Dating Structural Changes in UK Monetary Policy

1 Introduction

The importance of having an accurate description of the historical evolution of monetary policy in a country is widely recognised in the macroeconomic literature. In particular, the timing with which this evolution took place is crucial information to be able to compare the performance of alternative monetary institutions, but also to understand the influence monetary policy can have on the economy. For the first objective, indeed, we need accurate dates to assign different periods to specific policy regimes in order to perform counterfactual simulations that highlights the role of a change in the stance of monetary policy. For the second objective, we need the same information to correctly isolate from the systematic policy function the exogenous policy shocks, which allow us to measure the response of the macroeconomic variables of interest to government intervention.

The UK history offers a unique setting to study how different institutional developments have shaped the monetary policy actions. In Table 1 we list a set of historical events that may have had an impact on the systematic policy decisions of the British monetary authorities. Also to reflect their different scope, we distinguish: important institutional events that are very likely to have influenced the overall monetary policy strategy (broad institutional changes); modifications that involved more operational aspects of the monetary policy decisions (technical modifications); and particular circumstances in the world economy (international events). This long list of events highlights one aspect that the existing empirical work has not fully addressed, which is the difficulty in dating, based only on a priori information, the periods in which the systematic actions of the monetary authorities were effectively different.

A great deal of empirical work has used simple policy instrument equations either to assess the importance of specific regime changes over certain historical periods, or to capture a generic and continuous random shift in the policy reaction function (see the literature review in the next section).^[1] Little attention has been devoted to three important aspects that instead we address in this paper. First, it is important to have an accurate estimate of the dates in which an effective change in the stance of monetary policy took place, along with a rigorous statistical assessment of such evidence. This evidence would help us to identify the relative importance of different institutional events in terms of their ability to influence the actual policymaking. Second, it is undoubtedly of interest to identify in a robust fashion which of the macroeconomic priorities of the monetary authorities did effectively shift at each policy regime change. Third, because a simple policy instrument equation is an empirical approximation to a complex decision-making process that makes use of a small number of parameters, we do not know a priori when such an empirical approximation is adequate and when it is not.

To investigate these three aspects of the historical evolution of the monetary policy in the UK, we focus on the systematic actions of the monetary authorities which are typically captured by a policy function describing the endogenous policy response to the dynamics of the economy, and we develop a flexible empirical approach that builds on a set of tools provided by the econometrics of structural stability suitably adapted to our research question. This approach delivers three main empirical findings: 1) many important historical events were not accompanied by a simultaneous shift in policy behaviour; 2) a clear evolution of the priorities of the monetary authorities emerges with inflation goals in particular becoming gradually more important from the 1970s up to early 2000s, and the output gap dominating after the 2008 financial crisis; 3) regime changes intended as shifts in the systematic actions of the monetary authorities explain a considerable share of the observed movements in the official interest rate, while disregarding such regime changes leads to an overestimation of the importance of policy inertia. The first two points in particular suggest that dating policy regimes based on individual institutional events can be misleading, and that instead a better characterization of the UK monetary policy is in terms of evolution in phases with the change in regime determined by the accumulation of many institutional events.

The empirical approach we propose consists of three steps. In the first step, we ask ourselves if there is any substantial evidence of a change in the long-run targets of the policy makers, and in their ability to influence private inflation expectations and short-term market interest rates. To this aim we perform stability tests on the mean of the variables representing the macroeconomic targets and on the term spread that distinguishes a short-term market rate from the policy rate under the direct control of the policy makers. In the second step, we explore the existence of changes in the systematic actions of the monetary authorities. With this objective we perform a comprehensive investigation of the parameter shifts in a standard monetary policy instrument rule, addressing in particular the problems produced by the presence of forward-looking variables and the possibility of structural changes in marginal distributions, two issues that can confound the whole estimation and inference on parameter change (Hansen 2000). To tackle these two issues, we employ the method developed by Hall, Han, and Boldea (2012) that hinges on a careful testing procedure to model multiple breaks at unknown location in the presence of correlation between regressors and the error term. Finally, in the third step, we characterize each estimated regime change by identifying which of the macroeconomic priorities did effectively undergo a revision. Here, we employ a modified version of the testing approach suggested by Inoue and Rossi (2011) to estimate the unstable subset of parameters.

