Randomized singular spectrum analysis for long time series

ABSTRACT

Singular spectrum analysis (SSA) is a relatively new method for time series analysis and comes as a non-parametric alternative to the classical methods. This methodology has proven to be effective in analysing non-stationary and complex time series since it is a non-parametric method and do not require the classical assumptions over the stationarity or over the normality of the residuals. Although SSA have proved to provide advantages over traditional methods, the challenges that arise when long time series are considered, make the standard SSA very demanding computationally and often not suitable. In this paper we propose the randomized SSA which is an alternative to SSA for long time series without losing the quality of the analysis. The SSA and the randomized SSA are compared in terms of quality of the model fit and forecasting, and computational time. This is done by using Monte Carlo simulations and real data about the daily prices of five of the major world commodities.     

A comparison of stepwise common singular spectrum analysis and horizontal multi-channel singular spectrum analysis

The problem of forecasting a time series by using information provided by a second time series is considered. Two multivariate extensions of Singular Spectrum Analysis (SSA) are compared in terms of forecast error: Horizontal Multi-channel SSA and Stepwise Common SSA. Different signal structures, defined in terms of trend, period, amplitude and phase, are investigated. In broad terms we find that neither Horizontal Multichannel SSA nor Stepwise Common SSA is best in all cases. Horizontal MSSA is outperformed particularly in cases where different trends are considered.     

Vector and recurrent singular spectrum analysis: which is better at forecasting?

Singular spectrum analysis (SSA) is an increasingly popular and widely adopted filtering and forecasting technique which is currently exploited in a variety of fields. Given its increasing application and superior performance in comparison to other methods, it is pertinent to study and distinguish between the two forecasting variations of SSA. These are referred to as Vector SSA (SSA-V) and Recurrent SSA (SSA-R). The general notion is that SSA-V is more robust and provides better forecasts than SSA-R. This is especially true when faced with time series which are non-stationary and asymmetric, or affected by unit root problems, outliers or structural breaks. However, currently there exists no empirical evidence for proving the above notions or suggesting that SSA-V is better than SSA-R. In this paper, we evaluate out-of-sample forecasting capabilities of the optimised SSA-V and SSA-R forecasting algorithms via a simulation study and an application to 100 real data sets with varying structures, to provide a statistically reliable answer to the question of which SSA algorithm is best for forecasting at both short and long run horizons based on several important criteria.     

On the Optimal Parameters for Reconstruction and Forecasting in Singular Spectrum Analysis

There are two main parameters in Singular Spectrum Analysis (SSA). The aim of this study is to determine whether the optimal values of these parameters are different for reconstruction and forecasting stages, and if those are worth the extra computational effort and time which they require. Here, we evaluate these issues using simulation study.     

Multivariate posterior singular spectrum analysis

A generalized, multivariate version of the Posterior Singular Spectrum Analysis (PSSA) method is described for the identification of credible features in multivariate time series. We combine Bayesian posterior modeling with multivariate SSA (MSSA) and infer the MSSA signal components with a credibility analysis of the posterior sample. The performance of multivariate PSSA (MPSSA) is compared to the single-variate PSSA with an artificial example and the potential of MPSSA is demonstrated with real data using NAO and SOI climate index series.     

A Review of Some Modern Approaches to the Problem of Trend Extraction

This article presents a review of some modern approaches to trend extraction for one-dimensional time series, which is one of the major tasks of time series analysis. The trend of a time series is usually defined as a smooth additive component which contains information about the time series global change, and we discuss this and other definitions of the trend. We do not aim to review all the novel approaches, but rather to observe the problem from different viewpoints and from different areas of expertise. The article contributes to understanding the concept of a trend and the problem of its extraction. We present an overview of advantages and disadvantages of the approaches under consideration, which are: the model-based approach (MBA), nonparametric linear filtering, singular spectrum analysis (SSA), and wavelets. The MBA assumes the specification of a stochastic time series model, which is usually either an autoregressive integrated moving average (ARIMA) model or a state space model. The nonparametric filtering methods do not require specification of model and are popular because of their simplicity in application. We discuss the Henderson, LOESS, and Hodrick–Prescott filters and their versions derived by exploiting the Reproducing Kernel Hilbert Space methodology. In addition to these prominent approaches, we consider SSA and wavelet methods. SSA is widespread in the geosciences; its algorithm is similar to that of principal components analysis, but SSA is applied to time series. Wavelet methods are the de facto standard for denoising in signal procession, and recent works revealed their potential in trend analysis.     

An algorithm for computing principal points with respect to a loss function in the unidimensional case

In Flury (1990) the k principal points of a random vector X are defned as the points p(1),..., p(k) minimizing EX–p(i)²; i=1,..., k. We extend this concept to that of k principal points with respect to a loss function L, and present an algorithm for their computation in the univariate case.     

