While Covid-19 caused an aggregate decline in economic activity, the severity of the impact varies highly across different parts of the economy. This pattern will lead to a permanent reallocation of capital and labor which will occur primarily across firms within industries according to both survey evidence (Barrero et al. 2020) and historical evidence (Davis and Haltiwanger 1992).
Figure 1 Value-weighted mean and cross-sectional inter-quartile range of US equity returns, all days in 2019 and days with large market-level moves in February-March 2020
The behaviour of the stock market in the early stages of the pandemic illustrates the vast differences in firm-level outcomes associated with the arrival of COVID-19. Figure 1 plots the average US equity market return against the inter-quartile range (IQR) of individual returns for all trading days in 2019, and for 20 ‘jump days’ in February and March 2020 on which average returns rose or fell by at least 2.5%. Colours indicate the reason journalists gave for these moves in next-day newspaper accounts (Baker et al. 2020). A clear pattern emerges whereby larger market-level moves correspond to greater variation in firm-level outcomes. On three separate days (16 March, 18 March, and 24 March) the IQR is more than 15 standard deviations greater than the average IQR value in 2019. In a recent paper (Davis et al. 2020), we explain and interpret the structure of firm-level returns, and provide insights into how the pandemic will reshape the economy.
Our primary data source is pre-pandemic 10-K reports, which are filed annually by publicly traded US firms. Since 2006, these 10-K reports have contained a Risk Factors (RF) section that provides a detailed description of known sources of future earnings uncertainty. We match 2,155 individual stock returns on 2020 jump days with corresponding RF texts from 2010 to 2016. The key idea is that pre-pandemic RF content, which explains returns on the jump days, reveals the forces that the market expects to impact post-pandemic earnings.
Explaining cross-sectional returns using risk-factor text
Implementing this idea requires a quantitative representation of the Risk Factors (RF) text, for which we use two popular but distinct text-analytic approaches: dictionary methods and supervised machine learning. This allows us to compare the methods’ strengths and weaknesses in a concrete setting, and we combine elements of both to simultaneously explain and interpret outcomes.
Dictionary methods count the frequency of certain terms that relate to concepts of interest. Our specific dictionaries come from Baker et al. (2019), and their construction relies heavily on domain expertise. There are 36 categories in all: 16 for economic conditions and 20 for government policy. The dictionary measures, together with controls for sector and firm financial variables, explain one-third of the average firm-level return on pandemic fallout days (those marked red in Figure 1). Categories associated with negative returns in reaction to bad pandemic news include ‘inflation’, ‘credit indicators’, ‘taxes’, and ‘transportation’. Categories associated with positive returns include ‘healthcare policy’ and ‘intellectual property’ where the latter is particularly relevant for pharmaceutical firms. Dictionaries also explain returns on jump days driven by other types of shocks but in different ways. For example, they explain one-quarter of the variation in returns on jump days driven by monetary policy news, with an important role for ‘inflation’, ‘interest rates’, and ‘real estate’ in driving positive returns.
Figure 2 R-squared achieved by dictionary and machine learning approaches
Note: 45-degree line is dashed.
While the dictionaries provide an informative characterisation of returns, they draw on a small fraction of RF content. Together they contain 244 unique terms, which make up 2.4% of the RF corpus (i.e. the collection of RF texts). But overall there are more than 18,000 unique terms in the corpus, and the dictionaries hence miss potentially relevant information. Incorporating the entire corpus into a returns model requires machine learning (ML) methods due to the scale of the data. We adopt the inverse regression model of Taddy (2013, 2015), which has seen successful recent applications in economics (e.g. Gentzkow et al. 2020). To compare approaches in terms of their ability to fit the returns data, we fit returns jump-day-by-jump-day using both methods and plot the resulting R-squared values in Figure 2. The top panel restricts the ML model to only operate on the 244 terms present in the dictionaries. In this case, both approaches have equivalent explanatory power. The bottom panel shows what happens when the ML model operates on the whole RF corpus. Now the ML approach yields a nearly uniform increase in goodness-of-fit by 20 percentage points.
