ABSTRACT
This paper studies how macroeconomic news affected the spreads of Italian sovereign bonds before and during the quantitative easing by the European Central Bank. Daily changes in the bond spreads are regressed on macroeconomic news shocks, where the news shocks are computed as the difference between the published data and the preceding private-sector forecasts. The analysis shows that macroeconomic news shocks had economically and statistically significant effects on Italian bond spreads in 2012–2014 before quantitative easing, while the effects were negligible afterwards although with a partial exception of a period in 2019 when the net asset purchases were paused.
Acknowledgements
The authors would like to thank three anonymous referees for their insightful comments. The authors would also like to thank Michael Ehrmann, Michael Funke, Pavlo Illiashenko, Enrique Neder, Alari Paulus, Tairi Rõõm and participants at numerous presentations for valuable feedback. Stephan Fahr and Liina Kulu helped with the data. Parts of the paper were written while Lenno Uusküla worked for the Bank of Estonia. The views expressed are those of the authors and not necessarily those of Luminor Bank, the Bank of Estonia or the Eurosystem.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Notes
1 Studies have pointed to three key channels through which quantitative easing can alter the expected returns or risk of sovereign bonds and thus their yield (Andrade and Ferroni Citation2021; Krishnamurthy and Vissing-Jorgensen Citation2011). The first is the market stabilization channel, which asserts that the liquidity provided by asset purchases eases illiquidity in financial markets. The second is the portfolio rebalancing channel, which contends that asset purchases may shift investments towards higher-risk assets. The third is the signalling channel, which asserts that asset purchases signal the intention to keep monetary policy accommodating for a period of time.
2 Besides the five shocks used in the paper we also considered including consumer price shocks, but left them out as the expected sign of the shock is ambiguous.
3 The risk of omitted variables bias is limited given that the news shocks depict market surprises and so are unlikely to be correlated with various macroeconomic and financial variables that may be used as control variables.
4 This corresponds to N = 2 in EquationEquation (1)(1)
(1) in Subsection 3.1.
5 Experimenting with the Newey-West correction left the results broadly unchanged.
6 The sum of the coefficients of the five lags of the dependent variable is close to 0. This suggests that the effect on the level of the spread of a news shock is not reversed over time, but remains persistent. This is in line with Altavilla et al (Citation2017, Citation2019). who find that news shocks have long-lasting effects on yields.
7 The data are available on the webpage https://sites.google.com/view/jingcynthiawu/shadow-rates.
8 The unemployment news is excluded because the estimated coefficient in the reference period is negative, while retail sales are excluded as there are some gaps in the data published by Bloomberg.
9 Some studies have examined possible spillover of news shocks from one country to another (Beetsma et al. Citation2013; Van Der Heijden, Beetsma, and Romp Citation2018). We sought to estimate the impact of non-Italian news on the Italian spread, but the non-Italian news contained many missing observations and the results were inconclusive.
10 De Santis (Citation2020) uses market news on quantitative easing to study the effects of quantitative easing on long-term sovereign yields in the euro area. The study finds sizable reductions in spreads before the official announcement was made.
11 This corresponds to N = 1 in EquationEquation (1)(1)
(1) in Subsection 2.1.
12 The abrupt and relatively large changes in the coefficient estimates of AGG5 in the samples containing data for 2013 and 2014 are partly the result of substantial volatility in ΔSPREAD; see . We have experimented with rolling regressions where extreme observations of ΔSPREAD are left out, and this results in less abrupt changes in the coefficient estimates of AGG5.
13 This corresponds to N = 4 in EquationEquation (1)(1)
(1) in Subsection 2.1.