Downloads & Free Reading Options - Results

We study online prediction of bounded stationary ergodic processes. To do so, we consider the setting of prediction of individual sequences and build a deterministic regression tree that performs asymptotically as well as the best L-Lipschitz constant predictors. Then, we show why the obtained regret bound entails the asymptotical optimality with respect to the class of bounded stationary ergodic processes.

In this paper we apply a new approach of the string theory to the real financial market. It is direct extension and application of the work [1] into prediction of prices. The models are constructed with an idea of prediction models based on the string invariants (PMBSI). The performance of PMBSI is compared to support vector machines (SVM) and artificial neural networks (ANN) on an artificial and a financial time series. Brief overview of the results and analysis is given. The first model is based on the correlation function as invariant and the second one is an application based on the deviations from the closed string/pattern form (PMBCS). We found the difference between these two approaches. The first model cannot predict the behavior of the forex market with good efficiency in comparison with the second one which is, in addition, able to make relevant profit per year.

In this paper we apply a new approach of the string theory to the real financial market. It is direct extension and application of the work [1] into prediction of prices. The models are constructed with an idea of prediction models based on the string invariants (PMBSI). The performance of PMBSI is compared to support vector machines (SVM) and artificial neural networks (ANN) on an artificial and a financial time series. Brief overview of the results and analysis is given. The first model is based on the correlation function as invariant and the second one is an application based on the deviations from the closed string/pattern form (PMBCS). We found the difference between these two approaches. The first model cannot predict the behavior of the forex market with good efficiency in comparison with the second one which is, in addition, able to make relevant profit per year.

We present a simple randomized procedure for the prediction of a binary sequence. The algorithm uses ideas from recent developments of the theory of the prediction of individual sequences. We show that if the sequence is a realization of a stationary and ergodic random process then the average number of mistakes converges, almost surely, to that of the optimum, given by the Bayes predictor. The desirable finite-sample properties of the predictor are illustrated by its performance for Markov processes. In such cases the predictor exhibits near optimal behavior even without knowing the order of the Markov process. Prediction with side information is also considered.

Forecasting a time series from multivariate predictors constitutes a challenging problem, especially using model-free approaches. Most techniques, such as nearest-neighbor prediction, quickly suffer from the curse of dimensionality and overfitting for more than a few predictors which has limited their application mostly to the univariate case. Therefore, selection strategies are needed that harness the available information as efficiently as possible. Since often the right combination of predictors matters, ideally all subsets of possible predictors should be tested for their predictive power, but the exponentially growing number of combinations makes such an approach computationally prohibitive. Here a prediction scheme that overcomes this strong limitation is introduced utilizing a causal pre-selection step which drastically reduces the number of possible predictors to the most predictive set of causal drivers making a globally optimal search scheme tractable. The information-theoretic optimality is derived and practical selection criteria are discussed. As demonstrated for multivariate nonlinear stochastic delay processes, the optimal scheme can even be less computationally expensive than commonly used sub-optimal schemes like forward selection. The method suggests a general framework to apply the optimal model-free approach to select variables and subsequently fit a model to further improve a prediction or learn statistical dependencies. The performance of this framework is illustrated on a climatological index of El Ni\~no Southern Oscillation.

In this thesis, a class of time series models for forecasting a hurricane's future position based on its previous positions and a generalized model of hurricane motion are examined and extended. Results of a literature review suggest that meteorological models continue to increase in complexity while few statistical approaches, such as linear regression, have been successfully applied. An exception is provided by a certain class of time series models that appear to forecast storms almost as well as current meteorological models without their tremendous complexity. A suggestion for enhancing the performance of these time series models is pursued through an examination of the forecast errors produced when these models are applied to historical storm tracks. The results uncover no patterns that can be exploited in developing an improved model and suggest that there are meteorological processes or factors at work beyond those that can be modeled with the available historical data base. The statistical structure of the time series approach is exploited to develop a practical method for determining prediction regions which probabilistically describe a hurricane's likely future position. The Monte Carlo approach used to develop these prediction ellipses is seen to be applicable for predicting areas subject to risk from hurricane landfall.

