
on Forecasting 
By:  Gargano, Antonio; Timmermann, Allan G 
Abstract:  We compare different approaches to accounting for parameter instability in the context of macroeconomic forecasting models that assume either small, frequent changes versus models whose parameters exhibit large, rare changes. An empirical outofsample forecasting exercise for U.S. GDP growth and inflation suggests that models that allow for parameter instability generate more accurate density forecasts than constantparameter models although they fail to produce better point forecasts. Model combinations deliver similar gains in predictive performance although they fail to improve on the predictive accuracy of the single best model which is a specification that allows for timevarying parameters and stochastic volatility. 
Keywords:  GDP growth; inflation; regime switching; stochastic volatility; timevarying parameters 
JEL:  C22 C53 
Date:  2016–06 
URL:  http://d.repec.org/n?u=RePEc:cpr:ceprdp:11355&r=for 
By:  Elliott, Graham; Timmermann, Allan G 
Abstract:  Practices used to address economic forecasting problems have undergone substantial changes over recent years. We review how such changes have influenced the ways in which a range of forecasting questions are being addressed. We also discuss the promises and challenges arising from access to big data. Finally, we review empirical evidence and experience accumulated from the use of forecasting methods to a range of economic and financial variables. 
Keywords:  Big Data; Forecast evaluation; Forecast models; Model Instability; Parameter Estimation 
Date:  2016–06 
URL:  http://d.repec.org/n?u=RePEc:cpr:ceprdp:11354&r=for 
By:  Michal Franta 
Abstract:  Iterated multistep forecasts are usually constructed assuming the same model in each forecasting iteration. In this paper, the model coefficients are allowed to change across forecasting iterations according to the insample prediction performance at a particular forecasting horizon. The technique can thus be viewed as a combination of iterated and direct forecasting. The superior point and density forecasting performance of this approach is demonstrated on a standard mediumscale vector autoregression employing variables used in the Smets and Wouters (2007) model of the US economy. The estimation of the model and forecasting are carried out in a Bayesian way on data covering the period 1959Q12016Q1. 
Keywords:  Bayesian estimation, direct forecasting, iterated forecasting, multistep forecasts, VAR 
JEL:  C11 C32 C53 
Date:  2016–06 
URL:  http://d.repec.org/n?u=RePEc:cnb:wpaper:2016/05&r=for 
By:  Brave, Scott (Federal Reserve Bank of Chicago); Butters, R. Andrew (Indiana University); Justiniano, Alejandro (Federal Reserve Bank of Chicago) 
Abstract:  Mixed frequency Bayesian vector autoregressions (MFBVARs) allow forecasters to incorporate a large number of mixed frequency indicators into forecasts of economic activity. This paper evaluates the forecast performance of MFBVARs relative to surveys of professional forecasters and investigates the influence of certain specification choices on this performance. We leverage a novel realtime dataset to conduct an outofsample forecasting exercise for U.S. real gross domestic product (GDP). MFBVARs are shown to provide an attractive alternative to surveys of professional forecasters for forecasting GDP growth. However, certain specification choices such as model size and prior selection can affect their relative performance. 
