published_or_final_version / abstract / toc / Statistics and Actuarial Science / Master / Master of Philosophy

In classic GARCH models for financial returns the innovations are usually assumed to be normally distributed. However, it is generally accepted that a non-normal innovation distribution is needed in order to account for the heavier tails often encountered in financial returns. Since the structure of the normal inverse Gaussian (NIG) distribution makes it an attractive alternative innovation distribution for this purpose, we extend the normal GARCH model by assuming that the innovations are NIG-distributed. We use the normal variance mixture interpretation of the NIG distribution to show that a NIG innovation may be interpreted as a normal innovation coupled with a multiplicative random impact factor adjustment of the ordinary GARCH volatility. We relate this new volatility estimate to realised volatility and suggest that the random impact factors are due to a news noise process influencing the underlying returns process. This GARCH model with NIG-distributed innovations leads to more accurate parameter estimates than the normal GARCH model. In order to obtain even more accurate parameter estimates, and since we expect an information gain if we use more data, we further extend the model to cater for high, low and close data, as well as full intraday data, instead of only daily returns. This is achieved by introducing the Brownian inverse Gaussian (BIG) process, which follows naturally from the unit inverse Gaussian distribution and standard Brownian motion. Fitting these models to empirical data, we find that the accuracy of the model fit increases as we move from the models assuming normally distributed innovations and allowing for only daily data to those assuming underlying BIG processes and allowing for full intraday data. However, we do encounter one problematic result, namely that there is empirical evidence of time dependence in the random impact factors. This means that the news noise processes, which we assumed to be independent over time, are indeed time dependent, as can actually be expected. In order to cater for this time dependence, we extend the model still further by allowing for autocorrelation in the random impact factors. The increased complexity that this extension introduces means that we can no longer rely on standard Maximum Likelihood methods, but have to turn to Simulated Maximum Likelihood methods, in conjunction with Efficient Importance Sampling and the Control Variate variance reduction technique, in order to obtain an approximation to the likelihood function and the parameter estimates. We find that this time dependent model assuming an underlying BIG process and catering for full intraday data fits generated data and empirical data very well, as long as enough intraday data is available. / Thesis (Ph.D. (Risk Analysis))--North-West University, Potchefstroom Campus, 2006.

Multi-task learning refers to learning multiple tasks simultaneously, in order to avoid tabula rasa learning and to share information between similar tasks during learning. We consider a multi-task Gaussian process regression model that learns related functions by inducing correlations between tasks directly. Using this model as a reference for three other multi-task models, we provide a broad unifying view of multi-task learning. This is possible because, unlike the other models, the multi-task Gaussian process model encodes task relatedness explicitly. Each multi-task learning model generally assumes that learning multiple tasks together is beneficial. We analyze how and the extent to which multi-task learning helps improve the generalization of supervised learning. Our analysis is conducted for the average-case on the multi-task Gaussian process model, and we concentrate mainly on the case of two tasks, called the primary task and the secondary task. The main parameters are the degree of relatedness ρ between the two tasks, and πS, the fraction of the total training observations from the secondary task. Among other results, we show that asymmetric multitask learning, where the secondary task is to help the learning of the primary task, can decrease a lower bound on the average generalization error by a factor of up to ρ2πS. When there are no observations for the primary task, there is also an intrinsic limit to which observations for the secondary task can help the primary task. For symmetric multi-task learning, where the two tasks are to help each other to learn, we find the learning to be characterized by the term πS(1 − πS)(1 − ρ2). As far as we are aware, our analysis contributes to an understanding of multi-task learning that is orthogonal to the existing PAC-based results on multi-task learning. For more than two tasks, we provide an understanding of the multi-task Gaussian process model through structures in the predictive means and variances given certain configurations of training observations. These results generalize existing ones in the geostatistics literature, and may have practical applications in that domain. We evaluate the multi-task Gaussian process model on the inverse dynamics problem for a robot manipulator. The inverse dynamics problem is to compute the torques needed at the joints to drive the manipulator along a given trajectory, and there are advantages to learning this function for adaptive control. A robot manipulator will often need to be controlled while holding different loads in its end effector, giving rise to a multi-context or multi-load learning problem, and we treat predicting the inverse dynamics for a context/load as a task. We view the learning of the inverse dynamics as a function approximation problem and place Gaussian process priors over the space of functions. We first show that this is effective for learning the inverse dynamics for a single context. Then, by placing independent Gaussian process priors over the latent functions of the inverse dynamics, we obtain a multi-task Gaussian process prior for handling multiple loads, where the inter-context similarity depends on the underlying inertial parameters of the manipulator. Experiments demonstrate that this multi-task formulation is effective in sharing information among the various loads, and generally improves performance over either learning only on single contexts or pooling the data over all contexts. In addition to the experimental results, one of the contributions of this study is showing that the multi-task Gaussian process model follows naturally from the physics of the inverse dynamics.