Our approach presents several advantages over existing techniques that have been used to capture regime changes in monetary policy. Firstly, we highlight how treating the break dates as unknown parameters to estimate represents a clear advantage over an analysis that imposes the time location of the regime change based on a priori information. This latter practice, indeed, is exposed to the risk of being arbitrary and misleading because we often do not know for sure when exactly the announced institutional changes were effectively enforced, as a result of delays and anticipations, and also which specific events were statistically important for the actual policy actions. Secondly, because it can reasonably be argued that policy regime changes are occasional events, it seems appropriate to allow policy coefficients to shift a finite number of time within a sample of typical size to capture the sequence of phases that characterized the evolution of policy over time. While it is important to avoid imposing constancy of coefficients over the whole sample, on the other hand there is no need to assume they vary at every observations, as in a time-varying-parameter (TVP) framework. The transition from one regime to the other might occur in a gradual fashion in some circumstances, but if regime changes can in general be approximately described as discrete occurrences there is no need to assume a monetary policy rule that is continuously drifting over time. Moreover, compared to TVP models our approach is better designed to identify the timing with which subsequent regimes was established, especially by providing a measure of the statistical evidence for the estimated dates of change. This feature makes our approach more informative about the likely historical events behind these changes, but also facilitates the partition of history in different time intervals characterized by a specific regime, something which is useful for counterfactual policy simulations. In particular, by treating the dates of the shifts as unknown, we are also able to let the data select the specific time periods when a policy rule effectively existed and when instead it does not provide a satisfactory description of the policy actions. Finally, compared to regime-switching frameworks we do not restrict the regimes to a finite set of recurrent states since this would not be an appropriate assumption when the aim is to capture the possibly involved history of the policy instrument setting over a period almost as long as half a century. Changes in the general monetary policy strategy are better described as possibly unique and occasional events.

The time span we consider starts from the termination of the Bretton Woods system, a drastic institutional event which we take as an evident regime change, to recent years when the monetary authorities have had to rely on unconventional policy tools in a Zero Lower Bound (ZLB) environment. Because the actual policy rate has been held almost constant at 0.5% since 2009Q1, we employ for this period the shadow rate series constructed by Wu and Xia (2016), which can be regarded as the natural extension of the actual official bank rate in representing the monetary policy stance during ZLB periods. The resulting policy rate series allows us to explore the timing with which a new regime were established during the onset of the financial crisis, as well as the possibility that more than one regime have characterized the ZLB policy environment. As this rather long time span covers many potential episodes of change in the policy rule, we inevitably have to deal with small intervals containing few observations. Hence, even though Monte Carlo studies have demonstrated the good small-sample performance of the econometric tools we employ, a limitation remains in our inference about structural changes due to our relying on asymptotic critical values.

The structure of the paper is the following. In the next Section we provide a brief overview of how the existing empirical literature has addressed the issue of a changing monetary policy. In Section 3 we illustrate our methodology. In Section 4 we present the results from applying this methodology to the UK. In Section 5 we discuss the estimates in light of the UK history, and we compare them with results from previous studies. Section 6 offers some concluding remarks. More details about the results, the testing procedure and our implementation strategy are presented in the Supplementary Material.

2 Existing Work on Monetary Policy Rules

The fact that the monetary authorities’ reaction function is likely to have been different across time has been typically addressed by presenting different estimates obtained after splitting the sample in specific segments that are justified by historical events. This is the approach pioneered by Clarida, Galì, and Gertler (2000), when they compare the responsiveness of the Federal Reserve under subsequent chairmen, with a particular focus on the period before and after the appointment of Volcker in 1979. To explore the influence of institutional changes of the Bank of England on its monetary policy, Adam, Cobham, and Girardin (2005) compare the coefficients of a forward-looking Taylor rule over subsequent subsamples delimitated by a series of events that are expected to be relevant for the monetary policy conduct. They observe a changing role played by domestic versus international factors in the interest rate setting, and a far stronger evidence of a substantial change in the policy reaction function after the introduction of operational independence in 1997, rather than with the adoption of the inflation targeting in 1992.

