Lag selection, forecasting and ARDL models

Chikumbe, SE
21st June, 2023

Econometrics II
Department of Economics, Kwame Nkrumah University

1 / 35
Recap of last week

• OLS assumptions
• Zero conditional mean

• No serial correlation

• Static models → yt = β 0 + β 1 xt + ut

• Distributed Lag (DL) models → yt = β 0 + β 1 xt + β 2 xt −1 + ut

• Autoregressive (AR) models → yt = β 0 + β 1 yt −1 + ut

2 / 35
Roadmap for this week

• AR and DL model lag selection

• Autoregressive Distributed Lag (ARDL) model

• One period ahead forecasting and forecast errors

3 / 35
Lag selection

4 / 35
Putting the p in AR(p)

• We’ve looked at AR(1), AR(2), AR(3) etc.

• How do we know the correct order of an AR model?

• Why is it important to choose the correct order?

• The order, p, directly influences the zero conditional mean assumption

5 / 35
AR order and the zero conditional mean assumption

• We implement an AR(2) model: yt = β 0 + β 1 yt −1 + β 2 yt −2 + ut

• We find the coefficients on β 1 and β 2 to be highly statistically significant

• Zero conditional mean: E (ut |yt −1 , yt −2 ) = 0

• Now we implement an AR(3) model: yt = β 0 + β 1 yt −1 + β 2 yt −2 + β 3 yt −3 + ut

• We find β 3 to also be highly statistically significant: E (ut |yt −1 , yt −2 ) ̸= 0

• Estimating the AR(3) however: E (ut |yt −1 , yt −2 , yt −3 ) = 0

6 / 35
Choosing the p in AR(p)

• How do we know the correct order of an AR model?

• Three approaches
1. F-statistic approach

2. Bayes Information Criterion (BIC)

3. Akaike Information Criterion (AIC)

7 / 35
F-statistic approach

• Start with a model with as many lags as possible

• Test the significance of the last lag

• Do this until the last lag becomes significant at the 95% confidence level

• Advantage: intuitive

• Drawback: it will suggest too many lags, some of the time

• Assume the true AR order is 5, so that the sixth coefficient is zero

• A test using the t-statistic will incorrectly reject the null 5% of the time

• → when the true value of p is five, this method will estimate p to be six 5% of the time.

8 / 35
Autoregressive Model of % ∆ ZAR/USD

Dependent Variable: % ∆ZAR/USD

AR(1) AR(2) AR(3)
% ∆ZAR/USD t −1 0.307∗∗∗ 0.343∗∗∗ 0.344∗∗∗
(0.051) (0.054) (0.054)

% ∆ZAR/USD t −2 −0.109∗∗ −0.116∗∗

(0.054) (0.057)

% ∆ZAR/USD t −3 0.023
(0.055)

Intercept 0.004∗∗ 0.005∗∗ 0.004∗∗

(0.002) (0.002) (0.002)

R2 0.094 0.106 0.105

9 / 35
Bayes Information Criterion (BIC)

 
SSR (p ) ln(T )
BIC (p ) = ln + (p + 1)
T T

• SSR (p ) → the sum of squared residuals of the AR(p) model

• SSR = ∑ni=1 (ûi )2 → how much variation is not explained by our regression

• (p + 1) → number of variables used, plus 1 for the intercept

ln(T )
• T → log of the number of observations / number of observations

10 / 35
Bayes Information Criterion (BIC)

 
SSR (p ) ln(T )
BIC (p ) = ln + (p + 1)
T T

• SSR decreases as we add more AR terms

ln(T )
• (p + 1) T → increases as we add more AR terms

• BIC trades off these two forces

• We are interested in finding p̂, the value of p which minimizes BIC (p ) among possible
choices of p

11 / 35
Bayes Information Criterion (BIC)

12 / 35
Akaike Information Criterion (AIC)
 
SSR (p ) 2
AIC (p ) = ln + (p + 1)
T T

• ln(T ) is replaced by 2, in the second term → second term is smaller in the AIC than
the BIC