In classic GARCH models for financial returns the innovations are usually assumed to be normally distributed. However, it is generally accepted that a non-normal innovation distribution is needed in order to account for the heavier tails often encountered in financial returns. Since the structure of the normal inverse Gaussian (NIG) distribution makes it an attractive alternative innovation distribution for this purpose, we extend the normal GARCH model by assuming that the innovations are NIG-distributed. We use the normal variance mixture interpretation of the NIG distribution to show that a NIG innovation may be interpreted as a normal innovation coupled with a multiplicative random impact factor adjustment of the ordinary GARCH volatility. We relate this new volatility estimate to realised volatility and suggest that the random impact factors are due to a news noise process influencing the underlying returns process. This GARCH model with NIG-distributed innovations leads to more accurate parameter estimates than the normal GARCH model. In order to obtain even more accurate parameter estimates, and since we expect an information gain if we use more data, we further extend the model to cater for high, low and close data, as well as full intraday data, instead of only daily returns. This is achieved by introducing the Brownian inverse Gaussian (BIG) process, which follows naturally from the unit inverse Gaussian distribution and standard Brownian motion. Fitting these models to empirical data, we find that the accuracy of the model fit increases as we move from the models assuming normally distributed innovations and allowing for only daily data to those assuming underlying BIG processes and allowing for full intraday data. However, we do encounter one problematic result, namely that there is empirical evidence of time dependence in the random impact factors. This means that the news noise processes, which we assumed to be independent over time, are indeed time dependent, as can actually be expected. In order to cater for this time dependence, we extend the model still further by allowing for autocorrelation in the random impact factors. The increased complexity that this extension introduces means that we can no longer rely on standard Maximum Likelihood methods, but have to turn to Simulated Maximum Likelihood methods, in conjunction with Efficient Importance Sampling and the Control Variate variance reduction technique, in order to obtain an approximation to the likelihood function and the parameter estimates. We find that this time dependent model assuming an underlying BIG process and catering for full intraday data fits generated data and empirical data very well, as long as enough intraday data is available. / Thesis (Ph.D. (Risk Analysis))--North-West University, Potchefstroom Campus, 2006.

<p>Cutting edge research problems require the use of complicated and computationally expensive computer models. I will present a practical overview of the design and analysis of computer experiments in high energy nuclear and astro phsyics. The aim of these experiments is to infer credible ranges for certain fundamental parameters of the underlying physical processes through the analysis of model output and experimental data.</p><p>To be truly useful computer models must be calibrated against experimental data. Gaining an understanding of the response of expensive models across the full range of inputs can be a slow and painful process. Gaussian Process emulators can be an efficient and informative surrogate for expensive computer models and prove to be an ideal mechanism for exploring the response of these models to variations in their inputs.</p><p>A sensitivity analysis can be performed on these model emulators to characterize and quantify the relationship between model input parameters and predicted observable properties. The result of this analysis provides the user with information about which parameters are most important and most likely to affect the prediction of a given observable. Sensitivity analysis allow us to identify what model parameters can be most efficiently constrained by the given observational data set.</p><p>In this thesis I describe a range of techniques for the calibration and exploration of the complex and expensive computer models so common in modern physics research. These statistical methods are illustrated with examples drawn from the fields of high energy nuclear physics and galaxy formation.</p> / Dissertation