Nevertheless, the common approach based on sample splitting is subject to two major criticisms: there is a risk that the exact dates used to define the subsamples are arbitrary; we cannot be certain that the policy function is actually stable within the selected segments. As a consequence, this crude way of recognizing a changing value for the coefficients of the policy reaction function has been replaced by more statistically rigorous methods to capture either time-varying coefficients or non-linear responses of the interest rate to macroeconomic indicators.^[2] A first available option is offered by Markov-switching models, a notable example of which is Sims and Zha (2006), who use alternative specifications of a multivariate model of the US economy to find that instability in the monetary policy rule had a far smaller influence on the macroeconomic fluctuations than the shifts in the error variance. Assenmacher-Wesche (2006) estimates a policy function for a set of three countries that includes the UK, employing a Markov-switching model where both the response coefficients and the error variance are subject to independent changes. An improved ability to describe the US history of monetary policy is shown by Castelnuovo, Greco, and Raggi (2014) to be the main feature of a policy equation where the inflation target changes, while parameters and the error variance are subject to a two-regime switching. Taylor and Davradakis (2006) fit a threshold model with two regimes triggered by the level of inflation to underline the existence of a non-linearity in the monetary policy conduct of the UK. The interest rate setting displays an asymmetry, in the sense that an aggressive response to inflation deviations emerges only after it has significantly overcome its target. By contrast, using instrumental variables to estimate a threshold model for a forward-looking Taylor rule, Koustas and Lamarche (2012) find that they are not able to reject the hypothesis of a linear policy rule for the UK.

Even though regime-switching models improve the fit of estimated policy equations, there are reasons to believe that regime changes tend to occur in a gradual fashion, rather than in the form of abrupt shifts and some evidence seems to confirm this picture (e.g. Canova and Gambetti 2009; Koop, Leon-Gonzalez, and Strachan 2009).^[3] Smooth Transition Regression Models have been used by Castro (2011) and Kharel, Martin, and Milas (2010) to study respectively the role played by a financial condition index in the UK and the European Monetary Union, and the response of the UK nominal interest rate to the real exchange rate. Boivin (2006) employs a Time-Varying Parameters (TVP) model to estimate a forward-looking version of the Taylor rule on US real-time data from 1969 to 1998, and discovers that the coefficients have been subject to changes that are gradual, in particular around the appointment of Volcker in 1979, and often non synchronised. He infers then that previous conflicting results on the same period must be due to the failure of simple models that impose a single regime change at a fixed date to capture the actual complex pattern of the monetary policy actions. Same econometric technique is applied by Trecroci and Vassalli (2010) to a group of countries that includes the UK, confirming the improved statistical performance of a time-varying Taylor rule over the classical fixed-parameter specification. After developing a procedure to address the problem of endogeneity in time-varying parameter models, Kim and Nelson (2006) employs it to estimate a monetary policy rule for the US, while Kishor (2012) applies the same approach to a set of countries that includes the UK.

3 Methodology

As anticipated in the Introduction, the empirical approach we propose consists of three steps. In the first step we look for evidence within our sample of possible shifts in the implicit long-run target of the monetary authorities, as well as changes in their procedure to control market short-term interest rates. In the second step, we perform a comprehensive investigation of possible changes in the systematic actions of the policy makers by examining a standard policy instrument rule. Finally, in the third step, we identify more clearly which policy objectives have shifted in correspondence with each estimated regime change.

In all stability tests we rely on the asymptotic distribution of the test statistics, which is inevitably an imperfect approximation of the actual finite-sample distribution, especially in small subsamples like the ones we end up with in the second step as a result of the relatively numerous episodes of structural change we obtain. This implies an inevitable risk of overfitting breaks due to size distortions, as well as underfitting breaks due to lack of power. It is important to mention this limitation of the inference we produce in the second step, although the econometrics papers developing these tests document satisfactory finite-sample performance. Moreover, to increase test power and decrease size distortion, we devote particular attention on some specific technical choices during the execution of the tests.^[4]

4 Results

We now present the results we obtained from implementing the methodology described in the previous section on UK data, with a focus on the main relevant technical aspects. An economic interpretation of these results, along with a discussion of the historical context, is offered in the subsequent Section 5. Several of the more technical details of the estimation are shown in the Supplementary Material.

