Forecasting near-term trends in the labour market – Bank Underground

Harvey Daniell and Andre Moreira

The latest developments in the labour market are often central to monetary policy decisions. We outline a framework for mapping labour market indicators to near-term employment and pay growth, drawing on established insights from the ‘nowcasting’ literature. The key benefits of our approach are: the ability to map a range of ‘soft’ and ‘hard’ indicators of different frequencies to quarterly official data; the empirical determination of how much weight to place on each indicator; and the ability to shift those weights flexibly as more data become available. This framework beats simple benchmark models in our labour market application.

Understanding the latest developments in the labour market is often key for monetary policy decisions. In May, for example, the Monetary Policy Committee linked further tightening in monetary policy to, among other things, the tightness of labour market conditions and the behaviour of wage growth.

However, official data on the labour market are published with a lag. To improve understanding of current conditions when setting policy, it’s necessary to extract the signal from more contemporaneous indicators to ‘nowcast’ – ie to predict current and near-term developments in – the labour market.  

What are the key insights of our approach?

1: Hard and soft indicators, of different frequencies, can be mapped individually and directly to the target labour market variable.

The growth of high-frequency data over the past decade or so means we have indicators of the labour market at quarterly, monthly, and even weekly frequency. These indicators can be mapped individually to a target variable, such as quarterly employment growth, using ‘mixed-data sampling’ (or ‘MIDAS’) techniques. MIDAS techniques are ‘essentially tightly parameterised, reduced form regressions that involve processes sampled at different frequencies‘.

These techniques have the advantage of being able to handle data of different frequency, avoiding the need to convert variables to the same frequency, such as reducing weekly data to quarterly. This means we avoid any loss of information (or use of additional assumptions) that transforming variables to the same frequency entails.

A further benefit of our particular approach is the ability to map each indicator individually to the variable of interest, before combining those into an overall model nowcast. In other words, we start by obtaining a different nowcast from each indicator, which is often of interest to policymakers, as it allows us to discuss features such as the dispersion across individual indicators.

As an example, Chart 1 illustrates the nowcasts for quarterly employment growth from MIDAS regressions using a range of labour market indicators. The grey lines illustrate the individual nowcasts, which track the target variable, in the blue line, fairly well, despite the volatility in the latter.

Chart 1: Indicator-based nowcasts for quarter on quarter employment growth (per cent) ^(a)

(a) Indicator-based nowcasts are based on Bank of England Agents’ scores, the Lloyds Business Barometer, ONS/HMRC PAYE payrolls growth, S&P Global/CIPS Purchasing Managers Index and KPMG/REC Report on Jobs.

2: The steer from different indicators can be combined into one overall view, where the weights are informed by the information content of the relevant indicator.

Different indicators often point to different nowcasts for the same variable, so it’s vital to know how much weight to place on each indicator. In our approach, we weight the steer from each indicator according to its relative forecast performance in the recent past – a common approach in the forecasting literature, which we find also works well in this application to the UK labour market. In other words, more accurate indicators receive a higher weight.

Chart 2 and Chart 3 plot a measure of forecast performance for a range of indicators of quarter-ahead employment and pay growth. They show that ‘soft’ indicators like the monthly employment balance of the S&P Global/CIPS Purchasing Managers Index and the monthly KPMG/REC Report on Jobs perform the best at predicting quarterly employment and pay growth, two quarters ahead. Crucially, though, the combined steer from all indicators outperforms relying on any individual indicator alone.

