Twenty years of the ECB Survey of Professional Forecasters

Prepared by Rupert de Vincent‑Humphreys, Ivelina Dimitrova, Elisabeth Falck and Lukas Henkel

Published as part of the ECB Economic Bulletin, Issue 1/2019.

It is twenty years since the ECB first launched its Survey of Professional Forecasters (SPF). The survey asks for point forecasts and probability distributions for HICP inflation, HICP inflation excluding energy, food, alcohol and tobacco, real GDP growth and the unemployment rate at six horizons, as well as point forecasts for wage growth, the exchange rate, the oil price and the ECB’s policy rate. All quantitative data collected in the survey are systematically published shortly after the completion of the survey. This makes the SPF the most long-standing, comprehensive and transparent survey of the aggregate euro area economy.

The past twenty years have seen a wide variety of economic conditions, including the Great Moderation, with relatively high economic growth and stable inflation, the financial crisis and, more recently, a prolonged period of subdued inflationary pressures. This article documents the evolution of the SPF through this changing economic landscape and what we have learned from it. The SPF remains as useful for economic analysis and as relevant to the monetary policy debate today as it was when it was first launched.

1 Introduction to the Survey of Professional Forecasters

Inflation expectations play a central role in the ECB’s economic and monetary analyses, particularly in view of its mandate to maintain price stability in the euro area. The ECB sets monetary policy with the aim of maintaining annual euro area HICP inflation at a rate below, but close to, 2% over the medium term. In this context, private agents’ inflation expectations can affect the economy, because they can influence economic decisions in areas such as saving, consumption and investment, as well as wage and price setting. The role of inflation expectations in determining actual wage and price inflation can be modelled in a forward-looking Phillips curve relationship. Similarly, financial market participants’ inflation expectations can directly influence the pricing of some financial instruments, such as nominal bonds, and thus directly affect the transmission of monetary policy to the real economy.^[1] In addition, inflation expectations also serve as a valuable cross‑check on the inflation outlook in the Eurosystem/ECB staff macroeconomic projections, which in turn inform monetary policy decisions. The ECB therefore closely monitors private agents’ inflation expectations using a range of sources, not least the results of its own quarterly Survey of Professional Forecasters (SPF).

The SPF has measured inflation expectations and other macroeconomic expectations since the beginning of monetary union. At the time of its launch, the SPF was the only gauge of private sector macroeconomic expectations for the euro area as a whole. The survey collects information on the expected rates of consumer price inflation, wage growth, real GDP growth and unemployment in the euro area at several horizons, ranging from the current year to the longer term. In addition, respondents provide expectations for exogenous variables underpinning their forecasts, such as the oil price and the exchange rate, and qualitative comments that enrich their quantitative forecasts. Thus the overall survey results provide a comprehensive depiction of experts’ aggregate assessment of the macroeconomic outlook.

Expectations are sampled at different horizons for different purposes. In the shorter term, expectations are collected for the current and the following two calendar years, as well as two rolling horizons, one year and two years ahead of the latest available data. The evolution of shorter-term expectations allows us to track how the professional forecasting community is assessing new information about the shocks hitting the economy, e.g. from the incoming data, as well as learning about the effect of a given shock. In this context, calendar year point forecasts can easily be compared with those published in other surveys or with Eurosystem/ECB staff macroeconomic projections. The rolling horizons are better suited to measuring how perceptions of risk and uncertainty evolve over time, because these abstract from the natural decline in uncertainty that tends to occur as the forecast horizon shrinks. Finally, longer-term expectations can reveal information about the perceived steady state of the economy. In particular, longer-term inflation expectations can tell us about confidence in meeting the inflation objective.

Probability distributions provide a quantitative assessment of risk and uncertainty. A distinguishing feature of the SPF is that for inflation, core inflation, unemployment and real GDP growth, expectations at all horizons, including the longer term, are collected not just in the form of point forecasts, but also probability distributions. This allows a quantification of forecast uncertainty and of whether forecasters consider the uncertainty to be broadly balanced around their point forecast or skewed towards the upside or the downside.

