ECON-C4210 - Econometrics II: Capstone

Lecture 9A: Time series I

Otto Toivanen

Toivanen ECON-C4210 Lecture 9A 1 / 59

Learning outcomes
• At the end of lecture 9A, you know

1 the definition of a time series

2 what lags (=lagged values of a variable) and differences are and how to create them

3 what autocovariance and autocorrelation are

4 what an autoregressive model of order p is.

5 what a deterministic and what a stochastic trend is

6 what stationarity means and what consequences it has

7 what a random walk is and

8 how to test for stationarity

Toivanen ECON-C4210 Lecture 9A 2 / 59

Why separate lectures on time series?

• Time series: values of a (set of) variable(s) Yt over time.

• Recall OLS assumptions.

• In particular, error terms are assumed to be i.i.d

• This (very) unlikely to hold with a time series: think of the COVID-19 employment shock.

Toivanen ECON-C4210 Lecture 9A 3 / 59

# employed in Finland

Toivanen ECON-C4210 Lecture 9A 4 / 59

Time series & economics

• What most lay people consider key economic data are time series:
1 Price indices (inflation).
2 Unemployment.
3 GDP.
4 Exports and imports.

Toivanen ECON-C4210 Lecture 9A 5 / 59

Any particular reason to study time series?

• Forecasting: what is the unemployment rate in 12 months?.

• Dynamic causal effects: Does a change in the central bank interest rate affect inflation 3
months / 12 months ahead?.
• Modeling risks in financial markets (volatility).

• Time series techniques have plenty of uses outside economics (climate modeling,
engineering systems, computer science).
• Even if not interested in the time series nature, need to take it into account.

Toivanen ECON-C4210 Lecture 9A 6 / 59

Basic concepts

• Yt = value of variable Y in period t.

• Time period = time unit you are using.

• Year, quarter, month, week, day, hour, minute, second, ...

• Example: U.S inflation & unemployment 1949 - 2024.

• Data Source: U.S.Bureau of Labor Statistics.

• Data set: {Y1 , ..., YT } are T observations on the time series variable Y .

• We will study time series that are consecutive, i.e., there are not breaks (=missing
observations) in the series.

Toivanen ECON-C4210 Lecture 9A 7 / 59

U.S. Price index
monthy ear cpi
”01-2015” 700.083
”02-2015” 703.122
”03-2015” 707.306
”04-2015” 708.746
”05-2015” 712.357
”06-2015” 714.855
”07-2015” 714.902
”08-2015” 713.89
”09-2015” 712.777
”10-2015” 712.458
”11-2015” 710.952
”12-2015” 708.524
”01-2016” 709.695
”02-2016” 710.278
”03-2016” 713.339
”04-2016” 716.719
”05-2016” 719.619
”06-2016” 721.982
”07-2016” 720.816
”08-2016” 721.476
”09-2016” 723.21
”10-2016” 724.113
Toivanen
”11-2016” 722.986
ECON-C4210 Lecture 9A 8 / 59
Price level

1000
800
mean cpi = 365
400 200
0 600

0 200 400 600 800 1000

1949m1 - 2024m2: vertical line 2000m1

Toivanen ECON-C4210 Lecture 9A 9 / 59

Lags, differences, autocorrelation

• Econometric software usually have ready-made operators to produce differences and

lags.
• First lag: LYt = Yt−1 ;

• p th lag: Lp Yt = Yt−p , p = 1, ....

• The first difference: ∆Yt = Yt − Yt−1 ;

• p th lagged difference ∆p Yt = Yt−p − Yt−p−1

• Notice that you ”lose” observations from the beginning of the series when you take lags
and/or differences.
• Stata time series operators.

Toivanen ECON-C4210 Lecture 9A 10 / 59

Lags, Differences, autocorrelation

• Autocovariance and autocorrelation:

1 cov (Yt , Yt−τ ) = τ th autocovariance.
cov (Yt ,Yt−τ )
2 ρ= √
var ( Yt )
= τ th autocorrelation.

