Forecasting & Nowcasting
Forecasting and Nowcasting
"Forecasting Output," with Simon Potter, in Handbook of Economic Forecasting, vol. 2, ed. A. Timmermann and G. Elliott, Elsevier/North Holland, 1-56, 2013. (Download Working Paper or Handbook)
Abstract: This chapter surveys the recent literature on output forecasting, and examines the real-time forecasting ability of several models for U.S. output growth. In particular, it evaluates the accuracy of short-term forecasts of linear and nonlinear structural and reduced-form models, and judgmental forecasts of output growth. Our emphasis is on using solely the information that was available at the time the forecast was being made, in order to reproduce the forecasting problem facing forecasters in real-time. We find that there is a large difference in forecast performance across business cycle phases. In particular, it is much harder to forecast output growth during recessions than during expansions. Simple linear and nonlinear autoregressive models have the best accuracy in forecasting output growth during expansions, although the dynamic stochastic general equilibrium model and the vector autoregressive model with financial variables do relatively well. On the other hand, we find that most models do poorly in forecasting output growth during recessions. The autoregressive model based on the nonlinear dynamic factor model that takes into account asymmetries between expansions and recessions displays the best real time forecast accuracy during recessions. Even though the Blue Chip forecasts are comparable, the dynamic factor Markov switching model has better accuracy, particularly with respect to the timing and depth of output fall during recessions in real time. The results suggest that there are large gains in considering separate forecasting models for normal times and models especially designed for periods of abrupt changes, such as during recessions and financial crises.
Abstract: This chapter surveys the recent literature on output forecasting, and examines the real-time forecasting ability of several models for U.S. output growth. In particular, it evaluates the accuracy of short-term forecasts of linear and nonlinear structural and reduced-form models, and judgmental forecasts of output growth. Our emphasis is on using solely the information that was available at the time the forecast was being made, in order to reproduce the forecasting problem facing forecasters in real-time. We find that there is a large difference in forecast performance across business cycle phases. In particular, it is much harder to forecast output growth during recessions than during expansions. Simple linear and nonlinear autoregressive models have the best accuracy in forecasting output growth during expansions, although the dynamic stochastic general equilibrium model and the vector autoregressive model with financial variables do relatively well. On the other hand, we find that most models do poorly in forecasting output growth during recessions. The autoregressive model based on the nonlinear dynamic factor model that takes into account asymmetries between expansions and recessions displays the best real time forecast accuracy during recessions. Even though the Blue Chip forecasts are comparable, the dynamic factor Markov switching model has better accuracy, particularly with respect to the timing and depth of output fall during recessions in real time. The results suggest that there are large gains in considering separate forecasting models for normal times and models especially designed for periods of abrupt changes, such as during recessions and financial crises.
"Real-Time Nowcasting of Nominal GDP with Structural Breaks", with W. Barnett and D. Leiva-Leon, Journal of Econometrics. Vol. 191, April, 312-324, 2016. (Download or JE)
Abstract: This paper provides early assessments of current U.S. Nominal GDP growth, which has been considered as a potential new monetary policy target. The nowcasts are computed using the exact amount of information that policy makers have available at the time predictions are made. However, real time information arrives at different frequencies and asynchronously, which poses the challenge of mixed frequencies, missing data, and ragged edges. This paper proposes a multivariate state space model that not only takes into account asynchronous information inflow it also allows for potential parameter instability (DYMIBREAK). We use small scale confirmatory factor analysis in which the candidate variables are selected based on their ability to forecast nominal GDP. The model is fully estimated in one step using a nonlinear Kalman filter, which is applied to obtain simultaneously both optimal inferences on the dynamic factor and parameters. Differently from principal component analysis, the proposed factor model captures the comovement rather than the variance underlying the variables. We compare the predictive ability of the model with other univariate and multivariate specifications. The results indicate that the proposed model containing information on real economic activity, inflation, interest rates, and Divisia monetary aggregates produces the most accurate real time nowcasts of nominal GDP growth.
Abstract: This paper provides early assessments of current U.S. Nominal GDP growth, which has been considered as a potential new monetary policy target. The nowcasts are computed using the exact amount of information that policy makers have available at the time predictions are made. However, real time information arrives at different frequencies and asynchronously, which poses the challenge of mixed frequencies, missing data, and ragged edges. This paper proposes a multivariate state space model that not only takes into account asynchronous information inflow it also allows for potential parameter instability (DYMIBREAK). We use small scale confirmatory factor analysis in which the candidate variables are selected based on their ability to forecast nominal GDP. The model is fully estimated in one step using a nonlinear Kalman filter, which is applied to obtain simultaneously both optimal inferences on the dynamic factor and parameters. Differently from principal component analysis, the proposed factor model captures the comovement rather than the variance underlying the variables. We compare the predictive ability of the model with other univariate and multivariate specifications. The results indicate that the proposed model containing information on real economic activity, inflation, interest rates, and Divisia monetary aggregates produces the most accurate real time nowcasts of nominal GDP growth.
"Mortgage Default Risk: New Evidence from Internet Search Queries", with S. Gabriel and C. Lutz, Journal of Urban Economics. Vol. 96, November, 91-111, 2016. (Download)
Abstract: We use Google search query data to develop a broad-based and real-time index of mortgage default risk. Unlike established indicators, our Mortgage Default Risk Index (MDRI) directly reflects households’concerns regarding their risk of mortgage default. The MDRI predicts housing returns, mortgage delinquency indicators, and subprime credit default swaps. These results persist both in- and out-of-sample and at multiple data frequencies. Together, research findings suggest internet search queries yield valuable new insights into household mortgage default risk.