A computer program is described and presented for calculating finite memory predictors and prediction variances for autoregressive moving average time series models. The Cholesky decomposition algorithm is used, and a number of simplifying results are described and implemented in the program. (Author)

A numerical and experimental study of time domain methods for modeling and parameter identification of structural systems is presented. Models are developed which can be used to predict the transient response of multiple- degree-of-freedom systems subjected to arbitrary input. The linear, discrete time transfer function is expressed in a form called the Autoregressive Moving Average (ARMA) model. The ARMA model is a minimum parameter model that may be parameterized with a minimum number of measured quantities. The ARMA model is contrasted to traditional models such as differential equation models and modal methods. The ARMA model indentification algorithms are also compared to time domain modal methods. Though it has proven useful for a number of low-order systems, the identification of the ARMA model is often hampered by the sensitivity of parameter estimates to measurement noise bias.

For oscillating time series, the prediction is often focused on the turning points. In order to predict the turning point magnitudes and times it is proposed to form the state space reconstruction only from the turning points and modify the local (nearest neighbor) model accordingly. The model on turning points gives optimal prediction at a lower dimensional state space than the optimal local model applied directly on the oscillating time series and is thus computationally more efficient. Monte Carlo simulations on different oscillating nonlinear systems showed that it gives better predictions of turning points and this is confirmed also for the time series of annual sunspots and total stress in a plastic deformation experiment.

Prediction of events is fundamental to both human and artificial agents. The main problem with previous prediction techniques is that they cannot predict events that have never been experienced before. This dissertation addresses the problem of predicting such novelty by developing algorithms and computational models inspired from recent cognitive science theories conceptual blending theory and event segmentation theory. We were able to show that prediction accuracy for event or state prediction can be significantly improved using these methods. The main contribution of this dissertation is a new class of prediction techniques inspired by conceptual blending that improves prediction accuracy overall and has the ability to predict even events that have never been experienced before. We also show that event segmentation theory, when integrated with these techniques, results in greater computational efficiency. We implemented the new prediction techniques, and more traditional alternatives such as Markov and Bayesian techniques, and compared their prediction accuracy quantitatively for three domains a role-playing game, intrusion-system alerts, and event prediction of maritime paths in a discrete-event simulator. Other contributions include two new unification algorithms that improve over a nave one, and an exploration of ways to maintain a minimum-size knowledge base without affecting prediction accuracy.

We establish rates of convergences in time series forecasting using the statistical learning approach based on oracle inequalities. A series of papers extends the oracle inequalities obtained for iid observations to time series under weak dependence conditions. Given a family of predictors and $n$ observations, oracle inequalities state that a predictor forecasts the series as well as the best predictor in the family up to a remainder term $\Delta_n$. Using the PAC-Bayesian approach, we establish under weak dependence conditions oracle inequalities with optimal rates of convergence. We extend previous results for the absolute loss function to any Lipschitz loss function with rates $\Delta_n\sim\sqrt{c(\Theta)/ n}$ where $c(\Theta)$ measures the complexity of the model. We apply the method for quantile loss functions to forecast the french GDP. Under additional conditions on the loss functions (satisfied by the quadratic loss function) and on the time series, we refine the rates of convergence to $\Delta_n \sim c(\Theta)/n$. We achieve for the first time these fast rates for uniformly mixing processes. These rates are known to be optimal in the iid case and for individual sequences. In particular, we generalize the results of Dalalyan and Tsybakov on sparse regression estimation to the case of autoregression.

In real-time monitoring of hospital patients, high-quality inference of patients' health status using all information available from clinical covariates and lab tests are essential to enable successful medical interventions and improve patient outcomes. In this work, we develop and explore a Bayesian nonparametric model based on Gaussian process (GP) regression for hospital patient monitoring. Our method, MedGP, incorporates 24 clinical and lab covariates and supports a rich reference data set from which the relationships between these observed covariates may be inferred and exploited for high-quality inference of patient state over time. To do this, we develop a highly structured sparse GP kernel to enable tractable computation over tens of thousands of time points while estimating correlations among clinical covariates, patients, and periodicity in high-dimensional time series measurements of physiological signals. We apply MedGP to data from hundreds of thousands of patients treated at the Hospital of the University of Pennsylvania. MedGP has a number of benefits over current methods, including (i) not requiring an alignment of the time series data, (ii) quantifying confidence intervals in the predictions, (iii) exploiting a vast and rich database of patients, and (iv) providing interpretable relationships among clinical covariates. We evaluate and compare results from MedGP on the task of online state prediction for three different patient subgroups.

