Note: Data+code for this are available here https://docs.google.com/spreadsheets/d/1WYgGMTHO43_oDoLz25WBoQFGrZt7ai64asgzh5ckjRQ/edit?usp=sharing and here https://github.com/paulgp/GoogleTrendsUINowcast

Many thanks to Dylan Piatt and Zach Swaziek for their help with this

Goal

Understanding changes in national and state-level initial unemployment insurance (UI) claims has value to markets, policymakers, and economists. Initial claims measure the number of Americans filing new claims for UI benefits is one of the most-sensitive, high-frequency official statistics used to detect changes in the labor market. However, official federal data on UI claims comes out at a weekly interval and at a lag. The U.S. Department of Labor aggregates reports state unemployment insurance systems for weekly release of advance estimates on Thursdays, which covers the prior Sunday to Saturday week. They revise it over the following week, so official estimates are released 12 days after each week ends. We aim to forecast official UI initial claims statistics.

Below, we forecast initial claims nationally and by state for the week ending Saturday, March 21. The official, advance estimates will be released Thursday, March 26. This looks to be the week with the largest number of initial claims and the largest rise in unemployment in U.S. history, due to widespread quarantines. But just how large will this shock be?

Many state agencies reported partial information to the press over the course of the week, due the staggering growth in UI claims. The first part of our approach gathers and harmonizes the reported numbers across press reports to calculate a estimated claims for as many states as possible. The Data section provides more details.

The second part of our approach imputes state’s UI claims harnessing data from Google Trends. We construct a dataset of the intensity of search for the term “file for unemployment” by state over time. We regress this measure on the set of states where we have constructed a growth rate in UI claims using news reports, and use this to impute the initial claims for all states. We do this in two ways: 1) using the daily change in UI rates for states where we have daily data, and 2) aggregating to the weekly level and using weekly predictions. In the current week (ending March 28), we forecast UI claims for the current week using the daily model and more-current Google trends data.

These models are new and unvalidated. So we present a few sensible, alternative versions and hope to learn over time how they perform and how they can be improved. We always welcome feedback ().

Summary of Results

For the week ending March 21, the models presented yield initial UI claims nationally between 3.4 and 3.84 million, depending on exact specification. To put this in context, 3.4 million Americans moving from employment to unemployment would imply a 2 percentage point increase in the unemployment rate in a single week, jumping by more than half from 3.5 percent to 5.5 percent. The range of our confidence intervals depends on assumptions, with our daily model yielding a range of 3.609 to 4.07 million around an estimate of 3.84 million, and our preferred weekly model yielding a range of 3.11 to 3.56 million around estimate of 3.34 million.

For the week of 3/22-3/28, we need to extrapolate Google Trends interest for the rest of the week. Google Trends interest shows that Sunday, March 22, the first day that will be included in the report, was 500 percent higher than the prior Sunday and Monday 250 percent higher than the prior Monday. Combining this information with our model and conservatively assuming that the search interest will remain flat at its current level through Saturday 3/28, implies a prediction of 4.7 million UI claims, with a 95% CI of 4.27 million and 5.13 million for the week of 3/22-3/28.

We predict large variation across states and the table below describes, for each state, the estimated claims level based only on extrapolation from news reports, the Google Trends change and the forecast claims level based on the model combining news reports and Google Trends information.