These results show that the entire gain in explanatory power from ML is due to its ability to operate on the full set of terms contained in RF texts. There remains a strong relationship between goodness-of-fit values achieved by the two approaches, a result we extend in several ways in Davis et al. (2020). Our findings suggest that the information contained in the non-dictionary terms refines and extends rather than substitutes for the signals captured by the dictionaries.
Targeted exposure construction for interpreting returns
The ML model greatly increases the ability of the RF texts to predict firm-level returns because it draws on a vastly larger set of terms. This very fact makes interpreting its output difficult, because the fitted ML model involves a huge number of estimated parameters. To overcome this challenge, we propose and implement an algorithm for targeted risk factor construction to gain insight into the specific RF content that drives returns on pandemic fallout days. First, we use the fitted inverse regression model to identify ‘seed’ terms that relate especially strongly to firm-level returns. We then find other terms that (1) explain returns similarly to the seed according to the ML model, and (2) co-occur with the same surrounding words in the RF texts, a standard notion of linguistic similarity. The resulting set of terms associated with the seed effectively defines a new risk exposure category. For example, starting with the seed word ‘tantalum’, we find related terms such as ‘tin’ and ‘tungsten’ to form a new term set that we label Raw Metals and Minerals. Proceeding in this manner, we obtain 38 categories for pandemic fallout days, which we use in place of the baseline, untargeted dictionaries to explain returns.
Table 1 highlights selected negative and positive exposures that are important drivers of returns on pandemic fallout days. For example, high exposures to traditional retail, restaurants, and travel – all of which are directly harmed by social distancing – drive negative returns on days with bad pandemic news. But we also find positive return exposures associated with substitution away from these activities, such as ecommerce and basic foodstuffs. Moreover, exposure to intermediate inputs affected by downstream demand shocks drives returns in both negative (e.g. for energy) and positive (e.g. for the technology supply chain) directions on days with bad pandemic news. Mortgages drive negative returns, but other financial activities like banking and investment funds drive positive returns. Most of these effects remain even when we control for very narrow industry codes, meaning that they capture firm-level variation within industries.
Table 1 A selection of targeted exposures for explaining returns on pandemic fallout days
Our results point to a large variety of firm-level risk exposures that drive returns in both positive and negative directions in response to the arrival of pandemic news, which provides insights into how the market expects reallocation activity to take place across firms. In time, we can assess how these exposures relate to changes in the real economy, but a reasonable conclusion is that the economic impact of COVID-19 will stretch far beyond the most obviously exposed sectors and firms.
More broadly, we show how researchers can use the rich textual data in regulatory filings to account for firm-level outcomes by combining machine learning and human judgment. This approach has many potential applications in other settings.
Baker, S R, N Bloom, S J Davis and K J Kost (2019), “Policy News and Stock Market Volatility”, NBER Working Paper Series No w25720.
Baker, S R, N Bloom, S J Davis, K J Kost, M C Sammon and T Viratyosin (2020), “The Unprecedented Stock Market Reaction to COVID-19”, The Review of Asset Pricing Studies, forthcoming.
Barrero, J M, N Bloom and S J Davis (2020), “COVID-19 Is Also a Reallocation Shock”, Brookings Papers on Economic Activity, forthcoming.
Davis, S J and J Haltiwanger (1992), “Gross Job Creation, Gross Job Destruction, and Employment Reallocation”, The Quarterly Journal of Economics 107(3):819– 863.
Davis, S, S Hansen and C Seminario-Amez (2020), “Firm-level Risk Exposures and Stock Returns in the Wake of COVID-19”, CEPR Discussion Paper 15314.
Gentzkow, M, J M Shapiro and M Tadd (2019b), “Measuring Group Differences in High-Dimensional Choices: Method and Application to Congressional Speech”, Econometrica 87(4):1307–1340.
Taddy, M (2013), “Multinomial Inverse Regression for Text Analysis”, Journal of the American Statistical Association 108(503):755–770.
Taddy, M (2015), “Distributed Multinomial Regression”, The Annals of Applied Statistics 9(3):1394–1414.