key: cord-0853772-e9h8q0by
authors: Larson, William D.; Sinclair, Tara M.
title: Nowcasting unemployment insurance claims in the time of COVID-19
date: 2021-01-11
journal: Int J Forecast
DOI: 10.1016/j.ijforecast.2021.01.001
sha: d99e6ba9c7c2d0cfb441ff78ce782637ee7d3aa1
doc_id: 853772
cord_uid: e9h8q0by

Near-term forecasts, also called nowcasts, are most challenging but also most important when the economy experiences an abrupt change. In this paper, we explore the performance of models with different information sets and data structures in order to best nowcast US initial unemployment claims in spring of 2020 in the midst of the COVID-19 pandemic. We show that the best model, particularly near the structural break in claims, is a state-level panel model that includes dummy variables to capture the variation in timing of state-of-emergency declarations. Autoregressive models perform poorly at first but catch up relatively quickly. The state-level panel model, exploiting the variation in timing of state-of-emergency declarations, also performs better than models including Google Trends. Our results suggest that in times of structural change there is a bias–variance tradeoff. Early on, simple approaches to exploit relevant information in the cross sectional dimension improve forecasts, but in later periods the efficiency of autoregressive models dominates.

Simple time series models often forecast well in normal times (see the long literature started by Nelson (1972) ), but in the midst of dramatic upheaval, * Corresponding author.

E-mail address: william.larson@fhfa.gov (W.D. Larson) . 1 The authors thank Dallas Phillips and Rebecca Sullivan for compiling declarations timing data; Lynn Fisher, Saty Patrabanch, Anju Vajja, Justin Contat, and Forrest Pafenberg for encouragement and support; Andrew Butters, Andrew Martinez, Constantin Burgi, and two anonymous referees for detailed comments; and seminar participants at the FHFA and GWU for valued comments and discussion.

2 Papers prepared by staff of the Federal Housing Finance Agency (FHFA) are preliminary products circulated to stimulate discussion and critical comment. The analysis and conclusions are those of the authors alone, and should not be represented or interpreted as conveying an official FHFA position, policy, analysis, opinion, or endorsement. Any errors or omissions are the sole responsibility of the authors. References to FHFA Working Papers (other than acknowledgment) should be cleared with the authors to protect the tentative character of these papers.

such as in the first few weeks of the response to COVID-19, these models do not adjust quickly to changing conditions (Castle, Clements, & Hendry, 2016) . These models exploit a long time series of an aggregate to provide precise, consistent estimates, while typically ignoring information on disaggregate components. We argue that in times of structural change, forecasters should attempt to exploit information within disaggregates, especially when there exists timing variation in the breaks among the disaggregate entities. In the case of the spring of 2020, we find that variation in state-level COVID-19related state-of-emergency declarations provides useful information for predicting the national quantity of initial unemployment insurance claims. The number of persons filing initial claims for unemployment insurance (UI) is an important indicator of current US economic conditions (Berge & Jordà, 2011; Lewis, Mertens, & Stock, 2020) . UI data are both timely and frequent, and constitute one of a limited number of macroeconomic indicators available at a weekly cadence. an important indicator used to evaluate the state of the economy and the economic toll of the pandemic and associated stay-at-home orders in near-real time. It also itself became a key variable to forecast, to predict the economic costs of the pandemic. 3 In this paper we present current-week forecasts (i.e. nowcasts) of the advance estimates of US national initial unemployment claims produced with different information sets and data structures. 4 We show that a small T panel model performs well shortly after a structural break if it has relevant information. The variation in emergency declaration date across the states provides this relevant information and outperforms models that include Google Trends data, as well as autoregressive models. We further show that autoregressive models do catch up within a few periods, but that they perform poorly in the crucial weeks directly after the dramatic increase in claims that came in mid-March of 2020.