4.2 Estimation of Policy Regimes

In this section we present first the results for the reduced-form equations of inflation and the output gap (first stage), then those for the policy rule equation (second stage). The strategy we follow for selecting the instruments is discussed in the Supplementary Material.^[23]

4.2.1 Reduced-Form

The outcome of the break inference on the inflation and the output gap equations is displayed in Tables 2 and 3, while detailed results on the estimation of the breaks can be found in the Supplementary Material. As for the inflation equation stability is unequivocally rejected at 1% by both the F and the Wald versions of the tests against a fixed number of breaks and by the UD max test. Our favoured method, where the sequential testing uses its own break dates, suggests the presence of one break at 1980Q2, which is estimated with fairly low uncertainty considering that the 90% confidence interval covers only 6 and 10 observations using respectively the OLS and White standard errors.^[24] The relatively high accuracy of the break date estimate is confirmed by plotting the F statistic for one break evaluated at each possible date (see the Supplementary Material). This graph exhibits, indeed, a sharp peak at 1980Q2, which also matches the timing of the shift in the mean of the marginal distribution of inflation, estimated at 1980Q2 (see Section 4.1).

Table 2:

Break interval, inflation eq.

Break date: 30 (1980Q2)
Std err	Confidence level
	90%	95%	97.5%	99%
White	28–32	27–33	26–34	24–36
LS	28–32	28–32	28–32	27–33

Table 3:

Break interval, output gap eq.

Break date: 31 (1980Q3)
Std err	Confidence level
	90%	95%	97.5%	99%
White	25–37	22–40	19–43	14–48
LS	28–34	26–36	25–37	23–39
Break date: 143 (2008Q2)
Std err	Confidence level
	90%	95%	97.5%	99%
White	141–143	141–143	141–143	140–144
LS	141–143	140–144	140–144	140–144

As for the output gap equation stability is unanimously rejected by both the F and Wald tests against a fixed number of breaks, and also taking its unweighted maximum. Our preferred criterion, the sequential testing starting with the UD max F at 1% significance level, suggests two breaks at 1980Q3 and 2008Q2. While the confidence intervals of the first break date are very wide, those associated with the second break date are extremely narrow, denoting a high accuracy in the identification of the timing of this episode of instability. The large uncertainty about the location of the break at 1980Q3 is confirmed by the graph of the F test statistic for one break over the whole sample (see the Supplementary Material), which is, indeed, beyond the 1% critical value for most of the observations until 1994Q2. Nevertheless, both estimated break dates correspond to local maxima of the F statistic.

We now have to select the specific set of reduced-form breaks that will be incorporated in the fitted values employed in the second stage of the HHB procedure. We decide to model a single common break at 1980Q2 since this break date was found for the inflation equation but is only one observation distant from the first break date in the output gap equation, it has a confidence interval that is entirely encompassed by the one calculated for the output gap, and the F test for one break in the output gap equation remains very large in a wide neighbourhood of its peak. Because we need to define a common partition for the break estimation in the second stage we impose the second break in the output gap equation, at 2008Q2, also on the inflation equation, without noticeable changes in the parameter estimates.

4.2.2 Structural Equation

As a consequence of the two structural breaks detected in the reduced-form system we have a partition of three stable segments that we have to examine individually in the second stage. The confidence intervals of the estimated break dates that are idiosyncratic to the structural equation are displayed in Table 4, while detailed results on the break estimation is presented in the Supplementary Material. Given that segments are small, we show the confidence intervals calculated using both the White and the Least Squares (LS) estimate of the covariance matrix.

Table 4:

Break interval, policy rule eq.

Break date: 23 (1978Q3)
Std err	Confidence level
	90%	95%	97.5%	99%
White	21–25	21–25	20–26	20–26
LS	21–25	21–25	20–26	19–27
Break date: 62 (1988Q2)
Std err	Confidence level
	90%	95%	97.5%	99%
White	60–64	60–64	60–64	59–65
LS	60–64	60–64	60–64	59–65
Break date: 74 (1991Q2)
Std err	Confidence level
	90%	95%	97.5%	99%
White	73–75	73–75	72–76	72–76
LS	73–75	72–76	72–76	72–76
Break date: 144 (2001Q2)
Std err	Confidence level
	90%	95%	97.5%	99%
White	112–116	112–116	111–117	110–118
LS	112–116	112–116	111–117	111–117
Break date: 155 (2011Q3)
Std err	Confidence level
	90%	95%	97.5%	99%
White	154–156	154–156	154–156	154–156
LS	154–156	154–156	153–157	153–157

