Discussion of Self-fulfilling Runs: Evidence from the U.S. Life Insurance Industry Amir Sufi Chicago Booth September 27, 2015
Big Picture Question liquid liabilities are potentially vulnerable to swift changes in investors beliefs about the actions of other investors when investors withdraw based on their beliefs and their action leads other investors to withdraw, then the original belief is verified and a self-fulfilling run has occurred such a run is in contrast to a fundamental-based run Is it possible to empirically demonstrate that an investor s decision to liquidate, conditional on fundamentals, is a direct function of beliefs about other investors liquidation decisions?
Institutional setting: XFABN securities St+1 [0, REijt+1] Figure 3: Timeline for XFABN elections Dιt Dιt+1 t + m t + m + 1 Q t St+1 Q t Q + t Q t+1 Dιt Current extension decision t + m Maturity date of Dιt spinoff Dιt+1 Next extension decision t + m + 1 Maturity date of Dιt+1 spinoff St+1 Fraction of other XFABN REijt Fraction of XFABN that are that are spunoff up for election Q t Maturing FABS during [t, t + m] Q + t Other predetermined maturing FABS Q t+1 Maturing FABS during [t + 1, t + m + 1] Q t Maturing FABS before the next election Figure 4: Run on Extendible FABN D ijt = γ 0 + γ 1 S ijt+1 + x jt β + ɛ ijt
Comment 1: We need an instrument! A fundamental shock at t = 0 will affect both S ij1 and D i0, and therefore induce a strong correlation between the two: OLS will be polluted The authors make a valiant attempt to include controls for fundamentals, but investors see more than econometrician sees The authors recognize need for exogenous shifter for S ij1 Explanation of instrument: RE ijt+1 is the maximum fraction of XFABN that can be converted into short-term fixed maturity bonds between an individual XFABN s i s election dates t and t + 1 Plausibly exogenous to fundamentals, but increases risk of run conditional on same fundamental shock (note instrument only works when there is a negative shock)
Variation in instrument Different XFABN for same insurer j have different lengths of election cycles (monthly election is median and most common) XFABN election cycles do not begin immediately after issuance many make investors wait a year or more before elections start Election cycles end before final maturity appendix example has elections ending in June 2016 with final maturity June 2017 My take: This is a really cool instrument, and the authors should be commended for finding this institutional environment that allows us to test such an important idea!
Q t+1 Maturing FABS during [t + 1, t + m + 1] Q t Maturing FABS before the next ele Comment 2: Focus on Summer 2007 Figure 4: Run on Extendible FABN Source: authors calculations based on data collected from Bloomberg Financial LLP.
Comment 2: Focus on Summer 2007 Table 2: Runs on Extendible FABN: Reduced Form Results This table summarizes the main reduced form results on the run on U.S. life insurers that occurred in the summer of 2007. The unit of observation is the election date t of an individual XFABN i issued by insurer j, and the sample extends from January 1, 2005 to December 31, 2010. The dependent variable Dijt is the fraction of XFABN i issued by insurer j that is converted into a fixed maturity bond at election date t. The main explanatory variables are Sijt+1 the fraction of all XFABN from insurer j that is converted between the current election date t and the next election date t + 1, and Qjt the fraction of XFABN from insurer j that were converted prior to election date t. Columns 2 through 7 include insurer fixed effects. Column 3 decomposes Qjt into a most recent and older component Qjt Sijt and Sijt, respectively. Column 4 includes the amount of fixed maturity FABS Q F jt ABS and Q F jt ABS that matures before or on the date at which an XFABN converted at date t is set to come due divided by total FABS. Column 5 includes the VIX and the amount of U.S. ABCP outstanding. Column 6 includes quarterly time fixed effects. Column 7 includes sponsoring insurer stock price, 5-year CDS, and 1-year EDF. Robust standard errors are reported in parentheses. ***, **, and * represent statistical significance at the 1%, 5%, and 10% level, respectively. 40 (1) (2) (3) (4) (5) (6) (7) Dep. var.: Dijt No Insurer Flexible FABS VIX Time Financials control Fixed Effect Queue Rollover risk & ABCP Fixed Effect Health Sijt+1 0.884*** 0.891*** 0.857*** 0.832*** 0.735*** 0.339** 0.476** (0.129) (0.144) (0.150) (0.149) (0.156) (0.158) (0.241) Qjt 0.000607*** 0.000618*** (0.000110) (0.000139) Qjt Sijt 0.000553*** 0.000481*** 0.000308 0.000383* 0.000417 (0.000170) (0.000175) (0.000209) (0.000204) (0.000818) Sijt 0.00921 0.00852 0.00959 0.00756* 0.103 (0.00563) (0.00556) (0.00597) (0.00456) (0.106) Q F jt ABS 0.349*** 0.330*** 0.470*** 0.825* (0.119) (0.116) (0.156) (0.422) Q F jt ABS -0.117-0.131-0.141-0.0808 (0.302) (0.295) (0.293) (0.617) VIX 0.00411*** -0.00428-0.00450 (0.00139) (0.00293) (0.00535) ABCP outstanding (USD bn) 1.75e-05-0.00105*** -0.00124** (3.17e-05) (0.000319) (0.000539) 5-Year CDS Spread (bps) -0.000214 (0.000635) 1-Year EDF (%) -0.00711 (0.106) Stock Price ($) -0.000490 (0.00317) Observations 921 921 921 921 921 921 383 Adjusted R-squared 0.172 0.187 0.191 0.202 0.219 0.300 0.365 FA provider FE N Y Y Y Y Y Y Quarter FE N N N N N Y Y Source: authors calculations based on data collected from Bloomberg Finance LP, Markit and Center for Research in Security Prices (CRSP) via Wharton Research Data Services (WRDS), Moody s Analytics: KMV, Federal Reserve Bank of St Louis, Federal Reserve Economic Data (FRED).
A more transparent specification? Why not focus on summer/fall of 2007, more of a first difference type approach? Ideal experiment: large set of XFABN with election date on say August 3rd, 2007 (D i0 ), with large amount of variation in RE ij1 Advantages of this approach: Focuses on source of variation driving first stage potentially more power Easier to conduct orthogonality tests: correlate REij1 with observables as of 2006, for example Can see results much more clearly in graphs and test for non-linearities
Comment 3: Timing still bothers me Ideal experiment would be several XFABN with election dates on exact same date, with lots of variation in RE ij Problem is, we cannot conduct this experiment because we likely don t have so many XFABN for different insurance companies with elections on exact same date we have a staggered panel data set Given the nature of runs, even days or hours can make a difference best the authors can do is use weekly fixed effects Runs on one insurance company inform investors in other insurance companies Lots of robustness tests on this issue, but I came away from paper still worrying about it
Concluding remarks Overall, I felt this was a really cool paper very fun to read (although institutional detail took some time to understand!) Authors should be highly commended: tons of data collection on a market we don t know much about Even though the particular market may not be that large, it helps us understand shadow banking system in general Jim Poterba once said to me: if you show summary statistics that are completely new to the literature, then you re in good shape I think the authors can do some work to make the empirical strategy more transparent and convincing