How forecasts evolve--the growth forecasts of the federal reserve and the Bank of England.

Allen, William A. ; Mills, Terence

We investigate how central bank forecasts of GDP growth evolve through time, and how they are adapted in the light of official estimates of actual GDP growth. Using data for 1988-2005, we find that the Federal Open Market Committee (FOMC) has typically adjusted its forecast for growth over the coming four quarters by about a third of the unexpected component of estimated growth in the four quarters most recently ended. We were unable to find any clear signs of systematic errors in the FOMC's forecasts. UK data for 1998-2005 suggest that the Bank of England Monetary Policy Committee (MPC) did not adjust its forecasts in this way, and that there were systematic forecast errors, but the evidence from the latter part of the period 2001-5 tentatively shows a behaviour pattern closer to that of the FOMC, with no clear signs of systematic errors.

Keywords: Forecasts; monetary policy; data revisions; vintages; Federal Reserve; Federal Open Market Committee; Bank of England, Monetary Policy Committee JEL classification: E47, E58, N12, N14

I. Introduction

It is widely understood that discretionary monetary policy actions taken now have their effects on the economy not immediately, but in the future. It follows that rational decisions about such actions must be based on a forecast of the future course of the economy. Therefore, forecasts are a vital part of discretionary policy-making: the quality of the policy depends greatly on the quality of the forecasts.

For that reason, it is of great interest to explore the forecasting process. Indeed, the evaluation of forecasts of economic variables has engendered considerable research in recent years, as the monographs by Clements and Hendry (1998) and Clements (2005) reveal. In particular, assessing the track records of the various institutions and agencies engaged in macroeconomic forecasting has provided interesting reading: Pain (1994), Pepper (1998) and Mills and Pepper (1999) being notable examples for the UK. The purpose of this paper is somewhat different from these earlier studies, as it examines and compares how forecasts of GDP evolve in the light of new information in the form of GDP outturns. The examination is based on published information about forecasts made in the US by the members of the Federal Open Market Committee, and in the UK by members of the Bank of England Monetary Policy Committee.

2. The United States

Twice a year, in February and July, the Federal Reserve Chairman delivers a monetary policy report to the United States Congress. The report includes a table of the economic projections of the Federal Reserve Governors and Reserve Bank Presidents. The projections are for nominal and real GDP, the PCE price index, and the civilian unemployment rate. The February reports provide projections just for the current year (in the case of nominal and real GDP and the PCE price index, the percentage change from the fourth quarter of the preceding year to the fourth quarter of the current year; in the case of unemployment, for the fourth quarter of the current year). (1) The July reports provide projections for both the current year and the following year. As an illustration, the forecast table from the July 2004 report is shown in table 1.

The table always reports both the range of the forecasts of the Federal Reserve Governors and Reserve Bank Presidents, and the 'central tendency', which is normally a narrower range. For the purposes this paper, we have taken the mid-point of the narrower 'central tendency' range to be the FOMC's forecast.

3. The United Kingdom

Each quarter, the Bank of England publishes an Inflation Report incorporating the Monetary Policy Committee's forecasts of CPI inflation and GDP. Alternative forecasts are provided on the alternative assumptions that interest rates remain unchanged at their current level throughout the forecast period, and that they change in line with market expectations as the Bank of England interprets them from market prices. The forecasts are quarterly and run for two years into the future (since August 2004, the forecasts based on the 'market interest rate' assumption have run three years into the future). For the purposes of this paper, we use the forecasts based on the assumption that interest rates remain unchanged at their current level. We doubt whether switching to the alternative assumption would make any significant difference to the results.

The Bank of England forecasts are in the form of a probability distribution of each variable at each forecast date. For the purposes of this paper, we use the mean forecast for each date.

As an illustration, the Bank of England GDP forecast made in August 2004, assuming interest rates remain unchanged at their current level, is shown in table 2, reproduced from Bank of England (2004).