• In a dataset with 1000 observations

• ln(1000) = 6.9 which is more than 3 times larger than 2

• → a smaller decrease in the SSR is needed to justify including another AR term

• In large samples, AIC will overestimate p̂ with nonzero probability

• Still useful, if you are concerned there are too few lags suggested by the BIC
13 / 35
Comparing the 3 approaches
Dependent Variable: % ∆ZAR/USD
AR(1) AR(2) AR(3) • F-Statistic approach → AR(2)
% ∆ZAR/USD t −1 0.307∗∗∗ 0.343∗∗∗ 0.344∗∗∗
(0.051) (0.054) (0.054) • AIC → AR(2)
% ∆ZAR/USD t −2 −0.109∗∗ −0.116∗∗
(0.054) (0.057) • BIC → AR(1)
% ∆ZAR/USD t −3 0.023
(0.055)

Intercept 0.004∗∗ 0.005∗∗ 0.004∗∗

(0.002) (0.002) (0.002)

R2 0.094 0.106 0.105

AIC -6.774 -6.782 -6.744
BIC -6.752 -6.749 -6.730

14 / 35
Autoregressive Distributed Lag (ARDL)
models

15 / 35
Types of time series models

• Static models
• yt = β 0 + β 1 xt + ut

• Distributed lag models (DL)

• yt = β 0 + β 1 xt + β 2 xt −1 + ut

• Autoregressive models (AR)

• yt = β 0 + β 1 yt −1 + ut

• Autoregressive distributed lag models (ARDL)

• yt = β 0 + β 1 yt −1 + β 2 xt −1 + ut

16 / 35
ARDL models

yt = β 0 + β 1 yt −1 + β 2 yt −2 + σ1 xt −1 + σ2 xt −2 + ut

• ARDL(p, q) where p = order of AR terms and q = order of DL terms

• The example above is therefore an ARDL(2,2) model

• Combines the benefits of AR and DL models

• Zero conditional mean assumption

• → at time t, the error term is independent of every explanatory variable (y & x), in every
period

17 / 35
ARDL models

yt = β 0 + β 1 yt −1 + β 2 yt −2 + σ1 xt −1 + σ2 xt −2 + ut

• A note on ARDL convention

• When using ARDL models, the convention is to typically exclude contemporaneous DL
terms (no xt )

• → does not allow for contemporaneous effects

• Why? In time series, the focus is often on forecasting, so values in time t are never known

• However, including xt is perfectly fine

18 / 35
ARDL example: ARDL Model of the %∆ ZAR/USD

Dependent Variable: %∆ ZAR/USD

ARDL(1,1) ARDL(2,2)
%∆ ZAR/USDt −1 0.211∗∗∗ 0.221∗∗∗
(0.074) (0.081)

%∆ ZAR/USDt −2 −0.132∗
(0.079)

%∆ ALSIt −1 −0.253∗∗∗ −0.261∗∗∗

(0.064) (0.065)

%∆ ALSIt −2 −0.072
(0.070)

Intercept 0.007∗∗ 0.008∗∗∗

(0.003) (0.003)

R2 0.155 0.175

19 / 35
ARDL models and lag selection
   
SSR (k ) ln(T ) SSR (k ) 2
BIC (k ) = ln + (k + 1) AIC (k ) = ln + (k + 1)
T T T T
• k = p+q

• This can result in many different models needing to be tested

• k = 4, with p = 1 and q = 3

• k = 4, with p = 2 and q = 2

• k = 4, with p = 3 and q = 1 etc.

• Convention (but not a requirement!): set p = q

• If you do this however, think about implications for zero conditional mean

20 / 35
ARDL lag selection example: ARDL Model of the %∆ ZAR/USD
ARDL(1,1) ARDL(2,2)
%∆ ZAR/USDt −1 0.211∗∗∗ 0.221∗∗∗
(0.074) (0.081)

%∆ ZAR/USDt −2 −0.132∗
(0.079)

%∆ ALSIt −1 −0.253∗∗∗ −0.261∗∗∗

(0.064) (0.065)

%∆ ALSIt −2 −0.072
(0.070)

Intercept 0.007∗∗ 0.008∗∗∗

(0.003) (0.003)

R2 0.155 0.175
AIC -6.681 -6.676
BIC -6.624 -6.582

21 / 35
Forecasting

22 / 35
Forecasting is a core part of time series econometrics

• “What is your best forecast of next month’s unemployment rate?”

• Critical for policy making, planning, investment decisions etc.

