Predicting Initial Unemployment Insurance Claims 

Paul Goldsmith-Pinkham, Elizabeth Pancotti, Aaron Sojourner

Thanks to Minneapolis Fed’s Opportunity and Inclusive Growth Institute

And Dylan Piatt and Zach Swaziek for their help

Goal: forecasting new uninsurance claims

Initial unemployment insurance (UI) claims count 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.

Goal: forecasting uninsurance claims

Initial unemployment insurance (UI) claims count 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.

Goal: forecasting new uninsurance claims

Initial unemployment insurance (UI) claims count 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.

Goal: forecasting new uninsurance claims

Initial unemployment insurance (UI) claims count 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.

How to forecast UI claims?

Difficulties:

  1. Training data arrives at weekly intervals
    • Limited number of data points
    • Low resolution for fast changes
  2. UI data arrives with a lag
    • Pure time series forecasting is already late

Solutions:

  1. Construct training data at higher frequency
    • Use news reports and states’ posted daily UI claims
  2. High frequency Google Trends data
    • (Nearly) immediate information on demand for UI filings

Measuring daily UI claims through news reports

Many states reported various UI claims statistics over the course of the recent weeks.

  • Example: “the State received 10,394 claims on Monday and Tuesday.”

Reconcile these differences in reporting dates by using a report-level dataset

  • each reported statement is treated as an observation

Example Source - Wisconsin Department of Labor

Example Source - Houston Business Journal

Report-level data

Texas

State Report ID Reported Claims Start of Reported Dates End of Reported Dates
TX 1 15,375 2020-03-15 2020-03-18
TX 2 22,261 2020-03-15 2020-03-21
TX 3 10,000 2020-03-16 2020-03-16
TX 4 19,000 2020-03-17 2020-03-17
TX 5 37,500 2020-03-22 2020-03-25

Report-level data available here:

Google Trends - “file for unemployment”

Google Trends - California and New York

Google Trends - Michigan and New York

Google Trends - an index of search interest

Problems:

  1. Locations + time need to be compared to one another (raw levels are not provided)
  2. Google Trends does not allow comparison of more than 5 locations over time
  3. Too low of volume is compressed to 0.

Solution:

  1. Focus on states to ensure sufficient data
    • Caveat: some states are still poorly measured (e.g. Alaska)
  2. Compare all states to a baseline state (California) and renormalize relative to that state

Code to do this for any search term: https://github.com/paulgp/GoogleTrendsUINowcast

Estimating Predictive Model

Goal is to predict UI claims per week.

We proceed in two steps:

  1. Estimation data constructed in report-level data, normalized to “daily” levels
  2. Estimate growth in UI claims at state level relative to baseline growth in Google Index changes \[\mathbb{E}\left(\text{UI Growth}_{it} \middle| \Delta\textrm{Google Index}_{it} \right) = \alpha + \beta \Delta\textrm{Google Index}_{it} \]

Estimating Predictive Model: Restrict to Early Data

Estimating Predictive Model

Number of different ways we can estimate this model

We have focused on simplest case: use data from just 3/15-3/22

Given more data now, can do additional estimation and compare

Estimating Model: Rolling Estimation

Will focus on using data from March 28th onwards

Estimating Model: Rolling Estimation

Will focus on using data from March 28th onwards

Prediction and validation

We scale our predicted growth rates back by baseline UI claims rates, and calculate weekly UI claims predictions.

Can do the same exercise cross-sectionally

Scale our predicted estimates by states’ February labor force

Can do the same exercise cross-sectionally

And combine our three weeks into one

Additional uses beyond forecasting

Estimate response in UI to shutdowns

Conclusion

Across the 3 weeks, 5-7 days ahead of each official report, we are able to predict UI claims.

Given first two days of 4/05-4/11, we predict 4.73 [4.21-5.25] million additional claims.

Additional uses beyond forecasting

Evaluating demand for UI?