Testing for IID Noise/White Noise: I

Transcription

1 Testing for IID Noise/White Noise: I want to be able to test null hypothesis time series {x t } or set of residuals {r t } is IID(0, 2 ) or WN(0, 2 ) there are many such tests, including informal test based on sample ACF (see overhead II 64) portmanteau test turning point test di erence-sign test rank test runs test will illustrate tests using all-star game time series {x t } & residuals {z t } from AR(1) model for detrended Lake Huron levels BD 35, CC 46, SS-149 IV 1

2 outcome of game NL win tie AL win First All-Star Baseball Game in Each Year (I 22) year IV 2

3 ACF Sample ACF for All-Star Game x t h (lag) IV 3

4 AR(1) Residuals z t = r t ˆrt 1 (III 15) z t year IV 4

5 Unit Lag Scatter Plot of AR(1) Residuals z t (III 16) z t IV 5 z t

6 ACF Sample ACF for AR(1) Residuals z t (III 17) h (lag) IV 6

7 Portmanteau Test: I can regard as formal version of informal test based on sample ACF, which makes use of fact that, when {X t } IID(µ, 2 ), then corresponding ˆ X (1),..., ˆ X (h) are approximately IID N (0, 1/n) for large sample sizes n (but: h/n must be small!) hence ˆ X (j) p 1/n = n 1/2ˆ X (j), j = 1,..., h, are approximately IID standard normal RVs (i.e., N (0, 1)) recall that, if Z 1, Z 2,..., Z are IID N (0, 1), then Z Z Z2 2 obeys a chi-square distribution with degrees of freedom BD 36; CC 183; SS 150 IV 7

8 Portmanteau Test: II hence, under null hypothesis {X t } IID(µ, 2 ), hx Q(h) n 1/2ˆ 2 hx X (j) = n ˆ 2 X (j) should obey a j=1 2 h distribution j=1 Q(h) will be large under alternative hypotheses, so can reject null hypothesis at level of significance if Q(h) > 2 1 (h), where 2 1 (h) is (1 )th quantile of 2 h distribution refinement is Ljung Box version of portmanteau test: hx ˆ 2 Q LB (h) = n(n + 2) X (j) n j (obeys j=1 2 h distribution to better approximation than Q(h) does) BD 36; CC 183; SS 150 IV 8

9 Portmanteau Test: III if we want to test residuals {r t } rather than a time series {x t }, then hx ˆ 2 Q LB (h) = n(n + 2) R (j) n j j=1 has the same form as before, but now obeys a 2 h p distribution, where p is the number of parameters estimated in forming {r t } as examples, let s look at Q LB (h) for all-star game time series {x t } and Lake Huron residuals {z t }, for which p = 3 arguably (have estimated intercept, slope and the AR(1) parameter ) will focus on tests with level = 0.05 red horizontal lines show (h p) blue dots show Q LB (h) CC 183, SS 150 IV 9

10 Q LB (h) and 95% quantile Portmanteau Tests of All-Star Game {x t } h (lag) IV 10

11 p values p-values for Portmanteau Tests of All-Star Game {x t } h (lag) IV 11

12 Q LB (h) and 95% quantile Portmanteau Tests of Lake Huron {z t } h (lag) IV 12

13 p values p-values for Portmanteau Tests of Lake Huron {z t } h (lag) IV 13

14 Turning Point Test: I time series {x t } is said to have a turning point at t if one of the following two patterns exist: 1. x t > x t 1 and x t > x t+1 2. x t < x t 1 and x t < x t+1 if X t 1, X t and X t+1 are IID RVs with a continuous distribution, the following six events are equally likely: X t 1 < X t < X t+1 (no turning point) X t 1 < X t+1 < X t (turning point) X t < X t 1 < X t+1 (turning point) X t < X t+1 < X t 1 (turning point) X t+1 < X t 1 < X t (turning point) X t+1 < X t < X t 1 (no turning point) BD 36, 37 IV 14

15 Turning Point Test: II let T be number of turning points in IID sequence of length n since probability of a turning point is 2/3 under IID hypothesis, µ T E{T } = 2(n 2)/3 can show that, under IID hypothesis, 2 16n 29 T var {T } = 90 large T µ T indicates time series is either fluctuating more rapidly than what would be expected from an IID sequence or there are fewer fluctuations than expected under IID hypothesis, T is approximately N (µ T, 2 T ), so can reject null at level if T µ T / T > 1 /2, where 1 /2 is 1 /2 quantile for standard normal distribution BD 36, 37 IV 15

16 Turning Point Test: III note: cannot apply to all-star {x t } (has a discrete distribution) as example, T = 60 for Lake Huron {z t } (see next overhead) since n = 97, get µ T. = 63.3 and 2 T. = 16.9 ( T. = 4.1), so T µ T T. = 0.81 p-value is 0.42, so fail to reject null hypothesis BD 36, 37 IV 16