Keywords:  Mixed frequency; Bayesian VAR; Realtime data; Nowcasting 
JEL:  C32 C53 E37 
Date:  2016–05–20 
URL:  http://d.repec.org/n?u=RePEc:fip:fedhwp:wp201605&r=for 
By:  Geert Dhaene; Jianbin Wu 
Abstract:  We introduce and evaluate mixedfrequency multivariate GARCH models for forecasting lowfrequency (weekly or monthly) multivariate volatility based on highfrequency intraday returns (at fiveminute intervals) and on the overnight returns. The lowfrequency conditional volatility matrix is modelled as a weighted sum of an intraday and an overnight component, driven by the intraday and the overnight returns, respectively. The components are specified as multivariate GARCH (1,1) models of the BEKK type, adapted to the mixedfrequency data setting. For the intraday component, the squared highfrequency returns enter the GARCH model through a parametrically specified mixeddata sampling (MIDAS) weight function or through the sum of the intraday realized volatilities. For the overnight component, the squared overnight returns enter the model with equal weights. Alternatively, the lowfrequency conditional volatility matrix may be modelled as a singlecomponent BEKKGARCH model where the overnight returns and the highfrequency returns enter through the weekly realized volatility (defined as the unweighted sum of squares of overnight and highfrequency returns), or where the overnight returns are simply ignored. All model variants may further be extended by allowing for a nonparametrically estimated slowlyvarying longrun volatility matrix. The proposed models are evaluated using fiveminute and overnight return data on four DJIA stocks (AXP, GE, HD, and IBM) from January 1988 to November 2014. The focus is on forecasting weekly volatilities (defined as the low frequency). The mixedfrequency GARCH models are found to systematically dominate the lowfrequency GARCH model in terms of insample fit and outofsample forecasting accuracy. They also exhibit much lower lowfrequency volatility persistence than the lowfrequency GARCH model. Among the mixedfrequency models, the lowfrequency persistence estimates decrease as the data frequency increases from daily to fiveminute frequency, and as overnight returns are included. That is, ignoring the available highfrequency information leads to spuriously high volatility persistence. Among the other findings are that the singlecomponent model variants perform worse than the twocomponent variants; that the overnight volatility component exhibits more persistence than the intraday component; and that MIDAS weighting performs better than not weighting at all (i.e., than realized volatility). 
Date:  2016–06 
URL:  http://d.repec.org/n?u=RePEc:ete:ceswps:544330&r=for 
By:  Jack Fosten (University of East Anglia) 
Abstract:  This paper provides an extension of DieboldMarianoWest (DMW) forecast accuracy tests to allow for factoraugmented models to be compared with nonnested benchmarks. The outof sample approach to forecast evaluation requires that both the factors and the forecasting model parameters are estimated in a rolling fashion, which poses several new challenges which we address in this paper. Firstly, we show the convergence rates of factors estimated in different rolling windows, and then give conditions under which the asymptotic distribution of the DMW test statistic is not affected by factor estimation error. Secondly, we draw attention to the issue of "signchanging" across rolling windows of factor estimates and factoraugmented model coefficients, caused by the lack of sign identification when using Principal Components Analysis to estimate the factors. We show that arbitrary signchanging does not affect the distribution of the DMW test statistic, but it does prohibit the construction of valid bootstrap critical values using existing procedures. We solve this problem by proposing a novel new normalization for rolling factor estimates, which has the effect of matching the sign of factors estimated in different rolling windows. We establish the firstorder validity of a simpletoimplement block bootstrap procedure and illustrate its properties using Monte Carlo simulations and an empirical application to forecasting U.S. CPI inflation. 
Keywords:  boostrap, diffusion index, factor model, predictive ability 
JEL:  C12 C22 C38 C53 
Date:  2016–01–28 
URL:  http://d.repec.org/n?u=RePEc:uea:ueaeco:2016_05&r=for 
By:  Chris McDonald; Craig Thamotheram; Shaun P. Vahey; Elizabeth C. Wakerly (Reserve Bank of New Zealand) 
Abstract:  We consider the fundamental issue of what makes a 'good' probability forecast for a central bank operating within an inflation targeting framework. We provide two examples in which the candidate forecasts comfortably outperform those from benchmark specifications by conventional statistical metrics such as root mean squared prediction errors and average logarithmic scores. Our assessment of economic significance uses an explicit loss function that relates economic value to a forecast communication problem for an inflation targeting central bank. We analyse the Bank of England's forecasts for inflation during the period in which the central bank operated within a strict inflation targeting framework in our first example. In our second example, we consider forecasts for inflation in New Zealand generated from vector autoregressions, when the central bank operated within a flexible inflation targeting framework. In both cases, the economic significance of the performance differential exhibits sensitivity to the parameters of the loss function and, for some values, the differentials are economically negligible. 