Other researchers have recognized the need for different tools to forecast UI claims in the time of COVID-19 beyond simple time series models. Aaronson, Brave, Butters, Sacks, and Seo (2020) use an event-study approach based on hurricanes in order to use Google Trends data to forecast claims. Goldsmith-Pinkham and Sojourner (2020) also use Google Trends at the state level to forecast claims. Our innovation is to use the timing of emergency declarations across the states in a panel framework. We compare our forecasts to those produced using Google Trends in both panel and time series frameworks as well as with autoregressive models. We show that all models miss the initial shock, but that in the early periods the information in the declaration-weeks panel leads to substantially better forecasts. Quickly thereafter, the time series models catch up. Google Trends models are typically second-or third-best in each time period, outperforming autoregressive models initially, but still underperforming declarations dummy variable models. 5 Our research also connects to the debate about forecasting the aggregate or aggregating forecasts (see discussion in , Hubrich (2011), Larson (2015) , and Heinisch and Scheufele (2018)). We find support for forecasting the individual states and then aggregating in the case when we have useful variation in the timing of the states, e.g. the emergency declaration dates. This is consistent with the argument from Castle, Fawcett, and Hendry (2011) , that we need relevant information available to forecast during breaks.

After the usefulness of that information is exhausted, however, simple autoregressive models perform best, with 3 The insured unemployed in the US is typically a subset of all unemployed, since some jobless do not qualify or apply for benefits. Furthermore, some that qualify for unemployment benefits in the pandemic may not be actively looking for work and therefore may be classified in surveys as out of the labor force rather than unemployed. 4 Wright (2019) emphasizes the important role of nowcasts in macroeconomic forecasting. 5 Google Trends has been shown to be useful in other contexts, including forecasting the US unemployment rate (D'Amuri & Marcucci, 2017) and nowcasting GDP, particularly when official data are not available (Ferrara & Simoni, 2019) . See Choi and Varian (2012) for discussion of using Google Trends for nowcasting several economic indicators, including unemployment claims.

little difference between direct national forecasts and aggregates of state forecasts. Research using Big Data for forecasting has emphasized the importance of a large number of time observations for forecasting, but in the face of structural breaks we may not have a long time series with consistent parameters (Bajari, Chernozhukov, Hortaçsu, & Suzuki, 2019) . We may, however, have useful information in the panel dimension. Since we have variation in the timing of the emergency declaration by state, we can exploit this information using dummy variables in a panel framework. We show that in this case the cross sectional information does improve forecasts in the periods near the structural break.

The remainder of our paper is as follows. First, we describe the UI claims, Google Trends, and disaster declaration series. Next, we describe the models used to generate pseudo-real-time forecasts, paying particular attention to information set assumptions regarding variables and timing and the structure of the data. Then, we turn to a discussion of our results, after which we conclude. Detailed model estimates and alternative forecasts can be found in the appendix.

Our target variable is the national total of the advance number of initial UI weekly claims under state programs, not seasonally adjusted (NSA). Fig. 1 , panel a reports this variable for each week since January 1998. The previous peak, near the end of the Great Recession, was just under 1 million people filing initial UI claims (NSA). In early 2020, the US economy was relatively strong, with around 215,000 weekly claims each week in February. In response to the COVID-19 pandemic and stay-at-home orders, claims quickly dwarfed the Great Recession high, rising almost 30-fold from this benchmark entry, with claims exceeding 6 million in a single week (panel b). Our analysis is focused on this crisis period and our forecasts begin for the week ending March 14th, 2020. Our objective throughout the paper is to forecast the advance release of the series presented in Fig. 1 .

Initial UI claims are released each Thursday morning at 8:30 a.m. Eastern Time by the US Department of Labor (DOL) and the latest data are the preliminary estimates for the week ending the Saturday five days prior. Each week our data set contains initial claims by state for a total of 52 entities in the panel: 50 US states plus the District of Columbia (DC) and Puerto Rico. 6 We refer to all the entities as states for simplicity. The objective of the forecasts will be the advance national totals, which are directly calculated as a sum of the 52 state entities. 7 We focus on NSA numbers throughout our analysis, 6 We do not model the US Virgin Islands, which typically has fewer than 100 claims per week.

