The third hypothesis on where to look for AI impacts on the labor market examines the effects on cyclical unemployment. The intuition behind testing at the cyclical level is that as employment fluctuates with output, analysts should observe a significant shift in the underlying relationship between the two. Expected scenarios could be inferred from changes in the slope coefficients as well as changes in the relationship’s residuals. Those scenarios are: the labor-output relation slope increased, decreased, showed no change, and/or a comparison with previously tested residuals may show the effect as a confounding. We focus on testing this latter scenario. Since we do not have clean proxies for AI, its effect -if any- must appear as a confounding in the errors. To do so, we estimate the labor-output relationship by fitting Ordinary Least Squares (OLS) regression models to Gross Domestic Product (GDP) and unemployment rate data across pre-COVID and post-COVID periods. We find that the residual statistics for the post-pandemic era differ significantly from those of the ten prior periods.
Our exploratory tests build on the intuition that statistically significant differences appear in residual distributions across historical windows. It is worth noting that we do not aim to fit a perfectly unbiased structural regression model, but rather to extract comparable residual metrics across forty years of data. Specifically, we collect the mean, skewness, kurtosis, and Jarque-Bera normality test statistics for the regression residuals across eleven distinct time windows. These metrics are then tested for statistical significance between the most recent period—which is assumed to capture the initial effects of AI—and all ten preceding periods.
The Model
We estimate an OLS model of the following form:
Annual Change in GDP(t) ~ Intercept + Slope(Unemployment(t-1)) + Residuals.
Consistent with our previous test on frictional unemployment, we exclude the COVID-19 pandemic shock. We assume the effect of labor on output lags behind GDP by one month. Both time series are confirmed to be approximately stationary based on Augmented Dickey-Fuller (ADF) tests.
The Data
To capture higher-frequency variance and match the monthly granularity of the unemployment rate, we use an alternative monthly GDP measure: the Brave-Butters-Kelley Real Gross Domestic Product (BBKMGDP). Analyzing recent data up to 2026 requires isolating COVID-19-related shocks. While excluding the pandemic years leaves only twelve quarterly observations when using traditional Bureau of Economic Analysis (BEA) GDP data, the monthly BBKMGDP series provides sufficient sample size. We fit the model across eleven periods from 1984 to 2026, as shown in Table 1.
Table 1: Residual Summary Statistics Across Time Windows
| Period | Residual Mean | Skewness | Excess Kurtosis | Jarque-Bera |
|---|---|---|---|---|
| 2023–2026 | -1.33e-16 | -0.578 | 0.104 | 2.744 |
| 2016–2019 | 1.27e-17 | -0.200 | -0.287 | 0.399 |
| 2013–2016 | 7.63e-17 | -0.597 | 1.065 | 6.108 |
| 2010–2013 | -4.13e-17 | -0.265 | -0.742 | 1.430 |
| 2003–2006 | -3.96e-17 | 0.001 | -0.767 | 0.901 |
| 2002–2003 | -1.39e-16 | -0.060 | -1.103 | 0.889 |
| 1997–2000 | 1.28e-17 | -0.077 | 0.343 | 0.525 |
| 1994–1997 | -1.29e-16 | -0.111 | -0.747 | 0.950 |
| 1991–1994 | -1.61e-16 | 0.011 | -0.849 | 1.145 |
| 1987–1991 | 2.32e-16 | -0.269 | 0.076 | 0.842 |
| 1984–1987 | -5.08e-17 | 0.224 | -0.557 | 0.836 |
Table 2: Hypothesis Test Results (2023–2026 vs. Prior Periods)
| Statistic | P-Value |
|---|---|
| Residual Mean | 0.015 |
| Skewness | 0.000 |
| Excess Kurtosis | 0.057 |
| Jarque-Bera | 0.032 |
We find that the residual statistics for the post-pandemic era differ significantly from those of the ten prior periods. We emphasize that econometric rigor in specifying a structural regression model is secondary here, as OLS is used primarily as a filtering mechanism to generate comparable residual series. Nevertheless, this test offers analysts preliminary empirical evidence on how AI may be altering cyclical labor market dynamics.
At Econometricus, we tackle these kind of questions and deliver quality mixed-methods research that achieves high academic and social media impact. We build cogent, data-rooted rationales that AI cannot match. Regarding data sources, we have more than fifteen years of experience leveraging major public datasets.
Are you ready to test your hypothesis with solid data and unbiased results?

Discover more from Econometricus.co
Subscribe to get the latest posts sent to your email.
Categories: Macroeconomics









































































1 reply »