Wasserstein Auto‐Regressive Models for Modeling Multivariate Distributional Time Series

Statistical parametric methodologies are widely employed in the analysis of time series of gas turbine sensor readings. These methodologies identify outliers as a consequence of excessive deviation from a statistically-based model, derived from available observations. Among parametric techniques, the k-σ methodology demonstrates its effectiveness in the analysis of stationary time series. Furthermore, the simplicity and the clarity of this approach justify its direct application to industry. On the other hand, the k-σ methodology usually proves to be unable to adapt to dynamic time series, since it identifies observations in a transient as outliers. As this limitation is caused by the nature of the methodology itself, two improved approaches are considered in this paper in addition to the standard k-σ methodology. The two proposed methodologies maintain the same rejection rule of the standard k-σ methodology, but differ in the portions of the time series from which statistical parameters (mean and standard deviation) are inferred. The first approach performs statistical inference by considering all observations prior to the current one, which are assumed reliable, plus a forward window containing a specified number of future observations. The second approach proposed in this paper is based on a moving window scheme. Simulated data are used to tune the parameters of the proposed improved methodologies and to prove their effectiveness in adapting to dynamic time series. The moving window approach is found to be the best on simulated data in terms of True Positive Rate (TPR), False Negative Rate (FNR) and False Positive Rate (FPR). Therefore, the performance of the moving window approach is further assessed towards both different simulated scenarios and field data taken on a gas turbine.

The autoregressive analysis of stationary multivariate time series includes construction of a time domain model in the form of a multivariate stochastic difference equation and its transformation into a spectral matrix to study frequency domain properties. The time series is treated as a linear system with one output process and one or several inputs. This approach provides an explicit time domain description of the time series as a function of the past of output and input components. In the bivariate case, the frequency domain results include spectral densities, coherent spectrum, coherence function, and gain and phase factors. The coherence function shows the degree of linear interdependence, the coherent spectrum presents the contribution of the input component to the output, the gain factor consists of amplification coefficients, and the phase factor describes lags between the time series. All these quantities are frequency dependent. The Granger causality concept and feedbacks within the system are discussed as a part of bivariate time and frequency domain analysis. Some comments are given regarding the software for parametric analysis of multivariate time series. The use of cross-correlation coefficient and regression equation for describing dependence between time series or for reconstructing them is shown to be incorrect.

Spectral analysis of multivariate geochemical time series

Results obtained with interindividual techniques in a representative sample of a population are not necessarily generalizable to the individual members of this population. In this article the specific condition is presented that must be satisfied to generalize from the interindividual level to the intraindividual level. A way to investigate whether this condition is satisfied is by means of multivariate time series analysis. More generally, time series analysis can be used to investigate psychological processes situated within individuals. In this article we consider a well established class of multivariate stationary time series models that may be used to study the intraindividual covariance structure. We demonstrate the application of some of these models with an empirical example consisting of state measurements of behavior associated with the Five Factor Model of Personality. We illustrate how one can investigate whether individuals are similar with respect to their intraindividual structure of variation, and whether this structure is similar to the structure of interindividual variation.

This work gives an overview of statistical analysis for some models for multivariate discrete-valued (MDV) time series. We present observation-driven models and models based on higher-order Markov chains. Several extensions are highlighted including non-stationarity, network autoregressions, conditional non-linear autoregressive models, robust estimation, random fields and spatio-temporal models.

Statistical analysis of time series. With compelling arguments we show that the Diffusion Entropy Analysis (DEA) is the only method of the literature of the Science of Complexity that correctly determines the scaling hidden within a time series reflecting a Complex Process. The time series is thought of as a source of fluctuations, and the DEA is based on the Shannon entropy of the diffusion process generated by these fluctuations. All traditional methods of scaling analysis, instead, are based on the variance of this diffusion process. The variance methods detect the real scaling only if the Gaussian assumption holds true. We call H the scaling exponent detected by the variance methods and d the real scaling exponent. If the time series is characterized by Fractional Brownian Motion, we have H¹d and the scaling can be safely determined, in this case, by using the variance methods. If, on the contrary, the time series is characterized, for example, by Lévy statistics, H ¹ d and the variance methods cannot be used to detect the true scaling. Lévy walk yields the relation d=1/(3-2H). In the case of Lévy flights, the variance diverges and the exponent H cannot be determined, whereas the scaling d exists and can be established by using the DEA. Therefore, only the joint use of two different scaling analysis methods, the variance scaling analysis and the DEA, can assess the real nature, Gauss or Lévy or something else, of a time series. Moreover, the DEA determines the information content, under the form of Shannon entropy, or of any other convenient entopic indicator, at each time step of the process that, given a sufficiently large number of data, is expected to become diffusion with scaling. This makes it possible to study the regime of transition from dynamics to thermodynamics, non-stationary regimes, and the saturation regime as well. First of all, the efficiency of the DEA is proved with theoretical arguments and with numerical work on artificial sequences. Then we apply the DEA to three different sets of real data, Genome sequences, hard x-ray solar flare waiting times and sequences of sociological interest. In all these cases the DEA makes new properties, overlooked by the standard method of analysis, emerge.