All three segments turn out to be undoubtedly unstable. Using our favourite criterion that applies the UD max F test statistic in the first step with 1% significance level, we find that the first segment (1973Q2–1980Q2) features one break at 1978Q3. The evidence for this instability is quite strong, considering that the same conclusion is reached across all four criteria, including the Wald version of the testing procedure, and that the confidence interval for the break location is fairly narrow. As for the second segment (1980Q3–2008Q2), we find three breaks at 1988Q2, 1991Q2 and 2001Q2, while the Wald version of the tests suggests an additional fourth break at 1989Q3. The evidence for the instability in 1991Q2 is very strong if we take into account that this is the first break found in the sequential procedure, with a huge value of the test statistic, and with a particularly tight confidence interval (with 95% probability the break date is within the year 1991). In the third segment (2008Q3–2016Q2), we obtain one break at 2011Q3, which is with 99% confidence within only three quarters. To assess the relevance of the two reduced-form breaks in the policy rule equation we apply the Wald test for common breaks constructed by HHB. We obtain that the break at 1980Q3 is not significant since the statistic is 8.16 with a p-value of 0.04, while the break at 2008Q2 is highly significant with a statistic that is 118.71 and a virtually zero p-value.

Overall, we observe that, though the sample period is characterized by numerous structural changes, the break inference establishes their location in time with a high level of accuracy (see Table 4 and Figure 1). The relevance of the estimated break dates as symptoms of distinct episodes of parameter instability is especially evident by the fact that no confidence interval features overlapping parts, neither using the widest option of 99%. Since both the Breusch-Pagan and the Breusch-Godfrey test statistics reject with virtually zero p-values, we refine our estimates of the policy rule equation by performing a GMM estimation, conditional on our estimated set of structural breaks.^[25] The GMM covariance matrix is obtained via the method of Newey-West to select the bandwidth combined with a Bartlett kernel. The results of this estimation are displayed in Table 5.

Figure 1:

Actual interest rate (black line) and predicted levels from the break model (blue line) and from the constant-coefficient model (red line), with estimated break dates indicated by vertical lines, along with their 90% confidence intervals.

Table 5:

GMM policy rule estimates.

Segment	Period	β 1	β 2	β 3
1	1973Q1–1978Q3	8.14	0.11	0.16
		(2.15)	(0.09)	(0.22)
2	1978Q4–1988Q2	8.02	0.44	−0.02
		(0.34)	(0.02)	(0.05)
3	1988Q3–1991Q2	4.09	1.27	0.78
		(1.27)	(0.17)	(0.10)
4	1991Q3–2001Q2	4.19	0.96	0.18
		(0.19)	(0.07)	(0.05)
5	2001Q3–2008Q2	2.96	0.63	0.08
		(0.30)	(0.23)	(0.11)
6	2008Q3–2011Q3	1.71	−0.11	1.25
		(0.48)	(0.14)	(0.06)
7	2011Q4–2016Q2	4.99	0.26	2.96
		(0.52)	(0.09)	(0.21)

1. In bold coefficients significant at 1%; standard errors in brackets.

We notice that until 1978Q4 there is no statistically significant policy rule, but thereafter all three coefficients of Eq. (3.2), the intercept β ₁, the inflation coefficient β ₂ and the output gap coefficient β ₃, are significant across all segments, with the only exception of β ₃ in segment 2 and 5 and β ₂ in segment 6. If we estimate the policy rule equation under the assumption of constant parameters, we obtain a significant value of 0.48 and −0.26 for respectively the inflation and the output gap coefficient. We highlight the size of the biasedness in such estimates, which is the consequence of ignoring the existent parameter instability, especially in the case of the output gap coefficient, which turns negative and significant, while in the model that allows for structural breaks a negative output gap coefficient occurs only in one segment and without being significant (see also Figure 1).^[26]

As a quantitative assessment of the importance of the shifts in the policy coefficients for the overall variability of the interest rate, we use the second stage residuals from the 2SLS estimation to decompose the total sum of squares in the explained sum of squares from the no-break model, ESS _c , which can be thought of as the variability in the interest rate that is explained by its average constant relationship with inflation and the output gap, the additional explanatory power produced by the break model, ESS _b –ESS _c , which represents the contribution to the interest rate variability of the shifts in the parameters of this relationship, and the residual TSS–ESS _b . What is remarkable in this accounting exercise is that not only does the explicit modelling of the shifts in parameters provide additional explanatory power, but it even outweighs substantially the overall fit provided by a stable policy rule equation. While the share of TSS explained by a constant policy rule is only 36%, modelling discrete changes in the parameters allows to explain an additional 61% of all the movements in the official interest rate over the whole period.^[27]