Chart 2: Forecast performance of indicators of employment growth, two-quarters ahead ^(a)

Chart 3: Forecast performance of indicators of pay growth, two-quarters ahead ^(a)

(a) ‘RMSE’ is root mean squared error, a standard measure of historical forecast performance. ‘BCC’ is British Chambers of Commerce quarterly economic survey. ‘Vacancies’ are ONS three-month vacancies growth. ‘Lloyds’ is Lloyds Business Barometer. ‘GDP’ is ONS three-month GDP growth. ‘Payrolls’ are ONS/HMRC PAYE payrolls and median pay growth. ‘REC’ is KPMG/REC Report on Jobs. ‘PMI’ is S&P Global/CIPS Purchasing Managers Index. ‘Agents’ are Bank of England Agents’ scores for employment and pay growth. ‘Job-to-job moves’ are ONS quarterly flows data. ‘Cycle’ indicator is sum of CPI inflation and the (inverse) unemployment gap from Bank of England Monetary Policy Reports.

It’s interesting to note that, within the range of individual nowcasts, soft indicators tend to perform much better than past observations of the ‘hard’ data – ie available monthly observations of the target ONS variable. Fundamentally, this is because the hard data are volatile, and are therefore not necessarily very informative even about their own trajectory. 

3: The weights placed on each indicator change over the data cycle, which is particularly important for the relative weights placed on ‘hard’ versus ‘soft’ indicators of the labour market.

When new data become available, the mapping between a given indicator and the target variable can also change, sometimes in quite an important way. So it’s crucial to re-specify and re-estimate the underlying relationships each time the information set changes. Acknowledging this, the model is also set up in a way that allows it to optimally shift the weight put on ‘hard’ information (such as ONS data) versus ‘soft’ information (such as business surveys) over the data cycle. This optimal weighing takes place in a simple second-stage OLS regression.

Taking a step back, updating the near-term outlook for an expanding information set is an important source of reducing forecast error, not least because monthly observations of the hard data, which the ONS also publishes, are a direct input to the calculation of the quarterly outturn.

Early in the data cycle, when no intra-quarter official data are available, the model tells us to place more weight on the timelier soft indicators such as the monthly business surveys. Once intra-quarter monthly hard data become available, however, these begin to receive a much higher weight.

Under this approach, the full model’s average forecast error progressively declines each time new data become available but, unsurprisingly, the improvement is particularly striking at the juncture where the hard indicators come in to supplement the steer from soft indicators. Chart 4 illustrates this point for a nowcast of employment growth.

Chart 4: Average forecast error for quarterly employment growth declines as more data become available

Viewed in this way, the framework can tell us both the marginal impact of new data on a point estimate of the nowcast, as well as the extent to which the new data also reduce uncertainty around that nowcast.

4: The approach significantly outperforms simple benchmark models for forecasting the labour market.

The key insights outlined above are important sources of reducing forecast error. For example, in a forecast for employment and pay growth two quarters ahead, the combined forecast in which a collective steer is taken from a range of indicators generally outperforms relying on any single indicator alone.

Moreover, the combined nowcast from the full model outperforms a simple autoregressive benchmark model – ie a simple model of quarterly employment/pay growth based on lagged employment/pay growth. This is generally true at all states of the data cycle, with the reduction in root-mean-squared errors increasing to around 65% prior to the release of the outturn.

Indeed, the model’s forecast performance at the quarter-ahead horizon is similar to the one-quarter ahead forecasts from successive Monetary Policy Reports. This reflects the fact that the approach outlined here formalises many of the existing heuristics that forecasters at the Bank employ, such as placing more weight on softer indicators early in the data cycle, and relying on a wide range of data.

Interestingly, performance at the one-year-ahead horizon of the combined model actually slightly outperforms successive Monetary Policy Reports. This is likely to reflect two factors: data-driven approaches to forecasting may outperform more structural models like those supporting the Monetary Policy Report at the one-year horizon; and Monetary Policy Report forecasts are conditional forecasts, whose forecast accuracy is only one requirement of the model among many other requirements.

Harvey Daniell and Andre Moriera work in the Bank’s Current Economic Conditions Division.

If you want to get in touch, please email us at bankunderground@bankofengland.co.uk or leave a comment below.

Comments will only appear once approved by a moderator, and are only published where a full name is supplied. Bank Underground is a blog for Bank of England staff to share views that challenge –or support – prevailing policy orthodoxies