22 - Introduction to Time Series Analysis. Lecture 22. 1....

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: Introduction to Time Series Analysis. Lecture 22. 1. Review: The smoothed periodogram. 2. Examples. 3. Parametric spectral density estimation. 1 Review: Periodogram The periodogram is defined as I() = |X()| 1 Xc () = n 1 Xs () = n 2 2 2 = Xc () + Xs (). n cos(2t)xt , t=1 n sin(2tj )xt . t=1 Under general conditions, Xc (j ), Xs (j ) are asymptotically independent and N (0, f (j )/2). Thus, EI(^(n) ) f (), but Var(I(^(n) )) f ()2 . 2 Review: Smoothed spectral estimators ^ f () = |j|Ln Wn (j)I(^(n) - j/n), where the spectral window function satisfies Ln , Ln /n 0, 2 Wn (j) 0, Wn (j) = Wn (-j), Wn (j) = 1, and Wn (j) 0. ^ Then f () f () (in the mean square sense), and asymptotically 2 ^ f (k ) f (k ) d , d where d = 2/ 2 Wn (j). 3 Introduction to Time Series Analysis. Lecture 22. 1. Review: The smoothed periodogram. 2. Examples. 3. Parametric spectral density estimation. 4 Example: Southern Oscillation Index Figure 4.4 in the text shows the periodogram of the SOI time series. The SOI is the scaled, standardized, mean-adjusted, difference between monthly average air pressure at sea level in Tahiti and Darwin: SOI = 10 PTahiti - PDarwin - x. For the time series in the text, n = 453 months. The periodogram has a large peak at = 0.084 cycles/sample. This corresponds to 0.084 cycles per month, or a period of 1/0.084 = 11.9 months. There are smaller peaks at 0.02: I(0.02) 1.0. The frequency = 0.02 corresponds to a period of 50 months, or 4.2 years. 5 Example: Southern Oscillation Index Consider the hypothesized El Ni~ o effect, at a period of around four years. n The approximate 95% confidence interval at this frequency, = 1/(4 12), is 2I() 2I() f () 2 2 (0.025) 2 (0.975) 2 2 0.64 2 0.64 f () 7.3778 0.0506 0.17 f () 25.5. The lower extreme of this confidence interval is around the noise baseline, so it is difficult to conclude much about the hypothesized El Ni~ o effect. n 6 Example: Southern Oscillation Index Figure 4.5 in the text shows the smoothed periodogram, with L = 9. Again, there is a large peak at = 0.08 cycles/month (period 12 months). There are smaller peaks at integer multiples of this frequency (harmonics). ^ There is also a peak at = 0.0215 (period 46.5 months): f (0.0215) 0.62. 7 Example: Southern Oscillation Index The approximate 95% confidence interval at the hypothesized El Ni~ o n frequency is ^ ^ 2Lf () 2Lf () 2 (0.025) f () 2 (0.975) 2L 2L 18 0.62 18 0.62 f () 31.526 8.231 0.354 f () 1.36. The lower extreme of this confidence interval is well above the noise baseline (the level of the spectral density if the signal were white and the energy were uniformly spread across frequencies). The text modifies the number of degrees of freedom slightly, to account for the fact that the signal is padded with zeros to make n a highly composite number, which simplifies the computation of the periodogram. 8 Choosing the bandwidth A common approach is to start with a large bandwidth, and look at the effect on the spectral estimates as it is reduced (`closing the window'). As the bandwidth becomes too small, the variance gets large and the spectral estimate becomes more jagged, with spurious peaks introduced. But if it is too small, the spectral estimate is excessively smoothed, and details of the shape of the spectrum are lost. The value of L = 9 chosen in the text for Figure 4.5 corresponds to a bandwidth of B = L/n = 9/480 = 0.01875 cycles per month. This means we are averaging over frequencies in a band of this width, so we are treating the spectral density as approximately constant over this bandwidth. Equivalently, we are not hoping to resolve frequencies more finely than about half of this bandwidth. 9 Simultaneous confidence intervals We derived the confidence intervals for f () assuming that was fixed. But in examining peaks, we might wish to choose after we've seen the data. If we want to make statements about the probability that k unlikely events E1 , . . . , Ek occur, we can use the Bonferroni inequality (also called the union bound): k k Pr i=1 Ei i=1 Pr{Ei }, and this probability is no more than k if Pr{Ei } = . For example, if Ei represents the event that f (i ) falls outside some confidence interval at level , then we can bound the probability that the spectral density is far from our estimates at any of the frequencies 1 , . . . , k . 