In this research the technology of complex Markov chains is applied to predict financial time series. The main distinction of complex or high-order Markov Chains and simple first-order ones is the existing of aftereffect or memory. The technology proposes prediction with the hierarchy of time discretization intervals and splicing procedure for the prediction results at the different frequency levels to the single prediction output time series. The hierarchy of time discretizations gives a possibility to use fractal properties of the given time series to make prediction on the different frequencies of the series. The prediction results for world's stock market indices is presented.

Time series prediction covers a vast field of every-day statistical applications in medical, environmental and economic domains. In this paper we develop nonparametric prediction strategies based on the combination of a set of 'experts' and show the universal consistency of these strategies under a minimum of conditions. We perform an in-depth analysis of real-world data sets and show that these nonparametric strategies are more flexible, faster and generally outperform ARMA methods in terms of normalized cumulative prediction error.

This thesis examines the topic of chaotic time series. An overview of chaos, dynamical systems, and traditional approaches to time series analysis is provided, followed by an examination of state space reconstruction. State space reconstruction is a nonlinear, deterministic approach whose goal is to use the immediate past behavior of the time series to reconstruct the current state of the system. The choice of delay parameter and embedding dimension are crucial to this reconstruction. Once the state space has been properly reconstructed, one can address the issue of whether apparently random data has come from a low- dimensional, chaotic (deterministic) source or from a random process. Specific techniques for making this determination include attractor reconstruction, estimation of fractal dimension and Lyapunov exponents, and short-term prediction. If the time series data appears to be from a low-dimensional chaotic source, then one can predict the continuation of the data in the short term. This is the inverse problem of dynamical systems. In this thesis, the technique of local fitting is used to accomplish the prediction. Finally, the issue of noisy data is treated, with the purpose of highlighting where further research may be beneficial

In this paper we show that different physiological states and pathological conditions may be characterized in terms of predictability of time series signals from the underlying biological system. In particular we consider systolic arterial pressure time series from healthy subjects and Chronic Heart Failure patients, undergoing paced respiration. We model time series by the regularized least squares approach and quantify predictability by the leave-one-out error. We find that the entrainment mechanism connected to paced breath, that renders the arterial blood pressure signal more regular, thus more predictable, is less effective in patients, and this effect correlates with the seriousness of the heart failure. The leave-one-out error separates controls from patients and, when all orders of nonlinearity are taken into account, alive patients from patients for which cardiac death occurred.

Multi-task learning employs shared representation of knowledge for learning multiple instances from the same or related problems. Time series prediction consists of several instances that are defined by the way they are broken down into fixed windows known as embedding dimension. Finding the optimal values for embedding dimension is a computationally intensive task. Therefore, we introduce a new category of problem called dynamic time series prediction that requires a trained model to give prediction when presented with different values of the embedding dimension. This can be seen a new class of time series prediction where dynamic prediction is needed. In this paper, we propose a co-evolutionary multi-task learning method that provides a synergy between multi-task learning and coevolution. This enables neural networks to retain modularity during training for building blocks of knowledge for different instances of the problem. The effectiveness of the proposed method is demonstrated using one-step-ahead chaotic time series problems. The results show that the proposed method can effectively be used for different instances of the related time series problems while providing improved generalisation performance.

