Leung Tat Wing. / Thesis (M.Phil.)--Chinese University of Hong Kong, 2006. / Includes bibliographical references (leaves 94-98). / Abstracts in English and Chinese. / Abstract --- p.i / Acknowledgement --- p.iv / Chapter 1 --- Introduction --- p.1 / Chapter 1.1 --- Unusual Pattern Discovery --- p.3 / Chapter 1.2 --- Trend Prediction --- p.4 / Chapter 1.3 --- Thesis Organization --- p.5 / Chapter 2 --- Unusual Pattern Discovery --- p.6 / Chapter 2.1 --- Introduction --- p.6 / Chapter 2.2 --- Related Work --- p.7 / Chapter 2.2.1 --- Time Series Discords --- p.7 / Chapter 2.2.2 --- Brute Force Algorithm --- p.8 / Chapter 2.2.3 --- Keogh et al.'s Algorithm --- p.10 / Chapter 2.2.4 --- Performance Analysis --- p.14 / Chapter 2.3 --- Proposed Approach --- p.18 / Chapter 2.3.1 --- Haar Transform --- p.20 / Chapter 2.3.2 --- Discretization --- p.22 / Chapter 2.3.3 --- Augmented Trie --- p.24 / Chapter 2.3.4 --- Approximating the Magic Outer Loop --- p.27 / Chapter 2.3.5 --- Approximating the Magic Inner Loop --- p.28 / Chapter 2.3.6 --- Experimental Result --- p.28 / Chapter 2.4 --- More on discord length --- p.42 / Chapter 2.4.1 --- Modified Haar Transform --- p.42 / Chapter 2.4.2 --- Fast Haar Transform Algorithm --- p.43 / Chapter 2.4.3 --- Relation between discord length and discord location --- p.45 / Chapter 2.5 --- Further Optimization --- p.47 / Chapter 2.5.1 --- Improved Inner Loop Heuristic --- p.50 / Chapter 2.5.2 --- Experimental Result --- p.52 / Chapter 2.6 --- Top K discords --- p.53 / Chapter 2.6.1 --- Utility of top K discords --- p.53 / Chapter 2.6.2 --- Algorithm --- p.58 / Chapter 2.6.3 --- Experimental Result --- p.62 / Chapter 2.7 --- Conclusion --- p.64 / Chapter 3 --- Trend Prediction --- p.69 / Chapter 3.1 --- Introduction --- p.69 / Chapter 3.2 --- Technical Analysis --- p.70 / Chapter 3.2.1 --- Relative Strength Index --- p.70 / Chapter 3.2.2 --- Chart Analysis --- p.70 / Chapter 3.2.3 --- Dow Theory --- p.71 / Chapter 3.2.4 --- Moving Average --- p.72 / Chapter 3.3 --- Proposed Algorithm --- p.79 / Chapter 3.3.1 --- Piecewise Linear Representation --- p.80 / Chapter 3.3.2 --- Prediction Tree --- p.82 / Chapter 3.3.3 --- Trend Prediction --- p.84 / Chapter 3.4 --- Experimental Results --- p.86 / Chapter 3.4.1 --- Experimental setup --- p.86 / Chapter 3.4.2 --- Experiment on accuracy --- p.87 / Chapter 3.4.3 --- Experiment on performance --- p.88 / Chapter 3.5 --- Conclusion --- p.90 / Chapter 4 --- Conclusion --- p.92 / Bibliography --- p.94

Not available / text

Ying-Qian Zhang. / Thesis (M.Phil.)--Chinese University of Hong Kong, 2001. / Includes bibliographical references (leaves 115-124). / Abstracts in English and Chinese. / Abstract --- p.i / Acknowledgement --- p.iii / Chapter 1 --- Introduction --- p.1 / Chapter 1.1 --- Background --- p.1 / Chapter 1.2 --- Objective --- p.2 / Chapter 1.3 --- Contributions --- p.3 / Chapter 1.4 --- Thesis Overview --- p.4 / Chapter 2 --- Literature Review --- p.6 / Chapter 2.1 --- Takens' Theorem --- p.6 / Chapter 2.2 --- Linear Models for Prediction --- p.7 / Chapter 2.2.1 --- Autoregressive Model --- p.7 / Chapter 2.2.2 --- Moving Average Model --- p.8 / Chapter 2.2.3 --- Autoregressive-moving Average Model --- p.9 / Chapter 2.2.4 --- Fitting a Linear Model to a Given Time Series --- p.9 / Chapter 2.2.5 --- State-space Reconstruction --- p.10 / Chapter 2.3 --- Neural Network Models for Time Series Processing --- p.11 / Chapter 2.3.1 --- Feed-forward