US: Is there rising wage growth on the horizon?

Mede-auteur: Wouter Verbeek van Berenschot B.V.

The spectre of inflation is haunting markets (figure 1) as strengthening global growth has raised concerns about a concomitant rise in price pressures. The surge in volatility really got a grip on markets after the release of the January Employment Report, which showed that average hourly earnings rose 2.7% (y-o-y) in December and 2.9% in January. This is the fastest pace of earnings growth since the financial crisis. The strengthening in wages was to some extent already visible in the (broader) Employment Cost Index, which showed that wages and salaries rose by 2.6% annualized in the fourth quarter of 2017 (figure 2). This rise is also backed by survey evidence: the January release of the NFIB small business survey points at increasing wage pressure. The share of firms planning to raise wages rose to 24%, whereas even 31% of firms reported to have increased wages in the last three months (figure 3).

The acceleration in wages has fueled market speculation that inflation is rising from the dead. This would fully put the pressure on the Fed to speed up its tightening cycle. In its January statement, the FOMC already included a phrase that firming economic conditions “would warrant further gradual increases in the fed funds rate”. Certainly in light of recent data, this could be interpreted as four hikes rather than the three currently envisaged in the FOMC’s economic projections.

The FOMC sees the below-target readings of the past year as transitory and is confident that core inflation will pick up in 2018. It is banking on the premise of the Phillips curve: if unemployment is low, then wage and price pressures are around the corner and a gradual tightening of monetary policy is warranted. Even though the labour market has already tightened significantly, we would argue that one relatively high reading in average hourly earnings is only scant evidence of a material uptick in wages or core inflation. In fact, we’ve seen previously that similar upticks have been revised. The Fed’s optimism may be challenged later this year if it indeed proved to be a blip and not the start of an upward trend. For that to conclude, more data is needed.

This Special is therefore not intended to change our house view on inflation or Fed policy, but we do see it useful to assess the possible impact of labour markets taking a more dramatic turn. We use a two-step approach. We will first assess labour market tightness using data on labour market flows within the framework of the so-called Beveridge curve, which measures the relationship between job openings and unemployment. In the second part of this Special, we employ a wage model to assess whether we could then expect more wage pressures in the near future.

The Beveridge curve is a useful tool to get an idea of how tight the labour market really is. It displays the historical combinations of the unemployment (supply of labour) rate and vacancies (demand for labour), both expressed as a percentage of the labour force. The cyclical changes in labour demand and supply are projected by movements along an ellipse curve as in figure 4. The most recent cycle starts in 2007 at the top of this curve, right before the start of the Great Recession. Lehman fell, vacancies dropped, layoffs accelerated, and within one-and-a-half year we have moved to the lower-right end of the ellipse. With an unemployment rate of 10.6% and a job openings rate of 1.7%, the US labour market hit rock bottom in October 2009 (i.e. the bottom-right quadrant). The path of recovery to current levels (i.e. the top-left quadrant) took eight long years. The latest observation is marked by the blue dot.

The Beveridge curve framework has three benefits over a simple assessment of the unemployment rate: 1) it looks at the combination of labour demand and supply, 2) it shows how labour market cycles behave over time and possibly says something about future dynamics, and 3) it can be linked to labour market flows, which provide additional information on the current situation versus a steady-state situation of the labour market.

Steady-state Beveridge curves

Let’s focus on the third. Even though there is all this new technology and advances in artificial intelligence that should – in theory! – improve the matching of labour supply and demand, Hobijn and Sahin (2013) have actually argued that the matching efficiency of the US labour market deteriorated due to the crisis. They ascribe this to economic factors related to the Great Recession, such as the decline in voluntary quits, the poor performance of the construction sector and the extension of the unemployment insurance. In support of their findings, we also see an outward shift of the steady state Beveridge curve in the wake of the crisis, yet it has to be said that the latest labour market readings are somewhat more in line with the pre-crisis steady-state Beveridge curve.

The central question is then how much gains in matching efficiency is the US economy able to realise in the next few years. For this looks to be necessary to extend the current economic cycle. However, with almost two decades of observations, it doesn’t look very likely that the labour market is able to break out of the solid patterns illustrated in figure 5. Given these patterns of job flows, we have calculated a possible future trajectory for the US labour market going forward (figure 6). The technicalities of these calculations are shown in the first part of the appendix. It is important to note that this trajectory is not our baseline forecast, but simply one based on the patterns seen in two previous labour market cycles. It’s not looking pretty, but this is what you typically get to see in the late stages of the economic cycle. For instance, we have stressed previously that our yield curve model predicts a recession probability of 31% at a 17-month horizon. Even though the recent steepening of the Treasury yield curve should produce somewhat lower probabilities now, it’s still markedly higher than what the NY Fed currently expects (see Marey, 2017).

