Weights and robustness of model-based seasonal adjustment

Model-based seasonal adjustment implicitly defines a set of weights at the ends of series as well as in the middle. Until now, with the exception of very simple models, the weights have been obtained numerically. In this paper we give the analytical expressions for the weights for both the structural and the ARIMA framework for a model which contains trend, seasonal and irregular component. In the final part of the paper we address the question of robustness of model-based seasonal adjustment. We analyse practically, using real time series, and theoretically, through the analysis of the shape of the weights, how the fitting of different specifications for the non-seasonal part affects the extraction of the seasonal component. © 1998 John Wiley & Sons, Ltd.     

Short term load forecasting in electric power systems: A comparison of ARMA models and extended wiener filtering

On-line prediction of electric load in the buses of the EHV grid of a power generation and transmission system is basic information required by on-line procedures for centralized advanced dispatching of power generation. This paper presents two alternative approaches to on-line short term forecasting of the residual component of the load obtained after the removal of the base load from a time series of total load. The first approach involves the use of stochastic ARMA models with time-varying coefficients. The second consists in the use of an extension of Wiener filtering due to Zadeh and Ragazzini. Real data representing a load process measured in an area of Northern Italy and simulated data reproducing a non-stationary process with known characteristics constitute the basis of a numerical comparison allowing one to determine under which conditions each method is more appropriate.     

Structural Time-series Modelling of Monetary Aggregates: A Case Study for Eleven European Countries

Monetary aggregates for eleven European countries are analysed using the structural time-series methodology, paying special attention to unit root issues. Estimation of the parameters of the models is carried out by applying the asymptotic least squares (ALS) procedure. A comparison with the maximum likelihood estimates obtained via the Kalman filter shows that ALS is an alternative to Kalman filter estimation. The empirical results show that for only a small number of series the four variance parameters of the basic structural model are strictly positive. For the majority of the series the variance of the irregular component is equal to 0.©1997 John Wiley & Sons, Ltd.     

Temporal aggregation and systematic sampling in structural time-series models

Given a structural time-series model specified at a basic time interval, this paper deals with the problems of forecasting efficiency and estimation accuracy generated when the data are collected at a timing interval which is a multiple of the time unit chosen to build the basic model. Results are presented for the simplest structural models, the trend plus error models, under the assumption that the parameters of the model are known. It is shown that the gains in forecasting efficiency and estimation accuracy for having data at finer intervals are considerable for both stock and flow variables with only one exception. No gain in forecasting efficiency is achieved in the case of a stock series that follows a random walk.     

Analysis of many short time sequences: Forecast improvements achieved by shrinkage

Credibility models in actuarial science deal with multiple short time series where each series represents claim amounts of different insurance groups. Commonly used credibility models imply shrinkage of group-specific estimates towards their average. In this paper we model the claim size yu in group i and at time t as the sum of three independent components: y_it = μ_r + δ_i + ?_it. The first component, μt = μ_t?1 + m_t, represents time-varying levels that are common to all groups. The second component, δ_i, represents random group offsets that are the same in all periods, and the third component represents independent measurement errors. In this paper we show how to obtain forecasts from this model and we discuss the nature of the forecasts, with particular emphasis on shrinkage. We also assess the forecast improvements that can be expected from such a model. Finally, we discuss an extension of the above model which also allows the group offsets to change over time. We assume that the offsets for different groups follow independent random walks.     

Short‐term forecasts of economic activity: Are fortnightly factors useful?

A short‐term mixed‐frequency model is proposed to estimate and forecast Italian economic activity fortnightly. We introduce a dynamic one‐factor model with three frequencies (quarterly, monthly, and fortnightly) by selecting indicators that show significant coincident and leading properties and are representative of both demand and supply. We conduct an out‐of‐sample forecasting exercise and compare the prediction errors of our model with those of alternative models that do not include fortnightly indicators. We find that high‐frequency indicators significantly improve the real‐time forecasts of Italian gross domestic product (GDP); this result suggests that models exploiting the information available at different lags and frequencies provide forecasting gains beyond those based on monthly variables alone. Moreover, the model provides a new fortnightly indicator of GDP, consistent with the official quarterly series.     

Convergence and the Constant Dynamic Linear Model

It is well known that, as calculated using the Kalman filter recurrence relationships, the posterior parameter variance and the adaptive vector of observable constant dynamic linear models converge to limiting values. However, most proofs are tortuous, some have subtle errors and some relate only to specific cases. An elegant probabilistic convergence proof demonstrates that the limit is independent of the initial parametric prior. The result is extended to a class of multivariate dynamic linear models. Finally the proof is shown to apply to many non-observable constant DLMs. © 1997 John Wiley & Sons, Ltd.     

Transformations and dynamic linear models

The dynamic linear model (DLM) with additive Gaussian errors provides a useful statistical tool that is easily implemented because of the simplicity of updating a normal model that has a natural conjugate prior. If the model is not linear or if it does not have additive Gaussian errors, then numerical methods are usually required to update the distributions of the unknown parameters. If the dimension of the parameter space is small, numerical methods are feasible. However, as the number of unknown parameters increases, the numerial methods rapidly grow in complexity and cost. This article addresses the situation where a state dependent transformation of the observations follows the DLM, but a priori the appropriate transformation is not known. The Box-Cox family, which is indexed by a single parameter, illustrates the methodology. A prior distribution is constructed over a grid of points for the transformation parameter. For each value of the grid the relevant parameter esitmates and forecasts are obtained for the transformed series. These quantities are then integrated by the current distribution of the transformation parameter. When a new observation becomes available, parallel Kalman filters are used to update the distributions of the unknown parameters and to compute the likelihood of the transformation parameter at each grid point. The distribution of the transformation parameter is then updated.     

Nonlinear predictor models

A complete dynamic model is introduced within the Bayesian framework. This includes the dynamic linear model and the normal discount Bayesian model as special cases and extends to some well-known models with nonlinear predictors. A number of practically important models are formulated and simple recurrence formulas, similar to those of Kalman, are used in the sequential estimation of the parameters. Finally, a number of practical examples and applications are given.     

Deletion diagnostics for transformations of time series

Deletion diagnostics are derived for the effect of individual observations on the estimated transformation of a time series. The paper uses the modified power transformation of Box and Cox to provide a parametric family of transformations. Inference about the transformation parameter is made through regression on a constructed variable. The effect of deletion of observations on residuals and on the estimate of the regression parameter are obtained. Index plots of the diagnostic quantities are shown to be highly informative. Structural time series modelling is used, so that the results readily extend to inference about regression on other explanatory variables.     

Disaggregation of annual flow data with multiplicative trends

Various methods based on smoothing or statistical criteria have been used for constructing disaggregated values compatible with observed annual totals. The present method is based on a time‐series model in a state space form and allows for a prescribed multiplicative trend. It is applied to US GNP data which have been used for comparing methods suggested for this purpose. The model can be extended to include quarterly series, related to the unknown disaggregated values. But as the estimation criteria are based on prediction errors of the aggregated values, the estimated form may not be optimal for reproducing high‐frequency variations of the disaggregated values. Copyright © 1999 John Wiley & Sons, Ltd.