Efficient bayesian learning in non-linear dynamic models

This paper demonstrates the practical application of recently developed techniques of efficient numerical analysis for dynamic models. The models presented share a common basic structural foundation but nevertheless cover a very large arena of possible applications, as will be shown.     

Finite sample forecast results for vector autoregressive moving average models

Using the 'standard' approach to forecasting in the vector autoregressive moving average model, we establish basic general results on exact finite sample forecasts and their mean squared error matrices. Comparison between the exact and conditional methods of initiating the finite sample forecast calculations is presented, and a few illustrative cases are given.     

Estimation procedures for structural time series models

A univariate structural time series model based on the traditional decomposition into trend, seasonal and irregular components is defined. A number of methods of computing maximum likelihood estimators are then considered. These include direct maximization of various time domain likelihood function. The asymptotic properties of the estimators are given and a comparison between the various methods in terms of computational efficiency and accuracy is made. The methods are then extended to models with explanatory variables.     

Forecasting the business cycle without using minimum autocorrelation factors

We introduce a forecasting technique based on multivariate ideas previously applied in remote sensing. The approach has the trivial but nonetheless fundamental purpose of dividing the information inherent in the time series into important and unimportant. Important information is used for forecasting purposes while the unimportant is discarded. Although related to vector autoregression, giving asymptotically the same estimates, there are reasons to believe that the approach gives better precision of parameter estimates for finite samples as well as more precise predictions.     

Moving average threshold heterogeneous autoregressive (MAT-HAR) models

We propose moving average threshold heterogeneous autoregressive (MAT-HAR) models as a novel combination of heterogeneous autoregression (HAR) and threshold autoregression (TAR). The MAT-HAR has multiple groups of lags of a target series, and a threshold term can appear in each group. The threshold is a moving average of lagged target series, which guarantees time-varying thresholds and simple estimation via least squares. We show via Monte Carlo simulations that the MAT-HAR has sharp in-sample and out-of-sample performance. An empirical application on the industrial production of Japan suggests that significant threshold effects exist, and the MAT-HAR has a higher forecast accuracy than the HAR.     

ARARMA models for time series analysis and forecasting

Methods of time series forecasting are proposed which can be applied automatically. However, they are not rote formulae, since they are based on a flexible philosophy which can provide several models for consideration. In addition it provides diverse diagnostics for qualitatively and quantitatively estimating how well one can forecast a series. The models considered are called ARARMA models (or ARAR models) because the model fitted to a long memory time series (t) is based on sophisticated time series analysis of AR (or ARMA) schemes (short memory models) fitted to residuals Y(t) obtained by parsimonious‘best lag’non-stationary autoregression. Both long range and short range forecasts are provided by an ARARMA model Section 1 explains the philosophy of our approach to time series model identification. Sections 2 and 3 attempt to relate our approach to some standard approaches to forecasting; exponential smoothing methods are developed from the point of view of prediction theory (section 2) and extended (section 3). ARARMA models are introduced (section 4). Methods of ARARMA model fitting are outlined (sections 5,6). Since‘the proof of the pudding is in the eating’, the methods proposed are illustrated (section 7) using the classic example of international airline passengers.     

Finite sample prediction and interpolation for ARIMA models with missing data

A transformation which allows Cholesky decomposition to be used to evaluate the exact likelihood function of an ARIMA model with missing data has recently been suggested. This method is extended to allow calculation of finite sample predictions of future observations. The output from the exact likelihood evaluation may also be used to estimate missing series values. Copyright © 1999 John Wiley & Sons, Ltd.     

