Jurado-Ludvigson-Ng (2015) Measuring Uncertainty

Summary

This paper builds a direct econometric measure of time-varying macroeconomic uncertainty, defined not as the volatility of economic series but as the common variation in the unforecastable component of a large number of them. For each series the authors first strip out everything predictable using a rich information set (factors extracted from hundreds of macro and financial variables), then model the conditional volatility of the resulting forecast error with a stochastic-volatility model, and finally aggregate these individual uncertainties into an economy-wide index. The resulting "JLN" index differs substantially from popular proxies (VIX, cross-sectional dispersion, news-based indices): genuine uncertainty episodes are far rarer, but when they occur they are larger and more persistently correlated with real activity — providing a benchmark against which theories of uncertainty-driven business cycles can be judged.

Key Claims

• Uncertainty ≠ volatility of the series. The $h$-step-ahead uncertainty of series $y_{jt}$ is the conditional volatility of its purely unforecastable future component, $U^y_{jt}(h)=\sqrt{\mathbb E\big[(y_{jt+h}-\mathbb E[y_{jt+h}\mid I_t])^2\mid I_t\big]}$. The forecastable part must be removed first — otherwise predictable movement is mistaken for uncertainty.
• Remove the forecastable part with factors. The conditional mean $\mathbb E[y_{jt+h}\mid I_t]$ is estimated from a factor-augmented (diffusion-index) forecasting model using common factors extracted from a large panel (≈100+ macro and financial series), so the information set $I_t$ is far richer than the series' own past.
• Model the forecast-error volatility. A parametric stochastic-volatility model is fitted to the forecast errors of both the predictors/factors and the idiosyncratic components, giving time-varying conditional variances that feed the uncertainty estimate.
• Aggregate to macro uncertainty. The economy-wide index is the (approximately common) average of the individual series' uncertainties, $U_t(h)=\text{plim}\,\frac{1}{N}\sum_j U^y_{jt}(h)$, computed at several horizons $h$.
• Distinct from proxies. The measure has large independent variation from VIX, profit/return/forecast dispersion, and news-based indices — implying much of those proxies' movement is not uncertainty. True uncertainty spikes are infrequent but big and persistent.

Concepts Introduced or Extended

• Macroeconomic Uncertainty
• Factor Model
• Dynamic Factor Model

Entities Mentioned

• Kyle Jurado
• Sydney C. Ludvigson
• Serena Ng

Quotes

"We define… $h$-period ahead uncertainty in the variable $y_{jt}$… to be the conditional volatility of the purely unforecastable component of the future value of the series."

"The proper measurement of uncertainty requires removing the forecastable component."

My Take

The paper's key methodological insight — that uncertainty is the volatility of the residual after optimal forecasting, not the volatility of the raw series — is deceptively simple and quietly demolishes the casual use of stock-market volatility as an uncertainty proxy: a series can be very volatile yet highly forecastable (hence low uncertainty), or calm yet unpredictable. Operationally it is a marriage of two tools the wiki already documents — the diffusion-index forecasting of Stock-Watson to purge the predictable part, and stochastic volatility to time-stamp the residual variance — aggregated across a big panel. The JLN index became a macro-finance benchmark for exactly the reason the paper argues: it isolates the object theory cares about. The honest caveats are that the measure inherits the factor model's specification and the "unforecastable from whose information set?" question, and that it is a filtered estimate, so its own uncertainty is rarely propagated downstream.