The respondents to the survey are expert economists working in either financial or non-financial institutions, using different forecasting methodologies. Approximately 55 responses are received each quarter, which is relatively high compared with other expert macroeconomic surveys for the euro area as a whole. The majority of respondents are from financial institutions, although a significant number of economic research institutions also contribute. Since 2008 a special survey conducted every five years has provided additional insight on how the expectations reported in the SPF are formed (see Box 1).

High data transparency and economic relevance have led to considerable research based on the SPF. All the quantitative data, including microdata, are published on the ECB’s website each quarter, along with the summary report. This has helped stimulate substantial academic research into the SPF and what we can learn from its results. For a selective summary of such research, see Box 2.

This article looks at twenty years of results from the SPF and what we have learned from them. Section 2 explores the extent to which co-movement of expectations for pairs of variables is informative of the underlying economic relationships. Sections 3 and 4 examine, respectively, the point forecast results and the risk parameters implied by the probability distribution functions (PDFs), while Section 5 considers what we can learn from the longer-term expectations. Section 6 concludes.

2 What do SPF results reveal about underlying economic relationships?

Forecasters typically form their expectations on the basis of economic concepts and relationships. Differences in perceived underlying relationships can thus be an important source of differences between forecasts and need to be considered in the assessment and communication of expectations. For instance, relationships such as the Phillips curve or Okun’s law are building blocks of many macroeconomic models and often shape the common economic thinking of the professional forecasting community. The ECB reviews and uses such relationships in its economic analysis,^[7] and it is important to understand whether SPF expectations are also generated in line with corresponding economic relationships. So what are the conventional ways of specifying these relationships and to what extent can they be tested using SPF data?

The Phillips curve connects price or wage movements with measures of economic slack, such as the output or unemployment gap. A widely used version is the New Keynesian Phillips curve that can be expressed in the following equation:

π t = c + κ u t - u - + γ E t π t + 1

where current inflation ( π t) is a function of the unemployment gap, defined as the difference between the unemployment rate u t and its structural rate u - (often represented by the non-accelerating inflation rate of unemployment, NAIRU), and a term for expected future inflation E t ( π t + 1 ). As well as the unemployment gap, economic slack can also be measured using the output gap, i.e. the difference between actual output and a measure of potential output. Some versions of the Phillips curve (especially earlier versions) do not include expected inflation. A similar concept exists for wages, where price inflation is replaced by a measure of wage inflation.

Okun’s law describes the relationship between the unemployment rate and GDP growth. A widely used version relates changes in the unemployment rate to real GDP growth, as in the following equation:

∆ u t = c + κ ∆ y t

where Okun’s coefficient κ is negative and describes the strength of the relationship. If real GDP growth is one percentage point higher, the unemployment rate falls by κ percentage points. Another widely used formulation relates the unemployment gap to the output gap.

The results of the 2018 special survey indicate that point forecasts for the key variables tend to be jointly determined. Participants had the opportunity to specify whether any joint determination of their forecasts was done in a formal manner, e.g. within one model, or more informally, e.g. through expert consideration of, and judgement applied to, model outputs. Across the different pairs of economic variables, on average over 80% of respondents indicated that their point forecasts were determined jointly, and predominantly in a more informal manner (see Chart 1). By contrast, only about 40% of respondents indicated that they prepare their PDFs jointly.

Phillips curve and Okun’s law relationships were primarily used to inform revisions to forecasts at medium-term horizons (from one to three years ahead). This is consistent with there being a greater role for idiosyncratic shocks in the short term (the next twelve months), whereas longer-term expectations would be more influenced by views on the structural parameters of the economy. Similarly, the proportion of forecasters indicating that they revised the HICP inflation PDFs jointly with either their GDP growth or unemployment rate PDFs was greatest at the medium-term horizon, although the overall shares of positive responses were lower than for the point forecasts.