10 Introduction to Time Series Analysis. Lecture 22. 1. Review: The smoothed periodogram. 2. Examples. 3. Parametric spectral density estimation. 11 Parametric versus nonparametric estimation Parametric estimation = estimate a model that is specified by a fixed number of parameters. Nonparametric estimation = estimate a model that is specified by a number of parameters that can grow as the sample grows. Thus, the smoothed periodogram estimates we have considered are nonparametric: the estimates of the spectral density can be parameterized by estimated values at each of the Fourier frequencies. As the sample size grows, the number of distinct frequency values increases. The time domain models we considered (linear processes) are parametric. For example, and ARMA(p,q) process can be completely specified with p + q + 1 parameters. 12 Parametric spectral estimation In parametric spectral estimation, we consider the class of spectral densities corresponding to ARMA models. Recall that, for a linear process Yt = (B)Wt , fy () = e2i 2 2 w . For an AR model, (B) = 1/(B), so {Yt } has the rational spectrum fy () = = 2 w | (e-2i )| 2 p p j=1 2 2 w |e-2i - pj | 2, where pj are the poles, or roots of the polynomial . 13 Parametric spectral estimation The typical approach to parametric spectral estimation is to use the ^ ^ ^2 maximum likelihood parameter estimates (1 , . . . , p , w ) for the parameters of an AR(p) model for the process, and then compute the spectral density for this estimated AR model: ^ fy () = w ^2 ^ (e-2i ) 2. 14 Parametric spectral estimation For large n, ^()) 2p f 2 (). Var(f n (There are results for the asymptotic distribution, but they are rather weak.) Notice the bias-variance trade-off in the parametric case: as we increase the number of parameters, p: The bias decreases; we can model more complex spectra. For example, with an AR(p), we cannot have more than p/2 spectral peaks in the interval (0, 1). (This is because each pair of complex conjugate poles contributes one factor and hence peak to the product.) The variance increases linearly with p. 15 ARMA spectral estimation Sometimes ARMA models are used instead: estimate the parameters of an ARMA(p,q) model and compute its spectral density (recall that (B) = (B)/(B)): ^ f () = w ^2 ^ (e-2i ) ^ (e-2i ) 2 . However, it is more common to use large AR models, rather than ARMA models. 16 Parametric versus nonparametric spectral estimation The main advantage of parametric spectral estimation over nonparametric is that it often gives better frequency resolution of a small number of peaks: To keep the variance down with a parametric estimate, we need to make sure that we do not try to estimate too many parameters. While this may affect the bias, even p = 2 allows a sharp peak at one frequency. In contrast, to keep the variance down with a nonparametric estimate, we need to make sure that the bandwidth is not too small. This corresponds to having a smooth spectral density estimate, so the frequency resolution is limited. This is especially important if there is more than one peak at nearby frequencies. The disadvantage is the inflexibility (bias) due to the use of the restricted class of ARMA models. 17 Parametric spectral estimation: Summary Given data x1 , x2 , . . . , xn , 2 1. Estimate the AR parameters 1 , . . . , p , w (for example, using Yule-Walker/least squares or maximum likelihood), and choose a suitable model order p (for example, using AICc = (n + p)/(n - p - 2) or BIC = p log n/n). ^ ^ ^2 2. Use the estimates 1 , . . . , p , w to compute the estimated spectral density: w ^2 ^ fy () = 2. ^ (e-2i ) 18 Introduction to Time Series Analysis. Lecture 22. 1. Review: The smoothed periodogram. 2. Examples. 3. Parametric spectral density estimation. 19 ...
View Full Document

This note was uploaded on 05/14/2011 for the course STAT 153 taught by Professor Staff during the Fall '08 term at Berkeley.

Ask a homework question - tutors are online