4. Assessing sensitivity to current events

It is reasonable to believe that the forecasts represent the current expectations of the policy-making bodies of the two central banks about future economic developments. In the case of the Federal Reserve, the forecasts are the central tendency of the forecasts of the members of the policymaking body, the Federal Open Market Committee. (2) In the case of the Bank of England, the forecasts are the responsibility of the Monetary Policy Committee; the published forecast is said to represent the 'centre of gravity' of their individual views.

Sensitivity to current events can be assessed by estimating how much the change in forecasts from one forecasting round to the next is influenced by outturns that become known in the time interval between the forecasts. It is important to base that assessment on the estimate of the latest outturn that was available at the time of the current forecast, and not on the most recent estimate available at the time of conducting the research, since they are usually not the same: initial estimates are commonly revised later. Fortunately, databases showing different 'vintages' of macroeconomic data estimates are available for both the US and the UK. (3)

In order to compare the behaviour of the two central banks, we need a test that uses only information published by both central banks. The Federal Reserve publishes less information about each GDP forecast than the Bank of England. (4) Thus, the binding constraint is the information set published by the Federal Reserve; we need a test which uses only information published by the Federal Reserve.

When compiling its February forecasts, the Federal Reserve has available to it an official estimate of GDP in the preceding quarter, i.e., the fourth quarter of the preceding year, whereas in July it has no estimate of second-quarter growth. Therefore, the test of the Federal Reserve's sensitivity to current economic events must allow for this important difference between February and July forecasts. The Bank of England forecasts are produced in February, May, August and November, when the Bank of England always has available to it an official estimate of GDP in the preceding quarter. So a direct comparison of the Federal Reserve and the Bank of England has to be based on the Federal Reserve's February forecasts.

The sensitivity of GDP growth forecasts to estimated outturns is assessed by estimating the following equation:

(1) F(N, N + 3) = [[beta].sub.0] + [[beta].sub.1]F(N - 2, N + 3) + [[beta].sub.2]G(N, N - 1) + [[beta].sub.3]F(N - 2, N - 1) + u(N, N + 3)

where F(M, N) is the forecast made in quarter M of GDP growth over the four quarters ending in quarter N (M [less than or equal to] N), and G(M, N) is the estimate made in quarter M of GDP growth over the four quarters ending in quarter N (M > N). u(N, N +3) is the equation residual.

In equation (1), the parameters can be interpreted as follows:

F(N,N+3), the dependent variable, is the current forecast of growth in the coming year. To be precise, it is the forecast made in quarter N of GDP growth in the four quarters ending in N+3. The reason why the horizon for the coming year is N+3 and not N+4 is that, at the time the forecast is made, no estimate of the level of GDP in the current quarter N is available: the latest quarter for which an estimate is available is N-1, or even N-2 if the forecast is made early in the quarter.

[[beta].sub.1] measures the influence on the current forecast of the forecast made two quarters earlier, in quarter N-2, of growth in the same four quarters ending in N+3; in other words, it measures the persistence of the forecasts;

[[beta].sub.2] measures the influence on the current forecast of the estimated outturn in the four quarters ended N-1; in other words, it measures the sensitivity of the forecasts to the latest estimated outturns;

[[beta].sub.3] measures the influence on the current forecast of the forecast for the four quarters ended N-1 that was made two quarters earlier, in quarter N-2. If [[beta].sub.2] + [[beta].sub.3] = 0, then the current forecast is influenced only by the unexpected component of the estimated outturn in the four quarters ended N-1, i.e., the component that was not forecast two quarters earlier.

The estimation results are shown in table 3. In the case of the US, experimentation with the data revealed that the relationship among the variables appeared to change during the period. (5) A regression over 1988-2005, the Greenspan period, produced coefficient estimates that were clearly different from those produced by a regression over the longer period 1980-2005. The evidence for a structural break in the UK equation is not conclusive: the p-value of the Chow test for a structural break in 2001Q1 is 0.28 (meaning that the probability that the test F-statistic would have come out as it did assuming no structural break is 0.28); there is no a priori reason to expect that there would have been a structural break at that time. The comments

below are based on the Greenspan period regression for the US; we comment on both the whole-period and post-2001 estimates for the UK.

