D'Agostino-Gambetti-Giannone (2013) Macroeconomic Forecasting and Structural Change

Summary

D'Agostino, Gambetti, and Giannone assess whether explicitly modeling structural change improves the accuracy of macroeconomic forecasts. They produce real-time, out-of-sample forecasts of US inflation, the unemployment rate, and a short-term interest rate using a Time-Varying Coefficients VAR with Stochastic Volatility (TV-VAR) — both the coefficients and the shock volatilities drift over time. The model generates accurate predictions for all three variables; in particular, for inflation it outperforms every competing model — fixed-coefficient VARs, time-varying univariate ARs, and the naïve random walk — in mean squared forecast error, and the advantage holds over the recent period in which inflation has been especially hard to forecast. (ECB Working Paper 1167, 2010; published in the Journal of Applied Econometrics 28(1): 82–101.)

Key Claims

• Model both drifting coefficients and drifting volatilities. The TV-VAR combines time-varying VAR coefficients (random-walk evolution) with multivariate stochastic volatility, in the Cogley–Sargent / Primiceri tradition, so it can track both changing dynamic relationships and changing shock variances.
• Structural change improves forecasts — especially for inflation. In real-time out-of-sample tests, the TV-VAR beats fixed-coefficient VARs, time-varying ARs, and the random walk for inflation (in MSFE), and forecasts unemployment and the interest rate accurately.
• Robust to the recent, hard-to-forecast period. The inflation gains persist in the post-1985 / Great-Moderation era, when many models struggle to beat the random walk for inflation.
• Two channels. Allowing structural change helps by (i) capturing and exploiting changes in macroeconomic relationships, and (ii) detecting and forecasting changes in long-run dynamics — e.g., the decline in trend inflation and unemployment since the mid-1980s.
• Time variation in volatility is essential. The stochastic-volatility component is important for calibrated density forecasts and for the point-forecast gains, not just the drifting coefficients.

Concepts Introduced or Extended

• Time-Varying Parameter VAR
• Stochastic Volatility
• Forecasting

Entities Mentioned

• Antonello D'Agostino
• Luca Gambetti
• Domenico Giannone

Quotes

"The aim of this paper is to assess whether explicitly modeling structural change increases the accuracy of macroeconomic forecasts... In particular for inflation the TV-VAR outperforms, in terms of mean square forecast error, all the competing models."

My Take

This is the applied case for putting time variation — in both coefficients and volatilities — to work for forecasting, and its cleanest result is on inflation, the variable where beating a random walk is notoriously hard. The finding that a TV-VAR wins precisely during the Great-Moderation and post-2000 period is the practical payoff of the Cogley–Sargent/Primiceri machinery: the same drifting-parameter, stochastic-volatility structure that those papers used to describe changing US inflation dynamics also forecasts them better out of sample. It complements the large-BVAR strand (common stochastic volatility) — small TV-VARs with rich time variation vs. large BVARs with parsimonious time variation — two routes to the same goal of forecasting through structural change.