Policy makers must judge the present state of the economy before the national accounts describe it. Nowcasting is the name given to forecasting the current or the just-completed quarter using information that arrives earlier and more often than the target itself. The approach formalised by Giannone et al. (2008) treats the nowcast as something to be updated each time a new release appears, and the methods that have followed are surveyed by Bańbura et al. (2013). Most applications use a large panel of monthly surveys, production indices and labour market data.

Many statistical systems cannot supply such a panel. Business surveys may not exist, industrial production may be published with a long delay, and monthly labour data may cover only part of the workforce. What these economies do have are records generated by the operation of infrastructure. Grid operators log electricity load continuously. Railways, ports, toll roads and customs authorities count wagons, containers, vehicles and declarations as they pass. These series are timely, frequent and not subject to sampling error, and interest in them grew sharply when conventional statistics were disrupted in 2020 and weekly activity indices built from high-frequency data came into regular use (Lewis et al., 2022).

This article sets out a procedure for turning electricity and freight indicators into a nowcast of quarterly output growth. It is written for analysts in central banks, finance ministries and statistical offices who have a short history of quarterly accounts and a small team. The procedure is proposed on the basis of published methods. It is not applied to any country's data here, and no forecasting performance is claimed for it.

Electricity is an input to almost all production and cannot be stored cheaply at scale, so consumption tracks activity closely in time. The relation is strongest for industry and weakest for households, whose demand is driven by temperature and by the calendar. The link between electricity and output also shifts slowly as the economy moves towards services, as equipment becomes more efficient, and as firms install their own generation, which removes load from the grid without any fall in production. Where supply is rationed, recorded consumption measures what the system could deliver and not what users wanted.

Freight indicators measure the movement of goods and therefore capture industry, construction, trade and agriculture more than services. Volume measures such as rail tonne-kilometres, port container throughput and heavy-vehicle toll counts relate directly to domestic activity. Freight rates are a different kind of series: they reflect the balance of demand and shipping capacity and have been used to proxy global demand, as in the index of real activity constructed by Kilian (2009). Customs declarations and electronic waybills, where the tax administration records them, extend the coverage to road haulage, which carries most domestic freight in many countries and is otherwise the least well measured mode. Volumes are affected by strikes, port congestion, weather and changes in the split between road and rail. Neither electricity nor freight says much about public administration or most private services, a limit on what any nowcast built from them can achieve.

The econometric problem is to relate a quarterly target to indicators observed monthly, weekly or daily, some of which are only partly available for the current quarter. Three families of models address it. Bridge equations aggregate each indicator to quarterly frequency, after extending it over the missing part of the quarter with an auxiliary forecast, and regress output growth on the aggregates (Baffigi et al., 2004). Mixed-data sampling (MIDAS) regressions keep the indicator at its original frequency and impose a parsimonious weighting function on its lags (Ghysels et al., 2007); Clements and Galvão (2008) added an autoregressive term and applied the model to output growth. When the frequency gap is small, as between months and quarters, the weights can be left unrestricted and estimated by least squares (Foroni et al., 2015).

The third family casts the problem in state space. Mariano and Murasawa (2003) showed how quarterly output can be treated as a monthly series with missing observations and linked to a common factor, and Bańbura and Modugno (2014) gave an estimation method for factor models with arbitrary patterns of missing data that also decomposes each revision of the nowcast into the contributions of individual releases. Comparisons of MIDAS with mixed-frequency vector autoregressions suggest that the approaches are complements, with relative performance depending on the horizon (Kuzin et al., 2011). Table 1 summarises the candidates considered in the procedure and the circumstances that favour each.

The procedure has seven steps, shown in Figure 1. They are described in the order in which an analyst would carry them out, and the loop back to the start reflects the fact that a nowcasting system is re-estimated whenever a new quarter of output data, or a revision to earlier quarters, is published.

A team of two or three analysts can run steps 1 to 6 with bridge and MIDAS models in open-source statistical software, and can add the factor model later. The larger demands are institutional. Grid and transport operators must agree to deliver data on a fixed schedule and to report changes in coverage, which is best secured by a written agreement. Where quarterly accounts are themselves compiled partly from electricity or freight data, as happens when direct measures of output are missing, the nowcast will appear more accurate than it is, because indicator and target share an input. The analyst should ask the statistical office how the early estimates are built before interpreting any evaluation.

The procedure has not been tested in this article, and several of its premises are empirical. The relation between electricity and output may be too unstable in economies undergoing rapid electrification or frequent rationing for any fixed-coefficient model to work. Freight volumes may lag output where inventories are large. Short samples limit both estimation and evaluation, and a model that performs well over a few dozen quarters may owe that to a single episode. The indicators cover goods production and trade far better than services, so that errors will be largest when services and industry diverge. Whether a combination built from these indicators improves on the autoregressive benchmark, and at which point in the quarter, can only be established country by country through the evaluation described in step 5.

Electricity and freight records give economies with thin monthly statistics a workable basis for nowcasting, provided the indicators are cleaned with care and the evaluation respects what was known at each date. The seven steps proposed here favour small models, an explicit benchmark, forecast combination and plain reporting of uncertainty. They describe a method that has yet to be validated, and the test of it lies in application to real-time data.

To be declared by the author(s) before publication.

To be declared by the author(s) before publication.