United States District Court, S.D. New York
February 13, 2015
FEDERAL HOUSING FINANCE AGENCY, Plaintiff,
NOMURA HOLDING AMERICA, INC., et al., Defendants.
Philippe Z. Selendy, Manisha M. Sheth, Nicholas F. Joseph, Calli Ray, QUINN EMANUEL URQUHART & SULLIVAN, LLP New York, NY, for plaintiff.
David B. Tulchin, Steven L. Holley, Bruce E. Clark, Bradley A. Harsch, Katherine J. Stoller, SULLIVAN & CROMWELL LLP New York, NY, Amanda F. Davidoff, Elizabeth A. Cassacy, SULLIVAN & CROMWELL LLP Washington, DC, Thomas C. Rice, David J. Woll, Andrew T. Frankel, Alan Turner, Craig S. Waldman, SIMPSON THACHER & BARTLETT LLP, New York, NY, for defendants.
MEMORANDUM OPINION & ORDER
DENISE COTE, District Judge.
This Memorandum Opinion addresses defendants' January 8 motion to exclude certain expert testimony to be offered at trial on behalf of plaintiff Federal Housing Finance Agency ("FHFA") by Charles D. Cowan ("Cowan"). FHFA intends Cowan to testify regarding his extrapolations of analyses performed by other FHFA experts, including John A. Kilpatrick ("Kilpatrick").
Familiarity with the background of this case, as described in the Court's January 28 Opinion denying defendants' motion to exclude Kilpatrick's testimony, is assumed, FHFA v. Nomura Holding Am., Inc., No. 11cv6201 (DLC), 2015 WL 353929 (S.D.N.Y. Jan. 28, 2015) ("Kilpatrick Opinion"). All defined terms used in that Opinion carry the same definitions here.
Kilpatrick was retained by FHFA to provide an opinion as to whether the original appraisals used to value the properties and generate LTV ratios associated with the sample of loans at issue here were accurate and credible. FHFA retained Cowan to extrapolate Kilpatrick's results to the population of loans as a whole.
Kilpatrick recalculated appraisal values for 672 loans using an Automated Valuation Model ("Greenfield AVM"). He identified a subset of 208 out of 672 loans that were one standard deviation above the mean of the appraisal values determined when the loans were originated; he labeled these appraisals "inflated" and evaluated them further using a Credibility Assessment Model ("CAM"). See Kilpatrick Opinion, 2015 WL 353929, at *2-3 (describing Kilpatrick's use of the Greenfield AVM and CAM).
After extrapolating from Kipatrick's Greenfield AVM data for all 672 loans using a Monte Carlo simulation, Cowan arrived at an extrapolated AVM estimate ("Average Simulated AVM") and recalculated LTV ratios from the Average Simulated AVM figures for each loan. Using the Average Simulated AVM figures, Cowan identified a subset of 40 out of 672 loans whose individual loan-to-value ratio was "understated" to a statistically significant degree.
Defendants complain about two of Cowan's 43 calculations. They assert that Cowan failed to abide by the results of two individual-level analyses when calculating properties for groups of loans - specifically, when calculating the "Average Inflation Rate" of appraisals (Table 6) and when displaying a breakdown of LTV ratios by range (Table 8 and Chart 1). They argue that, for the Average Inflation Rate, Cowan should only have used an Average Simulated AVM figure for an appraisal if that Average Simulated AVM figure deviated from the original appraisal value by more than one standard deviation. For the LTV ratio comparison, they argue that Cowan should only have used Average Simulated AVM figures for his own subset of 40 and retained the original LTV ratios for the remainder of the loans.
It is telling that defendants have advanced no expert opinion in support of their argument; rather, they have only produced expert calculations conducted at their behest. Their argument seeks to impose limitations on Cowan's statistical analysis that have no basis in statistics. What defendants demand - using Average Simulated AVM values only for subsets while assuming the other original appraisal figures to be accurate - would not preserve the integrity of Cowan's aggregate analysis. It would defeat its purpose.
It is a basic principle of statistics that examining data in the aggregate may permit detection of patterns or relationships not necessarily detectable on an individual basis at statistically significant levels. When an expert undertakes an aggregate analysis, no general rule requires that expert to accept the results of previous individual-level analyses. Defendants' claim regarding Cowan's Table 6 of "Average Inflation Levels" well illustrates the point. Individual appraisals that were not labeled "inflated" by Kilpatrick may still be higher than the appraisal values generated by the Greenfield AVM - and the difference between averages may be statistically significant for the groups. This general tendency is what Cowan was examining. Moreover, the defendants' separate complaint about Cowan's use of the term "inflation" in the title of Table 6 is without merit. The fact that Kilpatrick and Cowan both used the term "inflated" in different contexts is of no consequence.
Defendants' complaint about Cowan's Table 8 and its associated chart displaying ranges of LTV ratios stands on similarly shaky ground. Cowan recalculated LTVs using Average Simulated AVM values for all 672 loans and determined that 40 loans had understated LTV ratios at a 95% confidence level. Then, separately, he tallied up the number of loans with Average Simulated LTV ratios in specific ranges (e.g., "Less than 75, " "Between 75 and 80"), converted those counts to percentages, and displayed them alongside the corresponding percentages from the Prospectus Supplements. The differences are plain: on the population loan tapes, 70% of LTVs had ratios of less than 80, 30% had ratios between 80 and 100, and 0% had ratios of over 100. By contrast, Cowan's sample, using Average Simulated AVM values to calculate LTVs, had 28% of LTVs at 80 or below, 62% of LTVs between 80 and 100, and 20% of LTVs at 100 or above. Cowan's Chart 1 is reproduced below.
Defendants argue that because only 40 of the 672 loans had LTV ratios deemed to be "understated" at a 95% confidence level, Cowan should only have used Average Simulated AVMs for those 40 and the original loan tape appraisal values for the other 632. Otherwise, they say, Table 8 and Chart 1 contradict Cowan's representation that every population-level extrapolation would have that confidence level. Indeed, they argue that Table 8 and Chart 1 currently have "no statistical significance whatsoever."
Cowan's Report does not claim otherwise. The Report makes no claims of population-level statistical significance about Table 8 and Chart 1, which is clearly indicated by the deliberate omission of upper and lower bounds and a margin of error from the table. Indeed, as defendants recognize, Table 8 and Chart 1 reflect only "simple percentage calculations" based on "loan-level averages, which do not permit the construction of confidence intervals for the population." Cowan was only required to use a 95% confidence interval when extrapolating to the population of loans. Apart from the calculation of the individual Average Simulated AVMs, there is no extrapolation involved here - just counting. Not providing a 95% confidence interval where defendants concede it would be impossible to do so is not a departure from his stated standard.
The defendants' January 8 motion to exclude portions of Cowan's testimony is denied.