Time series prediction has widespread application, ranging from predicting the stock market to trying to predict future locations of scud missiles. Recent work by Sauer and Casdagli has developed into the embedology theorem, which sets forth the procedures for state space manipulation and reconstruction for time series prediction. This includes embedding the time series into a higher dimensional space in order to form an attractor, a structure defined by the embedded vectors. Embedology is combined with neural technologies in an effort to create a more accurate prediction algorithm. These algorithms consist of embedology, neural networks, Euclidean space nearest neighbors, and spectral estimation techniques in an effort to surpass the prediction accuracy of conventional methods. Local linear training methods are also examined through the use of the nearest neighbors as the training set for a neural network. Fusion methodologies are also examined in an attempt to combine several algorithms in order to increase prediction accuracy. The results of these experiments determine that the neural network algorithms have the best individual prediction accuracies, and both fusion methodologies can determine the best performance. The performance of the nearest neighbor trained neural network validates the applicability of the local linear training set.

The purpose of this study was to combine knowledge of teacher demographic data with time-series forecasting methods to predict teacher turnover. Moving averages and exponential smoothing were used to forecast discrete time series. The study used data collected from the 22 largest school districts in Iowa, designated as FACT schools. Predictions were made for the 1968-69 school year and then validated with the actual data for that school year. Data were organized into 27 types, which were suggested by a literature review. Results of the study show that teachers with less than four years experience showed the highest amount of turnover, with married women under 30 having the highest rate, unmarried females under 30 being second, married males and unmarried males being third and fourth, respectively. (DB)

A Human-In-The-Loop simulation was conducted in January of 2013 in the Airspace Operations Laboratory at NASA's Ames Research Center. The simulation airspace included two en route sectors feeding the northwest corner of Atlanta's Terminal Radar Approach Control. The focus of this paper is on how uncertainties in the study's trajectory predictions impacted the controllers ability to perform their duties. Of particular interest is how the controllers interacted with the delay information displayed in the meter list and data block while managing the arrival flows. Due to wind forecasts with 30-knot over-predictions and 30-knot under-predictions, delay value computations included errors of similar magnitude, albeit in opposite directions. However, when performing their duties in the presence of these errors, did the controllers issue clearances of similar magnitude, albeit in opposite directions?

A computer program is described and presented for calculating finite memory predictors and prediction variances for autoregressive moving average time series models. The Cholesky decomposition algorithm is used, and a number of simplifying results are described and implemented in the program. (Author)

A computer program is described and presented for calculating finite memory predictors and prediction variances for autoregressive moving average time series models. The Cholesky decomposition algorithm is used, and a number of simplifying results are described and implemented in the program. (Author)

This thesis uses multiple layer perceptrons (MLP) neural networks and Kohonen clustering networks to predict and assign confidence to nonlinear time series classifications. The nonlinear time series used for analysis is the Standard and Poor's 100 (S&P 100) index. The target prediction is classification of the daily index change. Financial indicators were evaluated to determine the most useful combination of features for input into the networks. After evaluation it was determined that net changes in the index over time and three short-term idicators result in better accuracy. A back-propagation trained MLP neural network was then trained with these features to get a daily classification prediction of up or down. Next, a Kohonen clustering network was trained to develop 30 different clusters.The predictions from the MLP network were labeled as correct or incorrect within each classification and counted in each category to determine a confidence for a given cluster. Test data was then run through both networks and predictions were assigned a confidence based on which cluster they belonged to. The results of these tests show that this method can improve the accuracy of predictions from 51% to 73%. Within a cluster accuracy is near 100% for some classifications. (KAR) p. 9-10

Thesis advisors, Christopher Frenzen, Philip Beaver

Thesis advisors, Christopher Frenzen, Philip Beaver

The variability of AVHRR calibration coefficient in time was analyzed using algorithms of linear and non-linear time series analysis. Specifically we have used the spline trend modeling, autoregressive process analysis, incremental neural network learning algorithm and redundancy functional testing. The analysis performed on available AVHRR data sets revealed that (1) the calibration data have nonlinear dependencies, (2) the calibration data depend strongly on the target temperature, (3) both calibration coefficients and the temperature time series can be modeled, in the first approximation, as autonomous dynamical systems, (4) the high frequency residuals of the analyzed data sets can be best modeled as an autoregressive process of the 10th degree. We have dealt with a nonlinear identification problem and the problem of noise filtering (data smoothing). The system identification and filtering are significant problems for AVHRR data sets. The algorithms outlined in this study can be used for the future EOS missions. Prediction and smoothing algorithms for time series of calibration data provide a functional characterization of the data. Those algorithms can be particularly useful when calibration data are incomplete or sparse.

