Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Earth and Planetary Sciences, 1980. / Supervised by Charles R. Denham and Richard P. Von Herzen. Vita. Includes bibliographical references.

Thesis (Ph. D.)--University of California, San Diego, 2009. / Title from first page of PDF file (viewed September 17, 2009). Available via ProQuest Digital Dissertations. Vita. Includes bibliographical references (p. 234-241).

Liu, Li. / Thesis (M.Phil.)--Chinese University of Hong Kong, 2009. / Includes bibliographical references (leaves 41-44). / Abstract also in Chinese. / Chapter 1 --- Introduction --- p.1 / Chapter 1.1 --- Empirical Likelihood --- p.1 / Chapter 1.2 --- Empirical Likelihood for Dependent Data --- p.4 / Chapter 1.2.1 --- Spectral Method --- p.5 / Chapter 1.2.2 --- Blockwise Method --- p.6 / Chapter 1.3 --- Edgeworth Expansions and Bartlett Correction --- p.9 / Chapter 1.3.1 --- Coverage Errors --- p.10 / Chapter 1.3.2 --- Edgeworth Expansions --- p.11 / Chapter 1.3.3 --- Bartlett Correction --- p.13 / Chapter 2 --- Bartlett Correction for EL --- p.16 / Chapter 2.1 --- Empirical Likelihood in Time Series --- p.16 / Chapter 2.2 --- Stochastic Expansions of EL in Time Series --- p.19 / Chapter 2.3 --- Edgeworth Expansions of EL in Time Series --- p.22 / Chapter 2.3.1 --- Validity of the Formal Edgeworth Expansions --- p.22 / Chapter 2.3.2 --- Cumulant Calculations --- p.24 / Chapter 2.4 --- Main Results --- p.30 / Chapter 3 --- Simulations --- p.32 / Chapter 3.1 --- Confidence Region --- p.33 / Chapter 3.2 --- Coverage Error of Confidence Regions --- p.35 / Chapter 4 --- Conclusion and Future Work --- p.38 / Bibliography --- p.41

Master of Science / Department of Computer Science / William H. Hsu / Cryptocurrencies are digital currencies that have garnered significant investor attention in the financial markets. The aim of this project is to predict the daily price, particularly the daily high and closing price, of the cryptocurrency Bitcoin. This plays a vital role in making trading decisions. There exist various factors which affect the price of Bitcoin, thereby making price prediction a complex and technically challenging task. To perform prediction, we trained temporal neural networks such as time-delay neural networks (TDNN) and recurrent neural networks (RNN) on historical time series – that is, past prices of Bitcoin over several years. Features such as the opening price, highest price, lowest price, closing price, and volume of a currency over several preceding quarters were taken into consideration so as to predict the highest and closing price of the next day. We designed and implemented TDNNs and RNNs using the NeuroSolutions artificial neural network (ANN) development environment to build predictive models and evaluated them by computing various measures such as the MSE (mean square error), NMSE (normalized mean square error), and r (Pearson’s correlation coefficient) on a continuation of the training data from each time series, held out for validation.

Orientador: Claudia Maria Bauzer Medeiros / Dissertação (mestrado) - Universidade Estadual de Campinas, Instituto de Computação / Made available in DSpace on 2018-08-23T21:42:10Z (GMT). No. of bitstreams: 1 Silva_FelipeHenriquesda_M.pdf: 3283921 bytes, checksum: 6875b168c728390c5cbeb2e32389cb99 (MD5) Previous issue date: 2013 / Resumo: Séries temporais são sequências de valores medidos em sucessivos instantes de tempo. Elas são usadas em diversos domínios, tais como agricultura, medicina e economia. A análise dessas séries é de extrema importância, fornecendo a especialistas a capacidade de identificar tendências e prever possíveis cenários. A fim de facilitar sua análise, especialistas frequentemente associam anotações com séries temporais. Tais anotações também podem ser usadas para correlacionar séries distintas, ou para procurar por séries específicas num banco de dados. Existem muitos desafios envolvidos no gerenciamento destas anotações - desde encontrar estruturas adequadas para associá-las com as séries, até organizar e recuperar séries através das anotações associadas a estas. Este trabalho contribui para o trabalho em gerenciamento de séries temporais. Suas principais contribuições são o projeto e desenvolvimento de um arcabouço para o gerenciamento de múltiplas anotações associadas com uma ou mais séries em um banco de dados. Este arcabouço também fornece meios para o controle de versão das anotações, de modo que os estados anteriores de uma anotação nunca sejam perdidos. Serial Annotator é uma aplicação desenvolvida para a plataforma Android. Ela foi usada para validar o arcabouço proposto e foi testada com dados reais envolvendo problemas do domínio agrícola / Abstract: Time series are sequences of values measured at successive time instants. They are used in several domains such as agriculture, medicine and economics. The analysis of these series is of utmost importance, providing experts the ability to identify trends and forecast possible scenarios. In order to facilitate their analyses, experts often associate annotations with time series. Such annotations can also be used to correlate distinct series, or look for specific series in a database. There are many challenges involved in managing annotations - from finding proper structures to associate them with series, to organizing and retrieving series based on annotations. This work contributes to the work in management of time series. Its main contributions are the design and development of a framework for the management of multiple annotations associated with one or multiple time series in a database. The framework also provides means for annotation versioning, so that previous states of an annotation are never lost. Serial Annotator is an application implemented for the Android smart phone platform. It has been used to validate the proposed framework and has been tested with real data involving agriculture problems / Mestrado / Ciência da Computação / Mestre em Ciência da Computação