7 We are focused on the initial claims under state programs, not the additional claims for the national Pandemic Unemployment Assistance from the CARES Act (information at https://www.dol.gov/coronavirus/ unemployment-insurance.

following the recommendation of Rinz (2020) , since the multiplicative seasonal adjustment procedure that is used for UI claims is likely misleading in the case of the magnitudes of changes in the sample we are analyzing. 8 Our pseudo-real-time data set is constructed using the historical weekly file available on the DOL website through the week ending March 7th, 2020, and then the press release PDF files. 9 All models are estimated on data available at the time the forecast was made. In our exercise, forecasts are assumed to be made immediately following the Thursday release for the next week's release. Unemployment claims data are generally revised once, meaning that the last observation used in each week's estimates will be preliminary data (''advance'' estimates in DOL terminology) that is updated the following week. 10 In our analysis, we develop several forecasting models that exploit different samples of data and variable transformations. For autoregressive models (both state and national), our dependent variable is the natural log of weekly UI claims, and our sample begins in January 1998 in order to exploit a large number of periods, satisfy ''Large T'' asymptotics, and ensure ergodicity. Augmented Dickey-Fuller tests confirm this series to be stationary despite the rising labor force over the period, with all tests rejecting the null of a unit root at the 1% level.

In models exploiting the panel dimension-''Large N'' models-our dependent variable is the normalized weekly US claims for each state relative to the last pre-crisis week, which we identified to be the week ending February 15th. We then reverse this normalization when we aggregate back to the national level to produce the forecast of US national NSA initial unemployment claims. Our panel sample period begins the week ending February 1st.

Our declaration date analysis exploits the variation of timing in emergency declarations by states so that we 8 The use of seasonally adjusted (SA) versus not seasonally adjusted (NSA) numbers has been the subject of some debate in the COVID-19 downturn. Seasonal adjustments to advance weekly UI claims are multiplicative, which we believe to be overestimates of the true seasonal effect. Accordingly, we use NSA numbers throughout the paper. For some context, the seasonal adjustment to the week ending March 28th was over 600,000 (6.4 million SA versus 5.8 million NSA). The prior year's UI claims numbers for the same period were, respectively, 211,000 and 180,000, giving a seasonal effect of 31,000. See https://www.dol.gov/sites/dolgov/files/OPA/newsreleases/ ui-claims/20200551.pdf for the April 2nd, 2020 press release (accessed 5/28/2020). This file has been downloaded and retained by the authors and is available upon request.

9 These files are available at https://oui.doleta.gov/unemploy/ DataDownloads.asp series ETA 539 and https://oui.doleta.gov/ unemploy/archive.asp, respectively. 10 For our real-time data set we assume the data are revised only once, so each week we update our data set with the ''advance'' estimates for the previous week along with revised estimates for the week before that. There was one clear data error where preliminary data for Connecticut for the week ending on May 9th was originally reported as 298,680, but was quickly corrected to 30,000 per the Wall Street Journal: https://www.wsj.com/articles/unemployment-claims-keep-pilingup-in-tri-state-area-11589471382. We used the corrected number as the preliminary value for Connecticut in this case and adjusted the national values appropriately.

can learn from the early states for predicting outcomes for later states, similar to the approach of Liu, Moon, and Schorfheide (2020). The event time is based on the week in which a state of emergency was declared regarding COVID-19 for that state. All states, as well as DC and Puerto Rico, declared a state of emergency within a fourweek period in February and March of 2020 (Table 1) . Washington state was the first to declare an emergency on February 29th and West Virginia was the last on March 16th. Fig. 2 , panels a through d, show normalized state claims grouped by their week of declaration, where 0 on the x-axis for each group is the week that the group declared their emergency. From these figures it is clear that all states experienced a substantial increase in claims after declaring a state of emergency. The last states to declare an emergency already had increases before their declaration week, but continued to experience further increases in the weeks following. It is important to note that we are using declarations of states of emergency, not stayat-home orders or other restrictions, which, as discussed in Cronin and Evans (2020) , explain only a small amount of the change in behavior in response to the pandemic.