We report on a technique based on multi-agent games which has potential use in the prediction of future movements of financial time-series. A third-party game is trained on a black-box time-series, and is then run into the future to extract next-step and multi-step predictions. In addition to the possibility of identifying profit opportunities, the technique may prove useful in the development of improved risk management strategies.

A computer program is described and presented for calculating finite memory predictors and prediction variances for autoregressive moving average time series models. The Cholesky decomposition algorithm is used, and a number of simplifying results are described and implemented in the program. (Author)

A computer program is described and presented for calculating finite memory predictors and prediction variances for autoregressive moving average time series models. The Cholesky decomposition algorithm is used, and a number of simplifying results are described and implemented in the program. (Author)

We address the problem of nonparametric estimation of characteristics for stationary and ergodic time series. We consider finite-alphabet time series and real-valued ones and the following four problems: i) estimation of the (limiting) probability (or estimation of the density for real-valued time series), ii) on-line prediction, iii) regression and iv) classification (or so-called problems with side information). We show that so-called archivers (or data compressors) can be used as a tool for solving these problems. In particular, firstly, it is proven that any so-called universal code (or universal data compressor) can be used as a basis for constructing asymptotically optimal methods for the above problems. (By definition, a universal code can "compress" any sequence generated by a stationary and ergodic source asymptotically till the Shannon entropy of the source.) And, secondly, we show experimentally that estimates, which are based on practically used methods of data compression, have a reasonable precision.

This thesis uses multiple layer perceptrons (MLP) neural networks and Kohonen clustering networks to predict and assign confidence to nonlinear time series classifications. The nonlinear time series used for analysis is the Standard and Poor's 100 (S&P 100) index. The target prediction is classification of the daily index change. Financial indicators were evaluated to determine the most useful combination of features for input into the networks. After evaluation it was determined that net changes in the index over time and three short-term indicators result in better accuracy. A back-propagation trained MLP neural network was then trained with these features to get a daily classification prediction of up or down. Next, a Kohonen clustering network was trained to develop 30 different clusters. The predictions from the MLP network were labeled as correct or incorrect within each classification and counted in each category to determine a confidence for a given cluster. Test data was then run through both networks and predictions were assigned a confidence based on which cluster they belonged to. The results of these tests show that this method can improve the accuracy of predictions from 51% to 73%. Within a cluster accuracy is near 100% for some classifications.

This article is from BMC Bioinformatics , volume 15 . Abstract Background: Time series models can play an important role in disease prediction. Incidence data can be used to predict the future occurrence of disease events. Developments in modeling approaches provide an opportunity to compare different time series models for predictive power. Results: We applied ARIMA and Random Forest time series models to incidence data of outbreaks of highly pathogenic avian influenza (H5N1) in Egypt, available through the online EMPRES-I system. We found that the Random Forest model outperformed the ARIMA model in predictive ability. Furthermore, we found that the Random Forest model is effective for predicting outbreaks of H5N1 in Egypt. Conclusions: Random Forest time series modeling provides enhanced predictive ability over existing time series models for the prediction of infectious disease outbreaks. This result, along with those showing the concordance between bird and human outbreaks (Rabinowitz et al. 2012), provides a new approach to predicting these dangerous outbreaks in bird populations based on existing, freely available data. Our analysis uncovers the time-series structure of outbreak severity for highly pathogenic avain influenza (H5N1) in Egypt.

In this paper we address the problem of predicting a time series using the ARMA (autoregressive moving average) model, under minimal assumptions on the noise terms. Using regret minimization techniques, we develop effective online learning algorithms for the prediction problem, without assuming that the noise terms are Gaussian, identically distributed or even independent. Furthermore, we show that our algorithm's performances asymptotically approaches the performance of the best ARMA model in hindsight.

