英国月度GDP预测：基于大数据方法和预测组合算法的自下而上部门建模证据

Nowcasting MonthlyUK GDP: Evidencefrom Bottom-UpSectoral Modelingwith Big DataMethods andForecastCombinationAlgorithms Paula Bejarano CarboRory MacqueenEfthymios Xylangouras NIESR Policy Paper 46June 2025 About the National Institute of Economic and Social Research The National Institute of Economic and Social Research is Britain's longest establishedindependent research institute, founded in 1938. The vision of our founders was to carry outresearch to improve understanding of the economic and social forces that affect people’s lives,and the ways in which policy can bring about change. Over eighty years later, this remainscentral to NIESR’s ethos. We continue to apply our expertise in both quantitative and qualitativemethods and our understanding of economic and social issues to current debates and toinfluence policy. The Institute is independent of all party-political interests. National Institute of Economic and Social Research2 Dean Trench StLondon SW1P 3HET: +44 (0)20 7222 7665E:enquiries@niesr.ac.ukniesr.ac.ukRegistered charity no. 306083 Policy papers are written by members of the National Institute of Economic and Social Research tospecifically address a public policy issue. These may be evidence submitted to a public orparliamentary enquiry, or policy research commissioned by a third-party organisation. In all circumstances NIESR has full editorial control of these papers. We will make all policy papersavailable to the public whether they have been supported by specific funding as a matter of course.Some papers may be subsequently developed into research papers Nowcasting Monthly UK GDP: Evidence fromBottom-Up Sectoral Modeling with Big DataMethods and Forecast CombinationAlgorithms Paula Bejarano Carbo1, Rory Macqueen2, and Efthymios Xylangouras Abstract We construct a bottom-up sectoral nowcasting model, in the spirit of NIESR’s GDP tracker, topredict monthly GDP growth one month in advance of the first official estimate. We use a largedataset of public and private sector variables, including newly-available real time indicators, asinputs to 28 nowcasting models, each of which is applied to all 20 SIC industries. These rangefrom simple univariate processes to high-dimensional ‘big data’ approaches. We then applyalgorithms trained on previous forecast errors to combine the nowcasts for each sector,generating pseudo out-of-sample GDP nowcasts for 57 months covering before, during andafter the Covid-19 shock. We find evidence that (a) our combination algorithms are betterthan single-model approaches during the initial pandemic shock and generally no worse in‘normal’ times, (b) our algorithmic sectoral nowcasts are more accurate in terms of RMSFEthan ONS first estimates for some industries, but (c) for nowcasting total GDP, humanjudgement seems impossible to beat, (d) a bottom-up ARMA(1,1) outperforms a top-downARMA(1,1), but that the reverse is true for the high dimensional Dynamic Lasso with PCAmodel, and (e) our combination algorithms produce highly correlated nowcasts, but thatforecast combination may be a more fruitful direction for future research than adding to orimproving the individual nowcasts Classification:C51, C52, C53, E17, E66 Keywords: Nowcasting, bottom-up forecasting, big data, forecast combination Contents 1Introduction12Data22.1GDP Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .22.1.1Revisions and Benchmarks. . . . . . . . . . . . . . . . . . . . . . . . .42.2Nowcasting Inputs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .63Model73.1Basic Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .73.2Statistical Models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .84Single-Model Results95Nowcast Combination Algorithms115.1Optimal Model Selection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .115.2Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .125.2.1GDP Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .125.2.2Sector Nowcasts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .156Conclusion18Appendix22AData22BPre-specified Variables for Each Sector (Xi)28CModels29 1Introduction ‘Bottom-up’ forecasting of UK monthly gross domestic product (GDP) by means of aggregatedsectoral forecasts now has at least two decades of history. The Office for National Statistics publishes GDP estimates on three bases - income, expenditureand output (or production) approaches - but we are limited to constructing bottom-up forecastsbased on components of the output-based measure of GDP, as this is presently the only measureavailable at monthly frequency. Previous research on monthly GDP estimation was done by colleagues and predecessors at theNational Institute of Economic and Social Research (NIESR). Mitchell, Smith, et al. (2005) pio-neered the estimation of monthly GDP aggregated up from the indices of production, agriculture,p