• Given dependence in time series data, AR models are the workhorse model for
forecasting

23 / 35
Forecasting is a core part of time series econometrics

Figure: National Treasury forecasts of active vs passive debt stablization policies

24 / 35
Forecasting is a core part of time series econometrics

• Contrast this with the cross sectional domain

• Example: predict test scores in a cross section vs predicting test scores in a time series

• Time series models can be used to forecast, even if none of the coefficients have a
causal interpretation

25 / 35
Forecast Error
• We want to forecast yt +1 based off an AR(1) model

• yt +1 = β 0 + β 1 yt

• Since t + 1 is into the future, the true values of β 0 and β 1 are unknown

• → we use the OLS estimators β̂ 0 and β̂ 1 from historical data as a proxy

• We then seek to estimate yt +1 conditional on information at time t

• ŷt +1|t = β̂ 0 + β̂ 1 yt

• Forecast error: yt +1 − ŷt +1|t

• The difference between the realized value of yt +1 and the forecasted value

26 / 35
A forecast is not a predicted value

• From lecture 2: the predicted value of y is ŷ

• Formally, ŷi = βˆ0 + βˆ1 xi

• Predicted values are calculated for for observations in the sample used to estimate the
regression

• Forecasts are made for a date that exists outside of the sample

27 / 35
The forecast error is not an OLS residual

• From lecture 2: the residual of a regression is the difference the actual value for y and
the predicted value for y , ŷ for observations in the sample

• Formally, ûi = yi − ŷi = yi − βˆ0 − βˆ1 xi

• The forecast error is the difference between the future value of y, which is not
contained in the sample, and the forecast of that future value

• “Out-of-sample” versus “In-sample”

28 / 35
Forecasting with an AR(1) model

%∆ERt = 0.004 + 0.311%∆ERt −1

• Our forecast equation becomes

%∆ERt +1|t = 0.004 + 0.311%∆ERt

• %∆ERt = −0.023

• → %∆ERt +1|t = 0.004 + 0.31 ∗ −0.023 = −0.003

29 / 35
Forecasting with an AR(1) model

→ %∆ERt +1|t = 0.004 + 0.31 ∗ −0.023 = −0.003

• You are now told that %∆ERt +1 = −0.056

ˆ t +1|t = −0.056 − (−0.003) = −0.053

• Forecast error: ERt +1 − ER

• Our forecast overpredicts the realized value by -0.053

30 / 35
Forecast uncertainty

• Forecast error can be divided into two parts

1. Uncertainty about the regression coefficients

2. Uncertainty about the future value of ut

• If there are few coefficients, (2) > (1)

• We’ll introduce the Root Mean Squared Forecast Error (RMSFE) that incorporates
both (1) and (2)

31 / 35
Root Mean Squared Forecast Error (RMSFE)

• Start with an ARDL(1,1): yt = β 0 + β 1 yt −1 + β 2 xt −1 + ut

• Forecast: Ŷt +1|t = βˆ0 + βˆ1 yt + βˆ2 xt

• Forecast error: yt +1 − ŷt +1|t = ut +1 − ( βˆ0 − β 0 ) + ( βˆ1 − β 1 )yt + ( βˆ2 − β 2 )xt

 

 
MSFE = E yt +1 − ŷt +1|t
= σu2 + var ( βˆ0 − β 0 ) + ( βˆ1 − β 1 )yt + ( βˆ2 − β 2 )xt
 

√
• The RMSFE is then simply MSFE

32 / 35
Root Mean Squared Forecast Error (RMSFE)

• Forecast error can be divided into two parts

1. Uncertainty about the regression coefficients

2. Uncertainty about the future value of ut

• If there are few coefficients, (2) > (1)

• We’ll introduce the Root Mean Squared Forecast Error (RMSFE) that incorporates
both (1) and (2)

• If (1) > (2), then the RMSFE ≈

p
var (ut )

33 / 35
Forecast interval

• Similar in spirit to a confidence interval

• One major difference1

• Confidence interval: coefficient ± 1.96 * standard error

• Confidence interval is justified by CLT and therefore holds for a wide range of
distributions of ut

• The forecast error contains

1 If
this sounds a bit tricky, revisit your Chapter 4 lectures with Safia to refresh your grasp of confidence
intervals, hypothesis tests and the central limit theorem
34 / 35
Forecasting interest rates

•
Figure: Fan chart of interest rate forecast from the SARB

35 / 35