Neural Networks --- p.11 / Chapter 2.3.2 --- Recurrent Neural Networks --- p.14 / Chapter 2.3.3 --- Training Algorithms for Recurrent Networks --- p.18 / Chapter 2.4 --- Combining Neural Networks and other approximation techniques --- p.22 / Chapter 3 --- ForeNet: Model and Representation --- p.24 / Chapter 3.1 --- Fourier Recursive Prediction Equation --- p.24 / Chapter 3.1.1 --- Fourier Analysis of Time Series --- p.25 / Chapter 3.1.2 --- Recursive Form --- p.25 / Chapter 3.2 --- Fourier Recurrent Neural Network Model (ForeNet) --- p.27 / Chapter 3.2.1 --- Neural Networks Representation --- p.28 / Chapter 3.2.2 --- Architecture of ForeNet --- p.29 / Chapter 4 --- ForeNet: Implementation --- p.32 / Chapter 4.1 --- Improvement on ForeNet --- p.33 / Chapter 4.1.1 --- Number of Hidden Neurons --- p.33 / Chapter 4.1.2 --- Real-valued Outputs --- p.34 / Chapter 4.2 --- Parameters Initialization --- p.37 / Chapter 4.3 --- Application of ForeNet: the Process of Time Series Prediction --- p.38 / Chapter 4.4 --- Some Implications --- p.39 / Chapter 5 --- ForeNet: Initialization --- p.40 / Chapter 5.1 --- Unfolded Form of ForeNet --- p.40 / Chapter 5.2 --- Coefficients Analysis --- p.43 / Chapter 5.2.1 --- "Analysis of the Coefficients Set, vn " --- p.43 / Chapter 5.2.2 --- "Analysis of the Coefficients Set, μn(d) " --- p.44 / Chapter 5.3 --- Experiments of ForeNet Initialization --- p.47 / Chapter 5.3.1 --- Objective and Experiment Setting --- p.47 / Chapter 5.3.2 --- Prediction of Sunspot Series --- p.49 / Chapter 5.3.3 --- Prediction of Mackey-Glass Series --- p.53 / Chapter 5.3.4 --- Prediction of Laser Data --- p.56 / Chapter 5.3.5 --- Three More Series --- p.59 / Chapter 5.4 --- Some Implications on the Proposed Initialization Method --- p.63 / Chapter 6 --- ForeNet: Learning Algorithms --- p.67 / Chapter 6.1 --- Complex Real Time Recurrent Learning (CRTRL) --- p.68 / Chapter 6.2 --- Batch-mode Learning --- p.70 / Chapter 6.3 --- Time Complexity --- p.71 / Chapter 6.4 --- Property Analysis and Experimental Results --- p.72 / Chapter 6.4.1 --- Efficient initialization:compared with random initialization --- p.74 / Chapter 6.4.2 --- Complex-valued network:compared with real-valued net- work --- p.78 / Chapter 6.4.3 --- Simple architecture:compared with ring-structure RNN . --- p.79 / Chapter 6.4.4 --- Linear model: compared with nonlinear ForeNet --- p.80 / Chapter 6.4.5 --- Small number of hidden units --- p.88 / Chapter 6.5 --- Comparison with Some Other Models --- p.89 / Chapter 6.5.1 --- Comparison with AR model --- p.91 / Chapter 6.5.2 --- Comparison with TDNN Networks and FIR Networks . --- p.93 / Chapter 6.5.3 --- Comparison to a few more results --- p.94 / Chapter 6.6 --- Summarization --- p.95 / Chapter 7 --- Learning and Prediction: On-Line Training --- p.98 / Chapter 7.1 --- On-Line Learning Algorithm --- p.98 / Chapter 7.1.1 --- Advantages and Disadvantages --- p.98 / Chapter 7.1.2 --- Training Process --- p.99 / Chapter 7.2 --- Experiments --- p.101 / Chapter 7.3 --- Predicting Stock Time Series --- p.105 / Chapter 8 --- Discussions and Conclusions --- p.109 / Chapter 8.1 --- Limitations of ForeNet --- p.109 / Chapter 8.2 --- Advantages of ForeNet --- p.111 / Chapter 8.3 --- Future Works --- p.112 / Bibliography --- p.115