In the US case, and in both UK cases, the estimated value of [[beta].sub.1] is not significantly different from unity, which suggests that in both central banks, there was a high degree of forecast persistence.

The estimation results suggest that the Federal Reserve was significantly influenced in its forecasting by the estimated outturn in the four quarters just ended. For the US, the estimate of [[beta].sub.2] is significantly positive, and roughly equal and opposite to the estimate of [[beta].sub.3]. Furthermore, the estimate of [[beta].sub.1] is close to unity while the estimate of [[beta].sub.0] is insignificant. Imposing the implied restrictions [[beta].sub.0] = 0, [[beta].sub.1] =1 and [[beta].sub.2] + [[beta].sub.3] = 0 led to the regression

F(N, N + 3) = F(N - 2, N + 3) + 0.32 (G(N, N - 1)
 (2.88)
 - F(N - 2, N - 1))

[R.sup.2] = 0.73

Regression standard error = 0.43

Durbin-Watson = 2.35

The restrictions are acceptable, as they produce a test statistic of F(3, 13) = 0.43, which has a probability value of 0.73. The fit of the regression has barely deteriorated with the imposition of the restrictions and the coefficient suggests that the Federal Reserve adjusted its forecast of GDP growth in the coming year by about a third of the unexpected component of the latest estimated outturn, other things being equal.

For the UK whole-period regressions, the negative sign of the estimate of [[beta].sub.2], taken at face value, suggests that the Bank of England's forecast of growth over the coming year was negatively influenced by the latest outturn. It is not obvious how to interpret this finding. It is possible that the MPC had a view about what the future level of output would be, in which case high recent growth would lead to a lower forecast of future growth. However, this interpretation is not supported by either the Inflation Report texts or by the published minutes of MPC meetings. Another interpretation, which we think is more plausible, is that the MPC often disbelieved the ONS' preliminary estimate of GDP growth in the latest quarter, and minimised its effect on the forecast by assuming that the measurement error that it thought it perceived would be reversed in the following quarter (i.e., the first forecast quarter). (6) However, the post-2001 coefficient estimates are somewhat more like the US estimates, though the estimates of the coefficients [[beta].sub.2] and [[beta].sub.3] are much less statistically significant than in the US case. If we impose the restrictions [[beta].sub.0] = 0, [[beta].sub.1] = 1 and [[beta].sub.2] + [[beta].sub.3] = 0, the resulting equation is:

F(N, N + 3) = F(N - 2,N + 3) + 0.21 (G(N, N - 1)
 (0.99)
 - F(N-2,N-1))

[R.sup.2] = 0.30

Regression standard error = 0.43

Durbin-Watson = 1.70

Another interesting difference between the US and UK estimates is that the US equation has a much higher [R.sup.2] than either of the UK equations, indicating that the Bank of England Monetary Policy Committee was more influenced by factors not included in equation 1 than was the Federal Open Market Committee.

Equation 1 cannot be estimated using data from the Federal Reserve's July forecasts because the required data are not available. However, the following equation can be estimated:

F(N,N+1)= [[gamma].sub.0] + [[gamma].sub.1]F(N-2,N+1) +[[gamma].sub.2]G(N, N - 2) + u(N, N + 1)

The estimated parameters are shown in table 4. The estimated values of [[gamma].sub.2] suggest that the estimated outturn for the latest year (i.e. the year ending in Q1 of the year of the forecast) probably did have some influence on the Federal Reserve's July forecast; the evidence is a bit stronger for the Greenspan period alone than for the whole period.

5. Are there systematic errors in the forecasts of either central bank?

The analysis above has simply established some facts about the forecasts produced by the two central banks, and identified some differences. However, it has no normative content. It does not make a case that either central bank is behaving optimally or sub-optimally; even though there are behavioural differences between the two, the differences may be wholly explainable by differences in the behaviour of the two economies, or differences in the quantity or quality of information available to the two central banks.

The purpose of the analysis in this section is to go further and try to detect systematic errors in the forecasts of the two central banks. Specifically, we investigate whether the pattern of forecast errors is statistically related to variables that were known at the time when the forecasts were made. lf there is a statistically significant relationship, then that is prima facie evidence that the forecasting procedure was capable of being improved.