With the volume of digital media available today, automatic music recommendation services have proven a useful tool for consumers, allowing them to better discover new and enjoyable music. Typically, this technology is based on collaborative filtering techniques, employing human-generated metadata to base recommendations. Recently, work in content-based recommendation systems have emerged in which the audio signal itself is analyzed for relevant musical information from which models are built that attempt to mimic human similarity judgments. The current state-of-the-art for content-based music recommendation uses a timbre model based on MFCCs calculated on short segments of tracks. These feature vectors are then modeled using GMMs (Gaussian mixture models). GMM modeling of frame-based MFCCs has been shown to perform fairly well on timbre similarity tasks. However, a common problem is that of hubs , in which a relative small number of songs falsely appear similar to many other songs, significantly decreasing the accuracy of similarity recommendations. In this thesis, we explore the origins of hubs in timbre-based modeling and propose several remedies. Specifically, we find that a process of model homogenization, in which certain components of a mixture model are systematically removed, improves performance as measured against several ground-truth similarity metrics. Extending the work of Aucouturier, we introduce several new methods of homogenization. On a subset of the uspop data set, model homogenization improves artist R-precision by a maximum of 3.5% and agreement to user collection co-occurrence data by 7.4%. We also find differences in the effectiveness of the various homogenization methods for hub reduction, with the proposed methods providing the best results. Further, we extend the modeling of frame-based MFCC features by using a kernel density estimation approach to non-parametric modeling. We find that such an approach significantly reduces the number of hubs (by 2.6% of the dataset) while improving agreement to ground-truth by 5% and slightly improving artist R-precision as compared with the standard parametric model. Finally, to test whether these principles hold for all musical data, we introduce an entirely new data set consisting of Indian classical music. We find that our results generalize here as well, suggesting that hubness is a general feature of timbre-based similarity music modeling and that the techniques presented to improve this modeling are effective for diverse types of music.

We present limit theory for tests of homogeneity for finite mixture models. More specifically, we derive the asymptotic distribution of certain random quantities used for testing that a mixture of two distributions is in fact just a single distribution. Our methods apply to cases where the mixture component distributions come from one of a wide class of one-parameter exponential families, both continous and discrete. We consider two random quantities, one related to testing simple hypotheses, the other composite hypotheses. For simple hypotheses we consider the maximum of the standardised score process, which is itself a test statistic. For composite hypotheses we consider the maximum of the efficient score process, which is itself not a statistic (it depends on the unknown true distribution) but is asymptotically equivalent to certain common test statistics in a certain sense. We show that we can approximate both quantities with the maximum of a certain Gaussian process depending on the sample size and the true distribution of the observations, which when suitably normalised has a limiting distribution of the Gumbel extreme value type. Although the limit theory is not practically useful for computing approximate p-values, we use Monte-Carlo simulations to show that another method suggested by the theory, involving using a Studentised version of the maximum-score statistic and simulating a Gaussian process to compute approximate p-values, is remarkably accurate and uses a fraction of the computing resources that a straight Monte-Carlo approximation would.

Thesis (Ph. D.)--University of Sydney, 2002. / Title from title screen (viewed Apr. 28, 2008). Submitted in fulfilment of the requirements for the degree of Doctor of Philosophy to the School of Mathematics and Statistics, Faculty of Science. Includes bibliography. Also available in print form.

Thesis (Ph.D.) OGI School of Science & Engineering at OHSU, June 2008. / Includes bibliographical references (leaves 138-145).

Thesis (Ph. D.)--Industrial and Systems Engineering, Georgia Institute of Technology, 2007. / Lu, Jye-Chyi, Committee Member ; Shaprio, Alexander, Committee Member ; Amemiya, Yasuo, Committee Co-Chair ; Wu, C. F. Jeff, Committee Chair ; Vengazhiyil, Roshan Joseph, Committee Member.