• These are population correlations, i.e., they describe the joint distribution of the
population.
• The sample autocovariance and autocorrelation are estimates of the population
equivalents.

Toivanen ECON-C4210 Lecture 9A 11 / 59

Correlation

. pwcorr cpi L.cpi L2.cpi L3.cpi, sig

cpi L.cpi L2.cpi L3.cpi

cpi 1.0000

L.cpi 1.0000 1.0000

0.0000

L2.cpi 0.9999 1.0000 1.0000

0.0000 0.0000

L3.cpi 0.9999 0.9999 1.0000 1.0000

0.0000 0.0000 0.0000

Toivanen ECON-C4210 Lecture 9A 12 / 59

Autocorrelation

• Many economic time series exhibit strong autocorrelation.

• What about (1st) differences?

∆Yt = Yt − Yt−1
• Note that if one uses logs of Y , then ∆ ln Yt ≈ (Yt − Yt−1 )/Yt−1 .

• → economists often used logs of time series data.

• Illustration using the price level and inflation as examples.

Toivanen ECON-C4210 Lecture 9A 13 / 59

Correlation

. pwcorr cpi L.cpi L2.cpi L3.cpi, sig

cpi L.cpi L2.cpi L3.cpi

cpi 1.0000

L.cpi 1.0000 1.0000

0.0000

L2.cpi 0.9999 1.0000 1.0000

0.0000 0.0000

L3.cpi 0.9999 0.9999 1.0000 1.0000

0.0000 0.0000 0.0000

Toivanen ECON-C4210 Lecture 9A 14 / 59

(Log) Price level
US cpi. 1967 = 100

6 7
6
mean lncpi =
5 4

0 200 400 600 800 1000

1949m1 - 2024m2: vertical line 2000m1

Toivanen ECON-C4210 Lecture 9A 15 / 59

Inflation = growth rate of price level
US monthly inflation
.02

.01
mean infl = 0.0028

-.01

-.02
0 200 400 600 800 1000
1949m1 - 2024m2: vertical line 2000m1

Toivanen ECON-C4210 Lecture 9A 16 / 59

Correlation

. pwcorr infl L.infl L2.infl L3.infl, sig

infl L.infl L2.infl L3.infl

infl 1.0000

L.infl 0.5897 1.0000

0.0000

L2.infl 0.3831 0.5894 1.0000

0.0000 0.0000

L3.infl 0.2954 0.3846 0.5908 1.0000

0.0000 0.0000 0.0000

Toivanen ECON-C4210 Lecture 9A 17 / 59

First order autoregression - AR(1)

Yt = β0 + β1 Yt−1 + ut

• Let’s try on price index and inflation.

Toivanen ECON-C4210 Lecture 9A 18 / 59

AR(1) - levels

.
. regr cpi L1.cpi if time_ind >= 13

Source SS df MS Number of obs = 890

F(1, 888) > 99999.00
Model 57164275.4 1 57164275.4 Prob > F = 0.0000
Residual 2382.49714 888 2.68299228 R-squared = 1.0000
Adj R-squared = 1.0000
Total 57166657.9 889 64304.4521 Root MSE = 1.638

cpi Coefficient Std. err. t P>|t| [95% conf. interval]

cpi
L1. 1.001957 .0002171 4615.86 0.000 1.001531 1.002383

_cons .2438662 .0970096 2.51 0.012 .0534714 .434261

. estat ic

Akaike's information criterion and Bayesian information criterion

Model N ll(null) ll(model) df AIC BIC

. 890 -6189.121 -1701.039 2 3406.078 3415.661

Toivanen ECON-C4210 Lecture 9A 19 / 59

AR(1) - differenced

. regr infl L1.infl if time_ind >= 13

Source SS df MS Number of obs = 890

F(1, 888) = 509.07
Model .004064181 1 .004064181 Prob > F = 0.0000
Residual .007089324 888 7.9835e-06 R-squared = 0.3644
Adj R-squared = 0.3637
Total .011153505 889 .000012546 Root MSE = .00283

infl Coefficient Std. err. t P>|t| [95% conf. interval]

infl
L1. .6019819 .0266804 22.56 0.000 .5496178 .6543459

_cons .0011637 .0001221 9.53 0.000 .000924 .0014034

. estat ic

Akaike's information criterion and Bayesian information criterion

Model N ll(null) ll(model) df AIC BIC

. 890 3759.959 3961.617 2 -7919.234 -7909.651

Toivanen ECON-C4210 Lecture 9A 20 / 59

p th - order autoregression - AR(p)

Yt = β0 + β1 Yt−1 + ... + βp Yt−p + ut

• Let’s try on price index and inflation.