Unemployment variation over the business cycles: a comparison of forecasting models

Asymmetry has been well documented in the business cycle literature. The asymmetric business cycle suggests that major macroeconomic series, such as a country's unemployment rate, are non‐linear and, therefore, the use of linear models to explain their behaviour and forecast their future values may not be appropriate. Many researchers have focused on providing evidence for the non‐linearity in the unemployment series. Only recently have there been some developments in applying non‐linear models to estimate and forecast unemployment rates. A major concern of non‐linear modelling is the model specification problem; it is very hard to test all possible non‐linear specifications, and to select the most appropriate specification for a particular model. Artificial neural network (ANN) models provide a solution to the difficulty of forecasting unemployment over the asymmetric business cycle. ANN models are non‐linear, do not rely upon the classical regression assumptions, are capable of learning the structure of all kinds of patterns in a data set with a specified degree of accuracy, and can then use this structure to forecast future values of the data. In this paper, we apply two ANN models, a back‐propagation model and a generalized regression neural network model to estimate and forecast post‐war aggregate unemployment rates in the USA, Canada, UK, France and Japan. We compare the out‐of‐sample forecast results obtained by the ANN models with those obtained by several linear and non‐linear times series models currently used in the literature. It is shown that the artificial neural network models are able to forecast the unemployment series as well as, and in some cases better than, the other univariate econometrics time series models in our test. Copyright © 2004 John Wiley & Sons, Ltd.     

Hierarchical bayes forecasts of multinomial dirichlet data applied to coupon redemptions

This paper considers forecasting count data from a multinomial Dirichlet distribution. The forecasting procedure implements hierarchical Bayes methods in order to develop a prior distribution for a new series of data. The methodology is applied to the redemption of cents-off promotional coupons. In a forecasting experiment, early forecasts of new series are similar to those from pooling all redemptions from previous coupon promotions. However, the hierarchical Bayes model provides realistic estimates of forecasting errors, while those for the pooled forecasts are consistently optimistic. As the current series evolves, the hierarchical Bayes forecasts adapt more rapidly and are more accurate than pooled forecasts.     

Forecasting European GNP data through common factor models and other procedures

In this paper we present an extensive study of annual GNP data for five European countries. We look for intercountry dependence and analyse how the different economies interact, using several univariate ARIMA and unobserved components models and a multivariate model for the GNP incorporating all the common information among the variables. We use a dynamic factor model to take account of the common dynamic structure of the variables. This common dynamic structure can be non‐stationary (i.e. common trends) or stationary (i.e. common cycles). Comparisons of the models are made in terms of the root mean square error (RMSE) for one‐step‐ahead forecasts. For this particular group of European countries, the factor model outperforms the remaining ones. Copyright © 2002 John Wiley & Sons, Ltd.     

A comparison of models for forecasting the discovery of hydrocarbon deposits

There are many geologic resource assessment techniques available that can be used to estimate the hydrocarbon potential of frontier basins. The techniques do not, however, produce discovery rate forecasts, which are critical to government and industry planners. Various methods of predicting the discovery rate in frontier basins are reviewed and appraised within the context of the drilling history of Canada's Scotian Shelf. The reviewed models include a mean historical discovery rate technique, the exponential decline model, a logistic curve technique and a class of models based on a sampling without replacement approach to prediction. Models are found to perform best over the longer forecast horizons, with those models based on the sampling without replacement technique tending to perform best.     

Highest-density forecast regions for nonlinear and non-normal time series models

Forecast regions are a common way to summarize forecast accuracy. They usually consist of an interval symmetric about the forecast mean. However, symmetric intervals may not be appropriate forecast regions when the forecast density is not symmetric and unimodal. With many modern time series models, such as those which are non-linear or have non-normal errors, the forecast densities are often asymmetric or multimodal. The problem of obtaining forecast regions in such cases is considered and it is proposed that highest-density forecast regions be used. A graphical method for presenting the results is discussed.     

A Comparison of Box—Jenkins and objective methods for determining the order of a non-seasonal ARMA Model

This paper evaluates different procedures for selecting the order of a non-seasonal ARMA model. Specifically, it compares the forecasting accuracy of models developed by the personalized Box-Jenkins (BJ) methodology with models chosen by numerous automatic procedures. The study uses real series modelled by experts (textbook authors) in the BJ approach. Our results show that many objective selection criteria provide structures equal or superior to the time-consuming BJ method. For the sets of data used in this study, we also examine the influence of parsimony in time-series forecasting. Defining what models are too large or too small is sensitive to the forecast horizon. Automatic techniques that select the best models for forecasting are similar in size to BJ models although they often disagree on model order.     

