Forecasting Initial Jobless Claims

[WITH CODE] Stepping into prediction markets with quantitative insights

Reviewed and updated 20 July 2026

Hello!

Due to my occupation, I am limited in my ability to trade in the financial markets. At the time, prediction markets had captured my interest for about 18 months, and I wanted to test whether a systematic edge might exist.

These markets were relatively illiquid. I wondered whether that might leave opportunities not present in larger, more efficient markets. That was the hypothesis motivating this project, not an established result.

In this post I started a series on trading prediction markets with you. The specific market examined here was Kalshi’s contracts on weekly initial jobless claims.

I showcased academic research on forecasting initial jobless claims, my own preliminary model, and forecasts for the 2 October 2025 Kalshi market.

As always, this is for educational purposes only and not financial advice. Let’s get into it.

---

Research on Forecasting Initial Jobless Claims

Forecasting jobless claims has always been a test of how well models can capture sudden shifts in the labor market, and the literature shows a steady layering of ideas. Back in 2009, researchers discovered that Google search data could act as a real-time barometer of unemployment.

People typing “jobs” or “unemployment office” into Google provided signals that led official government statistics by a week. In Choi and Varian’s exercise, adding Google Trends reduced rolling 24-week out-of-sample mean absolute error by 15.74 percent in their long-term specification and 12.90 percent in their shorter recession specification. It was an early indication that internet activity could offer a faster read on labor market stress than traditional releases.

But there was a catch. Not all search spikes reflected actual job loss, and sometimes people were just following the news. A later study addressed this problem with a clever natural experiment: hurricanes. These disasters created sharp, localized jumps in both layoffs and unemployment-related searches, helping distinguish true job loss signals from background noise. Calibrating Google Trends against hurricane-driven shocks produced a more reliable model. The authors report that it largely predicted the evolution of claims in the 2008-09 and 2020 recessions, outperformed their comparison models, and delivered well-calibrated out-of-sample uncertainty estimates during the pandemic.

That brings us to the COVID-19 crisis, the ultimate stress test. In the first weeks of March 2020, UI claims exploded beyond any historical precedent. Here, even Google-based models struggled. The most effective forecasts came not from search data but from exploiting the staggered timing of state-level emergency declarations.

States that declared early provided a preview of what would hit the rest of the country, and this panel-based approach produced the best “nowcasts” during the crucial first weeks. Only later did simple autoregressive models catch up.

Taken together, these studies show the arc of progress: Google Trends provided a timely early signal, hurricane-calibrated models sharpened the signal, and disaggregated state-level data performed best among the compared models near the structural break.

When the economy is stable, high-frequency data like Google searches can enhance forecasts, but when the ground shifts suddenly, unique event-driven or local information can be useful in certain instances.

Here, I used only prior initial jobless claims as independent variables. At the time, I planned to make the model more complex and add other data sources, such as Google Trends.

---

Analysis & A Preliminary Model

I am going to keep this section relatively concise, as I know this post will be quite long. I spent a significant amount of time testing linear regression and autoregressive models, and I’ll share more detail on model choices and optimization in the next post as I improve the framework.

Using FRED data back to 1967, I loaded seasonally adjusted and non-seasonally adjusted initial claims and continuing claims, cleaned the series, and then built exploratory plots and correlation tables. The goal was to see how current initial claims relate to their own lags and to lagged continuing claims. The historical chart extends through 20 September 2025 for initial claims. Because the analysis does not identify a release-vintage dataset, it should not be interpreted as a point-in-time backtest.

Initial and Continuing Jobless Claims Over Time