Time Series Analysis and Classification Tarek Medkour January 31, 2024 Part 1: Time Series Analysis Introduction Time Series Models Spectral Analysis Time-Frequency Representation Multivariate Time Series Part 2: Time Series Classification Pattern Recognition and Detection Feature Extraction and Selection Models and Representation Learning Data Enhancement and Preprocessings Change-Point and Anomaly Detection Introduction and Examples 0.4 ● ● ● 0.2 ● 0.0 ● ● ● ● ● ● ●●●● ● ● ● ●● −0.2 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ●● ● ● ● ● ● ● ● ● 1860 1880 ● ●● ● ● ●● ● ●● ● ● ●●● ● 1900 ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●●● −0.4 Global temperature deviations ● ● ●● ● ● ● 1920 Year 1940 1960 1980 2000 Introduction and Examples 1700 ● ●● ● ● 1600 ● ● ● ● ● ● ● ● ● 1500 ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● 1300 Amount of milk produced ●● 1400 ● ●● ● ● ● ● ● ● ● ● 1994 1996 1998 2000 Year 2002 2004 2006 Introduction and Examples ● ● ● ● 200 190 180 170 CREF stock values 210 220 ●● ● ●● ● ●● ● ● ●● ● ●●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●● ●●●● ● ● ● ● ● ● ●● ● ● ● ●●●● ● ●● ● ● ● ● ● ● ●● ● ● ●●●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ●●● ● ● ●● ● ● ● ●● ●● ● ● ● ●● ●● ● ●● ●● ● ●● ● ●● ●● ●● ●● ● ●● ●● ●● ● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ●●● ● ● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ●● ●●● ●● ●● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ●● ●● ● ● ● ●●● ● ● ●● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ●●● ● ● ● ●● ● ● ●● ● ● ●● ● ●● ● ● ●● ● ● ●● ●● ● ● ● ●● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●●● ●● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● 0 100 200 300 Time 400 500 Introduction and Examples ● ● 200 ● ● ● ● ● ● ● 150 ● ● ● ● ● ● ●● ● 100 ●● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● 50 Number of Home Runs ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●●● 1920 1940 1960 Time 1980 ● ● ● ● ● ● ● ●● ● 2000 ● ● Introduction and Examples 40 ● ● ● ● ● ● ● 30 ● ● ● ● ● ● ● 25 ● ● ● ● ● ● 20 ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 15 ● ● ● ● ● ●● 10 ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● 5 Number of earthquakes (7.0 or greater) 35 ● 1900 1920 1940 1960 Time 1980 2000 Introduction and Examples ● ● ● ●● 25000 ● ● ● ●● ● ● ● ● ● ● ●● ●●● ●● ● ● ● ● ● 20000 ● ● 15000 ● ● ● 10000 ● ● ● ● ● ● 5000 USC Columbia fall enrollment ● ●●● ● ● ● ● ●● ● ●●●● 1900 1910 1920 1930 Time 1940 1950 ●●● 30 35 Introduction and Examples ● ● ● ●● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● 25 20 Star brightness 15 10 5 ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ●● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ●● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ●● ● ●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ●● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ●● ● ● ●● ●● ● ●● ● ●● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● 0 ● ● ●● ● ● 0 100 200 300 Time 400 500 600 Introduction and Examples Introduction and Examples 1500 ● ● ● ● ●● ● ● ● ●●● ● ● ●● 1450 ● ● 1350 1400 ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● 1300 ● ● ●● ● ● 1250 SP500 Index ● 1900 ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ●●●●● ● ●● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ●● ● ● ● ●● ● ● ●● ● ● ●● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● 1950 ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ● 2000 2050 Year 2100 2150 Introduction and Examples ● ●● 60 ● ●● ● ● ● ●● ● ● ● ● 40 ●● ● ● ● ● ● ● ● ●● ● ●● ● ●● ●● ●● ● ● ● ● ● ● ● ●● ●● ●● ● ●● ●●● ●● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● 30 ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ●● ● ● ● ●●● ●● ● ● ●● ● ● ● ●●●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ●● ● ●●● ● ●● ●● ● ●● ●● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ●● ● ●● ● ● ● ●● ● ● ● ● ● ● ● 20 ● 10 Oil prices 50 ● ● ● 1990 1995 Year 2000 2005 3000000 