This thesis provides a description of how a neural network can be trained to learn the order inherent in chaotic time series data and then use that knowledge to predict future time series values. It examines the meaning of chaotic time series data, and explores in detail the Glass-Mackey nonlinear differential delay equation as a typical source of such data. An efficient weight update algorithm is derived, and its two-dimensional performance is examined graphically. A predictor network which incorporates this algorithm is constructed and used to predict chaotic data. The network was able to predict chaotic data. Prediction was more accurate for data having a low fractal dimension than for high-dimensional data. Lengthy computer run times than for high-dimensional data. Lengthy computer run times were found essential for adequate network training. Keywords: Sine waves, Ada programming language.

The problem of extracting as much information as possible from a sequence of observations of a stationary stochastic process $X_0,X_1,...X_n$ has been considered by many authors from different points of view. It has long been known through the work of D. Bailey that no universal estimator for $\textbf{P}(X_{n+1}|X_0,X_1,...X_n)$ can be found which converges to the true estimator almost surely. Despite this result, for restricted classes of processes, or for sequences of estimators along stopping times, universal estimators can be found. We present here a survey of some of the recent work that has been done along these lines.

Linear autoregressive models serve as basic representations of discrete time stochastic processes. Different attempts have been made to provide non-linear versions of the basic autoregressive process, including different versions based on kernel methods. Motivated by the powerful framework of Hilbert space embeddings of distributions, in this paper we apply this methodology for the kernel embedding of an autoregressive process of order $p$. By doing so, we provide a non-linear version of an autoregressive process, that shows increased performance over the linear model in highly complex time series. We use the method proposed for one-step ahead forecasting of different time-series, and compare its performance against other non-linear methods.

When estimating models with pooled cross-section and time-series data (e.g. estimating demand equations for all 50 states) one has to decide whether or not to pool the data. The usual procedure is to first test for the overall homogeneity (equality) of the coefficients. If this hypothesis is not rejected, then a single equation is estimated with pooled data. If the hypothesis is rejected, further hypothesis testing may be necessary. For example, if the model contains more than one coefficient the equality constraint may be rejected for only a subset of the coefficients. In this case the data is pooled and dummy variables are used with the subset of coefficients for which the equality constraint does not hold. There are at least three problems with this procedure of pooling (or not pooling) after some preliminary tests of significance. First, as noted in Maddala, it raises problems about the inference from the pooled model. Second, there is the related question of what significance level to use when deciding whether or not to pool. Third, the choice of estimates to select from is quite limited. That is, one must pick either the pooled or the non-pooled estimate, even if these two estimates are very different. The problems suggest that an alternative (or hybrid) method of handling pooled cross-section and time-series data is needed. The purpose of this paper is to propose such a method.

This thesis examines the topic of chaotic time series. An overview of chaos, dynamical systems, and traditional approaches to time series analysis is provided, followed by an examination of state space reconstruction. State space reconstruction is a nonlinear, deterministic approach whose goal is to use the immediate past behavior of the time series to reconstruct the current state of the system. The choice of delay parameter and embedding dimension are crucial to this reconstruction. Once the state space has been properly reconstructed, one can address the issue of whether apparently random data has come from a low-dimensional, chaotic (deterministic) source or from a random process. Specific techniques for making this determination include attractor reconstruction, estimation of fractal dimension and Lyapunov exponents, and short-term prediction. If the time series data appears to be from a low-dimensional chaotic source, then one can predict the continuation of the data in the short term. This is the inverse problem of dynamical systems. In this thesis, the technique of local fitting is used to accomplish the prediction. Finally, the issue of noisy data is treated, with the purpose of highlighting where further research may be beneficial.