Google Trends data are used as an index of the relative search volume on Google for the keyword ''unemployment'' for the US and for each of the 50 states plus DC and Puerto Rico. 11 The Google Trends API only allows five comparison locations per search, so, following (Goldsmith-Pinkham & Sojourner, 2020), we include California in each of our rounds of data collection and renormalize each state relative to California. This approach allows us to compare across time and states. In order to get a long daily time series for the national trend, we follow a similar approach, pulling each six-month time period along with the initial six months. Because Google Trends data are based on a sample, we averaged over 10 different samples to reduce the noise (Stephens-Davidowitz & Varian, 2014) .

It is important to note that the models that include Google Trends are true nowcasts in the sense that they are using information available into the target week. Google Trends data are available with a delay of approximately 36 h, so for forecasts made on Thursday morning for the data that will be released the following week, we already have Google Trends data through Monday of the reference week. In order to use all information available at the time of our forecast, we create two different Google Trends variables. We include the average of the two days that are available in the current week, which we callG t . We 11 Aaronson et al. (2020) also use the keyword ''unemployment''. We also considered ''file for unemployment'' following Goldsmith-Pinkham and Sojourner (2020) with little impact on the results. One interesting finding is that including both keywords helps the national forecast and seems to be a good way to manage sampling error, but we leave deeper exploration of this issue to further research. Borup, Rapach, and Schütte (2020) considered a large number of keywords associated with unemployment and found that their high-dimensional set of Google Trends terms improved nowcasts of UI claims over a benchmark AR model. also include a separate variable that is the average of the previous week, where the dates line up with the latest available UI claims numbers, which we call G t−1 . 12 12 We also considered a single variable of the latest 7 days available, but it performed worse in our forecasting models. We also considered excluding G t−1 , but that produced a worse forecast as well. We find Similar to the UI claims, the Google Trends data are normalized by their average value from the week ending February 15th. Since there are many cases of zeros in particularly for the national Google Trends forecasts that including both G t and G t−1 produces the best forecast in terms of MAE, but with increased variance. Google Trends, we first add 1 to all values, and then the Trends are divided by the average of 2/9 through 2/15 in the case of G and by the average value of 2/9 and 2/10 in the case ofG. 13 Google Trends data are available from January of 2014. We use the full series for the national time series model. For the panel models we start in February of 2020, consistent with our other panel models.

Our forecast target is weekly US national initial unemployment claims with a focus on the first release, which is announced the Thursday of the week following the reference week. The nomenclature for the horizon of our forecast is tricky. Models that do not include Google Trends or disaster emergency declarations data can plausibly be considered one-week-ahead forecasts, since they 13 The national series of normalized claims and Google Trends can be found in the appendix, Figure A. 1. are based on data through the previous week. However, the UI claims data on the previous week are only released on Thursday morning of the current week, which means we cannot produce the forecasts until well into the current week. Thus if we name the horizon based on when we are able to make the forecast, it would be a currentweek forecast, also known as a nowcast. We estimate each of our models with data available on Thursday morning of the reference week. Thus the Google Trends data are available through Monday of that week, and declarations for the current week made after Wednesday are assumed not to have been made at the time of the forecast. Following Banbura, Giannone, Modugno, and Reichlin (2013), we use the term ''nowcasts'' for all models.

We consider several combinations of data structures and information sets. Data structures considered include US state-level panel models, state time series models that do not exploit the panel dimension, and national time series models that do not exploit any disaggregate information. For the state-level data, we sum up to the national level for evaluation of the forecast. For information sets, we use single-series autoregressive models, Google Trends at the state and national levels, depending on the model, and dummy variables to convert the event time of the emergency declaration by state into calendar time. In the main results section, we report a total of six different models, where in the case of state-level models we aggregate back to the US national values for the final forecast, which we describe below in greater detail.