Introduction and Examples ● ● ● ● ● 2500000 ● ● ● 2000000 ● ● ● ● 1500000 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 1000000 Algerian Wheat Production ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● 1960 ● 1970 1980 1990 Date ● ● ● 2000 2010 Introduction and Examples Terminology A time series is a sequence of ordered data. The ordering refers generally to time, but other orderings could be envisioned (e.g., over space, etc.). In this class, we will be concerned exclusively with time series that are measured on a single continuous random variable Y equally spaced in discrete time; that is, we will have a single realization of Y at each second, hour, day, month, year, etc. Introduction and Examples Ubiquity Time series data arise in a variety of fields. Here are just a few examples. In business, we observe daily stock prices, weekly interest rates, quarterly sales, monthly supply figures, annual earnings, etc. In agriculture, we observe annual yields (e.g., crop production), daily crop prices, annual herd sizes, etc. In engineering, we observe electric signals, voltage measurements, etc. In natural sciences, we observe chemical yields, turbulence in ocean waves, earth tectonic plate positions, etc. Introduction and Examples In medicine, we observe ECG or EEG measurements on patients, drug concentrations, blood pressure readings, etc. In epidemiology, we observe the number of flu cases per day, the number of health-care clinic visits per week, annual tuberculosis counts, etc. In meteorology, we observe daily high temperatures, annual rainfall, hourly wind speeds, earthquake frequency, etc. In social sciences, we observe annual birth and death rates, accident frequencies, crime rates, school enrollments, etc. Introduction and Examples Justification The purpose of time series analysis is twofold: 1 to model the stochastic (random) mechanism that gives rise to the series of data 2 to predict (forecast) the future values of the series based on the previous history Introduction and Examples Notes The analysis of time series data calls for a new way of thinking when compared to other statistical methods courses. Essentially, we get to see only a single measurement from a population (at time t) instead of a sample of measurements at a fixed point in time (cross-sectional data). Introduction and Examples The special feature of time series data is that they are not independent! Instead, observations are correlated through time. Correlated data are generally more difficult to analyze. Statistical theory in the absence of independence becomes markedly more difficult. Most classical statistical methods (e.g., regression, analysis of variance, etc.) assume that observations are statistically independent. Introduction and Examples For example, in the simple linear regression model Yi = β0 + β1 xi + ϵi , we typically assume that the ϵ error terms are independent and identically distributed (iid) normal random variables with mean 0 and constant variance. Introduction and Examples There can be additional trends or seasonal variation patterns (seasonality) that may be difficult to identify and model. The data may be highly non-normal in appearance and be possibly contaminated by outliers Introduction and Examples Modeling I Our goal in this part of the course is to build (and use) uni-variate time series models for data. This breaks down into different parts. 1 Model specification (identification) Consider different classes of time series models for stationary processes. Use descriptive statistics, graphical displays, subject matter knowledge, etc. to make sensible candidate selections. Abide by the Principle of Parsimony. 