An Example Of The Use Of Andrews’ Plots To Detect Time Variations In Model Parameters And To Detect Outlying Observations

Andrews (1972) introduced a method of plotting highdimensional data in two dimensions. This method is exploited as a graphical tool for the examination of changes over time in the parameters of a time series model. An example using a Fourier series model is given to illustrate the method. It is also shown how outlying observations in the data can be found.

AMS (MOS) Subject Classifivations: 62M10, 62H30     

Detecting and identifying interventions with the Whittle spectral approach in a long memory panel data model

This article provides a procedure for the detection and identification of outliers in the spectral domain where the Whittle maximum likelihood estimator of the panel data model proposed by Chen [W.D. Chen, Testing for spurious regression in a panel data model with the individual number and time length growing, J. Appl. Stat. 33(88) (2006b), pp. 759–772] is implemented. We extend the approach of Chang and co-workers [I. Chang, G.C. Tiao, and C. Chen, Estimation of time series parameters in the presence of outliers, Technometrics 30 (2) (1988), pp. 193–204] to the spectral domain and through the Whittle approach we can quickly detect and identify the type of outliers. A fixed effects panel data model is used, in which the remainder disturbance is assumed to be a fractional autoregressive integrated moving-average (ARFIMA) process and the likelihood ratio criterion is obtained directly through the modified inverse Fourier transform. This saves much time, especially when the estimated model implements a huge data-set.

Through Monte Carlo experiments, the consistency of the estimator is examined by growing the individual number N and time length T, in which the long memory remainder disturbances are contaminated with two types of outliers: additive outlier and innovation outlier. From the power tests, we see that the estimators are quite successful and powerful.

In the empirical study, we apply the model on Taiwan's computer motherboard industry. Weekly data from 1 January 2000 to 31 October 2006 of nine familiar companies are used. The proposed model has a smaller mean square error and shows more distinctive aggressive properties than the raw data model does.     

Stochastic modelling of regional annual rainfall anomalies in East Africa

ARIMA (p, d, q) models were fitted to areal annual rainfall of two homogeneous regions in East Africa with rainfall records extending between the period 1922–80. The areal estimates of the regional rainfall were derived from the time series of the first eigenvector, which was significantly dominant at each of the two regions. The first eigenvector accounted for about 80% of the total rainfall variance in each region.

The class of ARIMA (p, d, q) models which best fitted the areal indices of relative wetness/dryness were the A R M A (3, 1) models. Tests of forecasting skill however indicated low skill in the forecasts given by these models. In all cases the models accounted for less than 50% of the total variance.

Spectral analysis of the indices time series indicated dominant quasi-periodic fluctuations around 2.2–2.8 years, 3–3.7 years, 5–6 years and 10–13 years. These spectral bands however accounted for very low proportion of the total rainfall variance.     

Distribution of the C statistic with applications to the sample mean of Poisson data

The C statistic, also known as the Cash statistic, is often used in astronomy for the analysis of low-count Poisson data. The main advantage of this statistic, compared to the more commonly used χ2 statistic, is its applicability without the need to combine data points. This feature has made the C statistic a very useful method to analyze Poisson data that have small (or even null) counts in each resolution element. One of the challenges of the C statistic is that its probability distribution, under the null hypothesis that the data follow a parent model, is not known exactly. This paper presents an effort towards improving our understanding of the C statistic by studying (a) the distribution of C statistic for a fully specified model, (b) the distribution of C_min resulting from a maximum-likelihood fit to a simple one-parameter constant model, i.e. a model that represents the sample mean of N Poisson measurements, and (c) the distribution of the associated ΔC statistic that is used for parameter estimation. The results confirm the expectation that, in the high-count limit, both C statistic and C_min have the same mean and variance as a χ2 statistic with same number of degrees of freedom. It is also found that, in the low-count regime, the expectation of the C statistic and C_min can be substantially lower than for a χ2 distribution. The paper makes use of recent X-ray observations of the astronomical source PG 1116+215 to illustrate the application of the C statistic to Poisson data.     

Markov (Set) chains application to predict mortality rates using extended Milevsky–Promislov generalized mortality models

The mortality rates ( μx,t) measure the frequency of deaths in a fixed: population and time interval. The ability to model and forecast μx,t allows determining, among others, fundamental characteristics of life expectancy tables, e.g. used to determine the amount of premium in life insurance, adequate to the risk of death. The article proposes a new method of modelling and forecasting μx,t, using the class of stochastic Milevsky–Promislov switch models with excitations. The excitations are modelled by second, fourth and sixth order polynomials of outputs from the non-Gaussian Linear Scalar Filter (nGLSF) model and taking into account the Markov (Set) chain. The Markov (Set) chain state space is defined based on even orders of the nGLSF polynomial. The model order determines the theoretical values of the death rates. The obtained results usually provide a more precise forecast of the mortality rates than the commonly used Lee–Carter model.     