These predictions in combination with the results presented above at least illustrate that once the economy gets closer to the peak of its cycle, the risks become tilted to the downside. But we would argue that also a simple trend-related forecast of the consumer survey data of the Conference Board predicts a turnaround of labour market conditions in the third quarter of this year. The survey asks consumers whether jobs are plentiful or hard to get. These data cover a much longer time span (1978-2017) than the jobs openings data and, consequently, this ARIMA model takes into account more labour market cycles (figure 7).

And what about the underemployment rate?

An argument against this admittedly cloudy outlook is that there are still more than 5 million people outside the labour force who would like to work (figure 8). During the Great Recession, many unemployed ceased their job hunts and decided to withdraw from the labour force. Now that the labour market has improved markedly, many of these discouraged workers have re-entered the labour force. Indeed, the amount of people outside the labour force that want a job has seen a rapid decline, which illustrates that labour market tightness is having a positive effect in resolving labour market slack present in the non-labour force. But even if we take this specific group into account, we still see that labour market tightness is closing in on historical peak values at the moment (figure 9), although a ‘glass half full’ kind of person would argue that such peak levels were persistent for several years.

In a ‘normal’ Phillips-curve framework, persistent labour market tightness should already have had an upward effect on wage growth and inflation. That hasn’t happened (or, at least not as evidently as before the Great Recession)_and suggests that the curve is flat, on a lower level, or both.

There is a lot of debate among economists as to what has caused the Phillips curve to break down. Summers argues here that the relative bargaining power of employers has increased as technological progress gave them the upper hand in keeping wages down. In contrast, Daly and his colleagues argue that during the recovery from the Great Recession high-wage baby boomers have retired, while low-wage workers have taken up full-time jobs (or multiple part-time jobs!) in return. This implies that compositional shifts rather than labour market slack has been weighing on wage growth. In this case, the flat Phillips curve is temporary and the curve should steepen in the upcoming years as such cyclical composition effects dissipate (see also here).

Baseline scenario

In order to predict what US wages might do, we have estimated a wage model for the US, based on the Wages and Salaries index that is part of the report on the well-known quarterly Employment Cost Index. The details of this model can be found in Appendix 3. We use two models to estimate an upper and a lower bound for US wage growth. In both cases, we do see a reversal in wage growth trajectory in late 2019. The most complete model (#3) predicts wages to peak at 3.0% year-on-year, whereas the alternative model (#4) suggests that wage growth could reach a rate of 3.3% (figure 10). This is somewhat lower than the peaks seen prior to the Great Recession, but a little more in line with the Fed’s 3-4% tolerance range.

Scenario 2: a repeat of the dotcom crisis

Finally, we also ran a scenario in which the dotcom crisis would repeat itself in terms of GDP growth losses. We decided to pick this crisis as the preceding economic situation shows resemblance with current conditions: asset prices were highly elevated and a correction looked more or less unavoidable. However, even with the recent rise in market volatility in mind, this simulation is to be seen as a risk-scenario.

The effects of the financial market crash spilled over to the real economy and consumer and business confidence hit rock bottom. The negative sentiment got a grip on economic output and GDP growth declined from 1.0% q-o-q in 2001Q1 to -0.6%, -1.0% and -0.9% in the subsequent quarters. Labour market conditions deteriorated as well: the unemployment rate rose from 4.2% in 2001Q1 to a peak of 6.3% in the second quarter of 2003. Meanwhile, wages continued to rise and only just declined in late 2002 – well after the first recession signals. If we translate these patterns (in terms of GDP losses, changes in unemployment, diminishing labour market tightness and slightly slowing inflation) to the current situation, we end up with a scenario for wage growth as illustrated in figure 11. Even in this scenario, wage growth remains above 2.0% in the upcoming twelve quarters. In contrast to the forecast in scenario 1, wage growth is then expected to stabilise at slightly lower levels in the upcoming two years. Just as after the dotcom crisis, the slow response in wage growth to a potential bust is caused by a continuing feedthrough of current labour market tightness on future wage growth developments.

