英国GDP月度预测的新证据

By Paula Bejarano Carbo, Rory Macqueen and Efthymios Xylangouras The Office for National Statistics (ONS) publishes a monthly estimate of gross domesticproduct (GDP) at a lag of around 40 days from the end of the month, reflecting the time ittakes to collate data on economic output. At the same time, policymakers and businesseshave an interest in knowing how the UK economy is performing as quickly as possible,especially during times of crisis. As a result, nowcasting – or forecasting in real time – GDPas accurately as possible is highly important. Since 2018, NIESR has produced its monthly GDP ‘tracker’ on the ONS GDP estimaterelease date, commenting on the latest data point and producing a ‘bottom-up’ forecast(i.e. constructed by aggregating sectoral forecasts) of economic output up to the end of thenext quarter (Kara et al., 2018). In Bejarano Carbo et al. (2025), we develop this further, inlight of new data series available since the pandemic, and the increasing demand for timelysectoral nowcasts. This article provides an overview of the methodology and results in Bejarano Carbo etal. (2025). We use a large dataset of public- and private-sector variables, including newlyavailable ‘real time indicators’, as inputs to 28 nowcasting models. Each of these modelsis estimated for 20 industrial sectors to generate sectoral GVA nowcasts, which are thenaggregated to produce a nowcast for overall GDP. We then apply empirical algorithms tocombine the sectoral nowcasts, hoping to generate better GDP nowcasts by putting moreweight on nowcasts from models which have been more accurate previously. By analysinghow the different models and model combination algorithms perform over our sample of57 months covering before, during and after the Covid-19 shock, we provide new evidenceon nowcasting UK GDP. Data Our forecast variable of interest is the month-on-month growth rate of UK gross valueadded (GVA). Our nowcast is constructed indirectly from the bottom up, by nowcastingmonth-on-month growth rates for each of the twenty constituent industrial sectors of GVA(e.g. agriculture, construction, manufacturing, wholesale and retail trade, etc.). In twentyindependent nowcasts, therefore, the target variable is the monthly growth rate of activityin the sector in question. We produce one-month-ahead pseudo-out-of-sample nowcasts – meaning that for each ofour 57 test period months from June 2019 to March 2024, we generate a forecast for theupcoming publication month based on data which would have been available one monthbefore first release. For example, the first estimate of January 2024 GDP was publishedon 14 March 2024; our pseudo out-of-sample nowcast is conditional on data available by13February 2024 (after the first estimate of December 2023 GDP was published). National Institute UK Economic Outlook – Summer 2025 The data that we condition these nowcasts on include lagged values of the target variablesas well as contemporaneous and lagged values of a wide range of external regressors. Thelatter are sourced from both public and private sector sources, and include variables relatedto business and consumer confidence, prices, financial markets, trade, weather, labourmarkets, and Covid-specific indicators. Models We estimate a total of 28 statistical models, described in detail in Bejarano Carbo et al.(2025). For each sector, we estimate: four univariate models, which produce nowcasts based onlyon lagged values of the forecast variables; two limited-information multivariate models,which produce nowcasts based on lagged values of the forecast variables and selectedexternal regressors, chosen based on our experience with the NIESR tracker and in otherforecasting roles; and one full-information model making use of our full dataset (167variables). For each sector, each of these seven nowcasts is made, firstly, without treatmentof outliers and then with, doubling the number of nowcasts to 14. We then consider bothrecursive estimation (in which all available data is used when estimating model parameters)and a rolling window estimation procedure (in which only the 36 most recent monthlyobservations are used), doubling our total number of nowcasts generated in each monthfor each sector to 28. Single-Model Results We evaluate how the above 28 models perform once aggregated across all sectors toderive a single-model GDP nowcast. From this exercise, we find that outlier treatment improves forecast accuracy, as measuredby Root Mean Squared Forecast Error (RMSFE), significantly during the Covid period(2020-2021) but has limited effects in the post-pandemic period (2022-2024). Amonguntreated models, univariate approaches perform poorly during Covid compared to big dataapproaches, but the simpler univariate models are among the best in the post-pandemicperiod. Forecast performance can be enhanced using a bottom-up approach relative to a top-downapproach for a simple univariate model. Specifically, the recursive bottom-up ARMA(1,1