The financial market effects of unwinding the Federal Reserve's balance sheet

The paper asks whether unwinding the Federal Reserve’s balance sheet simply reverses quantitative easing (QE), or whether balance-sheet reduction (“QT”) produces financial-market effects through different channels than expansion did. It restricts attention to financial-market effects (yields, spreads, financial conditions) rather than macroeconomic outcomes like output or inflation, uses only U.S. data, and studies the Fed’s 2014-2019 normalization episode with the 2013 taper episode as a comparison point. The analysis is purely empirical – a high-frequency event study combined with structural VAR analysis – with no DSGE model (Abstract; Section 1, p. 1).

The paper splits the normalization episode into “Full Reinvestment” (September 2014-September 2017), when the Fed fully reinvested maturing securities so reserves fell without a commensurate decline in overall balance-sheet size, and “Asset Runoff” (October 2017-August 2019), when a pre-specified monthly cap let securities mature without replacement, so reserves and SOMA holdings both declined by roughly $650 billion. This split is not incidental – it structures nearly every empirical exercise in the paper: the reduced-form liquidity-effect windows, the taper-vs-unwind announcement classification, and the before/after comparison in the TVP-VAR (Section 1, pp. 2-3; Section 2, p. 3).

Regressing the federal-funds/IOR spread on log reserves, the estimated coefficient is -0.086 (s.e. 0.009) during the 2009Q1-2014Q3 expansion versus -0.299 (s.e. 0.015) during the 2014Q4-2019Q3 normalization – the magnitude of the liquidity effect roughly triples (in fact, is about three-and-a-half times larger) during QT relative to QE, visible in a 208-week rolling regression that steepens from 2017 onward (Section 2, p. 3; Fig. 1). The authors attribute the asymmetry to the zero lower bound: during QE, the ZLB placed a natural floor on how far the FF-IOR spread could fall as reserves grew, but no symmetric ceiling constrained the spread from rising as reserves fell during QT; the decline also started from a lower base level of reserves, amplifying the reserve-to-rate pass-through (Section 2, p. 4).

Across a two-day window around 9 QE announcement dates versus 11 QT announcement dates (sample: January 2008-December 2018, 4,016 daily observations), QE produced large, statistically significant cumulative declines (-1.49pp on the 10-year Treasury, -3.70pp on MBS, -1.00pp on 8-quarter Eurodollar futures), while QT produced no significant cumulative change in any of the three (0.28, 0.46, and 0.41pp respectively, none significant). Splitting the 11 QT dates further, the 2 taper announcements (Bernanke’s May and June 2013 remarks) had statistically significant effects across the yield curve (e.g., 0.29pp on the 10-year, significant at 1%) and across MBS, BBB corporate, the dollar index, equities, and Eurodollar futures, while the 9 unwind announcements (May 2014-September 2017) were insignificant across every one of these series (Tables 1, 4, 5, pp. 6-8).

The paper attributes the asymmetry to three factors: the FOMC deliberately framed balance-sheet normalization as separate from the interest-rate path (Yellen’s 2017 remark that it would be “like watching paint dry”), communication around the unwind was vague (the June 2017 Policy Normalization Principles and Plans said only that holdings would “decline in a gradual and predictable manner”), and QT itself was largely anticipated since Bernanke had signaled eventual unwinding as early as 2008, leaving less news content in the actual announcements. Consistent with this, the 0.32pp significant upward revision to 8-quarter Eurodollar futures around taper announcements reflects revised policy-rate expectations that the unwind announcements did not trigger (Section 3, pp. 8-9).

The SVAR imposes a block lower-triangular impact matrix that orders reserves first among the five variables (log reserves, the SOFR-IOR spread, the FF-IOR spread, a financial-conditions or asset-price variable, and log SOMA holdings), so that innovations to reserves are interpreted as “reserve supply shocks” originating from the Fed’s balance-sheet unwind rather than from Treasury bill issuance. The ordering rests on an institutional argument: the Fed did not adjust reserves week-to-week in response to money-market conditions, making weekly reserve supply effectively inelastic, and reserve changes from Treasury bill auctions settle with a one-week delay rather than contemporaneously, so same-week reserve innovations cannot reflect bill-supply effects (Section 4.1, pp. 11-12).

The Bayesian TVP-VAR (estimated October 2015-October 2019) finds that negative reserve-supply shocks produced no significant response in the Goldman Sachs Financial Conditions Index during Full Reinvestment, but a significant tightening response that becomes significant in early 2018 and persists through 2019 during Asset Runoff; over the same comparison, the FF-IOR spread’s response to reserve shocks increases roughly threefold in magnitude, with non-overlapping error bands before and after late 2018 indicating the difference is statistically meaningful; and the SOMA-holdings response goes from small and imprecisely estimated during Full Reinvestment to economically important and precisely estimated during Asset Runoff. The congruence between the reserve, spread, and SOMA responses supports interpreting the shock as reflecting genuine balance-sheet reduction rather than noise (Section 4.3, pp. 14-16; Figs. 6-8).

Estimating a separate constant-parameter Bayesian SVAR for each asset over the Asset Runoff sample (October 2017-August 2019), a negative one-standard-deviation reserve-supply shock raises the 10-year Treasury yield by roughly 4 basis points within the first 2-4 weeks, and the VAR-implied term premium (shown alongside the yield response) absorbs nearly all of that increase – suggesting the rise is driven by a term-premium increase rather than by higher expected short-term rates. Other assets show a similar pattern of near-term tightening: the repo spread (SOFR-IOR) peaks at about +4bp at week 1 before decaying to about +1bp by week 4; the FF-IOR spread peaks at about +0.4bp; the GS-FCI peaks at about +0.05 index units around weeks 6-8 and is statistically significant; BBB corporate yields peak at about +3bp and decay more slowly, remaining near +2bp at 28 weeks; MBS yields peak at about +4bp similarly to Treasuries; the broad dollar index peaks at about +0.3% around week 8 and is more persistent; and the S&P 500 shows only a small (about -0.5% peak), imprecisely estimated, statistically insignificant decline (Section 5, pp. 17-19; Fig. 9).

A counterfactual simulation comparing realized reserve-supply shocks to a scenario of continued full reinvestment implies a peak effect of approximately 40 basis points on the 10-year Treasury yield (occurring in mid-2018) and an average effect of approximately 8 basis points over the entire Asset Runoff period (October 2017-August 2019) – a magnitude the authors note, via a Williams (2014) conversion, is comparable to roughly a 40 basis point change in the federal funds rate (Section 5, p. 18). The paper flags several limitations: the sample deliberately excludes the September 2019 repo-rate spike due to endogeneity concerns about aggregate reserve dynamics, so results may understate the ultimate tightening as reserves approached their lower bound; Treasury bills outstanding rose nearly 30% during Asset Runoff, a potential confounder that the reserves-first VAR ordering addresses only via the institutional settlement-timing argument rather than an exogenous instrument; the results are U.S.-specific and the authors caution against directly transferring them given “important differences in the structure and global role of financial markets across economies”; the paper studies only financial-market effects, leaving real-economy and inflation transmission to future work; and because QT was not entirely unexpected, the event-study estimates may understate true announcement effects due to anticipation (Section 4.1, p. 12; Section 3, p. 9, footnote 4; Section 6, p. 20).