Nonlinear Structure in Regression Residuals

Slides:



Advertisements
Similar presentations
Rachel T. Johnson Douglas C. Montgomery Bradley Jones
Advertisements

Principal Component Analysis Based on L1-Norm Maximization Nojun Kwak IEEE Transactions on Pattern Analysis and Machine Intelligence, 2008.
From
FTP Biostatistics II Model parameter estimations: Confronting models with measurements.
Unit Roots & Forecasting
Simple Linear Regression
Chapter 2: Pattern Recognition
Curve-Fitting Regression
Bootstrap in Finance Esther Ruiz and Maria Rosa Nieto (A. Rodríguez, J. Romo and L. Pascual) Department of Statistics UNIVERSIDAD CARLOS III DE MADRID.
A First Peek at the Extremogram: a Correlogram of Extremes 1. Introduction The Autocorrelation function (ACF) is widely used as a tool for measuring Serial.
Atul Singh Junior Undergraduate CSE, IIT Kanpur.  Dimension reduction is a technique which is used to represent a high dimensional data in a more compact.
1 An Introduction to Nonparametric Regression Ning Li March 15 th, 2004 Biostatistics 277.
14 Vector Autoregressions, Unit Roots, and Cointegration.
Introduction to Regression Analysis, Chapter 13,
1 Simple Linear Regression 1. review of least squares procedure 2. inference for least squares lines.
Principles of the Global Positioning System Lecture 10 Prof. Thomas Herring Room A;
Chapter 15 Forecasting Copyright © 2011 Pearson Addison-Wesley. All rights reserved. Slides by Niels-Hugo Blunch Washington and Lee University.
Data Mining Techniques
Hypothesis Testing in Linear Regression Analysis
1 Least squares procedure Inference for least squares lines Simple Linear Regression.
COMMON EVALUATION FINAL PROJECT Vira Oleksyuk ECE 8110: Introduction to machine Learning and Pattern Recognition.
1 Recurrence analysis and Fisheries Fisheries as a complex systems Traditional science operates on the assumption that natural systems like fish populations.
The Examination of Residuals. Examination of Residuals The fitting of models to data is done using an iterative approach. The first step is to fit a simple.
Various topics Petter Mostad Overview Epidemiology Study types / data types Econometrics Time series data More about sampling –Estimation.
Applications of Neural Networks in Time-Series Analysis Adam Maus Computer Science Department Mentor: Doctor Sprott Physics Department.
Chapter 4 Linear Regression 1. Introduction Managerial decisions are often based on the relationship between two or more variables. For example, after.
Regression Analysis Week 8 DIAGNOSTIC AND REMEDIAL MEASURES Residuals The main purpose examining residuals Diagnostic for Residuals Test involving residuals.
1 Multiple Regression A single numerical response variable, Y. Multiple numerical explanatory variables, X 1, X 2,…, X k.
ECE 8443 – Pattern Recognition LECTURE 10: HETEROSCEDASTIC LINEAR DISCRIMINANT ANALYSIS AND INDEPENDENT COMPONENT ANALYSIS Objectives: Generalization of.
Low-Dimensional Chaotic Signal Characterization Using Approximate Entropy Soundararajan Ezekiel Matthew Lang Computer Science Department Indiana University.
ECE-7000: Nonlinear Dynamical Systems Overfitting and model costs Overfitting  The more free parameters a model has, the better it can be adapted.
Correlation & Regression Analysis
Chapter 8: Simple Linear Regression Yang Zhenlin.
Controlling Chaos Journal presentation by Vaibhav Madhok.
ECE-7000: Nonlinear Dynamical Systems 3. Phase Space Methods 3.1 Determinism: Uniqueness in phase space We Assume that the system is linear stochastic.
Time Series Analysis PART II. Econometric Forecasting Forecasting is an important part of econometric analysis, for some people probably the most important.
Economics 173 Business Statistics Lecture 18 Fall, 2001 Professor J. Petry
Simple Linear Regression and Correlation (Continue..,) Reference: Chapter 17 of Statistics for Management and Economics, 7 th Edition, Gerald Keller. 1.
MODEL DIAGNOSTICS By Eni Sumarminingsih, Ssi, MM.
CPH Dr. Charnigo Chap. 11 Notes Figure 11.2 provides a diagram which shows, at a glance, what a neural network does. Inputs X 1, X 2,.., X P are.
Part 3: Estimation of Parameters. Estimation of Parameters Most of the time, we have random samples but not the densities given. If the parametric form.
Nonlinear time series analysis Delay embedding technique –reconstructs dynamics from single variable Correlation sum –provides an estimator of fractal.
1 Simple Linear Regression Chapter Introduction In Chapters 17 to 19 we examine the relationship between interval variables via a mathematical.
1 C.A.L. Bailer-Jones. Machine Learning. Data exploration and dimensionality reduction Machine learning, pattern recognition and statistical data modelling.
Multiple Random Variables and Joint Distributions
Chapter 4 Basic Estimation Techniques
Chapter 7. Classification and Prediction
The Cournot duopoly Kopel Model
Regression Analysis AGEC 784.
Inference for Least Squares Lines
Linear Regression CSC 600: Data Mining Class 12.
LECTURE 11: Advanced Discriminant Analysis
3.1 Examples of Demand Functions
Chapter 6: Autoregressive Integrated Moving Average (ARIMA) Models
Roberto Battiti, Mauro Brunato
Autocorrelation.
Prepared by Lee Revere and John Large
The normal distribution
STOCHASTIC HYDROLOGY Random Processes
Spatial Data Analysis: Intro to Spatial Statistical Concepts
Spatial Data Analysis: Intro to Spatial Statistical Concepts
Estimating the number of components with defects post-release that showed no defects in testing C. Stringfellow A. Andrews C. Wohlin H. Peterson Jeremy.
Testing for Multifractality and Multiplicativity using Surrogates
Statistical Data Analysis
Product moment correlation
Chapter 8 Supplement Forecasting.
Parametric Methods Berlin Chen, 2005 References:
Lecturer Dr. Veronika Alhanaqtah
Autocorrelation.
Chap 7: Seasonal ARIMA Models
Presentation transcript:

Nonlinear Structure in Regression Residuals Michael McCullougha, Thomas L. Marshb, and Ron C. Mittelhammerc aResearch Associate(contact author), bAssociate Professor, and cDirector, School of Economic Sciences, Washington State University; PO Box 646210, Pullman, WA 99164, United States.

Overview Motivation Previous research Primer on phase space reconstruction Nonlinear structure in regression residuals Simulation Application: S&P 500 Final observations 1 Simulations from a contaminated deterministic system are performed illustrating the ability to discern nonlineararities from noise. 2 An application to the S&P 500 bridges phase space reconstruction’s original use and regression diagnostics.

Motivation This paper investigates phase space reconstruction as a diagnostic tool for determining the structure of nonlinear processes in regression residuals. Outcomes will be used to create phase portraits for qualitative analysis. In effect, this approach is analogous to simple scatter plots in linear models.

Previous Research Maasoumi and Racine, J Econometrics 2002 Uses an entropy measure of distance to examine the predictability of stock market returns. Granger, Maasoumi, and Racine, JTSA 2004 Further reviews the performance of the metric entropy measure under different circumstances. Previous phase space reconstruction research focused primarily on the detection of chaos. Information and Entropy Economics suggest methods to detect dependence in regression residuals.

Phase Space Reconstruction For any given event, n outcomes are observed and denoted by the time series vector Xt = [x(t), x(t-1), …, x(t-n)]’ with associated th lag vector Xt- = [x(t-), x(t-1-), …, x(t-n-)]’ Takens’ Theorem (1981) embeds a single time series onto a phase space that reproduces the entire structure of the system. The embedding is a diffeomorphism which creates a matrix of time delayed vectors Y = [Xt, Xt-, …, Xt- ] of dimension [(n-)  ].

Phase Space Reconstruction The Method of Delays requires two parameter estimates: An optimal time lag; . The goal is to find the  which first minimizes redundancy between time delay vectors Xt and Xt-. A minimum embedding dimension; . This statistic estimates the minimum dimension at which the entire system dynamics may be appropriately represented. This technique is essentially nonparametric in nature.