Tong Ka Yau Dennis. / Thesis (M.Phil.)--Chinese University of Hong Kong, 2004. / Includes bibliographical references (leaves 77-81). / Abstracts in English and Chinese. / Chapter 1 --- Introduction --- p.1 / Chapter 1.1 --- Motivations --- p.1 / Chapter 1.2 --- Progress on the Problem --- p.2 / Chapter 1.3 --- Our Contributions --- p.3 / Chapter 1.4 --- Thesis Organization --- p.4 / Chapter 2 --- Background on Retiming --- p.5 / Chapter 2.1 --- Introduction --- p.5 / Chapter 2.2 --- Preliminaries --- p.7 / Chapter 2.3 --- Retiming Problem --- p.9 / Chapter 3 --- Literature Review on Retiming --- p.10 / Chapter 3.1 --- Introduction --- p.10 / Chapter 3.2 --- The First Retiming Paper --- p.11 / Chapter 3.2.1 --- """Retiming Synchronous Circuitry""" --- p.11 / Chapter 3.3 --- Important Extensions of the Basic Retiming Algorithm --- p.14 / Chapter 3.3.1 --- """A Fresh Look at Retiming via Clock Skew Optimization""" --- p.14 / Chapter 3.3.2 --- """An Improved Algorithm for Minimum-Area Retiming""" --- p.16 / Chapter 3.3.3 --- """Efficient Implementation of Retiming""" --- p.17 / Chapter 3.4 --- Retiming in Physical Design Stages --- p.19 / Chapter 3.4.1 --- """Physical Planning with Retiming""" --- p.19 / Chapter 3.4.2 --- """Simultaneous Circuit Partitioning/Clustering with Re- timing for Performance Optimization" --- p.20 / Chapter 3.4.3 --- """Performance Driven Multi-level and Multiway Parti- tioning with Retiming" --- p.22 / Chapter 3.5 --- Retiming with More Sophisticated Timing Models --- p.23 / Chapter 3.5.1 --- """Retiming with Non-zero Clock Skew, Variable Register, and Interconnect Delay""" --- p.23 / Chapter 3.5.2 --- """Placement Driven Retiming with a Coupled Edge Tim- ing Model""" --- p.24 / Chapter 3.6 --- Post-Retiming Register Placement --- p.26 / Chapter 3.6.1 --- """Layout Driven Retiming Using the Coupled Edge Tim- ing Model""" --- p.26 / Chapter 3.6.2 --- """Integrating Logic Retiming and Register Placement""" --- p.27 / Chapter 4 --- Retiming with Gate and Wire Delay [2] --- p.29 / Chapter 4.1 --- Introduction --- p.29 / Chapter 4.2 --- Problem Formulation --- p.30 / Chapter 4.3 --- Optimal Approach [2] --- p.31 / Chapter 4.3.1 --- Original Mathematical Framework for Retiming --- p.31 / Chapter 4.3.2 --- A Modified Optimal Approach --- p.33 / Chapter 4.4 --- Near-Optimal Fast Approach [2] --- p.37 / Chapter 4.4.1 --- Considering Wire Delay Only --- p.38 / Chapter 4.4.2 --- Considering Both Gate and Wire Delay --- p.42 / Chapter 4.4.3 --- Computational Complexity --- p.43 / Chapter 4.4.4 --- Experimental Results --- p.44 / Chapter 4.5 --- Lin's Optimal Approach [23] --- p.47 / Chapter 4.5.1 --- Theoretical Results --- p.47 / Chapter 4.5.2 --- Algorithm Description --- p.51 / Chapter 4.5.3 --- Computational Complexity --- p.52 / Chapter 4.5.4 --- Experimental Results --- p.52 / Chapter 4.6 --- Summary --- p.54 / Chapter 5 --- Register Insertion in Placement [36] --- p.55 / Chapter 5.1 --- Introduction --- p.55 / Chapter 5.2 --- Problem Formulation --- p.57 / Chapter 5.3 --- Placement of Registers After Retiming --- p.60 / Chapter 5.3.1 --- Topology Finding --- p.60 / Chapter 5.3.2 --- Register Placement --- p.69 / Chapter 5.4 --- Experimental Results --- p.71 / Chapter 5.5 --- Summary --- p.74 / Chapter 6 --- Conclusion --- p.75 / Bibliography --- p.77