D.Phil. / During the last few years the number of known variable stars which show periodic light level changes has grown by several tens of thousands. The aim of the research reported here was to extend the suite of statistical methods available for the analysis of periodic variable star time series. Solution techniques for five problems are discussed. The first is an automated method for detecting periodic variable stars from a database containing of the order of 100 000 time series of observations. Typically only 100-200 brightness measurements of each star were obtained, spread irregularly over an interval of about 3 years. The proposed method is based on a signal to noise ratio. Percentiles for the statistic are found by studying randomisations of a large number of the observed time series. It is shown that the percentiles depend strongly on the number of observations in a given dataset, and the dependence is calibrated empirically. The estimation of the frequency, amplitude and phase of a sinusoid from observations contaminated by correlated noise is the second problem considered. The study of the observational noise properties of nearly 200 real datasets of the relevant type is reported: noise can almost always be characterised as a random walk with superposed white noise. A scheme for obtaining weighted nonlinear least squares estimates of the parameters of interest, as well as standard errors of these estimates, is described. Simulation results are presented for both complete and incomplete data, and an application to real observations is also shown. In the third topic discussed it is assumed that contemporaneous measurements of the light in-tensity of a pulsating star is obtained in several colours. There is strong theoretical interest in a comparison of the amplitudes and phases of the variations in the different colours. A general scheme for calculating the covariance matrix of the estimated amplitude ratios and phase differences is described. The first step is to fit a time series model to the residuals after subtracting the best-fitting sinusoid from the observations. The residuals are then crosscorrelated to study the interdependence between the errors in the different colours. Once the multivariate time series structure can be modelled, the covariance matrix can be found by bootstrapping. An illustrative application is described in detail. The times between successive instances of maximum brightness, or the times between successive brightness minima, serve as estimates for the periods of the so-called "long period variables" (stars with pulsation periods of the order of months). The times between successive maxima (or minima) vary stochastically, and are also subject to measurement errors, which poses a problem for tests for systematic period changes — the topic of the fourth problem studied. A simple statistical model for the times between successive maxima, or minima, of such stars is used to calculate the auto-correlation properties of a new time series, which is non-stationary in its variance. The new series consists of an alternation of cycle lengths based on respectively the times between maxima, and those between minima of the light curve. Two different approaches to calculating the theoretical spectrum of the non-stationary time series, as required in the proposed statistical hypothesis test, are given. Illustrative applications complete the relevant chapter.

In this thesis, we investigate the estimation problem and inference problem for the complex models. Two major categories of complex models are emphasized by us, one is generalized linear models, the other is time series models. For the generalized linear models, we consider one fundamental problem about sure screening for interaction terms in ultra-high dimensional feature space; for time series models, an important model assumption about Markov property is considered by us. The first part of this thesis illustrates the significant interaction pursuit problem for ultra-high dimensional models with two-way interaction effects. We propose a simple sure screening procedure (SSI) to detect significant interactions between the explanatory variables and the response variable in the high or ultra-high dimensional generalized linear regression models. Sure screening method is a simple, but powerful tool for the first step of feature selection or variable selection for ultra-high dimensional data. We investigate the sure screening properties of the proposal method from theoretical insight. Furthermore, we indicate that our proposed method can control the false discovery rate at a reasonable size, so the regularized variable selection methods can be easily applied to get more accurate feature selection in the following model selection procedures. Moreover, from the viewpoint of computational efficiency, we suggest a much more efficient algorithm-discretized SSI (DSSI) to realize our proposed sure screening method in practice. And we also investigate the properties of these two algorithms SSI and DSSI in simulation studies and apply them to some real data analyses for illustration. For the second part, our concern is the testing of the Markov property in time series processes. Markovian assumption plays an extremely important role in time series analysis and is also a fundamental assumption in economic and financial models. However, few existing research mainly focused on how to test the Markov properties for the time series processes. Therefore, for the Markovian assumption, we propose a new test procedure to check if the time series with beta-mixing possesses the Markov property. Our test is based on the Conditional Distance Covariance (CDCov). We investigate the theoretical properties of the proposed method. The asymptotic distribution of the proposed test statistic under the null hypothesis is obtained, and the power of the test procedure under local alternative hypothesizes have been studied. Simulation studies are conducted to demonstrate the finite sample performance of our test.

