Asymmetric Okun Models and Unemployment Forecast Risk

Abstract

This study evaluates whether asymmetric output growth improves one-quarter-ahead unemployment forecasts and whether any gains persist during major economic disruptions. Quarterly United States data spanning 1948–2024 were analyzed using autoregressive, linear Okun, and asymmetric Okun specifications. Predictive performance was assessed through 160 recursive pseudo-out-of-sample forecasts for 1985–2024, using root mean squared error and mean absolute error, alongside pandemic-related sensitivity analyses. All forecasts used revised historical data, not real-time information sets. Before 2020, the asymmetric model achieved an RMSE of 0.1779 percentage points, outperforming the linear model (0.1886) and autoregressive benchmark (0.1897). Across the full evaluation period, however, asymmetric RMSE increased to 1.1769, exceeding the linear model’s 1.0502. Excluding 2020–2021 forecast targets restored its advantage over the linear model but yielded virtually identical accuracy to the autoregressive benchmark. Retrospectively omitting pandemic observations from estimation improved post-2021 performance, indicating substantial sensitivity to extreme shocks. The results show that statistically detectable output–unemployment asymmetry does not ensure robust predictive improvement. Forecasting applications should retain parsimonious benchmarks and evaluate model stability across contrasting macroeconomic regimes.