由於特徵尺寸不斷縮小，集成電路在生產過程中的工藝偏差在運行環境中溫度和電壓等參數的波動以及在使用過程中的老化等效應越來越嚴重，導致芯片的時序行為出現很大的不確定性。多數情況下，芯片的關鍵路徑會不時出現時序錯誤。加入更多的時序餘量不是一種很好的解決方案，因為這種保守的設計方法會抵消工藝進步帶來的性能上的好處。這就為設計一個時序可靠的系統提出了極大的挑戰，其中的一些關鍵問題包括：(一)如何有效地分配有限的功率預算去優化那些正爆炸式增加的關鍵路徑的時序性能；(二)如何產生能夠捕捉準確的最壞情況時延的高品質測試向量；(三)為了能夠取得更好的功耗和性能上的平衡，我們將不得不允許芯片在使用過程中出現一些頻率很低的時序錯誤。隨之而來的問題是如何做到在線的檢錯和糾錯。 / 為了解決上述問題，我們首先發明了一種新的技術用於識別所謂的虛假路徑，該方法使我們能夠發現比傳統方法更多的虛假路徑。當將所提取的虛假路徑集成到靜態時序分析工具里以後，我們可以得到更為準確的時序分析結果，同時也能節省本來用於優化這些路徑的成本。接著，考慮到現有的延時自動向量生成(ATPG) 方法會產生功能模式下無法出現的測試向量，這種向量可能會造成測試過程中在被激活的路徑周圍出現過多(或過少)的電源噪聲(PSN) ，從而導致測試過度或者測試不足情況。為此，我們提出了一種新的偽功能ATPG工具。通過同時考慮功能約束以及電路的物理佈局信息，我們使用類似ATPG 的算法產生狀態跳變使其能最大化已激活的路徑周圍的PSN影響。最後，基於近似電路的原理，我們提出了一種新的在線原位校正技術，即InTimeFix，用於糾正時序錯誤。由於實現近似電路的綜合僅需要簡單的電路結構分析，因此該技術能夠很容易的擴展到大型電路設計上去。 / With technology scaling, integrated circuits (ICs) suffer from increasing process, voltage, and temperature (PVT) variations and aging effects. In most cases, these reliability threats manifest themselves as timing errors on speed-paths (i.e., critical or near-critical paths) of the circuit. Embedding a large design guard band to prevent timing errors to occur is not an attractive solution, since this conservative design methodology diminishes the benefit of technology scaling. This creates several challenges on build a reliable systems, and the key problems include (i) how to optimize circuit’s timing performance with limited power budget for explosively increased potential speed-paths; (ii) how to generate high quality delay test pattern to capture ICs’ accurate worst-case delay; (iii) to have better power and performance tradeoff, we have to accept some infrequent timing errors in circuit’s the usage phase. Therefore, the question is how to achieve online timing error resilience. / To address the above issues, we first develop a novel technique to identify so-called false paths, which facilitate us to find much more false paths than conventional methods. By integrating our identified false paths into static timing analysis tool, we are able to achieve more accurate timing information and also save the cost used to optimize false paths. Then, due to the fact that existing delay automated test pattern generation (ATPG) methods may generate test patterns that are functionally-unreachable, and such patterns may incur excessive (or limited) power supply noise (PSN) on sensitized paths in test mode, thus leading to over-testing or under-testing of the circuits, we propose a novel pseudo-functional ATPG tool. By taking both circuit layout information and functional constrains into account, we use ATPG like algorithm to justify transitions that pose the maximized functional PSN effects on sensitized critical paths. Finally, we propose a novel in-situ correction technique to mask timing errors, namely InTimeFix, by introducing redundant approximation circuit with more timing slack for speed-paths into the design. The synthesis of the approximation circuit relies on simple structural analysis of the original circuit, which is easily scalable to large IC designs. / Detailed summary in vernacular field only. / Detailed summary in vernacular field only. / Yuan, Feng. / Thesis (Ph.D.)--Chinese University of Hong Kong, 2012. / Includes bibliographical references (leaves 88-100). / Abstract also in Chinese. / Abstract --- p.i / Acknowledgement --- p.iv / Chapter 1 --- Introduction --- p.1 / Chapter 1.1 --- Challenges to Solve Timing Uncertainty Problem --- p.2 / Chapter 1.2 --- Contributions and Thesis Outline --- p.5 / Chapter 2 --- Background --- p.7 / Chapter 2.1 --- Sources of Timing Uncertainty --- p.7 / Chapter 2.1.1 --- Process Variation --- p.7 / Chapter 2.1.2 --- Runtime Environment Fluctuation --- p.9 / Chapter 2.1.3 --- Aging Effect --- p.10 / Chapter 2.2 --- Technical Flow to Solve Timing