Indeed, these macroeconomic relationships can also be seen in the aggregate data. A scatter plot of aggregate point forecasts from the SPF shows a negative relationship for two-year-ahead forecasts of inflation and the unemployment rate, which is consistent with a Phillips curve relationship (see Chart 2). However, the correlation between the two variables significantly weakened in the post-crisis period, which is consistent with the evidence presented in Box 3.^[8] Moreover, aggregate two-year-ahead forecasts of real GDP growth and changes in the two-year-ahead forecasts for unemployment exhibit a negative correlation, which is in line with Okun’s law. However, the negative correlation was not statistically significant before 2008 and has strengthened considerably in the post-crisis period. This observation is also present in realised data, which show a strong increase in the correlation between GDP growth and unemployment in the post-crisis period. The effect has been especially strong since the start of the recovery and points towards a possible structural change in the unemployment-GDP relationship.^[9]

3 How have SPF point forecasts evolved over time?

HICP inflation and real GDP growth outturns have been affected by strong and persistent shocks over the last ten years. In the first ten years of the SPF, HICP inflation was mostly a little over 2%, and was modestly, although persistently under-predicted by the aggregate forecast of the SPF, while aggregate real GDP growth forecasts either over-predicted or under-predicted the actual outturns. The last ten years have seen much greater swings in these variables and have been more characterised by overestimation of both HICP inflation and real GDP growth (see Charts 3 and 4).

SPF forecasts have typically exhibited strong reversion to trend. Despite the magnitudes of the gyrations in both HICP inflation and real GDP growth, SPF forecasts have tended immediately to revert back towards a long-run trend (exceptions being real GDP growth expectations in 2008 and 2011, when the sharp fall in the growth rate was expected to continue before recovering). In general, the further the data were from their perceived trend, as proxied by the longer-term expectation, the stronger was the movement expected back towards that trend. In other words, there is a strong negative relationship between deviation from trend, and the change expected over the next two years (see Chart 5). This relationship holds more tightly for inflation, which could be interpreted positively as an indication of the strength of the inflation anchor.

As well as the point forecasts themselves, the constellations of forecast errors can be informative about potential changes in underlying economic relationships. For example, a historically unusual pattern of forecast errors for wage growth and unemployment beginning in 2013 may point to a structural break in euro area labour market dynamics. Until 2013, SPF forecast errors for wage growth tended to mirror those for unemployment: episodes of weaker-than-expected wage growth coincided with a higher-than-expected unemployment rate and vice versa. After 2013, however, that pattern changed and both wage growth and the unemployment rate were jointly overestimated (see Chart 6). Similarly, Eurosystem/ECB staff projections of compensation per employee growth during this period consistently over-predicted, while forecasts of employment growth under-predicted.^[11] This might suggest that, even though the amount of slack in the labour market (as measured by the unemployment rate) turned out to be less than expected, other factors kept wage growth subdued.^[12]

4 How have SPF probability distributions performed over time?

PDFs allow survey respondents to express their views on the complete range of outcomes and are thus a valuable complement to point forecasts. Forecasters assign probabilities to outcomes of each variable at each horizon in an interval half a percentage point wide.^[18] Each quarter the SPF reports the aggregate probability distribution, i.e. the average of all probabilities reported for each interval. These distributions, at both the individual and the aggregate level, allow a quantitative assessment of perceived risk and uncertainty, adding the extra dimension of information to the point forecasts.

The aggregate probability distributions for the individual variables have seen dynamic and often large movements over the years (see Chart 7). In particular, there are some notable differences in the PDFs reported in the first ten years and the latest ten years of the SPF, which broadly correspond to the pre- and post-crisis periods. These reflect not only differences in the location of the PDFs, echoing differences in the point forecasts, but also differences in the shape of the PDFs. For example, even though at the two-year horizon the perceived deflation risk has vanished, the probability assigned to relatively low inflation outturns (<1.5%) is still quite elevated, as the PDFs have a more negative skew.

