Normally Distributed Seasonal Unit Root Tests D. A. Dickey North Carolina State University Note: this presentation is based on the paper “Normally Distributed Seasonal Unit Root Tests” authored by D. A. Dickey in the book Economic Time Series: Modeling and Seasonality edited by Bell, Holan, and McElroy published by CRC press, 2012 Model: Seasonal AR(1) Yt = r Yt-s + et, et is White Noise rˆ r Yt s et / Yt 2 s Goal: Test H0: r=1 Yt Yi,j=Ymonth, year J Yr. 1 Yr. 2 | Y1,1 Y2,1 F M A M J J A S O N D (s=12) Y1,2 Y1,3 Y1,4 Y1,5 Y1,6 Y1,7 Y1,8 Y1,9 Y1,11 Y1,11 Y1,12=Y1,s Y2,2 Y2,3 Y2,4 Y2,5 Y2,6 Y2,7 Y2,8 Y2,9 Y2,21 Y2,11 Y2,12=Y2,s | Yr. m Ym,1 Ym,2 Ym,3 Ym,4 Ym,5 Ym,6 Ym,7 Ym,8 Ym,9 Ym,21 Ym,11 Ym,12=Ym,s Yt = r Yt-s + et Yt – Yt-s = (r-1) Yt-s + ei,j Yi,j - Yi,j-1 = (r-1) Yi,j-1 + ei,j Previous work: Yt = r Yt-s + et Yi,j = r Yi,j-1 + ei,j Yi,j - Yi,j-1 = (r-1) Yi,j-1 + ei,j OLS : s s m s m 2 s rˆ 1 Y j 1,i e j ,i / Y j 1,i Ni / Di i 1 i 1 j 1 i 1 i 1 j 2 Ni / 2 0 m(m 1)(m 2) / 3 m(m 1) / 2 ~ , 2 2 Di / m(m 1) / 2 m(m 1)(m 2) / 3 m(m 1)(m m 1) / 3 Dickey & Zhang (2011, J. Korean Stat. Soc.) Under H0: (1) S large CLT t stat NORMAL (0,1) (O(s-1/2) mean adjustment helpful ) (2) Known O(s-1/2) adjustments to mean (same) for k periodic regressors added (k<<s) (3) MSE2 n.b.: Does not apply to seasonal dummy variables Known mean 0 OLS : s s m s m 2 s rˆ r Yi , j 1ei , j / Yi , j 1 Ni / Di i 1 i 1 j 1 i 1 i 1 j 2 Ni / 2 0 (m 2) / 3 1/ 2 ~ , m m 1 2 2 m ( m 1) / 2 ( m 2) / 3 ( m m 1) / 3 D / i N 0 , D0 *****Add seasonal dummy variables:***** OLS : s s s m s m 2 rˆ r (Yi , j 1 Yi , )ei , j / Yi , j 1 Yi , N i / Di i 1 i 1 j 1 i 1 i 1 j 2 (m 2) m6 12 Ni / 2 2 ~ , m 2 2 m Di / m(m 2) 6 6 N 0 , D0 m 6 2 m(2m 4m 9) 90 w.o.l.o.g. Assume 2 = 1 Notation: E{Ni}=N0 E{Di}=D0 (different for mean 0 versus seasonal means) MSE=Mean Square Error = (Total SSq – Model SSq)/df MSE in seasonal means case is (regressing deseasonalized differences on deseasonalized lag levels) s m 2 sN 2 N2 2 ei , j ei / s (m 2) ˆ D (m 2) D i 1 j 2 N 02 N2 p 2 2 ˆ only for mean 0 case !!! s (m 2) D m 2 D0 2 OLS : s s s m s m 2 rˆ r (Yi , j 1 Yi , )ei , j / Yi , j 1 Yi , N i / Di i 1 i 1 j 1 i 1 i 1 j 2 Standard error [(X’X)-1(MSE)]1/2 = t statistic, seasonal means model: f ( N , D, ˆ 2 ) MSE / ( sD) N D MSE / ( sD) sN 2 N2 D ˆ m 2 D 3 2 1 3(m 2) p 2 f N , D , ( 1) 0 0 s 2 2m 3 2 N2 D ˆ m 2 D N No Mean f ( N , D, ˆ 2 p ) 0 s Seasonal Means m2 2 m( m 2) D0 6 N0 Taylor Series, seasonal means : N i / 2 N 0 0 VN VND ~ , 2 D V V D / 0 D ND i s N 2 N2 ˆ D m 2 D sN 0 N 02 D0 1 ( m 2) D 0 g N , D, ˆ 2 R (N0=(m-2)/2<0) 2 D 1 s ˆ 2 N m 2 1/2 3s(m 2) 1 g N , D, ˆ 2 Op 2m 3 s g N , D, ˆ 2 1 D0 1 (1) 2 2 N0 m 2 3 2 A s N N 0 B s D D0 C s ˆ 2 1 1 D 1 (1) 02 2 N0 m 2 m6 12 m 2 Cov Ni , Di , ˆ (m 2) 6 1 (m 2) 1 D0 1 (1) 4 N02 m 2 3 3 2 A s N N B s D D C s ˆ 1 2 0 m 6 m 2m 2 4m 9 90 m 6(m 2) 0 2 D0 8m 1 A (1) 2 3 N 3 m 2 0 (m 2) 4 m B 1 (1) 2 2 N m 2 0 6(m 2) 2 D0 2m 2 C 2 m 2 N0 3 m 2 V11 V12 V13 A 3 2 64m 168m 108m) A B C V21 V22 V23 B 3 80 m 1. 