We investigate the evidence for structural breaks in the parameters of autoregressive models of U.S. post-war macroeconomic time series. There is substantial model uncertainty associated with such models, including uncertainty related to lag selection, the number of structural changes, and the specific parameters that change at each break date. We develop a feasible approach to Bayesian Model Averaging (BMA), where the model space encompasses each of these sources of uncertainty. This BMA procedure performs very well in Monte Carlo simulations calibrated to match relevant macroeconomic time series. We then apply the BMA approach to a cross-section of U.S. macroeconomic variables measuring inflation, production growth, and labor market conditions, finding substantial evidence for structural breaks in all of these series. For most series there are multiple structural breaks detected. We find pervasive evidence for at least one, and often multiple, breaks in conditional variance parameters, and for price inflation series we find strong evidence of changes in persistence. We find little evidence for changes in trend growth rates of production series, or in the natural rate of unemployment. For most series there is substantial uncertainty along one or more dimension of model specification, calling into question the common practice of basing inference on a single selected structural break model.
This paper investigates the informational content of regular revisions to real GDP growth and its components. We perform a real-time forecasting exercise for the advance estimate of real GDP growth using dynamic regression models that include revisions to GDP and its components. Echoing other work in the literature, we find little evidence that including aggregate GDP growth revisions improves forecast accuracy relative to an AR(1) baseline model; however, models that include revisions to components of GDP improve forecast accuracy. The first revision to consumption is particularly relevant in that every model that includes the revision outperforms the baseline model. The improvements are quite sizable with top-performing models forecasting 0.2 percentage points closer to the advance estimate of growth, and a large subset of models improves RMSFE by greater than 5%. We use Bayesian model averaging to underscore that our results are driven by the informational content of revisions. The posterior probability of models with the first revision to consumption is significantly higher than our baseline model, despite strong priors that the latter should be the preferred forecasting model.
This paper investigates the nature of the Federal Open Market Committee's (FOMC's) interest rate rule, with a focus on which variables have been relevant to the FOMC over the past 40 years. I consider a large number of potential variables, including alternate measures of inflation, aggregate real activity, and sectoral variables. Based on inclusion probabilities derived from Bayesian Model Averaging (BMA) over a sample from 1970-2007, I find that the FOMC responds to changes in unemployment rather than to changes in GDP growth. Additionally, I find that the FOMC reacts not only to inflation and aggregate output, but also to measures of sectoral activity, such as changes in commodity prices. Finally, I find that using BMA improves out-of-sample forecasting performance over baseline Taylor-type interest rate rules.
Many researchers and economic commentators believe that the Federal Open Market Committee (FOMC) has changed interest rate policy over time. Mathematically, FOMC policy is typically described as a linear interest rate rule, in which the FOMC changes interest rates based on measure of the inflation gap and a measure of the output gap. In this context, a change in policy is associated with a change in the coefficients of this rule. I introduce a new econometric model in order to determine if there have been changes in these coefficients. I call this model a Markov Switching Stochastic Search Variable Selection (MS-SSVS) model, as it builds on the work of George and McCulloch (1993) who introduced SSVS in linear models, and also on the work of Hamilton (1989), Kim and Nelson (1999), Fruwirth-Schnatter (2006) and others who have popularized the use of Markov-Switching models. This model embeds linear regression and Markov-Switching models as special cases, as it allows regression coefficients to be freely estimated, restricted to be identical across regimes, or restricted to be equal to zero. Because it builds these features into the likelihood function of a single model, estimation is fast and does not involve the approximation of marginal likelihoods. I find that there is very little evidence of coefficient change, with the data suggesting only a change in response to the unemployment gap.
Works in Progress
Break or No Break? Identifying Structural Breaks using Classical, Bayesian, and Machine Learning Approaches (with Yamin Ahmad and Ming Chien Lo)
What Predicts Informality? A Bayesian Model Averaging Approach (with Tyler Schipper)