• Let’s set p = 4.

Toivanen ECON-C4210 Lecture 9A 21 / 59

AR(4) - levels: price index

. regr cpi L(1/4).cpi if time_ind >= 13

Source SS df MS Number of obs = 890

F(4, 885) > 99999.00
Model 57165017.9 4 14291254.5 Prob > F = 0.0000
Residual 1639.97387 885 1.85307782 R-squared = 1.0000
Adj R-squared = 1.0000
Total 57166657.9 889 64304.4521 Root MSE = 1.3613

cpi Coefficient Std. err. t P>|t| [95% conf. interval]

cpi
L1. 1.642382 .0335556 48.95 0.000 1.576524 1.70824
L2. -.8545365 .063973 -13.36 0.000 -.980093 -.72898
L3. .2786968 .064333 4.33 0.000 .1524338 .4049599
L4. -.0655585 .0338885 -1.93 0.053 -.1320697 .0009528

_cons .1284434 .08106 1.58 0.113 -.0306489 .2875356

. estat ic

Akaike's information criterion and Bayesian information criterion

Model N ll(null) ll(model) df AIC BIC

. 890 -6189.121 -1534.846 5 3079.691 3103.647

Toivanen ECON-C4210 Lecture 9A 22 / 59

AR(4) - differenced: ”inflation”

. regr infl L(1/4).infl if time_ind >= 13

Source SS df MS Number of obs = 890

F(4, 885) = 133.34
Model .004194081 4 .00104852 Prob > F = 0.0000
Residual .006959424 885 7.8638e-06 R-squared = 0.3760
Adj R-squared = 0.3732
Total .011153505 889 .000012546 Root MSE = .0028

infl Coefficient Std. err. t P>|t| [95% conf. interval]

infl
L1. .5733712 .0332686 17.23 0.000 .5080767 .6386658
L2. -.0092169 .0383522 -0.24 0.810 -.0844888 .066055
L3. .0042502 .0381294 0.11 0.911 -.0705844 .0790848
L4. .1120095 .0331366 3.38 0.001 .046974 .1770449

_cons .0009373 .0001365 6.87 0.000 .0006694 .0012051

. estat ic

Akaike's information criterion and Bayesian information criterion

Model N ll(null) ll(model) df AIC BIC

. 890 3759.959 3969.846 5 -7929.693 -7905.737

Toivanen ECON-C4210 Lecture 9A 23 / 59

How many lags?

• Decide through testing.

• F-tests?

• Bayes information criterion: minp BIC (p) = ln SSR(p)

T + (p + 1) lnTT

• Akaike information criterion: minp AIC (p) = ln SSR(p)

T + (p + 1) T2
• AIC inconsistent, but yields more lags if (when) ln T > 2.

• Just like in machine learning, BIC and AIC utilize the bias-variance tradeoff.

• Again similar to machine learning, the objective here is prediction (forecasting).

Toivanen ECON-C4210 Lecture 9A 24 / 59

How many lags?

• Through AIC / BIC testing you get a model with possibly biased coefficients, but good
forecasts.
• In Problem Set 4 (or 5), you will search for the optimal # of lags using BIC and AIC.

Toivanen ECON-C4210 Lecture 9A 25 / 59

Example of AICBIC: U.S. cpi model AR(1) - AR(4)

U.S cpi 1949 - 2024

#lags AIC BIC
1 3406.078 3415.661
2 3105.941 3120.314
3 3081.447 3100.612
4 3079.691 3103.64

Toivanen ECON-C4210 Lecture 9A 26 / 59

Stationarity

• As we have seen, an autoregression in levels and an autoregression in differences behave

very differently.
• This takes us to the concept of stationarity.