5 V V V C 33 31 32 Approximate variance of in seasonal means case COMPARISON No Mean Model 2 m2 E{ } 3 s m(m 1) Var{ } 1 Seasonal Means Model 3s m 2 E{ } 2m 3 64m Var{ } 3 168m2 108m) 80 m 1.5 3 0.8 m Calculation “recipe” for Seasonal Means Model (1) Regress Yt-Yt-s on seasonal dummies and Yt-s. Get = t test for Yt-s 3s m 2 2m 3 (2) Compute E{ } (3) Compute 64m Var{ } (4) Compare Z 3 168m2 108m) 80 m 1.5 3 E to N(0,1) to get p-value. Var ( ) Alternative approach: Expand around (N, D) only, run large (1/2 million) simulation fixup for small m. Result for variance: m 64m2 48m 108 0.8777 4.4705 1 2.2215 0.0853 3 3/2 m 1.5 m 1.5 s m 1.5 80 m 1.5 Similar empirical adjustments to mean: 0.0505 3s m 2 1 0.2118 0.3704 2 m 3 m 1.5 s s Compare limit (sinfinity) variance formulas: Taylor 3 variable versus Taylor 2 variable with and without adjustments 1 million replicates s=12, m=6 Unadjusted (N,D) only (10 seconds run time) Reference normal variance from empirical adjustment from 3 variable Taylor: Notes: Graphs use sample means (both expansions give same mean approximation) 3 variable Taylor variance 1.1556 closer to simulated statistics’ variance 1.2310 than is empirical adjusted if no s adjustment used. With the finite s part in the empirically adjusted formula, that formula gives 1.2024 The choice m=6 gave the biggest vertical gap between the limit (s) variance formulas. In previous graph. NEXT: Sequence with m=6, s =4, 6, 12, 24 THEN: Sequence with s=4, m=6, 8, 10, 20, 100 Histogram Stats (reps = 1,000,000) formula 2 2formula m skew 4 6 0.34 0.65 2.59 2.51 1.39 1.29 6 6 0.26 0.35 3.06 2.99 1.31 1.25 12 6 24 6 0.18 0.15 4.16 4.12 0.13 0.07 5.77 5.74 1.23 1.20 1.19 1.17 52 6 0.09 0.03 8.40 8.38 1.17 1.14 s skew m kurt s kurt formula 2 2formula 4 6 0.34 0.65 2.59 2.51 1.39 1.29 4 8 0.19 0.33 2.60 2.54 1.16 1.11 4 10 4 20 0.11 0.22 2.60 2.56 0.02 0.08 2.61 2.59 1.05 1.03 0.90 0.89 4 100 0.10 0.04 2.62 2.62 0.80 0.79 4 1000 0.12 0.04 2.62 2.62 0.78 0.76 (formulas from book) Higher order models (seasonal multiplicative form) (1 r B s )(Yt ) Zt ( ( B))Zt et ( AR( p)) ( ( B))(1 r B s )(Yt ) et Suggested estimation (Dickey, Hasza, Fuller (1984)) (1) (under H0:r=1) Regress Dt = Yt-Yt-s on p lags of D AR(p) and residual rt (2) Filter Y using AR(p) model for D. Ft = filtered Yt (3) Regress rt on Ft-s & p lags of D (t test on Ft-s is Dickey, D. A., D. P. Hasza, and W. A. Fuller (1984). “Testing for Unit Roots in Seasonal Time Series”, Journal of the American Statistical Association, 79, 355-367. Example 1: Oil Imports (from book, s=12, m=36) Levels First Differences & Seasonal Means Variable Intercept Ft-12 filter12 D1 D2 D3 D4 D5 D6 D7 D8 DF Parameter Estimate 1 1 1 1 1 1 1 1 1 1 -163.62422 -0.81594 0.01853 -0.00511 0.00016148 0.02891 0.02369 0.00623 -0.01025 -0.04440 t Value -0.20 -16.19 0.48 -0.11 0.00 0.58 0.48 0.13 -0.22 -1.13 Pr > |t| 0.8442 <.0001 0.6349 0.9128 0.9974 0.5617 0.6344 0.8998 0.8268 0.2604 +--------------------------------------------------------------+ | Formulas from Economic Time Series Modeling and Seasonality | | pg. 398 (Bell, Holan McElroy eds.) | | | | s = 12 m = 36 | | Tau = -16.19 Mean = -4.2118 variance = 0.8377 | | | | Tau ~ N(-4.2118,0.8377) | | | | Z=(-16.19-(-4.2118))/sqrt(0.8377) | | | | Pr{Z <-13.09 } = 0.0000 | | | +--------------------------------------------------------------+ Maximum Likelihood Estimation Estimate 515.88809 Standard Error 4154.0 t Value 0.12 Approx Pr > |t| 0.9012 Lag 0 AR1,1 0.17664 0.05113 3.45 0.0006 12 amt 0 AR2,1 AR2,2 AR2,3 AR2,4 AR2,5 AR2,6 AR2,7 AR2,8 -0.67807 -0.40409 -0.19310 -0.23710 -0.20400 -0.17344 -0.18688 -0.07802 0.04966 0.05916 0.06194 0.06232 0.06233 0.06259 0.06021 0.05122 -13.65 -6.83 -3.12 -3.80 -3.27 -2.77 -3.10 -1.52 <.0001 <.0001 0.0018 0.0001 0.0011 0.0056 0.0019 0.1277 1 2 3 4 5 6 7 8 amt amt amt amt amt amt amt amt 0 0 0 0 0 0 0 0 NUM1 NUM2 NUM3 NUM4 NUM5 NUM6 NUM7 NUM8 NUM9 NUM10 NUM11 6522.6 -25856.9 16307.1 6196.3 5578.6 2777.3 3490.6 3538.0 -13966.8 5738.4 -11305.0 7137.8 5920.4 5505.8 6060.5 6051.0 5744.2 6036.5 6079.1 5499.4 5910.3 7131.6 0.91 -4.37 2.96 1.02 0.92 0.48 0.58 0.58 -2.54 0.97 -1.59 0.3608 <.0001 0.0031 0.3066 0.3566 0.6287 0.5631 0.5606 0.0111 0.3316 0.1129 0 0 0 0 0 0 0 0 0 0 0 month1 month2 month3 month4 month5 month6 month7 month8 month9 month10 month11 0 0 0 0 0 0 0 0 0 0 0 Parameter MU Variable amt Shift 0 Autocorrelation Check of Residuals To Lag 6 12 18 24 30 36 42 48 ChiSquare . 0.94 7.50 12.23 15.66 24.97 31.62 40.27 DF 0 3 9 15 21 27 33 39 Pr > ChiSq . 0.8169 0.5851 0.6615 0.7883 0.5761 0.5358 0.4140 --------------------Autocorrelations-------------------0.001 -0.003 -0.007 0.004 -0.001 -0.010 0.016 0.006 -0.009 -0.009 0.038 0.002 -0.048 0.076 0.016 -0.054 -0.052 -0.029 0.017 0.059 -0.048 -0.022 0.059 -0.019 0.014 0.022 0.023 -0.002 0.079 0.000 -0.002 0.061 -0.007 0.097 0.018 0.080 -0.019 0.024 -0.033 -0.074 -0.075 -0.030 0.002 -0.008 -0.097 0.051 0.024 0.072 Example 2: Airline Series from Box & Jenkins Original Scale Logarithmic Scale Log Passengers (1,12) with lags at 1, 12, 23 Parameter MU AR1,1 AR1,2 AR1,3 Estimate Standard