The spectre of inflation is haunting markets and the Fed is for once risking to get behind the curve. Even after the recent volatility in equity and rates markets, the OIS curve is already closing in on the Fed’s median projections. A significant pick up in US wage growth would fully put the pressure on the Fed to speed up its tightening cycle and there is an actual risk that the market will already do this for the Fed by virtue of a steepening money market curve (see figure 12).

Even though we know that economic models have their limitations, the ones that are presented in this Special indicate that such a firming of wage growth may be well under way. Our two models predict a wage growth to peak at 3.0 to 3.3% in the next two years. While this is higher than current levels, it’s on the low end of the Fed’s preferred range of 3 to 4%. At the same time, these models also show that we’re already getting close to the peak of the economic cycle. This means that the risks to economic growth, and eventually wage growth and inflation, are becoming increasingly tilted to the downside.

Appendix 1: steady-state Beveridge curve

The steady-state Beveridge curve is related to search and matching theory (see for an overview Petrongolo and Pissarides, 2001). The labour market is in a steady state when the growth rate of the labour force g^(lf) is equal to the growth rate of employment g^(e):

The growth rate of the labour force is equal

in which η_t is the number of people entering the labour market during period χ_t as a fraction of the number of people in the labour force at the beginning of period t and c_t representing the number of people who exit the labour force during period t divided by the number of people in the labour force at the beginning of period t.

For the growth rate of employment we estimate the following equation (see Erken, Van Loon and Verbeek (2015) for the derivation):

where v_t is defined as the number of job openings in period t as a ratio of the labour force. The hiring rate (h_t) measures all people finding a job as percentage of all job openings during. The separation rate is defined as all separations (i.e. quits and layoffs) during period t as a fraction of the employed labour force.

Substitution of equation (A.2) and (A.1) gives

Next, we link labour market inflow and outflow to our indicator for labour market tightness – the vacancy-unemployment (v/u) ratio. and estimate four gross flows that determine the growth rates of the labour force and employment (Barnichon et al., 2012). Thus, we estimate for :

Taking the exponent of equation (A.5) and substituting the estimated parameters, assuming ε_t = 0 and replacing the unemployment rate u_t with the unknown steady-state employment rate u* gives:

(A.6 to A.9) can be substituted in (A.5) giving:

Equation (A.10) is used to calculate the steady-state levels for the unemployment rate (u*). These growth rates are, subsequently, determined by the estimated parameters from the rate equations depicted in Table A.1. Since equation (A.10) has no explicit solution, the steady-state Beveridge curve is determined by solving equation (A.10) for u* at different points of v_t.

Parameter estimations

In order solve equation (A.10), we will have to empirically estimate the equations (A.6) to (A.9). We use three databases of the Bureau of Labor Statistics (BLS). First, we take vacancy data from the Job Openings & Labor Turnover Survey (JOLTS) database, labour market flows data from the Current Population Survey (CPS) and unemployment levels data Current Employment Statistics (CES). Table A.1 shows the estimation results for three samples with the Great Recession being the pivoting point: the pre-crisis sample (2000m12-2008m10), the post-crisis sample (2008m11-2017m11) and the total sample (2000m12-2017m11).

In all estimates, the separation rate (s_t) is negatively correlated with the v/u-ratio. The intuition is straightforward: if labour tightness increases (v/u-ratio increases), there are less layoffs. The hire rate (h_t) is also negatively correlated with labour market tightness. Again, the effect is what we expect: if the labour market tightens, it becomes more difficult for employers to fill job openings and hence the number of hires per vacancy (i.e. the hire rate) drops.

In our estimates, labour market exits (χ_t) is negatively correlated with labour market tightness, which captures the so-called ‘discouraged worker effect’. As labour market prospects deteriorate and the v/u ratio decreases, job finding opportunities become slim and search costs increase, which consequently discourages people to actively search for work. In this case, the exit rate increase, and older discouraged workers choose early retirement, whereas many discouraged youngsters go back to school.

Labour market entry flows (η_t) are also negatively correlated with labour market tightness, which is counterintuitive in the sense that people enter the labour market when job perspectives deteriorate (v/u ratio declines). Nevertheless, it is in line with the so-called ‘added worker effect’, which picks up the entry of household members who were previously not active on the labour market and start looking for work in downturns to minimise the negative effects on household income (Lundberg 1985).