2 Model fitting Introduction and Examples Modeling II Once a candidate model is chosen, estimate the parameters in the model. We will use least squares and/or maximum likelihood to do this. 3 Model diagnostics Use statistical inference and graphical displays to check how well the model fits the data. This part of the analysis may suggest the candidate model is inadequate and may point to more appropriate models. Time series and stochastic processes Terminology The sequence of random variables {Yt : t = 0, 1, 2, . . . }, or more simply denoted by {Yt }, is called a stochastic process. It is a collection of random variables indexed by time t; that is, Y0 = value of the process at time t = 0 Y1 = value of the process at time t = 1 Y2 = value of the process at time t = 2 .. . Yn = value of the process at time t = n. The subscripts are important because they refer to which time period the value of Y is being measured. Time series and stochastic processes A stochastic process can be described as a statistical phenomenon that evolves through time according to a set of probabilistic laws. A complete probabilistic time series model for {Yt } would specify all of the joint distributions of random vectors Y = {Y1 , Y2 , . . . , Yn }, for all n = 1, 2, . . . , : P(Y1 ≤ y1 , Y2 ≤ y2 , . . . , Yn ≤ yn ), for all y = (y1 , y2 , . . . , yn ) and n = 1, 2, . . . . Time series and stochastic processes This specification is not generally needed in practice. In this course, we specify only the first and second-order moments; i.e., expectations of the form E(Yt ) and E(Yt Yt−k ), for k = 0, 1, 2, . . . , and t = 0, 1, 2, . . . . Much of the important information in most time series processes is captured in these first and second moments (or, equivalently, in the means, variances, and covariances). Means, Variances, and Covariances Terminology For the stochastic process {Yt : t = 0, 1, 2, . . . ,}, the mean function is defined as µt = E(Yt ), for t = 0, 1, 2, . . . ,. That is, µt is the theoretical (or population) mean for the series at time t. Means, Variances, and Covariances The autocovariance function is defined as λt,s = cov(Yt , Ys ) = E(Yt Ys ) − E(Yt )E(Ys ), for t, s = 0, 1, 2, . . .. The autocorrelation function is given by p p ρt,s = corr(Yt , Ys ) = cov(Yt , Ys )/ var(Yt )var(Ys ) = λt,s / λt,t λs,s . Means, Variances, and Covariances Values of ρt,s near ± 1 ⇒ strong linear dependence between Yt and Ys . Values of ρt,s near 0 ⇒ weak linear dependence between Yt and Ys . Values of ρt,s = 0 ⇒ Yt and Ys are uncorrelated. Some examples of Stochastic Processes White Noise Example 1 A stochastic process {et : t = 0, 1, 2, . . . ,} is called a white noise process if it is a sequence of independent and identically distributed (iid) random variables with E(et ) = µe , var(et ) = σe2 . Some examples of Stochastic Processes 3 White Noise ● ● ● ● ● 2 ● ● ● ● ● ● ● 1 ● ●● ● ● ● ● ● ● ● ● −1 ● ● ●● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0 ● ● ● ● ● ● ● ● ● ● ● ● ● −2 Simulated white noise process ● ● ●● ● 0 50 100 Time ● 150 Some examples of Stochastic Processes White Noise Both µe and σe2 are constant (free of t). It is often assumed that µe = 0; that is, {et } is a zero mean process. A slightly less restrictive definition would require that the et are uncorrelated (not independent). However, under normality; i.e., et˜ iid N(0, σe2 ), this distinction becomes irrelevant (for linear time series models). Some examples of Stochastic Processes White Noise Auto-covariance Function For t = s, cov(et , es ) = cov(et , et ) = var(et ) = σe2 . For t ̸= s, cov(et , es ) = 0, because the et ’s are independent. Thus, the autocovariance function of et is ( σe2 , for |t − s| = 0 λt,s = 0, for |t − s| = ̸ 0 Some examples of Stochastic Processes White Noise Autocorrelation Function For t = s, ρt,s = corr(et , es ) = corr(et , et ) = λt,t / p λt,t λt,t = 1. for t ̸= s, p ρt,s = corr(et , es ) = λt,s / λt,t λs,s = 0. Thus, the autocorrelation function is ( 1, for |t − s| = 0 ρt,s = 0, for |t − s| = ̸ 0 Some examples of Stochastic Processes White Noise Remark A white noise process, by itself, is rather uninteresting for modeling real data. However, white noise processes still play a crucial role in the analysis of time series data! Some examples of Stochastic Processes White Noise Time series processes Yt