• Stationarity is an important characteristic of time series.

• It affects regression analysis.

• It affects in particular time series analysis with multiple/explanatory variables.

Toivanen ECON-C4210 Lecture 9A 27 / 59

Stationarity

• Definition (for a single time series): A time series Yt is stationary if its probability
distribution does not change over time, that is, if the joint distribution of
(Ys+1 , Ys+2 , ..., Ys+T ) does not depend on s.
• Otherwise Yt is nonstationary.

• Stationarity requries that in a probabilistic sense, the future is like the past.

Toivanen ECON-C4210 Lecture 9A 28 / 59

What is the fuzz about (non)stationarity

• Autoregressive coefficients are biased towards zero.

• It can be shown that with a random walk (which we will study more shortly) which is
nonstationary, the OLS coefficient of the lagged dependent variable (Yt−1 )

5.3
E[β1 ] ≈ 1 −
T

Toivanen ECON-C4210 Lecture 9A 29 / 59

Demonstration with artificial data

Stata code
1
2 s e t o b s 1000
3 gen t i m e = n
4 t s s e t time
5 gen u = 3 ∗ invnorm ( uniform ( ) )
6 gen y = u
7 replace y = y [ n − 1] + u i f time > 1
8
9 regr y L. y i f time < T
10 /∗ v a r y T = 2 6 , 5 1 , 1 5 1 , 1000(+1) ∗/

Toivanen ECON-C4210 Lecture 9A 30 / 59

The Monte Carlo data

time u y u+y[t-1]
1 -3.2819 -3.2819
2 1.1012 -2.1807 -2.1807
3 0.4362 -1.7445 -1.7445
4 0.7973 -0.9471 -0.9471
5 1.4382 0.4911 0.4911
6 -3.7009 -3.2098 -3.2098
7 0.9043 -2.3055 -2.3055
8 -4.6377 -6.9433 -6.9433
9 0.4167 -6.5265 -6.5265
10 3.3998 -3.1267 -3.1267

Toivanen ECON-C4210 Lecture 9A 31 / 59

Results

Toivanen ECON-C4210 Lecture 9A 32 / 59

What is the fuzz about (non)stationarity

• In an autoregressive model, the coefficients are biased.

• Also, the t-statistics will have non-normal distributions.

• Maybe more consequentially, you get bad forecasts.

Toivanen ECON-C4210 Lecture 9A 33 / 59

What causes nonstationarity

• Two important sources of nonstationarity:

1 trends
2 structural breaks (next lecture)

Toivanen ECON-C4210 Lecture 9A 34 / 59

Example of a trend: Price level

1000
800
mean cpi = 365
400 200
0 600

0 200 400 600 800 1000

1949m1 - 2024m2: vertical line 2000m1

Toivanen ECON-C4210 Lecture 9A 35 / 59

Trends

• So a trend is a long-term movement in the data.

• A deterministic trend is a non-random function of time, e.g., Yt = t (i.e., time itself).

• A stochastic trend is random and varies over time.

• An important special case of a stochastic trend is a random walk.

Toivanen ECON-C4210 Lecture 9A 36 / 59

Examples of trends: Finnish CPI and GDP per capita

Toivanen ECON-C4210 Lecture 9A 37 / 59

Random walk

Yt = Yt−1 + ut

• The value of Yt today is in expectation the same as what it actually was yesterday.

• Random walk = today’s value of Yt is equal to where you were yesterday + a (random)
step ut of unknown direction and length.
• More generally, the best prediction p periods into the future is Yt .

Toivanen ECON-C4210 Lecture 9A 38 / 59

Random walk

• Let’s think what happens to the (expected) value of a random walk:

period Yt ut
1 u1 u1
2 u2 + u1 u2
3 u3 + u2 + u1 u3
. . .
. . .
P
l u
i i u l
. . .
. . .
• The value of a random walk in period t is the sum of the shocks to the series until and
including period t.

