Understanding Society in Real-time: A Joint Nowcasting and Disaggregation Approach to Economic Modelling

本项目旨在与国家统计局（ONS）合作开发方法，通过使用统计模型实时吸收信息，帮助我们跟踪社会变化。考虑两个例子，我们试图提高我们对经济理解的分辨率：1)我们在地区层面观察GDP，但每年只观察一次——我们希望对每个季度和每个地区的GDP进行估计。2)我们以季度为单位观察所有服务行业的贸易流量总和——我们希望以月为单位估计分类“部门”的贸易。这两种情况都要求我们以我们能够观察到的更高分辨率估计感兴趣的时间序列，前者增加了时间分辨率，后者增加了粒度（聚合到子部门）和时间（季度到月）分辨率。至关重要的是，我们希望实时估计行为，这样当收集到任何相关数据时，输出估计将被更新。这类经济预测至关重要，并在政策制定决策中发挥作用，例如，针对基础设施融资、调整监管或改变利率等经济工具。这些任务要求我们结合围绕分解和临近预测的方法范式，然而现有的方法（例如Banburra等人，2010;Proietti, 2006； Koop等人，2020）倾向于解决其中一个问题，而不是同时解决两个问题。此外，考虑到目前可用的数据源范围很广（数百个），我们很可能拥有比我们希望描述的输出序列更高分辨率的相关信息，例如，我们可能拥有可以提供每日级别信息的金融时间序列，或者可以显示高时间和空间分辨率的区域数据集，例如道路交通计数。一个重要的挑战是如何以适当的方式将这些数据纳入统计模型。

IntroductionThis project aims to develop methodology in partnership with the Office for National Statistics(ONS) to help us track changes in society by using statistical models to assimilate information in areal-time fashion.Consider two examples where we try to enhance the resolution of our economic understanding:1) We observe GDP at a regional level, but only once per year-we wish to produceestimates for GDP each quarter and for each region.2) We observe trade flows aggregated across all services at a quarterly level-we wish toestimate trade for disaggregate "sectors" at a monthly level. Both these cases require us to estimate a time-series of interest at a higher resolution that we cannatively observe, the former increasing the temporal resolution, and the latter increasing both thegranularity (aggregate-to-sub-sector) and temporal (quarterly-to-monthly) resolution. Crucially,we desire to estimate the behaviour in real-time such that as any relevant data is collected, theoutput estimate will be updated. Economic forecasts of this kind are vital and find impact whenfeeding into policy making decisions, e.g., targeting infrastructure funding, adjusting regulation, oraltering economic instruments such as interest rates.These tasks require us to combine methodological paradigms surrounding disaggregation, andnowcasting-however existing methods (e.g. Banburra et al., 2010; Proietti, 2006; Koop et al.,2020) tend to tackle one or the other problem, not both jointly. Furthermore, considering thewide range (hundreds) of data-sources nowadays availiable, it is likely that we may have relevantinformation availiable to us at a resolution higher than the output series we wish to describe, e.g.,we may have financial time-series that can give information on a daily level, or regional datasetssuch as road-traffic counts that can exhibit both high temporal and spatial resolution. Animportant challenge is how to incorporate such data into statistical models in an appropriatemanner.