Variance estimation for multivariate dynamic linear models

The problem of estimating unknown observational variances in multivariate dynamic linear models is considered. Conjugate procedures are possible for univariate models and also for special very restrictive common components models but they are not generally applicable. However, for clarity of operation and in order to avoid numerical integration, it is desirable to have conjugacy or approximate conjugacy. Such an approximate procedure is proposed based upon a simple analytic approximation. It is exact for the sub-class of conjugate models and improves on a previous procedure based upon the Robust filter.     

A multi-country dynamic factor model with stochastic volatility for euro area business cycle analysis

This paper develops a dynamic factor model that uses euro area country-specific information on output and inflation to estimate an area-wide measure of the output gap. Our model assumes that output and inflation can be decomposed into country-specific stochastic trends and a common cyclical component. Comovement in the trends is introduced by imposing a factor structure on the shocks to the latent states. We moreover introduce flexible stochastic volatility specifications to control for heteroscedasticity in the measurement errors and innovations to the latent states. Carefully specified shrinkage priors allow for pushing the model towards a homoscedastic specification, if supported by the data. Our measure of the output gap closely tracks other commonly adopted measures, with small differences in magnitudes and timing. To assess whether the model-based output gap helps in forecasting inflation, we perform an out-of-sample forecasting exercise. The findings indicate that our approach yields superior inflation forecasts, both in terms of point and density predictions.     

Selection and estimation of component models for seasonal time series

We present a method for investigating the evolution of trend and seasonality in an observed time series. A general model is fitted to a residual spectrum, using components to represent the seasonality. We show graphically how well the fitted spectrum captures the evidence for evolving seasonality associated with the different seasonal frequencies. We apply the method to model two time series and illustrate the resulting forecasts and seasonal adjustment for one series. Copyright © 2000 John Wiley & Sons, Ltd.     

A forecasting procedure for nonlinear autoregressive time series models

Forecasting for nonlinear time series is an important topic in time series analysis. Existing numerical algorithms for multi‐step‐ahead forecasting ignore accuracy checking, alternative Monte Carlo methods are also computationally very demanding and their accuracy is difficult to control too. In this paper a numerical forecasting procedure for nonlinear autoregressive time series models is proposed. The forecasting procedure can be used to obtain approximate m‐step‐ahead predictive probability density functions, predictive distribution functions, predictive mean and variance, etc. for a range of nonlinear autoregressive time series models. Examples in the paper show that the forecasting procedure works very well both in terms of the accuracy of the results and in the ability to deal with different nonlinear autoregressive time series models. Copyright © 2005 John Wiley & Sons, Ltd.     

Covariance estimation for multivariate conditionally Gaussian dynamic linear models

In multivariate time series, estimation of the covariance matrix of observation innovations plays an important role in forecasting as it enables computation of standardized forecast error vectors as well as the computation of confidence bounds of forecasts. We develop an online, non‐iterative Bayesian algorithm for estimation and forecasting. It is empirically found that, for a range of simulated time series, the proposed covariance estimator has good performance converging to the true values of the unknown observation covariance matrix. Over a simulated time series, the new method approximates the correct estimates, produced by a non‐sequential Monte Carlo simulation procedure, which is used here as the gold standard. The special, but important, vector autoregressive (VAR) and time‐varying VAR models are illustrated by considering London metal exchange data consisting of spot prices of aluminium, copper, lead and zinc. Copyright © 2007 John Wiley & Sons, Ltd.     

Building neural network models for time series: a statistical approach

This paper is concerned with modelling time series by single hidden layer feedforward neural network models. A coherent modelling strategy based on statistical inference is presented. Variable selection is carried out using simple existing techniques. The problem of selecting the number of hidden units is solved by sequentially applying Lagrange multiplier type tests, with the aim of avoiding the estimation of unidentified models. Misspecification tests are derived for evaluating an estimated neural network model. All the tests are entirely based on auxiliary regressions and are easily implemented. A small‐sample simulation experiment is carried out to show how the proposed modelling strategy works and how the misspecification tests behave in small samples. Two applications to real time series, one univariate and the other multivariate, are considered as well. Sets of one‐step‐ahead forecasts are constructed and forecast accuracy is compared with that of other nonlinear models applied to the same series. Copyright © 2006 John Wiley & Sons, Ltd.