The Federal Reserve's Large-Scale Asset Purchase Programmes: Rationale and Effects

The paper asks through which channels Federal Reserve large-scale asset purchases (LSAPs) lower longer-term US Treasury yields, and how large each channel’s contribution is, estimating all channels jointly within a single unified empirical framework rather than testing them one at a time. It uses CUSIP-level (individual-security) Treasury data to construct maturity-specific measures of scarcity and aggregate duration risk, and regresses yields and their term-premium components on those measures over a pre-LSAP sample, then applies the estimated coefficients to the actual scale of the LSAP programmes to back out dollar-for-basis-point effects.

The paper considers an expectations/signalling channel, a scarcity (preferred-habitat) channel, and a duration channel. The expectations/signalling channel holds that LSAPs convey information about the future path of short-term policy rates, changing only the expectations component of long rates — it relies entirely on the expectations hypothesis and requires no imperfect asset substitution. The scarcity (preferred-habitat) channel, building on Culbertson (1957), Modigliani and Sutch (1966, 1967), Wallace (1967), and formalized by Vayanos and Vila (2009), holds that some investors have preferred maturities, so a Fed purchase at a specific maturity reduces the local supply available to private investors and depresses yields at nearby maturities via imperfect substitution; the authors note that this departure from the representative-agent framework is what overturns the Eggertsson and Woodford (2003) result that monetary policy’s effect on longer-term rates is limited to the expectations channel. The duration channel, also grounded in Vayanos-Vila (2009), holds that removing aggregate duration risk from the market lowers the term premium borne by risk-averse arbitrageurs, generating yield effects across much of the maturity spectrum rather than only at the purchased maturity.

On August 10, 2010, the FOMC’s 2:15pm announcement of agency-payment reinvestment raised prices roughly 0.6% across four 30-year bonds of differing remaining maturity (about 9.5, 10.0, 14.3, and 14.5 years) in a roughly equal way, consistent with an expectations/duration effect common across maturities; the FRBNY’s 2:45pm follow-up specifying that purchases would concentrate in the 2-10 year sector then reversed only about 20% of the earlier gains for bonds of 10 years or less but about 65-66% of the gains for bonds of more than 14 years. The authors interpret the contrasting size of the reversal across the two maturity groups as suggesting that “roughly two-thirds of the decline in the 14-year Treasury yield is attributable to the anticipation of a reduction in supply around that maturity” — i.e., the scarcity channel — since only the longer-maturity bonds, which were excluded from the announced 2-10 year purchase concentration, gave back most of their initial gain.

The analysis uses weekly CUSIP-level Treasury security data — total amount outstanding, System Open Market Account (SOMA) holdings, and cumulative Treasury buybacks, all at par value — over December 2002 to October 2008, with privately held debt in each CUSIP computed as total outstanding minus SOMA holdings minus buybacks. From this the authors build PHNT (privately held nominal Treasuries as a fraction of total Treasury debt outstanding, in a given maturity bucket), ADR (aggregate duration risk, the PHNT-weighted average modified duration across all CUSIPs), and DG (a duration gap equal to ADR minus the duration of the on-the-run 10-year note, constructed to mitigate simultaneity from the mechanical dependence of duration on yields). The baseline specification regresses each maturity’s yield on PHNT for the relevant maturity bucket (2-10 years for yields of 10 years or less; 14-30 years for longer yields), the duration gap DG, and a lagged yield-curve-slope control, estimated with Newey-West standard errors (4-week window); both PHNT and DG come in positive and statistically significant for maturities from 7 to 30 years, with adjusted R-squared of 0.46-0.70. The pre-LSAP sample is a deliberate choice: including the LSAP period itself would introduce endogeneity, because the Fed tended to purchase securities precisely when yields were rising, which would spuriously reverse the expected sign of the scarcity-yield relationship.

