Agustín Maravall

An Application of Nonlinear Time Series Forecasting

By means of a real application, it is seen how ARIMA forecasts can be improved when nonlinearities are present. The autocorrelation function (ACF) of the squared residuals provides a convenient tool to check the linearity assumption. Once nonlinearity has been detected, parsimonious bilinear processes seem rather adequate to model it. The detection of nonlinearity and the forecast improvement appear to be rather robust with respect to changes in the linear and bilinear specification. Finally, what bilinear models seem to capture are periods of atypical behavior or sequences of outliers.

Measuring business cycles in economic time series

1 Introduction and Brief Summary.- 2 A Brief Review of Applied Time Series Analysis.- 2.1 Some Basic Concepts.- 2.2 Stochastic Processes and Stationarity.- 2.3 Differencing.- 2.4 Linear Stationary Process, Wold Representation. and Auto-correlation Function.- 2.5 The Spectrum.- 2.6 Linear Filters and Their Squared Gain.- 3 ARIMA Models and Signal Extraction.- 3.1 ARIMA Models.- 3.2 Modeling Strategy, Diagnostics and Inference.- 3.2.1 Identification.- 3.2.2 Estimation and Diagnostics.- 3.2.3 Inference.- 3.2.4 A Particular Class of Models.- 3.3 Preadjustment.- 3.4 Unobserved Components and Signal Extraction.- 3.5 ARIMA-Model-Based Decomposition of a Time Series.- 3.6 Short-Term and Long-Term Trends.- 4 Detrending and the Hodrick-Prescott Filter.- 4.1 The Hodrick-Prescott Filter: Equivalent Representations.- 4.2 Basic Characteristics of the Hodrick-Prescott Filter.- 4.3 Some Criticisms and Discussion of the Hodrick-Prescott Filter.- 4.4 The Hodrick-Prescott Filter as a Wiener-Kolmogorov Filter.- 4.4.1 An Alternative Representation.- 4.4.2 Derivation of the Filter.- 4.4.3 The Algorithm.- 4.4.4 A Note on Computation.- 5 Some Basic Limitations of the Hodrick-Prescott Filter.- 5.1 Endpoint Estimation and Revisions.- 5.1.1 Preliminary Estimation and Revisions.- 5.1.2 An Example.- 5.2 Spurious Results.- 5.2.1 Spurious Crosscorrelation.- 5.2.2 Spurious Autocorrelation Calibration.- 5.2.3 Spurious Periodic Cycle.- 5.3 Noisy Cyclical Signal.- 6 Improving the Hodrick-Prescott Filter.- 6.1 Reducing Revisions.- 6.2 Improving the Cyclical Signal.- 7 Hodrick-Prescott Filtering Within a Model-Based Approach.- 7.1 A Simple Model-Based Algorithm.- 7.2 A Complete Model-Based Method Spuriousness Reconsidered.- 7.3 Some Comments on Model-Based Diagnostics and Inference.- 7.4 MMSE Estimation of the Cycle: A Paradox.- References.- Author Index.

On Structural Time Series Models and the Characterization of Components

This article analyzes certain properties of a class of recently proposed structural time series models in which particular structures are imposed upon the unobserved components of an observed time series. It is shown how the overall model can be expected to fit series, such as those for which the X-11 or Airline models are appropriate. As for the components, identification of the model is achieved by assigning a certain amount of white noise variation to the trend and seasonal components. It is shown that the structural approach can be modified to avoid trend and seasonal components contaminated by noise.

Temporal aggregation, systematic sampling, and the Hodrick-Prescott filter

The time aggregation properties of the Hodrick-Prescott (HP) filter, which decomposes a time series into trend and cycle, are analyzed for the case of annual, quarterly, and monthly data. Aggregation of the disaggregate components cannot be obtained as the exact result from direct application of an HP filter to the aggregate series. Employing several criteria, HP decompositions for different levels of aggregation that provide similar results can be found. The aggregation is guided by the principle that the period associated with the frequency for which the filter gain is 12 should not be altered. This criterion is intuitive and easy to apply. It is shown that it is approximated, to the first order, by an already proposed empirical rule and that alternative, more complex criteria yield similar results. Moreover, the values of the smoothing parameter of the HP filter that provide results which are approximately consistent under aggregation are considerably robust with respect to the ARIMA model of the series. Aggregation is found to perform better for the case of temporal aggregation than for systematic sampling. The desirability of exact aggregation consistency is investigated. A clarification concerning the supposed spuriousness of the cycles obtained by the HP filter is discussed.

Minimum Mean Squared Error Estimation of the Noise in Unobserved Component Models

In model-based estimation of unobserved components, the minimum mean squared error estimator of the noise component is different from white noise. In this article, some of the differences are analyzed. It is seen how the variance of the component is always underestimated, and the smaller the noise variance, the larger the underestimation. Estimators of small-variance noise components will also have large autocorrelations. Finally, in the context of an application, the sample autocorrelation function of the estimated noise is seen to perform well as a diagnostic tool, even when the variance is small and the series is of relatively short length.

An application of the TRAMO-SEATS automatic procedure; direct versus indirect adjustment

The ARIMA-model-based methodology of programs TRAMO and SEATS is applied for seasonal adjustment and trend-cycle estimation of the exports, imports, and balance of trade Japanese series. The programs are used in an automatic mode, and the results are analyzed. It is shown how the SEATS output can be of help when discriminating among competing models. Finally, the example is used to discuss the important problem of the choice between direct and indirect adjustment of an aggregate. It is concluded that, because aggregation has a strong effect on the spectral shape of the series, and because seasonal adjustment is a non-linear transformation of the original series, direct adjustment is preferable, even at the cost of destroying identities between the original series.

Estimation Error and the Specification of Unobserved Component Models.

The paper deals with the problem of identifying stochastic unobserved two-component models, as in seasonal adjustment or trend-cycle decompositions. Solutions based on the properties of the unobserved component estimation error are considered, and analytical expressions for the variances and covariances of the different types of estimation errors (errors in the final, preliminary, and concurrent estimator and in the forecast) are obtained for any admissible decomposition.

A note on minimum mean squared error estimation of signals with unit roots

Abstract Using an ARIMA parametrization, this note provides a very simple proof of how the Wiener-Kolmogorov-Whittle filter to estimate signals in time series can be extended to the nonstationary case. The proof is valid for any number and type of unit roots (not simply those implied by differencing) in both the signal and the overall model.

Revisions in ARIMA Signal Extraction

Abstract The problem of decomposing an observed series, assumed to follow an ARIMA process, into signal plus noise is considered. It is well known that the preliminary estimates of the signal will be subject to revisions as more data become available. For a general ARIMA process, the revision in the concurrent estimate of the signal is seen to follow a stationary ARMA process, easily derived from the overall series model. The results are extended to non-concurrent preliminary estimates. Finally, it is found that, except for a scale factor, the revisions are the same for all admissible decompositions and the canonical decomposition maximizes the variance of the revision.

A note on identification of multivariate time-series models

Abstract We analyze some of the difficulties of identifying multivariate time-series models with feedback, when carried through the two-stage procedure as in Haugh-Box (1977). We find that if the procedure is simplified, as in Jenkins (1979), the identified model can be seriously misspecified. When the procedure is applied correctly, it becomes extremely complicated, as in Granger-Newbold (1977). In such cases, there is a serious risk of overparametrization. The complication is mostly caused by the underlying structure of the multivariate model for the univariate innovations which can be considerably more complicated than the multivariate time-series model of interest.