For this purpose, we estimate the following equation:

(3) F(N, N + 3) - G(2005Q2, N + 3) = [[delta].sub.0] + [[delta].sub.1]F(N - 2, N + 3) + [[delta].sub.2]G(N, N - 1) + [[delta].sub.3]F(N-2,N-1) + u(N, N + 3)

where G(2005Q2, N+3) is the latest estimate (as of the time of completing the research reported in this paper in April 2005) of GDP growth in the four quarters ending in quarter N+3. The constant term is intended to detect any systematic bias uncorrelated with the other explanatory variables.

The dependent variable F(N,N + 3) - G(2005Q2, N + 3) is the forecast error calculated using the latest estimate of the outturn for GDP growth available at the time of completing this research. If the forecasting procedure were optimal, then the [delta] coefficients would all be zero, (7) and the equation would have no explanatory power, since each of the independent variables was known at the time of the forecast (i.e., in quarter N). Furthermore, there should be no evidence of autocorrelation in the residuals, as this would imply a systematic pattern in the forecast errors.

The estimation results are shown in table 5. Tests of the null hypothesis that the set of [delta] coefficients are jointly zero produced statistics with probability values of 0.55 and 0.36 for the two US regressions, so that there does not appear to be conclusive evidence of systematic errors in the Federal Reserve's forecasts of US GDP. On the other hand, such a test for the UK over the full period 1998-2004 clearly rejects the null hypothesis and, moreover, there is strong evidence from the Durbin-Watson statistic of autocorrelation in the residuals of the UK equation. We thus conclude that there were systematic errors in the Monetary Policy Committee's forecasts of UK GDP, and that the autocorrelation in the residuals of the equation suggests a degree of persistence in these errors. (8,9)

However, when equation (3) is estimated over the shorter period beginning in 2001Q1, the evidence of systematic errors becomes much less clear: the test of the null hypothesis that the set of [delta] coefficients are jointly zero produces a probability value of 0.38. However, the power of the test over such a short data period will tend to be rather limited.

We can rearrange equation (3), using the estimated coefficients for the whole-period UK regression, to produce an adjusted forecast equation:

Adjusted forecast = F(N, N+ 3) + 4.35 - 1.64F(N - 2, N - 3) + 0.41G(N, N - 1) - 0.48F(N - 2, N- 1)

The adjusted forecasts differ from the published ones in that they take:

(i) less account (in fact negative account, since the implied coefficient is 1.11 - 1.64, i.e. minus 0.53) of the forecast of GDP growth in the forthcoming four quarters that was published six months earlier; and

(ii) more account of the error in the forecast for the latest elapsed four quarters as calculated from the preliminary estimate of GDP published by the Office for National Statistics, even though that estimate is subject to revision.

Not surprisingly, it is possible to calculate that forecasts adjusted in this way would have been more accurate than the forecasts that were published; the root mean square error of the adjusted forecasts made over the period February 1998-February 2004 inclusive would have been 0.68 percentage points, compared to 1.08 percentage points for the published forecasts. Further examination of the data suggests that the first of these factors is much more important than the second; the gains in accuracy from de-emphasising previous forecasts would have been much greater than the gains from taking more account of recent estimated outturns.

The evidence of systematic forecast errors is significant, and suggests that there were imperfections in the forecasting process. However, we wish to emphasise that there is nothing optimal about the adjustments to the UK forecasts described above. For one thing, they are not based on a comprehensive examination of the Bank of England's forecasting record (for example, we have examined only that part of the Bank of England's record that is readily comparable with that of the Federal Reserve). More fundamentally, the adjustments are calculated ex post, and reflect the interaction between the imperfections in the forecasting process and the specific circumstances of the period. There is no reason to expect that the same adjustments would improve the accuracy of forecasts produced by the same process in a different period. Further research would be required to identify exactly what the imperfections in the forecasting process were.

