The widespread digitisation of data has opened up the possibility of monitoring the economy at a higher frequency and along previously unexplored dimensions. Central banks have responded to this opportunity by investing resources in the collection and evaluation of these new and relatively unfamiliar sources of data. The Federal Reserve Board and the Bank of Italy, for instance, have joined forces to host an annual conference on non-traditional data, machine learning, and natural language processing in macroeconomics, where such projects will be presented.
During the onset of the COVID-19 pandemic, attempts to measure economic activity at a high frequency gained a renewed sense of urgency, as monthly and quarterly measures proved insufficient for a crisis that was unfolding on a daily basis. Economists at both public agencies and within academia have since responded with the production of high frequency indicators. For instance, Lewis et al. (2020) constructed a weekly estimate of GDP growth, building on Stock and Watson (2002) and an earlier effort by the Council of Economic Advisers (2013). Chetty et al. (2020) and Andersen et al. (2020) made use of novel, high-frequency data sources to track economic activity at different levels of aggregation. Other projects, such as the Finnish Situation Room, expanded access to granular, high-frequency public data.1
In March, Sweden’s central bank responded to the pandemic by initiating an internal project to collect and analyse economic indicators on a daily basis. To extend this effort, its Research Division announced an experimental project that opened participation to the general public (including academics and private companies). This involved the creation of a web application that allowed real-time data sharing and visualisation with external participants.
Rather than requesting sensitive microdata and dedicating staff to learn its underlying structure and documentation, the Riksbank instead decentralised the process by opening up participation. Private companies (which know their own data well and have skilled data scientists on staff) could decide what to contribute and in which form to provide it. Academics could produce model simulations, indices, or scraped data, and contribute them to a site where they could be updated at an arbitrarily high frequency and would then be available to policymakers.
Daily retail price and volume indices
One of the earliest corporate contributors to the project was PriceRunner, a retail comparison site that currently includes listings for 2.3 million products from 6,400 stores. PriceRunner offered to produce daily price and volume indices for 15 major retail categories and to make the results available to the public through the project’s website.
PriceRunner’s detailed data enabled them to compute daily updates of rolling volume weights, making it possible to capture the impact of changes in the composition of consumption on prices. This could be particularly important during a pandemic, where social distancing and supply chain disruptions could induce non-trivial deviations in household consumption patterns.
Figure 1 shows a plot of the daily volume index produced by PriceRunner for 15 retail categories. Volumes are based on visitor click-outs to merchant sites, which serve as a proxy for item purchases. The effect of the pandemic on the composition of consumption appears to have emerged in March 2020.
Figure 1 Daily Volume Indices for 15 Retail Categories
Real-time analysis of COVID-19 social media content
The first contributor to provide ‘real-time’ updates was the company Peltarion, which is building a platform to enable organisations to operationalise the latest AI techniques at scale. Making use of streaming Twitter data and the state-of-the-art ‘BERT’ model, Peltarion produced real-time measures of the toxicity, negativity, and volume of COVID-19 related content on Twitter. Whenever a user loads or refreshes a chart on the platform, the server submits a query to Peltarion’s platform, retrieving the latest update for each measure.
The combination of streaming Twitter data and real-time model output delivery makes it possible to monitor activity in the discussion of COVID-19 as it develops. Furthermore, Peltarion’s participation made it possible to use state-of-the-art methods from machine learning without needing to develop the capacity in-house.
Figure 2 Real-Time Toxicity of COVID-19 Related Tweets
Weekly restaurant bookings data
BokaBord, a digital restaurant table booking service, offered to contribute weekly indices of table bookings in Sweden at the national level and for its three largest cities: Stockholm, Malmö, and Gothenburg. The visualisation they constructed shows normalised indices for all years between 2017 and 2020, along with the mean of the indices. The index for 2020 (visualised in Figure 3) shows a clear departure from previous years in March, followed by a recovery that begins in April 2020. By mid-July 2020, bookings appear to have fully converged to their pre-pandemic trend.