Toivanen ECON-C4210 Lecture 9A 39 / 59

Random walk
• Let’s think what happens to the variance of a random walk (assuming
cov (ut , ut−1 ) = 0):

period var (Yt ) var (u)

1 var (u) var (u)
2 var (u) + var (u) var (u)
3 var (u) + var (u) + var (u) var (u)
. . .
. . .
l l × var (u) var (u)
. . .
. . .
• The variance of a random walk in period t is the sum of the period-specific variances.

• Thus the variance of a random walk is increasing over time.

Toivanen ECON-C4210 Lecture 9A 40 / 59

Random walk with drift

Yt = β0 + Yt−1 + ut

• The value of Yt today is in expectation the same as what it actually was yesterday +β0 .

• Random walk = today’s value of Yt where you were yesterday +β0 + a (random) step ut
of unknown direction and length.
• More generally, the best forecast p periods ahead is

E[Yt+p |Yt ] = β0 p + Yt

Toivanen ECON-C4210 Lecture 9A 41 / 59

Random walk is a special case of AR(1)

Yt = β0 + β1 Yt−1 + ut

• If β1 = 1, Yt is nonstationary.

• If β1 < 1, Yt is stationary.

• For AR(p), p > 1, similar but more complicated statements apply.

Toivanen ECON-C4210 Lecture 9A 42 / 59

Nonstationarity, unit root & stochastic trend

• If β1 = 1, Yt is nonstationary.

• If β1 = 1, Yt has a unit root.

• If β1 = 1, Yt has a stochastic trend.

• Solving equation 1 − β1 z = 0 yields the root z = 1/β1 . If β1 = 1, z = 1; hence the

terminology ”unit root”.

Toivanen ECON-C4210 Lecture 9A 43 / 59

Testing for a unit root

• First thing to do: Plot the data to visually inspect whether there is a (stochastic) trend.

• More formally, let us study the AR(1) model:

Yt = β0 + β1 Yt−1 + ut

• H0 : β1 = 1, Yt is nonstationary / there is a unit root.

• H1 : β1 < 1, Yt is stationary / does not have a unit root.

• BUT: how to test when we know that it does not make sense to directly estimate the
above equation unless Yt is stationary, i.e., unless we know the answer to our question?

Toivanen ECON-C4210 Lecture 9A 44 / 59

Testing for a unit root

• Need a trick:
Yt = β0 + β1 Yt−1 + ut

Yt = β0 + β1 Yt−1 + Yt−1 − Yt−1 + ut

Yt − Yt−1 = β0 + (β1 − 1)Yt−1 + ut

∆Yt = β0 + δYt−1 + ut

Toivanen ECON-C4210 Lecture 9A 45 / 59

Testing for a unit root: The Dickey-Fuller test (AR(1))

∆Yt = β0 + δYt−1 + ut

• We regress the first difference of Yt on the lagged level of Yt .

• Pay attention to δ = β1 − 1, the coefficient of Yt−1 .

• We know that β1 ≤ 1 → δ ≤ 0.

• We want to test the H0 : δ = 0 against the one-sided alternative H1 : δ < 0.

Toivanen ECON-C4210 Lecture 9A 46 / 59

The augmented Dickey-Fuller test

∆Yt = β0 + δYt−1 + γ1 ∆Yt−1 ... + γp ∆Yt−p + ut

• We can add lagged differences to allow for higher order autocorrelation.

• H0 : δ = 0 against

• H1 : δ < 0.

Toivanen ECON-C4210 Lecture 9A 47 / 59

The augmented Dickey-Fuller test

• How do we get the equation for the augmented Dickey-Fuller test? Example with an
AR(2):
Yt = β0 + β1 Yt−1 + β2 Yt−2 + ut
Yt = β0 + (β1 + β2 )Yt−1 − β2 Yt−1 + β2 Yt−2 + ut
Yt = β0 + (β1 + β2 )Yt−1 − β2 [Yt−1 − Yt−2 ] + ut
Yt − Yt−1 = β0 − Yt−1 + (β1 + β2 )Yt−1 − β2 [Yt−1 − Yt−2 ] + ut
∆Yt = β0 + (β1 + β2 − 1)Yt−1 − β2 ∆Yt−1 + ut
∆Yt = β0 + δYt−1 + γ1 ∆Yt−1 + ut
• β1 z + β2 z − 1 = 0, then β1 + β2 = 1 and hence δ = 0, too.