The Methods for Approximation of Principal Points for Binary Distributions on the Basis of Submodularity

Principal points for binary distributions are able to be defined based on Flury’s principal points (1990 Flury, B.A. (1990). Principal points. Biometrica 77:33–41.[Crossref], [Web of Science ®] , [Google Scholar]). However, finding principal points for binary distributions is hard in a straightforward manner. In this article, a method for approximating principal points for binary distributions is proposed by formulating it as an uncapacitated location problem. Moreover, it is shown that the problem of finding principal points can be solved with the aid of submodular functions. It leads to a solution whose value is at least (1 ? 1/e) times the optimal value.     

Signal extraction goodness-of-fit diagnostic tests under model parameter uncertainty: Formulations and empirical evaluation

We present a time-domain goodness-of-fit (gof) diagnostic test that is based on signal-extraction variances for nonstationary time series. This diagnostic test extends the time-domain gof statistic of Maravall (2003 Maravall, A. (2003). A class of diagnostics in the ARIMA-model-based decomposition of a time series. Memorandum, Bank of Spain. Available at http://www.bde.es/servicio/software/tramo/diagnosticsamb.pdf [Google Scholar]) by taking into account the effects of model parameter uncertainty, utilizing theoretical results of McElroy and Holan (2009 McElroy, T., Holan, S. (2009). A local spectral approach for assessing time series model misspeci?cation. Journal of Multivariate Analysis 100:604–621.[Crossref], [Web of Science ®] , [Google Scholar]). We demonstrate that omitting this correction results in a severely undersized statistic. Adequate size and power are obtained in Monte Carlo studies for fairly short time series (10 to 15 years of monthly data). Our Monte Carlo studies of finite sample size and power consider different combinations of both signal and noise components using seasonal, trend, and irregular component models obtained via canonical decomposition. Details of the implementation appropriate for SARIMA models are given. We apply the gof diagnostic test statistics to several U.S. Census Bureau time series. The results generally corroborate the output of the automatic model selection procedure of the X-12-ARIMA software, which in contrast to our diagnostic test statistic does not involve hypothesis testing. We conclude that these diagnostic test statistics are a useful supplementary model-checking tool for practitioners engaged in the task of model-based seasonal adjustment.     

Critical values improvement for the standard normal homogeneity test by combining Monte Carlo and regression approaches

The distribution of the test statistics of homogeneity tests is often unknown, requiring the estimation of the critical values through Monte Carlo (MC) simulations. The computation of the critical values at low α, especially when the distribution of the statistics changes with the series length (sample cardinality), requires a considerable number of simulations to achieve a reasonable precision of the estimates (i.e. 10⁶ simulations or more for each series length). If, in addition, the test requires a noteworthy computational effort, the estimation of the critical values may need unacceptably long runtimes.

To overcome the problem, the paper proposes a regression-based refinement of an initial MC estimate of the critical values, also allowing an approximation of the achieved improvement. Moreover, the paper presents an application of the method to two tests: SNHT (standard normal homogeneity test, widely used in climatology), and SNH2T (a version of SNHT showing a squared numerical complexity). For both, the paper reports the critical values for α ranging between 0.1 and 0.0001 (useful for the p-value estimation), and the series length ranging from 10 (widely adopted size in climatological change-point detection literature) to 70,000 elements (nearly the length of a daily data time series 200 years long), estimated with coefficients of variation within 0.22%. For SNHT, a comparison of our results with approximated, theoretically derived, critical values is also performed; we suggest adopting those values for the series exceeding 70,000 elements.     

Trends cycles and seasons: Econometric methods of signal extraction

Alternative methods of trend extraction and of seasonal adjustment are described that operate in the time domain and in the frequency domain.

The time-domain methods that are implemented in the TRAMO–SEATS and the STAMP programs are compared. An abbreviated time-domain method of seasonal adjustment that is implemented in the IDEOLOG program is also presented. Finite-sample versions of the Wiener–Kolmogorov filter are described that can be used to implement the methods in a common way.

The frequency-domain method, which is also implemented in the IDEOLOG program, employs an ideal frequency selective filter that depends on identifying the ordinates of the Fourier transform of a detrended data sequence that should lie in the pass band of the filter and those that should lie in its stop band. Filters of this nature can be used both for extracting a low-frequency cyclical component of the data and for extracting the seasonal component.