Uncertainty Problem --- p.10 / Chapter 2.3 --- False Path --- p.12 / Chapter 2.3.1 --- Path Sensitization Criteria --- p.12 / Chapter 2.3.2 --- False Path Aware Timing Analysis --- p.13 / Chapter 2.4 --- Manufacturing Testing --- p.14 / Chapter 2.4.1 --- Functional Testing vs. Structural Testing --- p.14 / Chapter 2.4.2 --- Scan-Based DfT --- p.15 / Chapter 2.4.3 --- Pseudo-Functional Testing --- p.17 / Chapter 2.5 --- Timing Error Tolerance --- p.19 / Chapter 2.5.1 --- Timing Error Detection --- p.19 / Chapter 2.5.2 --- Timing Error Recover --- p.20 / Chapter 3 --- Timing-Independent False Path Identification --- p.23 / Chapter 3.1 --- Introduction --- p.23 / Chapter 3.2 --- Preliminaries and Motivation --- p.26 / Chapter 3.2.1 --- Motivation --- p.27 / Chapter 3.3 --- False Path Examination Considering Illegal States --- p.28 / Chapter 3.3.1 --- Path Sensitization Criterion --- p.28 / Chapter 3.3.2 --- Path-Aware Illegal State Identification --- p.30 / Chapter 3.3.3 --- Proposed Examination Procedure --- p.31 / Chapter 3.4 --- False Path Identification --- p.32 / Chapter 3.4.1 --- Overall Flow --- p.34 / Chapter 3.4.2 --- Static Implication Learning --- p.35 / Chapter 3.4.3 --- Suspicious Node Extraction --- p.36 / Chapter 3.4.4 --- S-Frontier Propagation --- p.37 / Chapter 3.5 --- Experimental Results --- p.38 / Chapter 3.6 --- Conclusion and Future Work --- p.42 / Chapter 4 --- PSN Aware Pseudo-Functional Delay Testing --- p.43 / Chapter 4.1 --- Introduction --- p.43 / Chapter 4.2 --- Preliminaries and Motivation --- p.45 / Chapter 4.2.1 --- Motivation --- p.46 / Chapter 4.3 --- Proposed Methodology --- p.48 / Chapter 4.4 --- Maximizing PSN Effects under Functional Constraints --- p.50 / Chapter 4.4.1 --- Pseudo-Functional Relevant Transitions Generation --- p.51 / Chapter 4.5 --- Experimental Results --- p.59 / Chapter 4.5.1 --- Experimental Setup --- p.59 / Chapter 4.5.2 --- Results and Discussion --- p.60 / Chapter 4.6 --- Conclusion --- p.64 / Chapter 5 --- In-Situ Timing Error Masking in Logic Circuits --- p.65 / Chapter 5.1 --- Introduction --- p.65 / Chapter 5.2 --- Prior Work and Motivation --- p.67 / Chapter 5.3 --- In-Situ Timing Error Masking with Approximate Logic --- p.69 / Chapter 5.3.1 --- Equivalent Circuit Construction with Approximate Logic --- p.70 / Chapter 5.3.2 --- Timing Error Masking with Approximate Logic --- p.72 / Chapter 5.4 --- Cost-Efficient Synthesis for InTimeFix --- p.75 / Chapter 5.4.1 --- Overall Flow --- p.76 / Chapter 5.4.2 --- Prime Critical Segment Extraction --- p.77 / Chapter 5.4.3 --- Prime Critical Segment Merging --- p.79 / Chapter 5.5 --- Experimental Results --- p.81 / Chapter 5.5.1 --- Experimental Setup --- p.81 / Chapter 5.5.2 --- Results and Discussion --- p.82 / Chapter 5.6 --- Conclusion --- p.85 / Chapter 6 --- Conclusion and Future Work --- p.86 / Bibliography --- p.100

Due to the rapid improvements in the information technology, high dimensional time series datasets are frequently encountered in a variety of fields such as macroeconomics, finance, neuroscience, and meteorology. Some examples in economics and finance include forecasting low frequency macroeconomic indicators, such as GDP or inflation rate, or financial asset returns using a large number of macroeconomic and financial time series and their lags as possible covariates. In these settings, the number of candidate predictors (pT) can be much larger than the number of samples (T), and accurate