Date:  2016–06 
URL:  http://d.repec.org/n?u=RePEc:nzb:nzbdps:2016/10&r=for 
By:  Tamal Datta Chaudhuri; Indranil Ghosh 
Abstract:  Any discussion on exchange rate movements and forecasting should include explanatory variables from both the current account and the capital account of the balance of payments. In this paper, we include such factors to forecast the value of the Indian rupee vis a vis the US Dollar. Further, factors reflecting political instability and lack of mechanism for enforcement of contracts that can affect both direct foreign investment and also portfolio investment, have been incorporated. The explanatory variables chosen are the 3 month Rupee Dollar futures exchange rate (FX4), NIFTY returns (NIFTYR), Dow Jones Industrial Average returns (DJIAR), Hang Seng returns (HSR), DAX returns (DR), crude oil price (COP), CBOE VIX (CV) and India VIX (IV). To forecast the exchange rate, we have used two different classes of frameworks namely, Artificial Neural Network (ANN) based models and Time Series Econometric models. Multilayer Feed Forward Neural Network (MLFFNN) and Nonlinear Autoregressive models with Exogenous Input (NARX) Neural Network are the approaches that we have used as ANN models. Generalized Autoregressive Conditional Heteroskedastic (GARCH) and Exponential Generalized Autoregressive Conditional Heteroskedastic (EGARCH) techniques are the ones that we have used as Time Series Econometric methods. Within our framework, our results indicate that, although the two different approaches are quite efficient in forecasting the exchange rate, MLFNN and NARX are the most efficient. 
Date:  2016–07 
URL:  http://d.repec.org/n?u=RePEc:arx:papers:1607.02093&r=for 
By:  Nick Taylor 
Abstract:  The benefits associated with modeling BoxCox transformed realised variance data are assessed. In particular, the quality of realised variance forecasts with and without this transformation applied are examined in an outofsample forecasting competition. Using various realised variance measures, data transformations, volatility models and assessment methods, and controlling for data mining issues, the results indicate that data transformations can be economically and statistically significant. Moreover, the quartic transformation appears to be the most effective in this regard. The conditions under which the effectiveness of using transformed data varies are identified. 
Keywords:  C22, C53, C58, G17. 
Date:  2016–06–10 
URL:  http://d.repec.org/n?u=RePEc:bri:accfin:16/4&r=for 
By:  Jack Fosten (University of East Anglia) 
Abstract:  This paper provides consistent information criteria for the selection of forecasting models which use a subset of both the idiosyncratic and common factor components of a big dataset. This hybrid model approach has been explored by recent empirical studies to relax the strictness of pure factoraugmented model approximations, but no formal model selection procedures have been developed. The main difference to previous factoraugmented model selection procedures is that we must account for estimation error in the idiosyncratic component as well as the factors. Our first contribution shows that this combined estimation error vanishes at a slower rate than in the case of pure factoraugmented models in circumstances in which N is of larger order than sqrt(T), where N and T are the crosssection and time series dimensions respectively. Under these circumstances we show that existing factoraugmented model selection criteria are inconsistent, and the standard BIC is inconsistent regardless of the relationship between N and T. Our main contribution solves this issue by proposing new information criteria which account for the additional source of estimation error, whose properties are explored through a Monte Carlo simulation study. We conclude with an empirical application to longhorizon exchange rate forecasting using a recently proposed model with countryspecific idiosyncratic components from a panel of global exchange rates. 
Keywords:  forecasting, factor model, model selection, information criteria, idiosyncratic 
JEL:  C13 C22 C38 C52 C53 
Date:  2016–03–14 
URL:  http://d.repec.org/n?u=RePEc:uea:ueaeco:2016_07&r=for 