generally contain two different types of variation: systematic variation (that we would like to capture and model; e.g., trends, seasonal components, etc.) random variation (that is just inherent background noise in the process). Some examples of Stochastic Processes White Noise Our goal as data analysts is to extract the systematic part of the variation in the data (and incorporate this into our model). If we do an adequate job of extracting the systematic part, then the only part left over should be random variation, which can be modeled as white noise. Some examples of Stochastic Processes Random Walk Example 2 Suppose that et is a zero mean white noise process with var(et ) = σe2 . Define Y1 = e1 Y2 = e1 + e2 ··· Yn = e1 + e2 + . . . + en . Some examples of Stochastic Processes Random Walk By this definition, note that we can write, for t > 1, Yt = Yt−1 + et , where E(et ) = 0 and var(et ) = σe2 . The process Yt is called a random walk process. Random walk processes are used to model stock prices, movements of molecules in gases and liquids, animal locations, etc. Some examples of Stochastic Processes Random Walk ● ●● ● 0 ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ●●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ●● ● ● ● ● ● ●● ● ● ●●● ●●● ● ●●●● ● ● ● ● ●●● ● ● ●● ●● −5 ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ●●●● ● ● ● ● −10 Simulated random walk process 5 ● ● ●● ● ● ● ● ● ● ● ● 0 50 100 Time 150 Some examples of Stochastic Processes Random Walk Mean Function The mean of Yt is µt = E(Yt ) = E(e1 + e2 + . . . + et ) = E(e1 ) + E(e2 ) + . . . + E(et ) = 0. That is, Yt is a zero mean process. Some examples of Stochastic Processes Random Walk Variance Function The variance of Yt is var(Yt ) = var(e1 + e2 + . . . + et ) = var(e1 ) + var(e2 ) + . . . + var(et ) = tσe2 , because var(e1 ) = var(e2 ) = . . .= var(et ) =σe2 , and cov(et , es ) = 0 for all t ̸= s. Some examples of Stochastic Processes Random Walk Autocovariance Function For t ≤ s, the autocovariance of Yt and Ys is λt,s = cov(Yt , Ys ) = cov(e1 + . . . + et , e1 + . . . + et + et+1 + . . . + es ) = cov(e1 + e2 + . . . + et , e1 + e2 + . . . + et ) +̇cov(e1 + e2 + . . . + et , et+1 + . . . + es ) Xt XX cov(ei , ej ) = cov(ei , ei ) + i=1 1≤i̸=j≤t Xt = var(ei ) = σe2 + σe2 + . . . + σe2 = σe2 . i=1 Because λt,s = λs,t , the autocovariance function for a random walk process is λt,s = tσe2 , for 1 ≤ t ≤ s. Some examples of Stochastic Processes Random Walk Autocorrelation Function For 1 ≤ t ≤ s, the autocorrelation function for a random walk process is r λt,s t tσe2 . = 2 2 = ρt,s = corr(Yt , Ys ) = p tσe sσe s λt,t λs, s Note that when t is closer to s, the autocorrelation ρt,s is closer to 1. That is, two observations Yt and Ys close together in time are likely to be close together, especially when t and s are both large (later on in the series). On the other hand, when t is far away from s (that is, for two points Yt and Ys far apart in time), the autocorrelation is closer to 0. Some examples of Stochastic Processes Random Walk Example 3 Suppose that et is a zero mean white noise process with var(et ) = σe2 . Define 1 Yt = (et + et−1 + et−2 ), 3 that is, Yt is a running (or moving) average of the white noise process (averaged across the most recent 3 time periods). Some examples of Stochastic Processes Random Walk 1.5 ● ● ● 1.0 ● ● ● ● 0.5 ● ● ● ● ● ● ● ●● ● ● ● ● ● 0.0 ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● −0.5 ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● 0 ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● −1.0 Simulated moving average process ● ● 50 100 Time 150 Some examples of Stochastic Processes Random Walk Mean Function The mean of Yt is 1 µt = E(Yt ) = E[ (et + et−1 + et−2 )] 3 1 = [E(et ) + E(et−1 ) + E(et−2 )] = 0, 3 because et is a zero-mean process. Yt is a zero mean process. Some examples of Stochastic Processes Random Walk Variance Function The variance of Yt is 1 var(Yt ) = var[ (et + et−1 + et−2 )] 3 1 = var(et + et−1 + et−2 ) 9 1 = [var(et ) + var(et−1 ) + var(et−2 )] 9 1 1 = (3var(et )) = σe2 , 9 3 because var(et ) = σe2 for all t and because et , et−1 , and et−2 are independent (all covariance terms are zero). Some examples of Stochastic Processes Random Walk Autocovariance Funcion We need to consider different cases. Case 1: If s = t, then 1 λt,s = λt,t = cov(Yt , Yt ) = var(Yt ) = σe2 . 