Figure 3 Weekly Restaurant Booking Index
Daily natural language processing on company press releases
Modulai, a data science and machine learning services company, offered to perform real-time natural language processing (NLP) on corporate press releases and to produce sectoral bankruptcy predictions. For the NLP project, Modulai is currently collecting all press releases from Nordic Nasdaq Exchange listed firms in real-time, classifying them according to relevance, and then processing them using a topic model. The model output, which consists of daily topic time series (Figure 4), is then pushed to the platform in real-time whenever a user loads or refreshes a chart.
Modulai is also currently configuring a service that will make use of their press release corpus and the state-of-the-art ‘BART’ model to allow the Riksbank (and other platform users) to submit queries, generating a time series (and real-time updates) for arbitrary topics. Figure 5 shows a demonstration of queries that could be submitted through the service. The top panel plots topic query results for “optimism” against the OMXS30 stock index. The bottom panel tracks content related to the pandemic, indicating a spike in March 2020, followed by a steady decline. By July 2020, pandemic content in press releases had returned to pre-crisis levels.
Figure 4 Press Release Topics from Nordic Nasdaq Listed Firms (Source: Modulai)
Figure 5 Example BART model topic queries
a) Query results for “optimism” against the OMXS30 stock index
b) Pandemic-related queries
Contributions from other public institutions and academia
In addition to private companies, academics and economists at other public institutions (including those outside of Sweden) were also invited to participate in the project. Two of the earliest contributions in this area were the Weekly Economic Index (WEI) for the US from Lewis et al. (2020), and a measure of the conditional variance of global macroeconomic risk, contributed by Kiss and Österholm (2020).
The WEI, plotted in the top panel of Figure 6, provides a weekly estimate of GDP growth, which appears to drop sharply during the Great Recession and the COVID-19 pandemic. The bottom panel shows the conditional variance of the residuals of the GEB Coincidence Indicator, which has historically risen sharply during recessions. COVID-19 appears to have generated a new high value for the conditional variance, substantially exceeding even the Great Recession’s value.
Figure 6 The Weekly Economic Index (top) and the Conditional Variance of the Global Coincidence Indicator (bottom)
Project status and future
The platform is available to the public here, and currently offers a variety of indicators for 18 categories of data. These include, for instance, weekly analysis of labour market-related Google searches, daily text analysis of US SEC filings by sector, daily house price indices for Sweden, daily electricity consumption data, as well as daily natural language processing of communication at more than 50 central banks.
Going forward, the project will continue to accept and solicit contributions from external participants. Any academic or private company who would like to visualise and periodically update data related to COVID-19 (or its impact on economic activity) may do so using the platform. This could include scraped data, indices, or model simulations to name but a few ideas. Finally, while the effort may be being organised in Sweden, all contributions are welcome, including those focused on the global economy and other countries.
Author’s note: The opinions expressed are those of the author and should not to be seen as the Riksbank’s view.
Andersen, A, E Hansen, N Johannesen and A Sheridan (2020), "Consumer responses to the COVID-19 crisis: Evidence from bank account transaction data", Covid Economics 7: 88-114.
Council of Economic Advisers (2013), "Economic Activity during the Government Shutdown and Debt Limit Brinkmanship", Report of the Executive Office of the President.
Chetty, R, J N Friedman, H Hendren and M Stepner (2020), "How Did COVID-19 and Stabilization Policies Affect Spending and Employment? A New Real-Time Economic Tracker Based on Private Sector Data", Opportunity Insights Working Paper.
Kiss, T and P Österholm (2020), "Corona, Crisis, and Conditional Heteroscedasticity", Örebro University Working Paper 2/2020.
Lewis, D, K Mertens and J H Stock (2020), "U.S. Economic Activity during the Early Weeks of the SARS-Cov-2 Outbreak", Federal Reserve Bank of New York Staff Reports, no. 920.
Stock, J H and M H Watson (2002), "Forecasting Using Principal Components from a Large Number of Predictors", Journal of the American Statistical Association 97(460): 1167-1179.
1 See here for a description of the Helsinki GSE’s efforts.