The Effects of Quantitative Easing on Interest Rates: Channels and Implications for Policy

Questions & answers

Q1. What is the paper’s central claim about how quantitative easing affects interest rates?

Quantitative easing does not operate through a single channel that moves all interest rates uniformly; instead it works through several distinct channels that affect particular assets differently, so that the effect on any given asset’s yield depends critically on which assets the central bank buys. The authors frame this as a corrective to the practice of judging QE’s success solely by its effect on Treasury yields: because a purchase of Treasuries and a purchase of MBS or agency debt operate through different channels, their effects on non-Treasury assets (mortgages, corporate bonds) can differ sharply even if their effects on Treasury yields look similar.

Q2. How do the authors decompose a bond’s yield, and what channels does this decomposition motivate?

The paper writes the real yield on a long-term, risky, illiquid asset as the expected safe short rate minus expected inflation, plus separate premia for duration risk, illiquidity, lack of safety, default risk, and prepayment risk. Each term in this decomposition motivates a corresponding channel through which asset purchases could move yields: (1) a signaling channel, in which QE credibly commits the Fed to keep the funds rate low longer (drawing on Eggertsson-Woodford 2003 and Clouse et al. 2000) and works through the expectations hypothesis, larger at intermediate than very long maturities; (2) a duration risk channel (following Vayanos-Vila 2009), in which removing duration from the market lowers the term premium on long-term nominal assets, with a larger effect for longer-duration securities; (3) a liquidity channel, in which QE raises the supply of highly liquid reserves relative to long-term securities, which would actually raise Treasury yields relative to less-liquid assets by reducing Treasuries’ liquidity premium; (4) a safety channel (building on the authors’ own 2010 work), in which a clientele demand for long-term, near-zero-default-risk assets means that QE purchases of safe assets (Treasuries, agencies, high-grade credit) lower their yields relative to less-safe assets, with Baa treated as the natural cutoff; (5) a prepayment risk channel specific to MBS purchases (following Gabaix-Krishnamurthy-Vigneron 2007); (6) a default risk channel, in which QE that stimulates the economy lowers default premia, measured via CDS; and (7) an inflation channel, in which expansionary QE raises inflation expectations (and possibly inflation uncertainty), producing larger declines in real than nominal rates.

Q3. What is the empirical strategy for separating these channels?

The identification strategy combines an event study of QE announcement dates with difference-in-differences comparisons across pairs of assets that share nearly all characteristics but differ along one dimension, supplemented by derivatives instruments that load predominantly on a single channel. For example, a long-term agency bond and a long-term Treasury bond of equal duration, safety, and default risk but differing liquidity isolate the liquidity channel; federal funds futures (3rd/6th/12th/24th-month contracts) proxy the signaling channel; CDS rates by rating proxy default risk; inflation swaps, TIPS, and swaption implied volatility proxy inflation expectations and uncertainty. The event study covers QE1’s five main announcement dates (Nov 25, Dec 1, Dec 16, 2008; Jan 28, Mar 18, 2009 – a subset of the eight dates used in Gagnon et al.) and QE2’s three dates (Aug 10, Sep 21, Nov 3, 2010), using one- and two-day yield changes (two days for less-liquid assets), with intraday Treasury yield and volume data used to confirm the announcements were the dominant news on those days.

Q4. What were the estimated effects of QE1, and which channels were operative?

Summed over the five QE1 event dates using two-day changes, Treasury yields fell 73 bp at 30 years, 107 bp at 10 years, and 74 bp at 5 years, while Fannie Mae agency yields fell by considerably more – 144-200 bp across maturities – and agency MBS yields fell 88-107 bp. The signaling channel is estimated to explain a meaningful share of this: QE1 shifted anticipated rate hikes about 6.3 months later, the 24-month fed funds futures rate fell 40 bp, and the implied signaling effect on longer yields is roughly 20-40 bp out to ten years (about 30 bp at five years), consistent with a CDS-adjusted decline in Baa corporate yields of 41 bp (long) and 25 bp (intermediate). The duration channel is found to work only partly: longer-maturity bonds generally fell more, but duration alone cannot explain why agency bonds fell the most of all instruments, nor the absence of a within-rating maturity effect among corporate bonds. The authors find the safety, prepayment (MBS), default (CDS-adjusted corporate), and inflation channels were all operative during QE1, and describe MBS purchases specifically as “crucial” for the decline in MBS yields and in corporate credit risk, possibly operating through a mortgage-refinancing boom.