Forecasters’ assessments of uncertainty seem to have increased permanently in 2009, across all variables and horizons. Aggregate uncertainty is typically measured by the width of the aggregate PDF. This, in turn, is determined by two factors: how uncertain each forecaster is, i.e. the width of the individual PDFs being aggregated, and how much forecasters disagree about the most likely outturn, i.e. the extent to which the individual distributions are centred differently. One striking feature of the SPF data is that, across all variables and horizons, there was a step increase in forecasters’ individual uncertainty at the time of the financial crisis (see Chart 8) which has persisted ever since. The effect of disagreement on aggregate uncertainty, on the other hand, has been much more varied across economic variables, forecast horizons and time.

Even after the 2009 step increase in forecasters’ uncertainty, the two-year-ahead PDFs underestimated subsequent data volatility. Over the last twenty years, the changes over two years in rates of inflation, real GDP growth and unemployment have tended to be much larger than the widths of the two-year-ahead PDFs. This can be seen from the concentration of the proportion of outturns in the first and last deciles (see Chart 9, right panel). For instance, over 30% of unemployment rate outturns came from the upper decile of the distribution expected two years previously. If the PDFs were good descriptors of the true, underlying distributions, then the same share of outturns (10%) would come from each of the deciles of the respective PDFs. Moreover, many of those tail outturns occurred after forecasters had widened their PDFs in 2009. This suggests that, while they believe economic developments to be intrinsically a little more uncertain than during the Great Moderation period before 2008, much of the volatility over the last ten years was driven by shocks which have been surprising in their magnitude, frequency or persistence.

Forecasters have actively adjusted their view on the balance of risks, as well as revising their point forecasts. The asymmetry of a forecast probability distribution is indicative of the balance of risks that the forecast embodies (see Box 5). Forecasters have adjusted their risk balances dynamically in response to the evolving economic outlook, especially for their expected inflation PDFs (see Chart 10). In 2006, as inflationary pressures appeared to be building, forecasters moved the balance of risks around their expectations for inflation in two years’ time strongly to the upside, but without much change in their point forecasts. In the survey in the first quarter of 2009, the aggregate point estimate fell by 0.3 percentage point (the largest quarter-on-quarter revision) and the balance of risk measure moved from +0.1 to -0.9. In the survey in the second quarter of 2009 the point forecast was revised down by only 0.1 percentage point, whereas the balance of risk measure moved further downwards very strongly to -2.7, before recovering over subsequent survey rounds, while the point forecast remained stable. In contrast, in the period 2013-2014, when HICP inflation was falling, point forecasts were progressively revised downwards, with little change in the balance of risks metric.

5 What can we learn from longer-term expectations?

Longer-term expectations in the SPF may provide information on professional forecasters views on the steady state of the economy. The steady state is important because this is what the economy reverts to once the effects of past and current shocks have died out and can evolve in line with structural trends in the economy. In the SPF, the longer-term horizon corresponds to a horizon about five years ahead. It might be reasonable to assume that most shocks hitting the economy are not that persistent, and so, by that horizon, the impact of any shock should have faded. If this were the case, then longer-term expectations should reflect only a view on the structural characteristics of the economy, such as the potential growth rate or the NAIRU.

However, the answers to the 2018 special survey suggest that the picture is more nuanced, i.e. that some shocks might be very persistent. SPF participants were asked to clarify how their long-term forecasts should be interpreted. While many respondents said that their long-term expectations did have a structural interpretation, even more said that this was only sometimes the case, and some suggested in their qualitative comments that the impact of some shocks may be more persistent, in which case five years would not be enough to return to the steady state (see Table 1). Indeed, the average longer-term unemployment rate expectations of those responding “no” to a NAIRU interpretation do appear to be more cyclical than for those responding “always”.