The overall validity of the model is confirmed by Hansen’s test for overidentifying restrictions, which, as expected, does not signal a substantial difference between the models with and without breaks (see the Supplementary Material). On the contrary, a considerable difference between the two models emerges in terms of residual autocorrelation. While the break model retains some of the significant autocorrelation, which prevents us from discarding the other two potential sources of interest rate persistence, there is a massive reduction in its magnitude compared to the no-break model if we look at the values of the Breusch-Godfrey F statistics. This is strongly confirmed by the fact that if we re-estimate the two models including as an additional regressor one lag of the dependent variable, the estimate of its coefficient falls by almost half once we control for the existent structural breaks, that is from 0.99 to 0.55, both with a p-value of 0.009. This result highlights how neglecting existent parameter instability in policy rule equations may translate into spurious evidence of policy inertia, or at least lead to substantially overstate its importance in the actual setting of the policy instrument. On the other hand, the fact that the testing procedure based on the autocorrelation-robust Wald statistic produces the same list of break dates, with the exception of one extra break in segment 2, suggests no substantial risk of spurious breaks induced by the autocorrelated errors. When we look for significant signs of piecewise heteroskedasticity, which might have confounded our break inference on the policy rule equation, we find that the Goldfeld-Quandt statistic features very large p-values in correspondence of four of the six estimated breaks, with a change in variance occurring only in the second and fifth segments, which suggests a very low risk that our estimated breaks are in fact spurious. Finally, we perform two robustness checks by running the break estimation after replacing the chosen output gap measure with an alternative one constructed by the OECD and using the RPIX as inflation index. The timing of the instability episodes remains largely unchanged, with only minor differences in the values of the policy coefficients, which confirms our findings about the evolution of the UK monetary policy (see Supplementary Material).

5 Discussion

A look at our estimation results in Table 5 reveals two key findings. First, many of the historical dates that we expected to be important, and that we included in Table 1, surprisingly do not come up as significant break points, but the overall evolution confirms existing accounts of the UK monetary policy. What emerges is the picture of a monetary policy that undergoes a sequence of phases over time, not strictly matching the timing of the official institutional events, and thus suggesting an important role played by anticipations, early transitions and delays in the actual enforcement of the institutional regime changes. Second, a policy instrument rule did not exist until the end of the 1970s, and since then a significant positive response to inflation has characterized monetary policy until 2016, with the only exception of the interval between 2008 and 2011. As can be seen from Figure 1, the fit of our estimated policy rule is indeed disastrous up to 1978Q3, but then it improves substantially during the 1980s, and it is able to capture very well the movements of the interest rate from 1988Q2 up to 2016Q2.

This rather sharp evidence confirms previous understandings of the UK history, and in particular the claim that exactly in that period a paradigm change about the transmission mechanism took place within academic and policy circles, which led to the abandonment of incomes policy and price regulation in favour of a monetary approach to the control of the inflation process (e.g. Batini and Nelson 2005).

This shift in the general monetary policy strategy is testified by a series of events such as the abolishment of the exchange rate controls in October 1979, the start of the Medium Term Financial Strategy (an explicit multi-year monetary target) in March 1980 and the end of the “corset” scheme to limit the banks’ lending in June 1980. The scope of the transformations taking place during this period is corroborated by the large downward shift in the mean of the inflation process we found in 1980Q2, along with a break in the reduced-form equation of both inflation and the output gap, which is accompanied in the subsequent quarter by a significant fall in the level of the interest rate spread. These pieces of evidence all confirm an increased credibility of the government policy in its effort to curb inflation, as well as a stronger ability of the monetary authorities to influence the market short-term interest rates.

As for the magnitude of the policy response to inflation (β ₂) and output gap (β ₃), it can be argued that the positive and significant inflation coefficient in the second segment reflects the restrictive macroeconomic strategy adopted to curb inflation, but otherwise the relatively low fit of the policy rule is the consequence of a series of events that affected the interest rate decisions beyond what could be expected from the mere consideration of the dynamics of prices and economic activity.^[29]