Q5. How did QE2’s effects differ from QE1’s, and why?

Because QE2 purchased only Treasuries (no MBS or agencies), its channel profile was narrower: signaling lowered 5-year yields by 11-16 bp and 10-year yields by 7-11 bp, safety lowered 10-year low-default-risk yields by a further 6-11 bp, and 10-year inflation expectations rose an estimated 14-16 bp, causing real rates to fall by more than nominal rates. The authors find no evidence of a prepayment channel (mechanically, since QE2 did not buy MBS), no duration channel (the signaling estimate already accounts for essentially all of the CDS-adjusted corporate bond decline, leaving no room for an additional duration effect), no liquidity channel (the Treasury-agency spread was unchanged, and liquidity premiums were themselves tiny – the 1-week versus 1-month bill spread was only about 2 bp), and no default channel (CDS rates actually rose during QE2, which the authors describe as puzzling and speculate may reflect markets inferring the economy was weaker than previously thought precisely because the Fed felt the need to do QE2). As a result, the authors conclude that QE2’s effect on the mortgage and lower-grade corporate rates most relevant to households “was thus through the signaling and inflation channels, rather than from a portfolio balance effect” (p. 255).

Q6. What does the safety-channel regression add, and how does it compare with the event-study estimates?

Using annual U.S. data from 1949 to 2008 and instrumenting log(long-term Treasury supply/GDP) with total Treasury supply/GDP (and its square and cube), the paper estimates a 2SLS slope of -0.83 (t = -5.83) on the Baa-Treasury spread and -0.32 (t = -3.02) on the Baa-Aaa spread, implying safety effects of roughly 4 bp (Baa-Aaa) to 11 bp (Baa-Treasury, an upper bound) for QE1’s reduction in long-term Treasury supply, and roughly 8-20 bp for QE2’s $511 billion (in ten-year-equivalents). These regression-implied magnitudes are considerably smaller than the safety effect the event study attributes to QE1, and the authors interpret this gap as evidence that QE1’s very large estimated safety effect reflects the unusually elevated, crisis-period demand for safe assets rather than an “average” historical relationship between Treasury supply and credit spreads. They also note that the Baa-Aaa spread likely understates the true safety effect (Aaa corporates are not as safe as Treasuries) while the Baa-Treasury spread likely overstates it (it embeds a liquidity premium), so the true safety effect lies somewhere between the two regression estimates.

Q7. What is the paper’s central mechanism claim, and how does it position the paper relative to other portfolio-balance accounts of QE?

The paper’s key structural claim is that the portfolio-balance channel operating substantially in both QE1 and QE2 works through a preferred-habitat safety channel – clientele demand for long-term, low-default-risk assets of particular maturities – and not through the duration-risk channel: “We do not find support for the operation of the duration risk channel. Instead, the role of duration appears to be through a preferred-habitat demand for particular maturities” (Conclusion, p. 263). This directly contrasts with accounts (such as Gagnon et al. 2011 and D’Amico et al. 2012) that emphasize duration/term-premium compression as the primary portfolio-balance mechanism. Under the authors’ account, QE that purchases safe assets bids up the price of the remaining safe assets and lowers their yields, but this effect does not automatically carry over to less-safe mortgage and lower-grade corporate yields except via the signaling and inflation channels (and, during QE1 specifically, the MBS prepayment and default-risk channels).

Q8. What about QE3, and what caveats do the authors attach to their own estimates?

QE3 (the September 21, 2011 announcement, discussed briefly) is found to operate mainly through an MBS risk-premium/portfolio-balance channel that is smaller than QE1’s (since markets were less stressed at the time), with a negligible signaling channel (24-month futures moved only -1 bp, likely because the August 9, 2011 calendar-based guidance had already used up the available room), while default risk rose and inflation expectations fell. On limitations, the authors caution that event-study estimates “are dependent on the dynamics of expectations through the event,” that there is “no direct way of precisely measuring” the change in QE expectations, and that QE1 occurred under “unusual market conditions, so that it is hard to extrapolate numbers from that period to more normal conditions” (Section IV, p. 257). They also note that omitting true event dates reduces the power of the channel-decomposition tests without necessarily biasing them, but could bias the estimated overall effect of QE either up or down (fn. 5, p. 226).