A computer program is described and presented for calculating finite memory predictors and prediction variances for autoregressive moving average time series models. The Cholesky decomposition algorithm is used, and a number of simplifying results are described and implemented in the program. (Author)

A computer subroutine is given for computing predictors of any specified horizon's ahead. (Author)

Prediction of events is fundamental to both human and artificial agents. The main problem with previous prediction techniques is that they cannot predict events that have never been experienced before. This dissertation addresses the problem of predicting such novelty by developing algorithms and computational models inspired from recent cognitive science theories: conceptual blending theory and event segmentation theory. We were able to show that prediction accuracy for event or state prediction can be significantly improved using these methods. The main contribution of this dissertation is a new class of prediction techniques inspired by conceptual blending that improves prediction accuracy overall and has the ability to predict even events that have never been experienced before. We also show that event segmentation theory, when integrated with these techniques, results in greater computational efficiency. We implemented the new prediction techniques, and more traditional alternatives such as Markov and Bayesian techniques, and compared their prediction accuracy quantitatively for three domains: a role-playing game, intrusion-system alerts, and event prediction of maritime paths in a discrete-event simulator. Other contributions include two new unification algorithms that improve over a na ve one, and an exploration of ways to maintain a minimum-size knowledge base without affecting prediction accuracy.

We consider $\alpha$-mixing observations and deal with the estimation of the conditional mode of a scalar response variable $Y$ given a random variable $X$ taking values in a semi-metric space. We provide a convergence rate in $L^p$ norm of the estimator. A useful and typical application to functional times series prediction is given.

A computer program is described and presented for calculating finite memory predictors and prediction variances for autoregressive moving average time series models. The Cholesky decomposition algorithm is used, and a number of simplifying results are described and implemented in the program. (Author)

A computer program is described and presented for calculating finite memory predictors and prediction variances for autoregressive moving average time series models. The Cholesky decomposition algorithm is used, and a number of simplifying results are described and implemented in the program. (Author)

Shang and Hyndman (2016) proposed grouped functional time series forecasting approach as a combination of individual forecasts using generalized least squares regression. We modify their methodology using generalized exponential smoothing technique for the most disaggregated series in order to obtain more robust predictor. We show some properties of our proposals using simulations and real data related to electricity demand prediction.

This article is from PLoS ONE , volume 9 . Abstract In this paper a novel application of a particular type of spiking neural network, a Polychronous Spiking Network, was used for financial time series prediction. It is argued that the inherent temporal capabilities of this type of network are suited to non-stationary data such as this. The performance of the spiking neural network was benchmarked against three systems: two “traditional”, rate-encoded, neural networks; a Multi-Layer Perceptron neural network and a Dynamic Ridge Polynomial neural network, and a standard Linear Predictor Coefficients model. For this comparison three non-stationary and noisy time series were used: IBM stock data; US/Euro exchange rate data, and the price of Brent crude oil. The experiments demonstrated favourable prediction results for the Spiking Neural Network in terms of Annualised Return and prediction error for 5-Step ahead predictions. These results were also supported by other relevant metrics such as Maximum Drawdown and Signal-To-Noise ratio. This work demonstrated the applicability of the Polychronous Spiking Network to financial data forecasting and this in turn indicates the potential of using such networks over traditional systems in difficult to manage non-stationary environments.

We tackle the issue of the blind prediction of a Gaussian time series. For this, we construct a projection operator build by plugging an empirical covariance estimation into a Schur complement decomposition of the projector. This operator is then used to compute the predictor. Rates of convergence of the estimates are given.

We present two approaches for next step linear prediction of long memory time series. The first is based on the truncation of the Wiener-Kolmogorov predictor by restricting the observations to the last $k$ terms, which are the only available values in practice. Part of the mean squared prediction error comes from the truncation, and another part comes from the parametric estimation of the parameters of the predictor. By contrast, the second approach is non-parametric. An AR($k$) model is fitted to the long memory time series and we study the error made with this misspecified model.

Results of analysis of long time variations of solar activities based on monthly Wolf number series (1749-2011 years) are represented. Shown the presence of stable oscillations in autocorrelation function for the Wolf numbers series with a period near 42.5 years. The statistical model of variations of autocorrelation function based on assumption of nonstationarity process are constructed. The question of using the model to predict the solar activity is discussed.

A computer program is described and presented for calculating finite memory predictors and prediction variances for autoregressive moving average time series models. The Cholesky decomposition algorithm is used, and a number of simplifying results are described and implemented in the program. (Author)