An estimated inflation coefficient that is greater than one between 1988Q3 and 1991Q2 captures the implementation of a very tight monetary policy in the face of an increasing inflation, whereas a response smaller than one in the subsequent period, 1991Q3–2001Q2, was arguably the consequence of an inflation rate that was falling throughout the whole segment, combined with an asymmetric response of the interest rate that was deliberately stronger with respect to increases than decreases in inflation. Since this fall in inflation started in 1991Q3, being the product of the past restrictive monetary policy but also of a more credible nominal anchor in place, the inflation targeting period from 1992Q3 appears de facto as a continuation of the same regime. Hence, both the adoption of the inflation targeting in 1992 and the established central bank independence in 1997 appear as part of a regime change that started earlier in time. The relatively more stable pattern for both inflation and the output gap is plausibly the reason why for the period 2001Q3–2008Q2 we obtain a smaller but still significant coefficient on inflation and a not significant output gap coefficient. The reverse in the statistical significance of the two coefficients starting from 2008Q3, with a strong and significant response only with respect to the output gap up to 2011Q2, reflects the priorities of the quantitative easing program, which starts officially in March 2009. Finally, the break we found in 2011Q3, which is accompanied by a massive increase in the response to output gap, from 1.25 to 2.96, as well as by the inflation coefficient becoming significant, matches the timing with which the second round of quantitative easing is announced as a means to ward off the risk of a double-dip recession. Indeed, this event translates into a sharp fall in the shadow rate from −2.53% to −3.70%. The fact that this announcement starts off a new regime rather than representing a single episode is confirmed by the numerous subsequent interventions in the same direction.^[30] It is also worth noting that since we obtained a reduced-form break at 2008Q2, the use of a shadow rate since 2009Q1 does not affect the break inference in the first part of the sample, where we use the actual official rate.

The Taylor principle ensuring a unique stationary equilibrium for the economy, as stated in Bullard and Mitra (2002), turns out to be satisfied only over the period 1988Q3–2001Q2. More precisely, it is clearly met in segment 3, given that the inflation coefficient is above 1, but also in segment 4 using a plausible range of values for the discount factor and the slope of the New Keynesian Phillips curve. This latter needs to be smaller than 0.045 if we use a discount factor of 0.99, which is not implausible for the UK (see e.g. estimates in Batini, Jackson, and Nickell 2005).

When we compare our results with those from previous studies, we notice that Clarida, Galì, and Gertler (1998) found that during the 1979–1990 period the response to inflation was 0.98, which is consistent with an average of the two values we obtain in segment 2 and 3. Comparing the results from a policy rule estimated in three a priori selected periods, Adam, Cobham, and Girardin (2005) obtained that from 1985 to 1990 the interest rate was exclusively determined by the US and German interest rates, and it is only with the introduction of the inflation targeting in 1992 that inflation and the output gap became significant factors, but they dominated the interest rate setting in the subsequent 1997–2002 period. While they interpret this outcome as a confirmation of the importance of the institutional change that attributed operational independence to the Bank of England in 1997, we find, on the contrary, that neither 1992 or 1997 are picked by the tests as significant break dates, suggesting an evolution of the effective systematic policy decisions that does not match exactly the historical timing of the institutional changes. In this respect our results are closer to Kishor (2012), who obtained from a TVP policy rule that the inflation coefficient reaches the highest level around 1990 and then slowly declines towards 1998, although his estimated response is greater than one already from 1983.

6 Conclusion

We studied the evolution of the UK monetary policy by developing a coherent empirical approach to estimate and accurately date its historical regime changes, building on a set of tools that have recently been advanced in the econometric literature on structural stability. The advantages of our approach compared to existing methods to capture regime changes are mainly two. It is specifically designed to provide a rigorous estimation of the timing with which policy regimes have been established, and it is able to identify which priorities of the policy makers have changed in the transition from one regime to another.

Our approach delivers a plausible description of the evolution of the UK monetary policy over time. We found that, in the almost 45 years following the collapse of the Bretton Woods system, regime changes have been frequent, they are responsible for a substantial share of the observed variation in the policy instrument, but their timing relative to the causing institutional change is far from obvious. As a result of an accumulation of institutional changes, the systematic policy actions of the monetary authorities underwent a sequence of phases, characterized by a different priority assigned to the main policy goals, but following a coherent evolution over time, with inflation becoming gradually more important from the 1970s up to early 2000s, and the output gap dominating after the 2008 financial crisis.

We showed how ignoring these frequent regime changes when estimating a policy rule produces highly biased parameter estimates and an overstatement of the importance of policy inertia. Overall, our results provide a clear warning to empirical studies that, for instance, build policy counterfactuals in an attempt to disentangle the effects of monetary policy on the economy, either taking for granted the timing with which institutional policy reforms exert their effects or imposing a generic continuous adjustment over time in the policy coefficients.