Error 0.0002871 -0.28601 -0.43072 0.30157 0.0022107 0.06862 0.07154 0.07270 t Value Approx Pr > |t| Lag 0.13 -4.17 -6.02 4.15 0.8967 <.0001 <.0001 <.0001 0 1 12 23 Autocorrelation Check of Residuals To Lag 6 12 18 24 ChiSquare 6.22 10.23 15.87 24.25 DF 3 9 15 21 Pr > ChiSq 0.1016 0.3319 0.3909 0.2809 --------------Autocorrelations----------------0.050 -0.066 -0.096 -0.105 0.112 0.075 -0.009 -0.045 0.137 -0.044 0.037 -0.063 -0.110 0.022 0.063 -0.094 0.100 0.045 -0.154 0.002 0.007 -0.014 0.015 -0.169 Parameter Estimates Variable Intercept filter12 D1 D12 D23 DF Parameter Estimate Standard Error t Value 1 1 1 1 1 -0.00095084 -0.42340 0.11925 0.24740 0.09924 0.00309 0.13897 0.08262 0.14225 0.07566 -0.31 -3.05 1.44 1.74 1.31 Pr > |t| 0.7587 0.0029 0.1517 0.0847 0.1923 +--------------------------------------------------------------+ | Formulas from Economic Time Series Modeling and Seasonality | | pg. 398 (Bell, Holan McElroy eds.) | | | | s = 12 m = 12 | | Tau = -3.05 Mean = -4.1404 variance = 0.9550 | | | | Tau ~ N(-4.1404,0.9550) | | | | Z = (-3.05-(-4.1404))/sqrt(0.9550) | | Pr{Z <1.1158 } = 0.8677 | | | +--------------------------------------------------------------+ Weekly Lower 48 States Natural Gas Working Underground Storage (Billion Cubic Feet) Example 3: Weekly Natural Gas Supplies (Energy Information Agency) November April Lag 1 model fits well for Natural Gas Series First and Span 52 Differences Conditional Least Squares Estimation Parameter MU AR1,1 Estimate Standard Error 0.23692 0.39305 2.51525 0.03101 t Value Approx Pr > |t| Lag 0.09 12.68 0.9250 <.0001 0 1 Autocorrelation Check of Residuals To Lag ChiSquare DF Pr > ChiSq 6 12 18 24 30 36 42 48 1.61 12.55 18.20 19.60 20.56 25.35 34.76 40.70 5 11 17 23 29 35 41 47 0.8997 0.3237 0.3764 0.6656 0.8747 0.8847 0.7432 0.7295 ------------------Autocorrelations----------------0.007 -0.051 0.018 0.003 0.013 0.023 -0.031 0.034 -0.026 0.000 0.059 -0.033 0.008 0.016 -0.082 0.054 0.029 0.034 0.020 -0.009 0.014 -0.011 -0.007 -0.030 -0.013 0.042 0.027 -0.013 0.008 -0.051 -0.042 0.005 -0.006 0.008 0.005 -0.010 -0.019 0.011 0.009 0.036 -0.003 0.081 0.036 -0.011 0.012 -0.040 -0.024 0.008 Natural Gas Example – OLS Regression Parameter Estimates Variable DF Parameter Estimate Standard Error t Value Pr > |t| Intercept 1 -0.27635 1.05426 -0.26 0.7933 filter52 1 -1.04170 0.03343 -31.16 <.0001 D1 1 0.00313 0.02138 0.15 0.8838 +----------------------------------------------------------+ | | | s = 52 m = 18 | | Tau = -31.16 Mean = -8.6969 variance = 0.8877 | | | | Tau ~ N(-8.6969,0.8877) | | | | Z = (-31.16 - (-8.6969))/sqrt(0.8877) = -23.84 | | | | Pr{Z <-23.84 } = 0.0000 | | | +----------------------------------------------------------+