Lecture 18. Time Series Nan Ye School of Mathematics and Physics University of Queensland 1 / 29
● ● 15 ● ● Quarterly Earnings per Share ● ● ● ● ● ● 10 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 5 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0 1960 1965 1970 1975 1980 Time Johnson & Johnson quarterly earnings per share (1960-I to 1984-IV) 2 / 29
0.6 ● ● ● ● ● ● ● Global Temperature Deviations ● ● 0.4 ● ● ● ● ● ● ● ● ● ● ● ● 0.2 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0.0 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● −0.2 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● −0.4 ● ● ● ● 1880 1900 1920 1940 1960 1980 2000 Time Yearly average global temperature (1880-2009) 3 / 29
4000 3000 speech 2000 1000 0 0 200 400 600 800 1000 Time Speech recording of the syllable aaa...hhh 4 / 29
0.05 0.00 NYSE Returns −0.05 −0.15 0 500 1000 1500 2000 Time Returns from NYSE from 2 Feb 1984 to 31 Dec 1991. 5 / 29
This Lecture • Nature of time series data • Time series modelling • Time series decomposition • Stationarity • Time domain models 6 / 29
Nature of Time Series Data Characteristics • A time series is often viewed as a collection of random variables { X t } indexed by time. • In a time series, measurements at adjacent time points are usually correlated. • As compared to other types of correlated data, such as clustered or longitudinal data, observations in a time series may explicitly depend on previous observations and/or errors. 7 / 29
Probabilistic description • We can describe a time series using the distribution of the random variables { X t } . • Frequently, we look at some summary statistics Mean function 𝜈 X ( t ) = E ( X t ) Autocovariance function 𝛿 X ( s , t ) = cov( X s , X t ) 𝛿 X ( s , t ) Autocorrelation function (ACF) 𝜍 X ( s , t ) = . √︁ 𝛿 X ( s , s ) 𝛿 X ( t , t ) • We often drop X from 𝜈 X , 𝛿 X and 𝜍 X when there is no ambiguity. 8 / 29
Time Series Modelling Chasing after stationarity • The objective of time series modelling is to develop compact representation of the time series, to facilitate tasks including interpretation, prediction, control, hypothesis testing and simulation. • Some form of time invariance is required to find regularity in data and extrapolate into future. • Stationarity is a basic form of time invariance, and much of time series modelling is about transforming times series so that the transformed time series is stationary. 9 / 29
Exploratory data analysis • Plotting the time series should be the first step before any formal modelling attempt. • This is useful for identifying important features for choosing an appropriate model. • For example, use the plots to look for the trends, presence of seasonal components or outliers. 10 / 29
Modelling paradigms • There are two main modelling paradigms. • The time domain approach views a time series as the description of the evolution of an entity, and focuses on capturing the dependence of current value on history. • The frequency domain approach views a time series as the superposition (addition) of periodic variations. 11 / 29
Time Series Decomposition An additive decomposition • A classical decomposition of a time series { X t } is X t = T t + S t + E t , where T t is the trend component, S t is the seasonal component (recurring variation with fixed period), E t is the error component. • The trend component and seasonal component can be deterministic or stochastic. • Sometimes, a cyclical component (recurring variation with non-fixed period) is included in the systematic component. 12 / 29
A multiplicative decomposition • A common multiplicative decomposition is X t = T t S t E t . • This is equivalent to first converting X t to the log scale and then perform an additive decomposition ln X t = ln T t + ln S t + ln E t 13 / 29