The analysis of a time series is a problem well known to statisticians. Neural networks form the basis of an entirely non-linear approach to the analysis of time series. It has been widely used in pattern recognition, classification and prediction. Recently, reviews from a statistical perspective were done by Cheng and Titterington (1994) and Ripley (1993). One of the most important properties of a neural network is its ability to learn. In neural network methodology, the data set is divided in three different sets, namely a training set, a cross-validation set, and a test set. The training set is used for training the network with the various available learning (optimisation) algorithms. Different algorithms will perform best on different problems. The advantages and limitations of different algorithms in respect of all training problems are discussed. In this dissertation the method of neural networks and that of ARlMA. models are discussed. The procedures of identification, estimation and evaluation of both models are investigated. Many of the standard techniques in statistics can be compared with neural network methodology, especially in applications with large data sets. Additional information available on two discs stored at the Africana section, Merensky Library. / Dissertation (MSc (Mathematical Statistics))--University of Pretoria, 2007. / Statistics / unrestricted

Global forests are experiencing a variety of stresses in response to climate change and human activities. The broad objective of this dissertation is to improve understanding of how temperate and boreal forests are changing by using remote sensing to develop new techniques for detecting change in forest ecosystems and to use these techniques to investigate patterns of change in North American forests. First, I developed and applied a temporal segmentation algorithm to an 11-year time series of MODIS data for a region in the Pacific Northwest of the USA. Through comparison with an existing forest disturbance map, I characterized how the severity and spatial scale of disturbances affect the ability of MODIS to detect these events. Results from these analyses showed that most disturbances occupying more than one-third of a MODIS pixel can be detected but that prior disturbance history and gridding artifacts complicate the signature of forest disturbance events in MODIS data. Second, I focused on boreal forests of Canada, where recent studies have used remote sensing to infer decreases in forest productivity. To investigate these trends, I collected 28 years of Landsat TM and ETM+ data for 11 sites spanning Canada's boreal forests. Using these data, I analyzed how sensor geometry and intra- and inter-sensor calibration influence detection of trends from Landsat time series. Results showed systematic patterns in Landsat time series that reflect sensor geometry and subtle issues related to inter-sensor calibration, including consistently higher red band reflectance values from TM data relative to ETM+ data. In the final chapter, I extended the analyses from my second chapter to explore patterns of change in Landsat time series at an expanded set of 46 sites. Trends in peak-summer values of vegetation indices from Landsat were summarized at the scale of MODIS pixels. Results showed that the magnitude and slope of observed trends reflect patterns in disturbance and land cover and that undisturbed forests in eastern sites showed subtle, but detectable, differences from patterns observed in western sites. Drier forests in western Canada show declining trends, while mostly increasing trends are observed for wetter eastern forests.

Ever since Granger first proposed the idea of quantitatively testing the causal relationship between data streams, the endeavor of accurately inferring the causality in data and using that information to predict the future has not stopped. Artificial Intelligence (AI), by utilizing the massive amounts of data, helps to solve complex problems, whether they include the diagnosis and detection of disease through medical imaging, email spam detection, or self-driving vehicles. Perhaps, this thesis will be trivial in ten years from now. AI has pushed humankind to reach the next technological level in technology. Nowadays, among most machine leaning inquiries, statistical relationships are determined using correlation measures. By feeding data into machine learning algorithms, computers update the algorithm’s parameters iteratively by extracting and mapping features to learning targets until the correlation increases to a significant level to cease the training process. However, with the increasing developments of powerful AI, there is really a shortage of exploring causality in data. It is almost self-evident that ”correlation is not causality." Sometimes, the strong correlation established between variables through machine learning can be absurd and meaningless. Providing insight into causality information through data, which most of the machine learning methods fall short to do, is of paramount importance. The subsequent chapters detail the four endeavors of studying causality in financial markets, earthquakes, animal/human brain signals, the predictivity of data sets. In Chapter 2, we further developed the concept of causality networks into a higher-order causality network. We applied these to financial data and tested their validity and ability to capture the system’s causal relationship. In next Chapter 3, We examined another type of time series-earthquakes. Violent seismic activities decimate people's lives and destroy entire cities and areas. This begs us to understand how earthquakes work and help us make reliably and evacuation-actionable predictions. The causal relationships of seismic activities in different areas are studied and established. Biological data, specifically brain signals, are time-series data and their causal pattern are explored and studied. Different human and mice brain signals are analyzed and clustered in Chapter 4 using their unique causal pattern to understand different brain cell activity. Finally, we realized that the causal pattern in the time series can be used to compress data. A causal compression ratio is invented and used as the data stream’s predictivity index. We describe this in Chapter 5.