estimation and prediction is made possible by relying on some form of dimension reduction. Given this ubiquity of time series data, it is surprising that few works on high dimensional statistics discuss the time series setting, and even fewer works have developed methods which utilize the unique features of time series data. This chapter consists of three chapters, and each one is self contained. The first chapter deals with high dimensional predictive regressions which are widely used in economics and finance. However, the theory and methodology is mainly developed assuming that the model is stationary with time invariant parameters. This is at odds with the prevalent evidence for parameter instability in economic time series. To remedy this, we present two L2 boosting algorithms for estimating high dimensional models in which the coefficients are modeled as functions evolving smoothly over time and the predictors are locally stationary. The first method uses componentwise local constant estimators as base learner, while the second relies on componentwise local linear estimators. We establish consistency of both methods, and address the practical issues of choosing the bandwidth for the base learners and the number of boosting iterations. In an extensive application to macroeconomic forecasting with many potential predictors, we find that the benefits to modeling time variation are substantial and are present across a wide range of economic series. Furthermore, these benefits increase with the forecast horizon and with the length of the time series available for estimation. This chapter is jointly written with Serena Ng. The second chapter deals with high dimensional non-linear time series models, and deals with the topic of variable screening/targeting predictors. Rather than assume a specific parametric model a priori, this chapter introduces several model free screening methods based on the partial distance correlation and developed specifically to deal with time dependent data. Methods are developed both for univariate models, such as nonlinear autoregressive models with exogenous predictors (NARX), and multivariate models such as linear or nonlinear VAR models. Sure screening properties are proved for our methods, which depend on the moment conditions, and the strength of dependence in the response and covariate processes, amongst other factors. Finite sample performance of our methods is shown through extensive simulation studies, and we show the effectiveness of our algorithms at forecasting US market returns. This chapter is jointly written with Yang Feng. The third chapter deals with variable selection for high dimensional linear stationary time series models. This chapter analyzes the theoretical properties of Sure Independence Screening (SIS), and its two stage combination with the adaptive Lasso, for high dimensional linear models with dependent and/or heavy tailed covariates and errors. We also introduce a generalized least squares screening (GLSS) procedure which utilizes the serial correlation present in the data. By utilizing this serial correlation when estimating our marginal effects, GLSS is shown to outperform SIS in many cases. For both procedures we prove two stage variable selection consistency when combined with the adaptive Lasso.

The research is based on financial time series modelling with special application / Thesis (M.Sc.) - University of KwaZulu-Natal, Pietermaritzburg, 2009.