The evidence from the post-2001 UK regressions must be regarded as tentative, because the data period is so short. (10) It is that, over that period, there were no systematic errors in the forecasts. This is interesting, because the estimates of equation (1) using the US data and the post-2001 UK data indicate roughly the same degree of sensitivity to recent data. Both central banks appear to have acted as though they were guided by a 'one-third adjustment' practice in forecasting GDP growth over the coming year, i.e., a practice of adjusting the six-month earlier forecast of GDP growth in the coming year by about a third of the unexpected component of the estimated outturn in the latest year for which data are available. This practice has the attractive feature that it appears not to have given rise to systematic errors in either country, though we stress that the evidence for the UK is very tentative.

6. Conclusions

Our conclusions are that both the FOMC and the MPC were heavily influenced in their forecasting of GDP growth by their own earlier forecasts, which seemed in both cases to be a starting point for their current forecasts. The FOMC was also influenced by the estimated outturn for the latest quarter and, in updating its previous forecast, appeared to incorporate about a third of the estimated error in its forecast of the preceding year's GDP growth into its forecast for the coming year. The evidence for the whole period of the MPC's existence, by contrast, is that it appeared if anything to react to faster estimated growth in the latest quarter by adjusting downwards its forecast of future growth. However, there are signs in the post-2001 evidence that, over that period, the MPC reacted in the same kind of way as the FOMC.

We found no evidence of systematic errors in the FOMC's forecasts. We did find such evidence in those of the MPC, though it was not apparent in the post-2001 period. Further research would be required in order to identify the imperfections in the forecasting process which were the source of the systematic errors.

Table 1. Federal Reserve's projections for 2004-5,
published in July 2004. Economic projections for 2004
and 2005, per cent

 Federal Reserve Governors
 and Reserve Bank presidents

Indicator Range Central tendency

2004
Change, fourth quarter
 to fourth quarter (a)
Nominal GDP 6-7 6 1/4-6 3/4
Real GDP 4-4 3/4 4 1/2-4 3/4
PCE price index
 excluding food and energy 1 1/2-2 1 3/4-2
Average level, fourth quarter
Civilian unemployment rate 5 1/4 - 5 1/2 5 1/4-5 1/2

2005
Change, fourth quarter
 to fourth quarter(a)
Nominal GDP 4 3/4-6 1/2 5 1/4-6
Real GDP 3 1/2-4 3 1/2-4
PCE price index
 excluding food and energy 1 1/2-2 1/2 1 1/2-2
Average level, fourth quarter
Civilian unemployment rate 5-5 1/2 5-5 1/4

Source: Federal Reserve Board (2004).

Note: (a) Change from average for fourth quarter of previous year to
average for fourth quarter of year indicated.

Table 2. Bank of England GDP growth forecasts published in August 2004.
GDP growth projections based on interest rates constant at 4.75% (a)

 2004Q3 2004Q4 2005Q1 2005Q2 2005Q3

Mode 3.69 3.55 3.70 3.45 3.03
Median 3.69 3.55 3.70 3.45 3.03
Mean 3.69 3.55 3.70 3.45 3.03
Uncertainty (b) 0.23 0.47 0.62 0.74 0.78
Skewness (c) 0.00 0.00 0.00 0.00 0.00

Probability that distribution will be above or below particular
values (d)
Pr.(<0%) <5% <5% <5% <5% <5%
Pr.(0%-1%) <5% <5% <5% <5% <5%
Pr.(1%-2%) <5% <5% <5% <5% 9%
Pr.(2%-3%) <5% 12% 13% 25% 39%
Pr.(3%-4%) 91% 71% 56% 50% 41%
Pr.(>4%) 9% 17% 31% 23% II%

 2005Q4 2006Q1 2006Q2 2006Q3

Mode 2.59 2.16 2.00 2.07
Median 2.59 2.16 2.00 2.07
Mean 2.59 2.16 2.00 2.07
Uncertainty (b) 0.84 0.93 1.00 1.00
Skewness (c) 0.00 0.00 0.00 0.00