The Method of Delays Estimating the optimal time lag; .  is estimated using an entropy measure of dependence called the mutual information function (Fraser and Swinney, PR:A 1986).

Estimating the minimum embedding dimension; . The Method of Delays Estimating the minimum embedding dimension; . Kennel and Brown (PR:A 1992) developed the False Nearest Neighbors. The False Nearest Neighbors technique uses Euclidean distances to determine if the vectors of Y are still “close” as the dimension of the phase space is increased.

The Method of Delays Estimating the minimum embedding dimension; . 

Simulation The simulation follows that in Chon et al (1997). The Ikeda Map is numerically generated and the deterministic variable Xt is isolated. Ikeda Map

Simulation Six variables are constructed by adding contamination, i,t to the deterministic component, Xt, post numerical iteration. Z1 is entirely deterministic and Z6 is entirely random. The simulation follows that in Chon et al (1997), except that we add noise post numerical iteration of the map and focus on the structural analysis of dependence whereas Chon et al (1997) included a noise term within the numerical iteration of the Ikeda map and focused on the detection of chaos. This allows us to separate the two components more distinctly, in particular e is independent of X.

Simulation: Zi(t) Traditional tests of dependence confirm the presence of a process in Z1-Z4.

Simulation Which one is deterministically generated, Z1, and which one is randomly generated, Z6?

Simulation: Z1(t) Z1(t) = Xn + n,1 n,1~N(0,0) The structure of the process is well defined and appears to follow a stable trajectory.

Simulation: Z3(t) n,3~N(0,0.125) Z3(t) = Xn + n,3 The contamination has a variance roughly equal to 25% of the deterministic process. While the general direction of the trajectory paths remain; the tightness, or clarity, of the paths has greatly diminished.

Simulation: Z5(t) n,5~N(0,0.5) Z5(t) = Xn + n,5 The phase portrait is close to that of Gaussian randomly generated data. Traditional methods of determining independence can no longer detect suspect behavior. The volatility of the contamination has overcome the deterministic component.

Simulation: Z6(t) n,6~N(0,0.5) Z6(t) = n,6 The “ball” of a normal distribution is the benchmark for comparison and detection of nonlinear processes in regression residuals.

Simulation Which one is deterministically generated, Z1, and which one is randomly generated, Z6? Z6, Randomly Generated Z1, Deterministically Generated

ARMA Fitted Z1(t) Residuals such as Z1 will only appear in the presence of a nonlinear process and cannot be controlled for using linear techniques.

Fitted Residual Reconstruction

Insights from Simulations If the first four time series of the simulation happened to be the residuals of an econometric model, model misspecification may have been concluded. Having that qualitative representation of the residual structure in the phase space reconstruction may enhance additional modeling. After the variance of the additive contamination increased above 25% of the variance of the deterministic component only general structure could be inferred. After the variance of the contamination increased passed 50% of the deterministic component the ability to distinguish structure from noise decreased rapidly.

Application: The S&P 500 Following Granger, Maasoumi, and Racine (JTSA 2004), regression residuals from a nonlinear time series model of S&P 500 from Jan. 3, 1996 to Dec. 22, 1997 are analyzed A simple ARIMA model is fit first: After inspection, however, the linear model does not fully capture all variability.

Application: Reconstructed Residuals

Application: The S&P 500 An Integrated Auto Regressive Model with ARCH and GARCH structured residuals is fit to account for the discovered dependence.

Application: Reconstructed Residuals

Application: Normal Generated Data

Final Observations Given the similarity to random noise, we can find no strong visual evidence of nonlinear structure. In cases when it is not sufficient to just detect dependence, phase space reconstruction can add valuable insight on the nature of the dependence. Indeed, phase space reconstruction can be thought of as a nonlinear scatter plot. The phase portrait may be used as a foundation for future model enhancement. This research differs from using phase space reconstruction in the detection of chaos or providing evidence of dependence using entropy metrics, by illuminating structure.

Questions?

Table 3: The average mutual information function for S&P 500 residual series.

Table 4: The percentage of false nearest neighbors for the S&P 500 residual series.