Appendix 2: expected trajectory

The deviation from the steady-state curve is given by . Deviations from the steady-state curve during the last two cycles (January 2001- September 2010) will provide the fundamentals for our expected trajectory. This implies that we use information on the labour market behaviour after the dotcom crisis of 2001 and the Great Recession in 2008/2009. Since the peak in unemployment differs between each crisis, we scale the observed unemployment rates linearly to range from the equilibrium unemployment (the lowest observed unemployment rate) to the expected return point û^r, which we define by the unemployment prediction where and , instead. Next, the following curve is fitted:

This curve has two constraints. First of all, the deviation of ν^t at the (forecasted) return point û^r is constrained at 0. Furthermore the deviation of ν^t at the steady-state unemployment u^eq is constrained at 0. Now we minimise the Sum of Squared Errors with respect to the parameters a, b and c, given these constraints. This optimisation problem is given by:

and solved using a GRG2 method. The corresponding Beveridge curve vacancy rate () is calculated and the expected deviation is added.

Appendix 3: US wage model

Starting point for our US wage model is the following a simple textbook Phillips curve:

where w represents nominal wages, c is a constant, p is the price level and u the unemployment rate. Blanchard and Katz (1999) derive the following equation:

where p^e are inflation expectations and y is the productivity level. We use (A.15) to arrive at our final model:

where p^c is core inflation (inflation minus food and energy items), u*is the non-accelerating wage rate of unemployment, θ is a measure for labour market tightness, T is a time trend, D^crisis is a recession dummy, t is quarter, Δ denotes the year-on-year mutation.

In order to estimate equation (A.15), we use data from various sources (Table A.2) on a quarterly basis. We use three information criterions (Hannan Quinn, Akaike and Schwarz) to determine the optimal lag structure. Ideally, we would like to use the v/u ratio as our tightness indicator, but series on job openings only start in 2000, which implies that we would have to estimate our model for a very short time period. Alternatively, we use data from a consumer survey of the Conference Board.

Table A.3 shows the estimation results of our model. Model (1) shows the basic estimation with all variables included. The independent variables all have a statistically significant impact and show the correct signs. The inflation gap (α₁) shows a positive impact, which indicates that a one percent rise of the lagged core inflation versus inflation expectations raises wage growth by 0.09ppts. The negative impact of (α₂), which denotes the error-correction term (ECM), is also logical: when real lagged wage levels are lower than productivity levels, higher wage growth will ensure that wages catch-up to these higher productivity levels. One percent higher labour productivity growth (α3) leads to higher wage growth by 0.12ppts. The unemployment gap (α₄) is negative: if unemployment vis-à-vis the non-accelerating wage rate of unemployment declines by 1%, the labour market become tighter and wage growth picks up by 0.56ppts. Labour market tightness measured by the plentiful jobs minus hard to find jobs index (α₅) is positively correlated with wage growth. Finally, there is a negative significant trend on wage growth (α₆), which might reflect less union bargaining power over time in setting wages. Other reasons for a negative trend in wage growth might be continuing technological progress resulting in higher capital gains, or the impact of globalization which has put wages in the US under continuing stress as a result of more foreign competition. The crisis dummy variable (α₇) is insignificant. Figure A.2 shows that our model has a very good fit, explaining 89% of variance from 1983 up to and including 2017.

One problems is that the unemployment gap and labour market tightness variable seem to work against each other. Therefore we ran a sensitivity regression in column (2) in Table A.3 with only the labour tightness variable of the Conference Board included. The model estimates remain stable, although the significance of the labour tightness variance jumps. A more severe problem, however, is the very low Durbin-Watson statistic, which indicates that we are dealing with a significant amount of serial correlation and biased estimates. A way to obviate serial correlation is the use of a lagged dependent variable, which is done in column (3)

The autoregressive (AR) term in the model has quite a lot of impact on the estimation results. First and foremost, the AR term solves the problem of autocorrelation as the Durbin-Watson statistics is now in the range of 2.0. The coefficient found for the autoregressive term is 0.68, which means that 68% of the wage growth in the previous quarter feeds into the current quarter. Moreover, most of the other variables still have a significant effect on wage growth and show the correct sign, although the magnitude of the impact is much lower than in the estimations that were subject to autocorrelation. The error-correction mechanism and the trend variable are the exception and do no show a statistically significant effect. If we remove the unemployment gap as a robustness test, we see that all variables again show a significant effect. The fit of model 3 is illustrated in figure A.2. We use the models (3) and (4) as lower and upper bounds to forecast wage growth in the US going forward.