Our main model is a panel model with dummy variables representing the distance week t is from state i's emergency declaration date. The variation in this information is what gives our panel good forecasting properties. 14 The panel regression is weighted based on covered employment for each state for the week ending March 7th.

14 Once we know the dates of all the emergency declarations, this model is equivalent to a panel structured around the week each state declared an emergency (event time), but we keep it structured around calendar time to ease aggregation and forecasting.

In this model, state differentials in declarations timing give us information about future UI claims behavior. The states are in four different groups based on the timing of their emergency declaration. Washington state was the first to declare an emergency on February 29th, which we classify as a ''week 1'' declaration. The remaining states all followed suit over the next three weeks (see Table 1 ) and are given labels of weeks 2 through 4. Our first nowcast was made on March 12th for the week ending March 14th.

The timing of our nowcasting exercise affects the availability of information on declaration dates. To illustrate, for the nowcast for the week ending March 14th, we use all information available up to March 12th, which includes the UI claims release that morning covering the week ending March 7th, and all declarations up to March 11th. All week 1 and week 2 declaration states have dummies set equal to 1 in their respective week. However, for week 3 states, only about half of the states would have dummy variables set equal to 1 for the week ending March 14th in this particular forecast vintage, because declarations after March 11th are not included. All other week 3 and week 4 declaration states have dummy variables set equal to 0 in all periods in this vintage. For the March 21st forecast vintage, the remaining week 3 states would have their dummy variables set equal to 1 for the week ending March 14th, and so on. In terms of notation, j is how many weeks state i is away from emergency declaration at time t. So, for instance, in the week ending March 14th, the dummy variable, d ijt for Washington, whose declaration date is February 29th is 1 for j = 2, and 0 otherwise. By the same logic, the dummy variable in the same time period for California (declaration date of March 4th) is 1 for j = 1, and 0 otherwise. We also include dummy variables for the two periods in advance of the declaration once the declaration date is known. These leading dummy variables serve two purposes: they absorb any pre-declaration explanatory power; and they serve a ''placebo'' role in inference related to declaration pre-trends, though that is not a focus of our model design. Thus we estimate J + 3 dummy variables, where J is the number of weeks since the week ending February 29th: two leads, one concurrent to the declaration date, and J lags.

We also estimate panel models that include Google Trends data for each state, whereG it denotes the average of the Sunday and Monday of the current week and G it−1 is the average for the entire prior week. 15,16

Normalized state claims are aggregated to national claims in several steps. First, we undo the normalization by multiplying estimated normalized claims by the February 15th value of state claims. Next, we aggregate to the national level by summing the 52 state estimates. Finally, because the estimator for the normalized data is not unbiased in terms of the national sum, we include a bias-adjustment term b that is equal to the percent error from the last in-sample predicted value. This is akin to intercept correcting a forecast based on the last estimated

Trends performed worse than a model with just declaration-week dummy variables. The results for this model are reported in Table A .5 in the appendix. We encourage further work on model selection in the vein of Hendry and Krolzig (2005) . 16 A remark is necessary concerning Pandemic Unemployment Assistance (PUA) search intensity and claims. We exclude PUA claims in our analysis, but Google searches for unemployment may be in part driven by PUA-related searches after the CARES Act was signed into law on March 27th. Google searches for the term ''PUA'' only slowly ramped up and PUA claims only began being tracked in large numbers beginning the week ending April 25th. It is possible there could be a break in the relationship between Google Trends data and regular unemployment claims due to the introduction of PUA during our sample. In this case, the Trends forecasts would over-estimate actual weekly non-PUA claims. We do find evidence of consistent positive bias in the state panel Google Trends forecasts beginning in the week ending April 4th, but the national Google Trends has a negative mean error overall, and a less clear pattern across the weeks, as can be seen in Table 2 . 