Abstract: We use Google search query data to develop a broad-based and real-time index of mortgage default risk. Unlike established indicators, our Mortgage Default Risk Index (MDRI) directly reflects households’concerns regarding their risk of mortgage default. The MDRI predicts housing returns, mortgage delinquency indicators, and subprime credit default swaps. These results persist both in- and out-of-sample and at multiple data frequencies. Together, research findings suggest internet search queries yield valuable new insights into household mortgage default risk.
"Forecasting Recessions Using the Yield Curve,” with S. Potter, Journal of Forecasting, 24, 2, 77-103, 2005. (Download Repec or ForecYield)
Abstract: We compare forecasts of recessions using four different specifications of the probit model: a time-invariant conditionally independent version, a business cycle specific conditionally independent model, a time-invariant probit with autocorrelated errors, and a business cycle specific probit with autocorrelated errors. ; The more sophisticated versions of the model take into account some of the potential underlying causes of the documented predictive instability of the yield curve. We find strong evidence in favor of the more sophisticated specification, which allows for multiple breakpoints across business cycles and autocorrelation. We also develop a new approach to the construction of real time forecasting of recession probabilities.
Abstract: We compare forecasts of recessions using four different specifications of the probit model: a time-invariant conditionally independent version, a business cycle specific conditionally independent model, a time-invariant probit with autocorrelated errors, and a business cycle specific probit with autocorrelated errors. ; The more sophisticated versions of the model take into account some of the potential underlying causes of the documented predictive instability of the yield curve. We find strong evidence in favor of the more sophisticated specification, which allows for multiple breakpoints across business cycles and autocorrelation. We also develop a new approach to the construction of real time forecasting of recession probabilities.
“Predicting Recessions: Evidence from the Yield Curve in the Presence of Structural Breaks,” with S. Potter, Economics Letters, Vol. 77, No. 2, 245-253, 2002. (Download Working Paper, Repec or PredicBreaks)
Abstract: We use a probit model of the term structure to examine the stability of recession forecasts under the presence of a structural break. We find strong evidence of a break, but with very uncertain location, which affects considerably recession predictions.
Abstract: We use a probit model of the term structure to examine the stability of recession forecasts under the presence of a structural break. We find strong evidence of a break, but with very uncertain location, which affects considerably recession predictions.
"Nonstationarities and Markov Switching Models," with Y. Su, in Recent Advances in Estimating Nonlinear Models, Springer, 123-148, 2013. (Download)
Abstract: This paper proposes a flexible model that allows for recent changes observed in the US business cycle in the last six decades. It proposes a Markov switching model with three Markov processes to characterize the dynamics of US output fluctuations. We consider the possibility that both the mean and the variance of growth rates of real GDP can have short run fluctuations in addition to the possibility of a long run permanent break. We find that, differently from several alternative specifications in the literature, the proposed flexible framework successfully represents all business cycle phases, including the Great Recession. In addition, we find that the volatility of US output fluctuations has both a long run pattern, characterized by a structural break in 1984, as well as business cycle dynamics, in which periods of high uncertainty are associated with NBER recessions.
Abstract: This paper proposes a flexible model that allows for recent changes observed in the US business cycle in the last six decades. It proposes a Markov switching model with three Markov processes to characterize the dynamics of US output fluctuations. We consider the possibility that both the mean and the variance of growth rates of real GDP can have short run fluctuations in addition to the possibility of a long run permanent break. We find that, differently from several alternative specifications in the literature, the proposed flexible framework successfully represents all business cycle phases, including the Great Recession. In addition, we find that the volatility of US output fluctuations has both a long run pattern, characterized by a structural break in 1984, as well as business cycle dynamics, in which periods of high uncertainty are associated with NBER recessions.
"Predicting Recessions in Brazil," with I. Morais, Latin American Meetings of the Econometric Society and Latin American and Caribbean Economic Association Meetings, Rio de Janeiro, Brazil, 2008. (Download ).
Abstract: This paper constructs leading indicators that anticipate turning points of the Brazilian business cycle. We propose a time-varying autoregressive probit model, which is composed of several economic series that display predictive power to anticipate the beginning or end of recessions. The Brazilian economy is characterized by several different policy regimes and instabilities that have potentially engendered breaks in its dynamics.The extended probit model is especially suited for this economy, since it takes into consideration parameter change across each cycle in addition to phase duration in the estimation of recession probabilities. We find that the extended probit model exhibits superior predictive performance over the standard probit model in several dimensions both in-sample and in an out-of-sample real time exercise.
Abstract: This paper constructs leading indicators that anticipate turning points of the Brazilian business cycle. We propose a time-varying autoregressive probit model, which is composed of several economic series that display predictive power to anticipate the beginning or end of recessions. The Brazilian economy is characterized by several different policy regimes and instabilities that have potentially engendered breaks in its dynamics.The extended probit model is especially suited for this economy, since it takes into consideration parameter change across each cycle in addition to phase duration in the estimation of recession probabilities. We find that the extended probit model exhibits superior predictive performance over the standard probit model in several dimensions both in-sample and in an out-of-sample real time exercise.