Stationarity Strict stationarity • A time series { X t } is strictly stationary if its probabilistic behavior is invariant to time shift. • To be precise, for any k , for any time points t 1 , . . . , t k , for any x 1 , . . . , x k , and for any 𝜀 , we have P ( X t 1 ≤ x 1 , . . . , X t k ≤ x k ) = P ( X t 1 + δ ≤ x 1 , . . . , X t k + δ ≤ x k ) 14 / 29
• Strict stationarity implies that the mean function 𝜈 ( t ) = E ( X t ) and the autocovariance function 𝛿 ( t , t + h ) = cov( X t , X t + h ) do not depend on t . • Strict stationarity is often too much to ask for, and usually not necessary for learning a model. 15 / 29
Stationarity • A time series { X t } is said to be (weakly) stationary if 𝜈 ( t ) and 𝛿 ( t , t + h ) do not depend on t . • The autocovariance and autocorrelation functions of a stationary time series can be more compactly written as 𝛿 ( h ) = 𝛿 ( t , t + h ) , 𝜍 ( h ) = 𝜍 ( t , t + h ) = 𝛿 ( h ) /𝛿 (0) . 16 / 29
• Randomness in a stationary time series is sufficiently constrained for finding out regularity in data. • A stationary time series has a trivial system component (constant mean). • Stationary time series are used as an important building block for the random component of more sophisticated models. 17 / 29
Time Domain Models White noise processes • A white noise process { 𝜗 t } is a collection of uncorrelated random variables with mean 0 and finite variance 𝜏 2 . • This is often denoted as 𝜗 t ∼ WN(0 , 𝜏 2 ). • The mean, autocovariance and autocorrelation functions are 𝜈 ( t ) = 0 {︄ 𝜏 2 , h = 0 , 𝛿 ( t , t + h ) = cov( 𝜗 t , 𝜗 t + h ) = 0 , h ̸ = 0 . {︄ 1 , h = 0 , 𝜍 ( t , t + h ) = 0 , h ̸ = 0 . • White noise processes are thus stationary, and they serve as an important building block for more sophisticated time series models. 18 / 29
WN(0, 1) 2 1 x 0 −1 −2 0 50 100 150 200 Time An example white noise series. 19 / 29
Random Walk • Consider the random walk X t = ∑︁ t i =1 𝜗 i , where 𝜗 t ∼ WN(0 , 𝜏 2 ). • The mean, autocovariance, and autocorrelation functions are 𝜈 ( t ) = 0 , 𝛿 ( t , t + h ) = t 𝜏 2 , t 𝜍 ( t , t + h ) = . √︁ t ( t + h ) • { X t } is not stationary. 20 / 29
Random walk 10 0 x −10 −20 0 50 100 150 200 Time Three random walk series from the same model. 21 / 29
Moving average process • { X t } is a moving average process of order 1, or MA(1), if X t = 𝜗 t + 𝜄𝜗 t − 1 , where 𝜗 t ∼ WN(0 , 𝜏 2 ). • The mean, autocovariance, and autocorrelation functions are 𝜈 ( t ) = 0 , ⎧ 𝜏 2 (1 + 𝜄 2 ) , h = 0 , ⎪ ⎨ 𝛿 ( t , t + h ) = 𝜏 2 𝜄, , h = ± 1 , ⎪ 0 , otherwise , ⎩ ⎧ 1 , h = 0 , ⎪ ⎨ 𝜍 ( t , t + h ) = 𝜄/ (1 + 𝜄 2 ) , . h = ± 1 , ⎪ 0 , otherwise , ⎩ • MA(1) is stationary. 22 / 29
MA ( 1 ) θ = 0.9 4 3 2 1 x 0 −1 −2 −3 0 50 100 150 200 Time Two MA(1) series from the same model. 23 / 29
Autoregressive process • { X t } is an autoregressive process of order 1, or AR(1), if X t = 𝜒 X t − 1 + 𝜗 t , where 𝜗 t ∼ WN(0 , 𝜏 2 ). • When AR(1) is stationary, the mean, autocovariance and autocorrelation functions are 𝜈 ( t ) = 0 , 𝛿 ( t , t + h ) = 𝜒 | h | 𝜏 2 1 − 𝜒 2 , 𝜍 ( t , t + h ) = 𝜒 | h | . 24 / 29
AR ( 1 ) φ = 0.9 10 5 x 0 −5 0 50 100 150 200 Time Two AR(1) series from the same model. 25 / 29
Linear processes • A linear process { X t } is a linear combination white noise variates 𝜗 t , that is, + ∞ ∑︂ X t = 𝜈 + 𝜔 i 𝜗 t − i , i = −∞ where 𝜗 t ∼ WN(0 , 𝜏 2 ). • The mean and covariance functions are 𝜈 ( t ) = 𝜈, ∞ ∑︂ 𝛿 ( t , t + h ) = 𝜏 2 𝜔 i 𝜔 i + h . i = −∞ 26 / 29
{︄ 1 , i = 0 , • White noise is a linear process with 𝜈 = 0 and 𝜔 i = i ̸ = 0 . . 0 , ⎧ 1 , i = 0 , ⎪ ⎨ • MA(1) is a linear process with 𝜈 = 0, and 𝜔 i = 𝜄, i = 1 , . ⎪ 0 , otherwise . ⎩ {︄ 𝜒 i , i ≥ 0 , • AR(1) is a linear process with 𝜈 = 0, and 𝜔 i = otherwise . . 0 , 27 / 29
ARMA • { X t } is ARMA( p , q ) if it is stationary and X t = 𝜒 1 X t − 1 + . . . + 𝜒 p X t − p + 𝜗 t + 𝜄 1 𝜗 t − 1 + 𝜄 q 𝜗 t − q , where 𝜒 p ̸ = 0, 𝜄 q ̸ = 0 and 𝜗 t ∼ WN(0 , 𝜏 2 ). • p and q are called the autoregressive and the moving average orders respectively. • AR(1) = ARMA(1 , 0), and MA(1) = ARMA(0 , 1). 28 / 29
Recommend
More recommend