The impact of perceived persistence of shocks is most visible in the evolution of long-term unemployment expectations. The sharp increases in actual unemployment and short-term unemployment expectations in the aftermaths of the financial and sovereign debt crises were also accompanied by significant upward revisions of long-term expectations (see Chart 11). Such a link between expectations across horizons can be explained using the concept of hysteresis, whereby temporary demand or supply-driven increases in unemployment can, for instance via duration effects, have persistent effects in terms of a higher equilibrium unemployment rate. In this respect, the survey results for the last few years suggest that forecasters also allow hysteresis to work symmetrically: the decline in short-term expectations since 2013 was accompanied by a decline in long-term expectations in roughly the same proportion as the previous upward movements. As a result, the long-term unemployment expectations display more cyclicality than would normally be expected from structural unemployment.

The extent to which longer-term inflation expectations remain anchored to the central bank’s target can indicate confidence in the central bank achieving its target. The degree of inflation anchoring can be measured in different ways. These are often based on the sensitivity of longer-term expectations to short-term developments, such as the element of surprise in data outturns, or shorter-term expectations. This was comprehensively assessed in the report of the Low Inflation Task Force.^[20]

Data from the quarterly survey suggest that euro area inflation expectations are well anchored. Over the last twenty years the longer-term inflation expectations in the SPF have been in the range of 1.8 to 2.0%, a range which could be interpreted as broadly consistent with the ECB’s objective of keeping inflation below, but close to, 2%. Moreover, longer-term expectations have been more stable than the average HICP inflation rate, which might otherwise have been taken as a guide to inform a naïve, backward-looking forecast (see Chart 12). In recent years, longer-term inflation expectations have been gradually and steadily recovering from the low recorded in the first quarter of 2015, i.e. the survey round prior to the announcement of the public sector purchase programme. In the fourth quarter of 2018 survey round, the longer-term expectation for HICP inflation stood at 1.9% and the longer-term expectation for HICP inflation excluding energy, food, alcohol and tobacco stood at 1.8%. However, forecasters’ balance of risks around longer-term inflation expectations still remains clearly skewed to the downside, as it has been since the sharp fall in annual inflation in the second half of 2008, although it has been improving.

The ECB’s inflation objective has remained the key factor informing SPF respondents’ longer-term expectations. The 2018 special survey repeated a question from 2013, asking survey participants what information they used to inform their longer-term inflation expectations. Relative to the previous survey, there were notable increases in the shares of respondents indicating the ECB’s inflation objective, trends in financial market based measures of inflation expectations and trends in wages, whereas there were notable declines in the shares of respondents indicating trends in monetary aggregates and other variables (see Chart 13). Given that trends in actual inflation, wages and longer-term market expectations have clearly weakened in recent years, the role of the inflation objective as an anchor of expectations has been brought to the fore.

6 Conclusions

The ECB’s SPF is the most long-standing, comprehensive and transparent survey of the aggregate euro area economy. In particular, relative to other surveys, the PDFs in the SPF allow a complete, quantitative evaluation of perceived risks to and uncertainty surrounding the outlook. This provides an additional valuable dimension to complement the point forecasts and form a more complete economic assessment of the information received in the quarterly survey, for example when considering the pace and nature of the economic normalisation following the financial crisis.

It is helpful as a cross-check not only of the ECB’s/Eurosystem’s own macroeconomic projection profiles but also of the fundamental economic relationships underpinning them. The twenty years of the SPF have generated a rich dataset which can be used to address topical economic questions and help inform monetary policy debate. For example, considering the SPF data for different variables together, we can deduce how the professional forecasting community assesses the relationships between key variables, such as growth and inflation or unemployment and wage growth, to be evolving. Thus, the SPF remains as useful for economic analysis and as relevant to monetary policy debate today as it was when it was first launched twenty years ago.