Orientadores: Fernando Jose Von Zuben, Clodoaldo Aparecido de Moraes Lima / Dissertação (mestrado) - Universidade Estadual de Campinas, Faculdade de Engenharia Eletrica e de Computação / Made available in DSpace on 2018-08-13T12:09:04Z (GMT). No. of bitstreams: 1 PumaVillanueva_WilfredoJaime_M.pdf: 2240734 bytes, checksum: a645fed1284c994e9fa5a37d9bce0ccb (MD5) Previous issue date: 2006 / Resumo: A capacidade de aproximação universal apresentada por redes neurais artificiais foi explorada nos últimos anos junto a problemas de classificação e regressão de dados, envolvendo técnicas de treinamento supervisionado. No entanto, as redes neurais resultantes podem produzir queda de desempenho frente a amostras de teste. Esta é a principal motivação para o emprego de comitês de máquinas, na forma de um ensemble ou uma mistura de especialistas. Um ensemble toma propostas de solução completas para um problema e se ocupa em selecionar e combinar essas propostas na obtenção de uma única resposta. Já numa mistura de especialistas, cada especialista é responsável por parte do problema e os especialistas, assim como o módulo que decide qual especialista irá atuar em cada caso, são sintetizados simultaneamente. A aplicação de comitês de máquinas em predição de séries temporais indica que esta estratégia pode conduzir a ganhos de desempenho, quando comparado ao uso de um único preditor e considerando vários casos de estudo. Ainda no contexto de predição, foram investigadas duas técnicas para seleção de variáveis, além de ser avaliado o desempenho de duas propostas de partição da série temporal em conjuntos de treinamento, validação e teste. Os resultados de teste de significância do ganho de desempenho permitem apontar uma técnica de seleção e uma proposta de partição como as mais indicadas / Abstract: The universal approximation capability presented by artificial neural networks has been explored in recent years to solve classification and regression problems, using the supervised learning framework. However, the resulting neural networks may present degradation of performance when the test dataset is considered. This is the main motivation for the use of committee machines, in the form of an ensemble or a mixture of experts. An ensemble takes full-solution proposals and tries to select and combine them toward a single response. In the realm of a mixture of experts, each expert is devoted to a parcel of the original problem, and the experts, together with the module that allocates the individual role for each expert, are synthesized simultaneously. The application of committee machines to time series prediction indicates that these machine learning strategies can promote improvement in performance, when compared to the use of a single predictor and taking several case studies. Still in the context of prediction, two techniques for variable selection have been investigated, and two proposals for the partition of the time series in training, validation, and test datasets have been compared. The results in terms of test of significance of the gain in performance clearly indicate the superiority of one of the selection techniques and one of the partition proposals / Mestrado / Engenharia de Computação / Mestre em Engenharia Elétrica

Made available in DSpace on 2014-06-11T19:26:16Z (GMT). No. of bitstreams: 0 Previous issue date: 2012-12-21Bitstream added on 2014-06-13T19:13:20Z : No. of bitstreams: 1 costa_af_me_bauru.pdf: 1184483 bytes, checksum: 1783e7ab1d2b2b0b7dab05babd2093e8 (MD5) / Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) / A previsão de séries temporais financeiras é uma das questões mais pesquisadas no campo das finanças, sobretudo, no que diz respeito ao mercado acionário e à análise de riscos. Para tanto, essas pesquisas envolvem desde modelos estatísticos e econométricos até modelos de inteligência artificial, como as redes neurais dinâmicas. Nesse sentido este estudo tem o propósito de desenvolver e aplicar dois modelos de redes neurais artificiais dinâmicas, a rede neural focada atrasada no tempo - FTDNN (focused time delay neural network) e a rede neural auto regressiva com entradas exógenas - NARX (nonlinear autoregressive network with exogenous inputs) para previsão de séries temporais financeiras, tendo como padrão de referência de desempenho mínimo um modelo estatístico tradicional do tipo ARMA-GARCH. Essa abordagem camparativa também considera três modalidades diferentes de transformação de dados na fase pré-processamento das redes: as diferenças de primeira ordem, os retornos logarítmicos e a transformação Box-Cox, buscando analisar o impacto de cada uma delas no desempenho preditivo das redes neurais. Também propõe uma abordagem neural para o processo de reversão dos dados previstos e uma métrica de erro capaz de verificar o desempenho preditivo das redes neurais e sua capacidade de captar tendências de curto prazo e eficiência negocial. Em sentido amplo, os resultados obtidos indicam que a rede NARX apresenta melhor desempenho preditivo que a rede FTDNN, sobretudo, no que diz respeito à captura de tendências; que a transformações de dados podem melhorar o nível de acurácia das previsões em ambas as redes e que a transformação por retornos logarítimos gera os melhores desempenhos. Quanto ao processo de reversão dos dados previstos para a escala da série original, o método neural