The variableb t consists of lagged values in the data, but is only available once the information set at time t is available. The panel models are then compared to state and national time series models. For state-level autoregressive models, we model the natural log of the state claims separately for each state i 18 : (3) We aggregate the state claims to the national level, first by exponentiating and adjusting for the variance of the residuals to correct for bias from Jensen's inequalitŷ

and then summing across the 52 states.

The national-level model with Google Trends is as follows, where the modeled variable is normalized US 17 For discussion of aggregation bias, see Baltagi (2008 ), Goodfriend (1992 , and Pesaran and Smith (1995) . Our approach is similar to a forecast error intercept correction, except that instead of taking the forecast error from the model used the previous period, we re-estimate the model this period and use the in-sample residual for the last observation, which was the forecast target last period. 18 We also estimated a state-level AR model with declaration-week dummy variables, but that model performed worse, see model 7 in the appendix, Table A.5. claims. When calculating the estimated US claims value, we use the same aggregation bias-correction method as the state model, but with national claims instead of state claims. 19

The national-level autoregressive model of order p is estimated in the same way as the states, with the level value calculated using the estimated residual variance to adjust for Jensen's inequality.

) = β 0 + p ∑ j=1 φ j ln(USClaims t−j ) + u t (5)

We report forecasts for 10 weeks from the week ending on March 14th, 2020, through the week ending May 16th, 2020. The forecasts are depicted in Fig. 3 along with the actual values of the advance data for the national UI claims, which are the target of the forecasts. Forecast percentage errors (defined as forecast minus actual as a share of actual) are reported in Table 2 . 20 The models reported line up with the equations in the previous section. Models 1 through 3 use state-level data, while models 4 and 5 use national-level data. Models 1 and 2 use a panel data structure, with our main focus being the panel declarations dummy variable (''declarations DV'') model. The remaining models use a time series data structure. 19 We also estimated a national-level model with Google Trends and AR terms. This model performed better than the Google Trends alone, but not as well as the AR model alone. We report forecasts from the combined model as model 10 in the appendix, For the autoregressive models, both state and national, we report results with p = 3. We include Google Trends in the panel framework for model 2 and in the national time series framework for model 4.

Focusing on the beginning of the sample, all models perform poorly in the first two weeks. The popular narrative surrounding UI claims highlights the spike for the week ending March 21st, when national claims jumped to a historic high of nearly three million from the previous week of just over 250,000. However, there was also a spike in the week ending March 14th, when claims rose from around 200,000 to 250,000. We can see from Table 2 that in the first week all models miss by double-digit percentage points, and in the second week the absolute forecast errors rise to 76%-92%. In these two periods, no model forecasts substantially better or worse than any other.

In the third week of our forecast analysis, the week ending March 28th, the panel models perform better than the other models, with the state Google Trends model missing by only 10% and the declarations DV model performing second best. The information contained in the declaration week dummy variables was fully populated in this period, with every state having declared a state of emergency by the time of the forecast. The experiences of Washington, California, New York, and some of the earliest-hit states were predictive of patterns exhibited in other states. The timing differentials were key in the forecasting variation in UI claims in states with later declarations. 21 Table 3 presents our parameter estimates for each vintage week for the state declaration models. We add an additional parameter each week as we move one week further away from the first emergency declaration. Parameter estimates improve from week to week as we get more variation from additional states. For example, if we look at the estimated parameter on the dummy variable when j = 2 we can follow across the row to see how the parameter varies as we get more information from more states. For the model estimated with data through March 12th for the week ending March 14th, there is only one observation that has a one for this variable: Washington in the latest period. Then the following week, the states in the week 2 group have a one for this variable in the last week, and Washington in the week before, so we get a more precise estimate-and so on for two more weeks until we get as much information as we can get from the states after 4 weeks. Then for week 5, we still have data revisions that affect the parameter estimates, but it is stable after that. A similar pattern appears for all the coefficients on the dummy variables.