Decomposing yields into a nominal term premium (TP), a real term premium (RTP), and an inflation risk premium (IRP) using an affine term-structure model augmented with TIPS, the authors find both PHNT and DG positive and statistically significant for TP and RTP at all maturities, while for the IRP only DG is consistently significant. At the 10-year maturity, the PHNT coefficient on the nominal term premium is 4.34 (s.e. 1.48) and the DG coefficient is 123.47 (s.e. 19.41); on the real term premium, PHNT is 3.63 (s.e. 0.92) and DG is 99.93 (s.e. 12.06); on the inflation risk premium, PHNT is an insignificant 0.71 (s.e. 0.75) while DG is 23.54 (s.e. 12.16). The authors note that “the bulk of the response is in the RTP component — as one would expect if preferred-habitat mechanisms are operative,” since preferred-habitat theory is a real, not nominal, phenomenon.

Yes — the PHNT and DG coefficients remain positive and statistically significant after adding controls for Treasury option-implied volatility, a flight-to-quality proxy (the weekly average intraday correlation between the 10-year yield change and S&P 500 returns), and the Aruoba-Diebold-Scotti (2009) business-conditions index. Among these additional controls, the flight-to-quality proxy has the highest explanatory power, but including it does not render either PHNT or DG insignificant. The authors’ preferred specification adds the swap-yield curve slope and the flight-to-quality proxy as controls; in that specification the PHNT coefficient on the nominal 10-year term premium is 5.79 (s.e. 1.45) and the DG coefficient is 107.43 (s.e. 15.84).

Applying the preferred-specification coefficients to the pre-LSAP relationship between a 1% change in PHNT and yields (roughly a 5-basis-point yield change per 1% of PHNT, where 1% of PHNT corresponded to about $64 billion at the end of the first LSAP), the authors estimate the first LSAP ($300 billion, concentrated in the 2-10 year sector) produced a scarcity effect of about 23 basis points for the 5-year yield and about 20 basis points for the 10-year yield, plus a duration effect of about 12 basis points from a 0.12-year reduction in average duration — a total effect of about 35 basis points. The second LSAP ($600 billion, also concentrated in the 2-10 year sector, at roughly $86 billion per 1% of PHNT) is estimated to have produced a scarcity effect of about 35 basis points and a duration effect of about 10 basis points (from a smaller 0.1-year duration reduction), for a total of about 45 basis points. Using the rule of thumb from Bernanke (2011) and Chung et al. (2012) that a 25-basis-point change in long-term rates is on average associated with a roughly 100-basis-point change in the federal funds rate, the authors describe the first LSAP as tantamount to a federal-funds-rate cut of about 140 basis points and the second LSAP as tantamount to a cut of about 180 basis points.

No — the authors state that their empirical results suggest LSAPs “do not operate solely or even primarily via the expectations channel.” The scarcity and duration coefficients remain statistically significant even after controlling for proxies of the expected path of short-term rates (the slope of the yield curve and swap rates), and this finding holds over a sample period that predates the zero lower bound. The authors conclude that “preferred-habitat elements appear to be a necessary model ingredient in obtaining a better understanding of monetary policy transmission even when short-term rates are away from the zero lower bound,” and argue more broadly that macroeconomic models in which LSAPs matter for long-term rates only by signalling future short-rate policy “do not adequately encompass the effects of LSAPs.”

The headline LSAP effect estimates are out-of-sample extrapolations: the underlying coefficients are estimated on pre-LSAP data (December 2002-October 2008) and then applied to the very different LSAP period, so any structural change in transmission during the financial crisis would not be captured. The pre-LSAP restriction is itself necessitated by endogeneity — PHNT and yields would be spuriously negatively correlated within the LSAP sample because the Fed bought securities during episodes of rising yields. CUSIP-level SOMA data are available only from December 2002, so the estimation window is short and spans only one full tightening cycle (2004-2006) and one easing cycle (2007-2008). The quantitative results are also scoped to nominal Treasury securities only; the paper does not directly quantify the separate effects of the agency debt and MBS purchased in the first LSAP round. Finally, the authors caution against directly comparing their duration-gap coefficients to those in Greenwood and Vayanos (2010) or Hamilton and Wu (2012), since those papers construct their duration variables differently (a level rather than a gap measure).