proposto foi bem sucedido apenas para a transformação Box-cox / Financial time-series forescasting is one of the most researched issues in finances, mainly with regard to the stock market and risk analysis. Therefore, these studies involve from statistical and econometric models up to artificial intelligence models, such as dynamic neural networks. In this sense, this study aims to develop and apply two models of dynamic artificial neural networks, FTDNN (focused time delay neural network) and NARX (nonlinear autoregressive network with exagenous inputs) for financial time series forescasting, with a traditional statistical model such as ARMAGARCH as the benchmark for minimum performance. The comparative approach also considers three diferent types of data transformation in the pre-processing phase of the networks: first order differences, logarithmics returns and Box-Cox transformation, and tries to analyze the impact of each on the predictive performance of the neural networks. It also proposes a neural approach to the process of reversing the predicted data set, and an error metric that could be able to verify the predictive performance of neural networks and its ability to capture short-term trends and negotiation efficiency. In a bropad sense, the reults indicate that the NARX network network performs beter than the FTDNN, especially with regard to capturing trends; that data transformations may improve the forescasting accuracy in both networks, and that the logarithmics returns transformation generates the best prediction performance. Regarding the process of reversing the predicted data for the scale of the original series, the neural method proposed succeeded only for Box-Cox transformation

Orientador: Antonio Fernando Crepaldi / Banca: Rogério Andrade Flauzino / Banca: José Alfredo Colovan Ulson / Resumo: A previsão de séries temporais financeiras é uma das questões mais pesquisadas no campo das finanças, sobretudo, no que diz respeito ao mercado acionário e à análise de riscos. Para tanto, essas pesquisas envolvem desde modelos estatísticos e econométricos até modelos de inteligência artificial, como as redes neurais dinâmicas. Nesse sentido este estudo tem o propósito de desenvolver e aplicar dois modelos de redes neurais artificiais dinâmicas, a rede neural focada atrasada no tempo - FTDNN (focused time delay neural network) e a rede neural auto regressiva com entradas exógenas - NARX (nonlinear autoregressive network with exogenous inputs) para previsão de séries temporais financeiras, tendo como padrão de referência de desempenho mínimo um modelo estatístico tradicional do tipo ARMA-GARCH. Essa abordagem camparativa também considera três modalidades diferentes de transformação de dados na fase pré-processamento das redes: as diferenças de primeira ordem, os retornos logarítmicos e a transformação Box-Cox, buscando analisar o impacto de cada uma delas no desempenho preditivo das redes neurais. Também propõe uma abordagem neural para o processo de reversão dos dados previstos e uma métrica de erro capaz de verificar o desempenho preditivo das redes neurais e sua capacidade de captar tendências de curto prazo e eficiência negocial. Em sentido amplo, os resultados obtidos indicam que a rede NARX apresenta melhor desempenho preditivo que a rede FTDNN, sobretudo, no que diz respeito à captura de tendências; que a transformações de dados podem melhorar o nível de acurácia das previsões em ambas as redes e que a transformação por retornos logarítimos gera os melhores desempenhos. Quanto ao processo de reversão dos dados previstos para a escala da série original, o método neural proposto foi bem sucedido apenas para a transformação Box-cox / Abstract: Financial time-series forescasting is one of the most researched issues in finances, mainly with regard to the stock market and risk analysis. Therefore, these studies involve from statistical and econometric models up to artificial intelligence models, such as dynamic neural networks. In this sense, this study aims to develop and apply two models of dynamic artificial neural networks, FTDNN (focused time delay neural network) and NARX (nonlinear autoregressive network with exagenous inputs) for financial time series forescasting, with a traditional statistical model such as ARMAGARCH as the benchmark for minimum performance. The comparative approach also considers three diferent types of data transformation in the pre-processing phase of the networks: first order differences, logarithmics returns and Box-Cox transformation, and tries to analyze the impact of each on the predictive performance of the neural networks. It also proposes a neural approach to the process of reversing the predicted data set, and an error metric that could be able to verify the predictive performance of neural networks and its ability to capture short-term trends and negotiation efficiency. In a bropad sense, the reults indicate that the NARX network network performs beter than the FTDNN, especially with regard to capturing trends; that data transformations may improve the forescasting accuracy in both networks, and that the logarithmics returns transformation generates the best prediction performance. Regarding the process of reversing the predicted data for the scale of the original series, the neural method proposed succeeded only for Box-Cox transformation / Mestre