Week 3/15-3/21
Daily Model
Weekly Model
State UI Claims From News Google Trends Change Advance Official UI Claims Predicted UI Claims Predicted UI Claims
AK 5,291 59 8,225 19,930 5,291
AL 9,347 22 9,490 28,071 9,347
AR 45 8,958 26,137 22,549
AZ 9 29,268 37,457 39,692
CA 599,250 28 186,809 658,504 599,250
CO 42,500 44 19,429 38,902 42,500
CT 60,444 37 25,098 72,520 60,444
DC 14,439 38 13,473 12,956 14,439
DE 17,944 64 10,720 12,173 17,944
FL 19 74,021 79,940 70,625
GA 25,495 14 11,746 63,110 25,495
HI 12,056 92 8,904 26,375 12,056
IA 67,116 49 41,890 47,386 67,116
ID 21 13,314 13,910 13,495
IL 120,889 49 114,663 181,021 120,889
IN 42,657 72 61,635 62,394 42,657
KS 12,065 36 23,687 26,448 12,065
KY 16,635 119 48,847 74,579 16,635
LA 35,172 77 72,620 45,844 35,172
MA 112,449 24 147,995 101,489 112,449
MD 30,600 68 41,882 64,529 30,600
ME 11,900 89 21,197 20,597 11,900
MI 123,205 96 129,298 161,260 123,205
MN 123,624 17 116,438 78,421 123,624
MO 29 40,508 48,920 45,303
MS 40 6,723 14,316 12,575
MT 13,147 92 14,704 16,728 13,147
NC 94,067 43 93,587 49,156 94,067
ND 27 5,968 6,964 6,150
NE 58 15,668 12,401 10,226
NH 26,953 109 21,878 23,111 26,953
NJ 85,000 22 155,454 114,658 85,000
NM 11,559 41 17,187 13,946 11,559
NV 129 93,036 120,296 86,575
NY 30 80,334 290,890 267,742
OH 148,750 93 187,784 197,788 148,750
OK 11,307 38 17,720 23,917 11,307
OR 93,975 47 22,824 73,849 93,975
PA 353,644 68 378,908 353,634 353,644
RI 40,865 113 35,436 37,501 40,865
SC 17,000 27 31,064 29,178 17,000
SD 26 1,703 2,618 2,465
TN 4,738 22 39,096 30,553 4,738
TX 116,167 15 155,657 145,449 116,167
UT 8 1,314 11,705 12,514
VA 42,500 18 46,885 45,480 42,500
VT 49 3,667 12,692 10,777
WA 22 133,478 114,943 110,887
WI 66,755 15 50,957 69,342 66,755
WV 4,303 48 3,435 20,347 4,303
WY 67 2,339 11,163 8,956

Data Sources

News Sources

For the week ending March 21, we are greatly helped by many states reporting various claims numbers over the course of the week. We gather and harmonize the various reported numbers across press reports to calculate an estimated “weekly” number. We found reports for 35 states and the District of Columbia, leaving 15 states without any reports. See this website for the data: https://docs.google.com/spreadsheets/d/1WYgGMTHO43_oDoLz25WBoQFGrZt7ai64asgzh5ckjRQ/edit?usp=sharing

Reports tend to describe the number of claims for a given set of dates (\(R\)) based on information from state officials. For our estimation below, in some models, we use the data sources at the daily frequency to exploit additional variation that is available in the Google Trends data. In other models, we manually extrapolate the claims for the whole week (\(C\)).

The weekly model relates states’ weekly averages in Google Trends and UI claims treating each as constant across the week, and differentiating only between week days and weekend days. The daily models use variation in Google Trends over the week to try to predict within-week-state changes in claims.

Extrapolation of reports to weekly claims

Our approach for manually extrapolating claims at the weekly level is as follows. Let \(D\) be the number of weekdays and \(E\) the number of weekend days represented in the report, \(C_D\) be the average number of claims on any weekdays, and \(C_E\) the average for a weekend day. Let their ratio be \(r \equiv C_E/C_D\), about which there is empirical uncertainty and variation across the few states in which it is observable. If observable, we use empirical \(r\) for the state. If not, we assume \(r=1/3\) and test for sensitivity.

Sets of reported dates come in three types. First, if only reports on weekdays are available, we compute a weekday rate and measure weekly claims as \(C = (5 + 2r) C_D\). Second, if weekend and weekdays are reported separately, \(C = 5\times C_D + 2\times C_E\). Third, if the report \(R\) contains information about total claims across both weekend and weekday dates but these are not separated, \(C = (5+2r) \times R/(D+Er)\).

Estimation

Finally, we consider the relationship between these two measures. There are two ways we can do this: exploiting the daily data, or focusing on weekly numbers. We discuss both below.

Daily Data

For 27 locations, we are able to use news sources to estimate daily UI claims, for a total of 89 state-day observations. With these claims data points, we construct a growth measure, relative to each state’s the average of initial claims in the four prior weeks, ending Saturdays 2/22-3/14. We consider the change in the Google Trends index between the most recent week (3/15-3/21) and the day of the week average from the last four weeks (2/22-3/14). We then plot these two measures to consider how correlated the change in Google Trends search intensity is with UI growth.