17 AR(1) Residuals z t = r t ˆrt 1 z t year IV 17

18 Di erence-sign Test: I let S be the number of times X t > X t 1, t = 2,..., n (equivalent to number of times X t X t 1 is positive) if {X t } is IID, then µ S E{S} = (n 1)/2 can shown that 2 S var {S} = (n + 1)/12 S is approximately equal in distribution to a N (µ S, for large n 2 S ) RV large positive (negative) value of S µ S indicates presence of increasing (decreasing) trend (doesn t have power against alternative of cyclic variations) reject null hypothesis at level if S µ S / S > 1 /2 BD 37 IV 18

19 Di erence-sign Test: II note: cannot apply to all-star {x t } (has a discrete distribution) as example, S = 48 for Lake Huron {z t } (see next overhead).. since n = 97, get µ S = 48 and = 8.2 ( S = 2.9), so 2 S S µ S S = 0 p-value is 1, so totally fail to reject null hypothesis!!! BD 37 IV 19

20 AR(1) Residuals z t = r t ˆrt 1 z t year IV 20

21 Rank Test: I given time series of length n, let P be the pairs (r, s) such that X r > X s and r > s total number of pairs such that r > s is n n(n 1) = 2 2 if {X t } is IID, each event of form X r > X s has probability of 1/2, so µ P E{P } = n(n 1)/4 can shown that 2 P var {P } = n(n 1)(2n + 5)/72 P is approximately equal in distribution to a N (µ P, for large n large positive (negative) value of P increasing (decreasing) trend 2 P ) RV µ P indicates presence of BD 37, 38 IV 21

22 Rank Test: II reject null hypothesis at level if P µ P / P > 1 /2 note: cannot apply to all-star {x t } (has a discrete distribution) as example, P = 2351 for Lake Huron {z t } since n = 97, get µ P = 2328 and P 2.. = ( P = 160.4), so P µ P. = 0.14 p-value is 0.89, so again fail to reject null hypothesis P BD 37 IV 22

23 test designed for Runs Test: I binary-valued {x t } (will assume x t is either 1 or 1) continuous-valued {x t }, with centering allowed; i.e., {x t where, e.g., c = x is sample mean of x t s residuals {r t } (sample mean r should be close to 0) run is defined to be set of consecutive values in series, all of which are above or below zero as an example, here is pattern of American League and National League victories over 78 years in all-star {x t }: c}, AAANANANAAANAAAANNNNANNAANNNNNNNNNNNNANNNNNNNNNNNANNANAAAAAANNNAAAAAAAAAAAANNN above has a total of 24 runs (12 each for both leagues) CC 46; SS 149 IV 23

24 Runs Test: II runs test is based on number of runs in series, conditional on number of observed values above and below zero (test does not require underlying probability of value being greater than zero to be 50%) let n + and n denote number of, respectively, positive and negative values in series (thus n + + n = n) let n r be number of runs under null hypothesis of IID, n r is approximately normally distributed with mean µ and variance 2 given by µ = 2n +n n + 1 and 2 = (µ 1)(µ 2) n 1 CC 46; SS 149 IV 24

25 Runs Test: III large positive (negative) value of n r µ indicates excessive choppiness (too much clumpiness) compared to what is reasonable under null hypothesis of IID reject null hypothesis at level if n r µ / > 1 /2 using all-star {x t } as example, n r = 24, n + = 36 & n = 42 get µ. = 39.8 and 2. = 19.0 (. = 4.4), so n r µ. = 3.6 p-value is , so null hypothesis is not tenable CC 46; SS 149 IV 25

26 Runs Test: IV for Lake Huron {z t }, have n r = 37, n + = 51 & n = 46 get µ = and 2.. = 23.9 ( = 4.9), so n r µ. = 2.5 p-value is 0.011, so would reject null hypothesis at = 0.05 level of significance, but would fail to reject (just barely!) at more stringent = 0.01 level number of runs smaller than to be expected, suggesting clumpiness in data (agrees with positive correlation at unit lag) CC 46; SS 149 IV 26

27 Testing for IID Noise/White Noise: III summary of IID tests for all-star {x t } informal test based on sample ACF: IID hypothesis untenable both ˆ X (1) & ˆ X (2) well outside bounds reasonable for IID noise portmanteau test: reject resoundingly for h = 1,..., 10 turning point test: inapplicable di erence-sign test: inapplicable rank test: inapplicable runs test: reject (p-value is ) conclusion: IID hypothesis for all-star {z t } untenable action item: bet on National League in 2013! IV 27

28 Testing for IID Noise/White Noise: IV summary of IID tests for Lake Huron {z t } informal test based on sample ACF: IID hypothesis seems tenable, but ˆ (1) outside of 95% CI portmanteau test: fail to reject at level = 0.5 for h 4 turning point test: fail to reject (p-value is 0.42) di erence-sign test: fail to reject (p-value is 1(!)) rank test: fail to reject (p-value is 0.89) runs test: reject (p-value is 0.011) IID hypothesis for Lake Huron z t s thus tenable, but some evidence for small but significant autocorrelation at least at lag h = 1 IV 28

29 Testing for IID Noise/White Noise: V not clear how to combine results of tests together (multiple comparison problem) need to keep in mind various tests have di ering strengths and weaknesses against particular alternative hypotheses two other tests: informal test based on fitting AR models (will discuss later) cumulative periodogram test (will discuss in Stat/EE 520) IV 29