3 Case 2: If s = t + 1, then λt,s = λt,t+1 = cov(Yt , Yt+1 ) 1 1 = cov[ (et + et−1 + et−2 ), (et+1 + et + et−1 )] 3 3 1 = [cov(et , et ) + cov(et−1 , et−1 )] 9 1 2 = [var(et ) + var(et−1 )] = σe2 . 9 9 Some examples of Stochastic Processes Random Walk Case 3: If s = t + 2, then λt,s = λt,t+2 = cov(Yt , Yt+2 ) 1 1 = cov[ (et + et−1 + et−2 ), (et+2 + et+1 + et )] 3 3 1 = cov(et , et ) 9 1 1 = var(et ) = σe2 . 9 9 Case 4: If s > t + 2, then λt,s = 0 because Yt and Ys will have no common white noise error terms. Because λt,s = λs,t , the autocovariance function can be written as 2 σe /3, |t − s| = 0 2σ 2 /9, |t − s| = 1 e λt,s = σe2 /9, |t − s| = 2 0, |t − s| > 2. Some examples of Stochastic Processes Random Walk Autocorrelation Function Recall that the autocorrelation function is ρt,s = corr(Yt , Ys ) = λt,s . λt,t λs,s Because λt,t = λs,s = σe2 /3, the autocorrelation function for this process is 1, |t − s| = 0 2/3, |t − s| = 1 ρt,s = 1/3, |t − s| = 2 0, |t − s| > 2. Some examples of Stochastic Processes Random Walk Notes Observations Yt and Ys that are 1 unit apart in time have the same autocorrelation regardless of the values of t and s. Observations Yt and Ys that are 2 units apart in time have the same autocorrelation regardless of the values of t and s. Observations Yt and Ys that are more than 2 units apart in time are uncorrelated Some examples of Stochastic Processes Autoregression Example 4 Suppose that et is a zero mean white noise process with var(et ) = σe2 . Consider the stochastic process defined by Yt = 0.75Yt−1 + et , that is, Yt is directly related to the (downweighted) previous value of the process Yt−1 and the random error et (a shock or innovation that occurs at time t). This is called an autoregressive model. Autoregression means regression on itself. Essentially, we can envision regressing Yt on Yt−1 . Note: We will postpone mean, variance, autocovariance, and autocorrelation calculations for this process until later when we discuss autoregressive models in more detail. Some examples of Stochastic Processes Autoregression Some examples of Stochastic Processes Seasonal Example 5 Many time series exhibit seasonal patterns that correspond to different weeks, months, years, etc. One way to describe seasonal patterns is to use models with deterministic parts which are trigonometric in nature. Suppose that et is a zero mean white noise process with var(et ) = σe2 . Consider the process defined by Yt = a sin(2πωt + ϕ) + et . Some examples of Stochastic Processes Seasonal In this model,a is the amplitude, ω is the frequency of oscillation, and ϕ controls the phase shift. With a = 2, ω = 1/52 (one cycle/52 time points), and ϕ = 0.6π, note that E(Yt ) = 2 sin(2πt/52 + 0.6π), since E(et ) = 0. Also, var(Yt ) = var(et ) = σe2 . Some examples of Stochastic Processes Seasonal 2 ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●●● ●● ●●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ●● ● ●● ●● ● ● ● ● ● ●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●●●● ● ●●● ● ● ●● ●●● ● ●●●● ●● ● ● ● ●● ●● ● ● ●● ● ● ● ●● ● ●●● ●● ● ● ● ●● ● ● ● −2 −2 −1 0 0 1 2 ● ● ● ●● ● ●●● ● ●● ● ● 0 50 50 100 150 ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ●●● ● ●● ● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● 50 100 Time 150 ● 5 ● 0 ● −5 4 ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ● ●●● ● ● ●● ● ● ● ●● ●● ● ●● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●●●● ● ● ●●● ●● ● ● ● ● ●●● ● ● ● ●● ● ●● ●● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● 2 0 −2 0 Time ● −4 150 Time ● ●● ● ●● ●● ● ● ● ● 0 100 ● ●● ● ●● ● ● ● ●● 0 ● 50 100 Time 150