In week 4, the declarations DV model reaches its peak in terms of benefit relative to other models. The forecast misses by only 4%, where no other model has an error less than 17%. In sum, in weeks 3 and 4, the information from the emergency declarations results in forecast improvements relative to the models without emergency declarations.

Google Trends and state declarations, and robust devices in the vein of Castle, Fawcett, and Hendry (2010) and Castle, Clements, and Hendry (2015) , but they typically performed worse than the models reported (see Table A .5). This period is crucial because the week ending April 4th was the week of peak claims (and up to this point in the history of the series) with over 6.2 million people filing initial unemployment claims across the country. The percentage forecast error does not fully capture the absolute forecast error in these periods; absolute forecast errors for this period are about 250,000 for the declarations DV model, compared to 1,000,000 to well over 3,000,000 for the alternative models. In the latter case, these errors far exceed the previous high in actual claims before COVID-19 of about 1,000,000 (see the appendix).

In the weeks ending on April 11th and 18th, the relative strength of the forecasting approaches becomes less clear, and it is debatable which forecasting approach is preferred. The trend in the forecast errors for the panel models is clearly pointing to plateauing performance. The state and national AR models, on the other hand, have similar forecast errors but they are showing monotonic improvements in forecasting performance since March 21st. The national Google Trends model continues to forecast poorly in this period.

In the week ending April 25th through the end of the sample on May 16th, the AR models are slightly preferred relative to the panel models, and are much better than the national Google Trends model, though the percent errors are generally smaller across all models and the absolute number of claims is also smaller. Consequently, the cost of choosing an inferior forecast in this period is much lower than in the earlier periods. Absolute forecast errors for the AR models hover between 1% and 6% while the panel models continue to average around 10% to 15%. This corresponds to absolute error differences between the two forecasting approaches of about 100,000 to 200,000.

Based on our analysis we can divide our sample into four distinct periods: (1) the first two periods when the crisis is ramping up and all models perform poorly; (2) the next periods when the panel model with declaration DVs stands out; (3) a period of ambiguity where models begin to converge; and (4) at the end when the AR models stand out as best. Fig. 4 illustrates this pattern by zooming in on the forecasts of the declarations DV model and the state autoregressive model. In the periods immediately following the onset of the COVID-19 outbreak, both the level of the claims and the percent errors are larger. Accordingly, over the full sample, the Harvey, Leybourne, and Newbold adjusted Diebold-Mariano (''DM'') test (Diebold & Mariano, 2002; Harvey, Leybourne, & Newbold, 1997) rejects equal predictive accuracy in favor of the declaration-dates panel model when comparing with any of the Google Trends models (models 2 and 4) at the 10% significance level. The AR models also perform worse over the full sample, but not statistically significantly so.

We believe these results have external validity beyond the experience in our specific setting. In general, forecasting UI claims in the time of COVID-19 involves a classic mean-variance tradeoff. Exploiting disaggregate information in the periods immediately following a break acts as a form of costly insurance. Early on, the information gained from the panel model is preferred; in later periods, the inefficiency of estimating individual dummy variable coefficients dominates. Fig. 5 illustrates this phenomenon at a state level. In the April 4th vintage, the actual values are almost uniformly above the state AR model forecasts. By the May 16th vintage, the bias is mostly gone and is no longer significant. The declarations DV model, on the other hand, is not significantly biased in either period, but it has a larger variance around the actual values, particularly for the week ending May 16th. When we observe these vintage coefficients from the declarations DV models in Table 3 , we can see the clear trend in the parameters when looking at descending rows within each column.

In the final period model, the coefficients trace out a hump-shaped curve with a tail that is decreasing at a decreasing rate. This is an inefficient way of modeling what is essentially the same type of decay found in an AR Table 2 . The figure shows nearly equal MAEs in the first week (week 1). In weeks 2 and 3, the Panel Declarations DV forecast substantially outperforms the State AR forecast. In weeks 4 and 5, both perform similarly. After week 5, the State AR forecast is consistently more accurate than the Declarations forecast. Table A.6. model. 22 Consequently, AR models outperform the panel DV model at the end of the sample.