Probability that distribution will be above or below particular
values (d)
Pr.(<0%) <5% <5% <5% <5%
Pr.(0%-1%) <5% 10% 13% 12%
Pr.(1%-2%) 21% 33% 34% 33%
Pr.(2%-3%) 44% 38% 34% 35%
Pr.(3%-4%) 27% 16% 14% 15%
Pr.(>4%) <5% <5% <5% <5%

Notes: (a) GDP at constant market prices. (b) The uncertainty measure
is that discussed in Britton, Fisher and Whitley (1998), in which it is
referred to as [sigma]. In the case of an unskewed distribution, it is
equal to the standard deviation. (c) There are various possible ways of
measuring the skew of a distribution. The skew measure shown here is
the difference between the mean and the mode. (d) These figures
represent the probabilities which the MPC assigns to CPI inflation or
GDP growth lying within a particular range at a specified time in the
future. Because of the difficulties in precisely quantifying
low-probability events, probabilities of less than 5% are not shown
in these tables.

Table 3. Estimated parameters of equation 1

Estimated parameters US US
(t-statistics shown in 1980-2005 1988-2005
parentheses)

[[beta].sub.0] -0.78 (1.70) -0.05 (0.11)
[[beta].sub.1] 1.05 (5.56) 0.93 (5.11)
[[beta].sub.2] -0.00 (0.01) 0.32 (2.58)
[[beta].sub.3] 0.20 (1.84) -0.26 (1.29)
Estimation period Feb 1980-Feb 2005 Feb 1988-Feb 2005
Number of observations 26 18
[R.sup.2] 0.80 0.74
Regression standard error 0.61 0.47
Durbin-Watson statistic 2.35 2.49

Estimated parameters UK UK
(t-statistics shown in 1998-2005 2001-2005
parentheses)

[[beta].sub.0] 0.21 (0.42) 0.53 (1.01)
[[beta].sub.1] 1.11 (5.28) 0.85 (2.05)
[[beta].sub.2] -0.23 (1.02) 0.28 (0.95)
[[beta].sub.3] 0.02 (0.11) -0.29 (1.19)
Estimation period Feb 1998-Feb 2005 Feb 2001-Feb 2005
Number of observations 29 17
[R.sup.2] 0.53 0.35
Regression standard error 0.53 0.46
Durbin-Watson statistic 1.38 1.75

Table 4. Estimated parameters of equation 2

Estimated parameters US US
(t-statistics shown 1980-2004 1988-2004
in parentheses)

[[gamma].sub.0] -0.48 (0.98) 0.38 (0.70)
[[gamma].sub.1] 1.11 (5.22) 0.69 (2.67)
[[gamma].sub.2] 0.11 (0.85) 0.19 (1.32)
Estimation period July 1980 July 1988
 -July 2004 -July 2004
Number of observations 25 17
[R.sup.2] 0.70 0.63
Regression standard error 1.05 0.63
Durbin-Watson statistic 2.28 2.46

Table 5. Estimated coefficients of equation 3

Estimated parameters US US
(t-statistics shown in 1980-2005 1988-2005
parentheses)

[[delta].sub.0] -0.66 (0.51) -3.25 (1.95)
[[delta].sub.1] -0.11 (0.20) 0.51 (0.90)
[[delta].sub.2] -0.01 (0.05) -0.35 (0.85)
[[delta].sub.3] 0.20 (0.64) 0.92 (1.29)
Estimation period Feb 1980-Feb 2004 Feb 1988-Feb 2004
Number of observations 25 17
[R.sup.2] 0.04 0.21
Regression standard error 1.64 1.42
Durbin-Watson statistic 1.88 1.89

Estimated parameters UK UK
(t-statistics shown in 1998-2005 2001-2005
parentheses)

[[delta].sub.0] -4.35 (5.51) 1.39 (0.74)
[[delta].sub.1] 1.64 (5.51) 0.05 (0.08)
[[delta].sub.2] -0.41 (1.23) -0.06 (0.12)
[[delta].sub.3] 0.48 (1.55) -0.54 (1.28)
Estimation period Feb 1998-Feb 2004 Feb 2001-Feb 2004
Number of observations 25 13
[R.sup.2] 0.61 0.21
Regression standard error 0.72 0.62
Durbin-Watson statistic 0.97 1.76

NOTES

(1) However, in February 2005, the Federal Reserve for the first time published projections for the following year. Note that up to and including July 1999, the forecast quoted the CPI, not the PCE index, and base-weighted, not chain linked, GDP estimates.

