differences, particularly moving left to right along the x-axis. The worst deal in the top left chart has a HomeVal WALTV of over 180, while its FHFA analog is just over 120. FHFA indexes generally overvalue underlying properties thereby undervaluing deal WALTVs.
Later we will see that this sixty point difference can have a significant affect with respect to bond performance expectations.
Can AVMs provide additional accuracy in estimating potential liquidation losses? We have addressed this question by comparing the actual sales prices of homes with their AVM estimates at the time of sale. Figure 2 illustrates the results.
The data underlying Figure 2 are based on recent actual liquidation sales roughly covering the time period of April 2010 through July 2010. The data set is made up of 19,928 loans from ABSNet Loan which liquidated from securitized deals during this period. Additionally, only those properties liquidated and valued at less than one million dollars were included.
In the statistical sense, the relationship between the sold-date AVMs (as used in HomeVal) and actual sales prices is fairly amazing. The regression R2 is 95.95, the slope coefficient is 1.074 with a t-statistic of 687.218 (the y-intercept has been constrained to zero). A regression with a slope coefficient of 1 and an R2 of 100 would symbolize a perfect 1:1 linear fit. In the world of econometric estimation, this result is unusual, powerful and informative.
In passing, liquidation date property values also were estimated using FHFA, CS, FHFA-CS Blend and FHFA-CS Model approaches. In every case, the constrained regression results using HomeVal AVMs on actual sold prices were superior in terms of goodness-of-fit (R2) and the extent to which the slope coefficient exceeded a value of one (i.e., 1.0). In other words, the HomeVal AVM versus sold price model was the closest to a perfect 1:1 linear fit of all approaches.
Finally we evaluate the differential impacts of liquidation loss implications associated with different property valuation approaches on bond yields and weighted average lives (WAL).
The Lewtan ticker for the worst deal (in terms of difference between WALTVs) in the top left chart of Figure 1 is BS.ABST.07.AQ1 (the Bloomberg ticker is BSABS 2007-AQ1 and a reference CUSIP is 07389VAA5). This is the deal with a HomeVal WALTV of just over 180 and an FHFA WALTV of just over 120.
For the July 2010 reporting period, 367 of the active 653 loans in this deal were in default. Default is here defined as any loan that is either in FC, or REO, or is at least 60 days delinquent, regardless of whether it is in FC or REO. Of this set of 367 loans, 324 are upside-down (i.e., current LTV>100) per HomeVal property valuations. Additionally, for this set of 324 upside-down borrowers, the aggregate difference between the outstanding principal balances and HomeVal property values is about -$43.8 million. This is the HomeVal estimate of the total dollar amount defaulted loans in this deal are upside-down. For example, if a borrower has an outstanding loan balance of $233,137.95 and an estimated property value of $203,000 (a current LTV of 1.15), then bond-holders are exposed to a potential liquidation loss of -$30,137.95, assuming the full $203,000 is recoverable (see Figure 2).
Conversely, the FHFA valuation approach shows there to be 290 upside-down borrowers, where the total FHFA-estimated dollar amount of default loss risk is -$20.5 million. For this particular deal, then, the implied principal losses based on HomeVal are about double that of FHFA. This is a quantified measure of the extent to which FHFA overvalues this deal's defaulted properties relative to HomeVal.
Table 1 gives times coverage ratios based on Bloomberg's pipeline delinquency method (or what was their method some 4 to 5 years ago), HomeVal, and FHFA methods. Times coverage is here defined as the dollar amount of current credit support divided by the dollar amount of current projected losses. The loss projection used for HomeVal and FHFA is the total dollar amount by which defaulted loans in a deal are upside-down.
When these HomeVal and FHFA default loss estimates, and these loss estimates only (with no prepay, default or other assumptions on any other loans), are fed into a cash-flow model for the deal, such that all losses are realized in month 1 (an unrealistic abstraction, for sure, but one which greatly simplifies and brings to full relief the point), recoveries occur in month 12, and no advancing, the results are more or less what should be expected. The senior tranches (A1, A2 and A3) pay down without or with ever so slight losses so the bond yields and WALs are comparable between the FHFA and HomeVal methods. Excess interest spread also is used to pay down principal, so it functions alongside subordination as a form of credit protection.
However, expected yields to maturity and weighted average lives on the mezzanine bonds, particularly the M1 and M2, can show vast differences due to the differential loss estimates. For example, the M1 would be completely written off based on the HomeVal loss projections, with a yield of -185.7% and a WAL of 1.07 years. Conversely, the M1 would suffer no write-downs under FHFA, would generate a 3.62% yield with a WAL of about 22.9 years.
Similarly, under HomeVal the M2 would be completely written down, with a yield of -184.4% and WAL of 1.07 years. But under FHFA, about 60% of its principal would be written down, its yield would be about 0.295%, and its WAL would be about 12.7 years. The M3 and M4 would be completely wiped out under both loss methods, and would therefore render identical bond performance statistics results. At least for this BS.ABST.07.AQ1 deal, these extensively different loss estimates have significant implications for remaining support-bond cash flows and bond performance statistics.
To summarize, ABSNet Loan HomeVal property valuations provide substantial valuation differences on underlying properties of RMBS transactions relative to FHFA and CS20 and combinations of the two.
The popular FHFA and CS20 indexes generally tend to overvalue properties while simultaneously undervaluing deal WALTVs.
AVMs on liquidated loans compare favorably with actual sales prices. And differing deal loss projections between valuation methods imply substantially different impacts on bond cash flows and, therefore, bond performance statistics.