In this paper we produced current-week forecasts (also known as nowcasts) for the national total of state initial unemployment insurance (UI) claims for ten weeks in the midst of the COVID-19 pandemic in the US. We considered different data structures and information sets and compared their performance. We found that in the weeks immediately following the jump in UI claims associated with the COVID-19 crisis, a panel model exploiting the time variation in states declaring a state of emergency performed remarkably well, with the lowest mean absolute error for the full sample across the competing models, and statistically significantly better than models that included Google Trends data available at the time the forecasts were made. Autoregressive models caught up in a few weeks and in the last weeks of the sample had the smallest absolute errors.

Prior research has emphasized the usefulness of simple autoregressive models in normal times, but has recommended much more complicated models for recessions, in order to incorporate sufficient relevant information to identify the change in regime (Chauvet & Potter, 2013) . Our findings support the view that simple autoregressive models miss dramatic changes, but we showed that it is possible in certain instances to exploit panel information if there is information on differences in timing in the cross section. 23 This allows us to still use a simple model for forecasting in a time of structural change. Our analysis also emphasizes that autoregressive models are again remarkably useful just a few periods after a break, consistent with the findings of Schorfheide and Song (2020), Primiceri and Tambalotti (2020) , and Lenza and Primiceri (2020) .

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Supplementary material related to this article can be found online at https://doi.org/10.1016/j.ijforecast.2021. 01.001. 22 This is also the type of result found in Pretis, Schneider, Smerdon, and Hendry (2016) , who observe patterns in the curve of a series of single indicators when modeling changes to annual temperature anomalies due to volcanic eruptions. They find that the time-period indicators can be fit to a simple energy balance model found in climate science. 23 The use of spatial timing differentials to identify and predict effects is common in the urban and regional economics literature. See, for instance, the literature that uses Wal-Mart store and distributioncenter diffusion from Arkansas for identification (Neumark, Zhang, & Ciccarella, 2008) .

Using the eye of the storm to predict the wave of Covid-19 UI claims

The impact of big data on firm performance: An empirical investigation

Forecasting with panel data

Nowcasting and the real-time data flow

Evaluating the classification of economic activity into recessions and expansions

Now-and backcasting initial claims with high-dimensional daily internet search-volume data

Robust approaches to forecasting

An overview of forecasting facing breaks

Forecasting with equilibrium-correction models during structural breaks

Forecasting breaks and forecasting during breaks

Nowcasting from disaggregates in the face of location shifts

Forecasting output

Predicting the present with google trends

Private precaution and public restrictions: What drives social distancing and industry foot traffic in the Covid-19 era

The predictive power of google searches in forecasting US unemployment

Comparing predictive accuracy

When are Google data useful to nowcast GDP? An approach via pre-selection and shrinkage: Banque de France Working Paper

Predicting initial unemployment insurance claims using google trends

Information-aggregation bias

Testing the equality of prediction mean squared errors

Bottom-up or direct? Forecasting german GDP in a data-rich environment

Combining disaggregate forecasts or combining disaggregate information to forecast an aggregate

The properties of automatic GETS modelling

On testing for unbiasedness and efficiency of forecasts

Forecasting an aggregate in the presence of structural breaks in the disaggregates: Discussion paper

How to estimate a VAR after

US economic activity during the early weeks of the SARS-Cov-2 outbreak: Discussion paper

The prediction performance of the FRB-MIT-PENN model of the US economy

The effects of Wal-Mart on local labor markets

Estimating long-run relationships from dynamic heterogeneous panels

Detecting volcanic eruptions in temperature reconstructions by designed break-indicator saturation

Macroeconomic forecasting in the time of COVID-19

Understanding unemployment insurance claims and other labor market data during the COVID-19

Real-time forecasting with a (standard) mixed-frequency VAR during a pandemic: Discussion paper

A hands-on guide to google data

Some observations on forecasting and policy