(2) The forecasts incorporate the views of the non-voting members of the Federal Open Market Committee, as well as the voting ones, but that is unlikely to make much difference to the central tendency.

(3) For the US, see Croushore and Stark (1999) and the regularly updated data available at http://www.phil.frb.org/econ/forecast/ reaindex.html. For the UK, see Castle and Ellis (2002) and the accompanying data available at http:// www.bankofengland.co.uk/Links/setframe.html. One of the authors has updated the real-time database to include all the data needed for this paper.

(4) On the other hand, the Federal Reserve publishes a forecast of unemployment, which the Bank of England does not.

(5) The probability value of the Chow test for a structural break in 1988 is 0.03.

(6) There have been some large revisions to the ONS' preliminary estimates of GDP growth, and the revisions were particularly large (and positive) in the case of the preliminary estimates of growth in the period from late 1998 to early 2000.

(7) Unless the forecasting process were capable of predicting data revisions which have not yet been made, in which case one or more of the S coefficients might be non-zero even if the forecasts were optimal. We neglect this possibility on the grounds that it is implausible.

(8) The Bank of England Monetary Policy Committee periodically reviews its own forecasting record. See, for example, the review in Bank of England (2004, pp. 50-51), which reports under-forecasting of output growth. It attributes the increase in the degree of under-forecasting since the previous review largely to data revisions.

(9) Our conclusion is therefore consistent with that of the House of Lords Select Committee on Economic Affairs, which commented on 'systematic and persistent errors in forecasting' (see House of Lords, 2004).

(10) Also, the likelihood that future data revisions will change the assessment of whether there were systematic errors in the forecasts is greater for the post-2001 period than for earlier periods.

REFERENCES

Bank of England (2004), Inflation Report, August, pp. 50-51.

Britton, E., Fisher, P. and Whitley, J. (1998), 'The Inflation Report projections: understanding the fan chart', Bank of England Quarterly Bulletin, February.

Castle, J. and Ellis, C. (2002), 'Building a real time database for GDP(E)', Bank of England Quarterly Bulletin, Spring, pp. 42-9.

Clements, M.P. (2005), Evaluating Econometric Forecasts of Economic and Financial Variables, Basingstoke, Palgrave Macmillan.

Clements, M.P. and Hendry, D.F. (1998), Forecasting Economic Time Series, Cambridge, Cambridge University Press.

Croushore, D. and Stark, T. (1999), 'A real-time data set for macroeconomists', Federal Reserve Bank of Philadelphia Working Paper 99-4.

Federal Reserve Board (2004), Monetary Policy Report submitted to the Congress on July 20, 2004, pursuant to section 2B of the Federal Reserve Act. Available at http://www.federalreserve.gov/ boarddocs/hh/2004/july/ReportSection1.htm .

House of Lords (2004), 3rd Report of Session 2003-04: Monetary and Fiscal Policy: Present Successes and Future Problems Volume I--Report, HL Paper 176-1, paragraph 46.

Mills, T.C. and Pepper, G.T. (1999), 'Assessing the forecasters: an analysis of the forecasting records of the Treasury, the London Business School and the National Institute', International Journal of Forecasting, 15, pp. 247-57.

Pain, N. (1994), 'Cointegration and forecast evaluation: some lessons from National Institute forecasts', Journal of Forecasting, 13, pp. 481-94.

Pepper, G.T. (1998), Inside Thatcher's Monetarist Revolution, Macmillan/Institute of Economic Affairs.

* Visiting Fellow, Faculty of Finance, Cass Business School, City University. e-mail: [email protected]. ** Professor of Applied Statistics and Econometrics, Loughborough University. e-mail: [email protected]. We are very grateful to those who commented on earlier versions of the paper, including Charles Goodhart, Geoffrey Wood and an anonymous referee. Of course, they don't necessarily agree with our conclusions.