5.1 By the mid-1920s the methodological advances discussed in the previous two chapters, namely periodogram analysis and the variate differencing method, appeared to be running out of steam, with few new applications appearing and increasing concern about the underlying assumptions of the techniques. At this point, three papers appeared in quick succession which transformed the subject and laid the foundations for modern approaches to the analysis of time series. Two of the papers, by Yule (1926) and Slutzky (1927), went a long way to establishing the basis for the theoretical analysis of stationary time series and, because of the enduring importance of their contributions, are consequently subjected to detailed scrutiny in this chapter, along with a subsequent and closely related paper by Working (1934). The third paper, also by Yule (Yule, 1927), attacked periodic time series in a new way and, in turn, provided the foundations for analysing oscillatory time series: this is our focus in Chapter 6.

Remote sensing plays a vital role in climate change studies by providing us with a record of 'snapshots' of the Earth's environment through time. Almost 30 years have past since the launch of the first routine and repetitive remote sensing systems. We are on the brink of being able to use remotely sensed imagery to define climatic normals using features of the Earth's surface environment as proxy indicators. Before we can do this, however, we need to develop new image analysis strategies--hypertemporal image analysis techniques--which we can use to measure variability and identify changes in long sequences of remotely sensed imagery. In this paper we explore one such technique: statistical time series analysis (TSA). We apply AutoRegressive-Moving Average (ARMA) time series models to a 108-month sequence of passive microwave images of Arctic sea ice concentrations. TSA is traditionally a manual technique applied to one or two temporal variables. Through this research we demonstrate how, by making two simple assumptions about the data, the approach can be automated to process the thousands of time series in a temporal image sequence. ARMA models are shown to be useful in deriving a spatial summary of the temporal characteristics of the modelled phenomenon, for identifying changes within that system, and for forecasting into the future with statistical confidence.

Time Series: State Space Methods

In this work, we are in the position to view a measurement of a physical observable as an experiment in the sense of probability theory. To every physical observable, a sample space called the spectrum of the observable is therefore available. We have investigated the Wasserstein spaces over the spectrums of a quantum observables, i.e. the space of all probability measures defined on the suitable σ-algebra of subsets of the spectrum equipped with certain metric. It has been shown here the existence of one-to-one correspondence between the set of all quantum states and the set of all distribution functions defined on the spectrum of an observable. Thanks to the correspondence, every quantum state can be represented by a probability measure associated with the corresponding distribution function. It has been shown that the Wasserstein space over the spectrum of an observable is homeomorphic to the Wasserstein space over the spectrum of another observable. Such homeomorphism is referred to as generalized Fourier transform. By using the work of von Renesse, it has been shown that the Newtonian dynamics of the system can be formulated in the Wasserstein space over the spectrum of each observable which is equivalent to the Schroedinger dynamics, starting from the solution of Newton equation of motion in the Wasserstein space over the spectrum of a certain observable.

Time Series Analysis Methods in Chemometrics

A fast LSH-based similarity search method for multivariate time series

SUMMARY We propose a method to specify an appropriate yet parsimonious vector autoregressive moving average (ARMA) model for a given multivariate time series. By considering contemporaneous linear transformations of the vector process, we introduce the concept of scalar component models within the vector ARMA framework (a) to reveal possibly hidden simplifying structures of the process, (b) to achieve parsimony in parameterization and (c) to identify the exchangeable models. The simplifying structures are of particular importance in the analysis of multivariate time series because they are often not obvious from the observed data but can be used to gain insights to the process under study. The analytical tool used to search for scalar component models is a canonical correlation analysis of vector processes and the proposed procedures are illustrated via two real examples.

Nowadays, great quantities of data are produced by a large and diverse family of sensors (e.g., remote sensors, biochemical sensors, wearable devices), which typically measure multiple variables over time, resulting in data streams that can be profitably organized as multivariate time-series. In practical scenarios, the speed at which such information is collected often makes the data labeling task uneasy and too expensive, so that limit the use of supervised approaches. For this reason, unsupervised and exploratory methods represent a fundamental tool to deal with the analysis of multivariate time series. In this paper we propose a deep-learning based framework for clustering multivariate time series data with varying lengths. Our framework, namely DeTSEC (Deep Time Series Embedding Clustering), includes two stages: firstly a recurrent autoencoder exploits attention and gating mechanisms to produce a preliminary embedding representation; then, a clustering refinement stage is introduced to stretch the embedding manifold towards the corresponding clusters. Experimental assessment on six real-world benchmarks coming from different domains has highlighted the effectiveness of our proposal.

Coupling of multivariate methods and time series analysis can be ueful for studying dynamics of aquatic communities. This is demonstratred with a data set from the pelagic area of an oligo-mesotrophic lake in Central Spain during 61 consecutive days of Autumn overturn. Abiotic variables, phytoplankton species and their total biomass were traced. Species abundance and specific biomass were considered as indices of community structure and resource partitioning, respectively. Abiotic and algal data sets were subjected to factor analyses of cases separately. Atmospheric forcing and nitrogen could be considered as the main (2) driving variables of the abiotic matrix. The coupling of motile abilities and cell size was associated to the main factors of the community structure matrix whereas phosphorus limitation and species responses to buoyancy represented the main factors of the biomass matrix. Coordinates of the two first factors could be used to mimic the trajectories in the data space. Significant short term lags (1–4 days) were found in most time series. Lagged responses of atmospheric forcing and nitrogen on phytoplankton community structure and resource partitioning at scales of 1–7 days were also shown. Overall phytoplankton biomass did not show significant delayed responses, thereby suggesting that it might be resulting from the interplay of other non-studied factors.