Seasonality

# Seasonality - ApEc 8212 Applied Econometrics - Lecture #25...

This preview shows pages 1–4. Sign up to view the full content.

ApEc 8212 Applied Econometrics -- Lecture #25 Time Series Analysis IV: Model Selection + Forecasts (Enders, Ch. 2, Sections 7-12) This lecture explains how to estimate ACFs and PACFs to choose the “best” ARMA(p, q) model for a time series. I. Sample Autocorrelations of Stationary Series The first step is to calculate ACF’s and PACF’s and their standard errors. The sample mean and variance of y are: y = (1/T) T t t y 2 ˆ σ = (1/T) T t 2 t ) y y ( The estimated ACF coefficients ρ s , denoted as r s , are: r s = = + = T 1 t 2 t T 1 s t s t t ) y y ( ) y y )( y y ( We want to check whether r s = 0. Box & Jenkins (1976, 1984) showed that, under the null hypothesis that ρ s = 0, and assuming that y t is stationary and ε t ~ N[0, σ ε 2 ]: 1

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Var(r s ) = 1/T for s = 1 = (1 + 2 Σ = 1 1 s j r j 2 )/T for s > 1 Asymptotically, r s will be normally distributed and have a mean of zero under the null that ρ s = 0. For PACF coefficients, under the null hypothesis φ ss = 0, the “approximate” variance of the estimate of φ ss is T -1 . Usually, if a ACF or PACF coefficient = 0, then all higher order coefficients also = 0. Don’t test each coefficient individually; even if all = 0, if your significance level is 5% there is a 1/20 chance that you reject the null. It is best to use a joint test. Box & Pierce proposed the ‘Q-statistic’: Q = T Σ = s k 1 r k 2 Under the null hypothesis that all s ACFs = 0, this statistic is asymptotically distributed as χ 2 (s). Ljung and Box modified this to give it better small sample performance: Q = T(T+2) Σ = s k 1 [r k 2 /(T-k)] Both tests can be applied to residuals from an estimated ARMA(p, q) model to see if they are “white noise”. 2
t and y in the r s formula with the estimated residuals. Of course, autocorrelated residuals contradict the white noise assumption. The degrees of freedom for the χ 2 test is lower (s - p - q); the residuals are from an estimated model, this “uses up” p+q degrees of freedom. Model Selection Criteria When estimating ARMA processes maybe you should, to be cautious, allow for many non-zero a’s and β ’s. Yet allowing for unnecessary a’s and β ’s will reduce the precision of the estimated a’s and β ’s that really are 0, reducing the precision of your forecasts. Two ‘model selection’ criteria are the Akaike Information Criterion (AIC) and the Schwartz Bayesian Criterion (SBC) : AIC = T × ln(residual sum of squares) + 2n SBC = T × ln(residual sum of squares) + n × ln(T) T is “effective sample size”, i.e. subtracting observations deleted due to lags, and n = parameters estimated. Choose the model with the smallest AIC or SBC. Both are increasing in the parameters (n). SBC puts a bigger penalty (ln(T) > 2 for T > 7). SBC has better asymptotic properties; AIC may have better small sample properties. 3

This preview has intentionally blurred sections. Sign up to view the full version.

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

## This note was uploaded on 02/24/2010 for the course ECON 570 taught by Professor Staff during the Fall '08 term at UNC.

### Page1 / 24

Seasonality - ApEc 8212 Applied Econometrics - Lecture #25...

This preview shows document pages 1 - 4. Sign up to view the full document.

View Full Document
Ask a homework question - tutors are online