Toivanen ECON-C4210 Lecture 9A 48 / 59

The augmented Dickey-Fuller test

• In the AR(1) case, if β1 = 1, i.e., there is a unit root,

∆Yt = β0 + ut

• In the AR(2) case, if β1 + β2 = 1, i.e., there is a unit root,

∆Yt = β0 + γ1 ∆Yt−1 + ut

Toivanen ECON-C4210 Lecture 9A 49 / 59

The augmented Dickey-Fuller test with deterministic trend

∆Yt = β0 + αt + δYt−1 + γ1 ∆Yt−1 ... + γp ∆Yt−p + ut

• H0 : δ = 0 against

• H1 : δ < 0.

Toivanen ECON-C4210 Lecture 9A 50 / 59

The augmented Dickey-Fuller test with deterministic trend

• What version to use - with or without deterministic trend?

• If no long-term growth (-/+), then the alternative is that Yt is stationary round a constant.
use the intercept-only version.
• If the time-series seems to exhibit long-term growth, then the alternative is that Yt is
stationary round a trend. → use the intercept plus deterministic trend version.
• Think back to Finnish CPI and GDP per capita - series.

Toivanen ECON-C4210 Lecture 9A 51 / 59

Examples of trends: Finnish CPI and GDP per capita

Toivanen ECON-C4210 Lecture 9A 52 / 59

The augmented Dickey-Fuller test

• The Dickey-Fuller test is one-sided.

• Why are we not worried about δ > 0, i.e., β1 > 1?

• Need special critical values.

Toivanen ECON-C4210 Lecture 9A 53 / 59

Demonstration #1: US price index in logs

. dfuller lncpi if time_ind >= 13 , regress

Dickey-Fuller test for unit root Number of obs = 890

Interpolated Dickey-Fuller
Test 1% Critical 5% Critical 10% Critical
Statistic Value Value Value

Z(t) -1.683 -3.430 -2.860 -2.570

MacKinnon approximate p-value for Z(t) = 0.4400

D.lncpi Coef. Std. Err. t P>|t| [95% Conf. Interval]

lncpi
L1. -.0002362 .0001404 -1.68 0.093 -.0005117 .0000393

_cons .0042164 .0007952 5.30 0.000 .0026557 .0057771

Toivanen ECON-C4210 Lecture 9A 54 / 59

DF test & critical values for different AR(p)

#lags Test stat 1% 5% 10%

1 -0.997 -3.43 -2.86 -2.57
2 -0.978 -3.43 -2.86 -2.57
3 -0.954 -3.43 -2.86 -2.57
4 -0.914 -3.43 -2.86 -2.57
5 -0.904 -3.43 -2.86 -2.57
6 -0.901 -3.43 -2.86 -2.57
7 -0.885 -3.43 -2.86 -2.57
8 -0.886 -3.43 -2.86 -2.57
9 -0.889 -3.43 -2.86 -2.57
10 -0.903 -3.43 -2.86 -2.57

Toivanen ECON-C4210 Lecture 9A 55 / 59

The augmented Dickey-Fuller test

• So there is a unit root... what to do?

• As suggested above, transform by differencing.

Toivanen ECON-C4210 Lecture 9A 56 / 59

Demonstration #2: U.S inflation = difference in log of price index
• Test results and critical values for different AR(p).

#lags Test stat 1% 5% 10%

1 -12.934 -3.43 -2.86 -2.57
2 -11.128 -3.43 -2.86 -2.57
3 -9.324 -3.43 -2.86 -2.57
4 -8.584 -3.43 -2.86 -2.57
5 -7.953 -3.43 -2.86 -2.57
6 -6.829 -3.43 -2.86 -2.57
7 -6.417 -3.43 -2.86 -2.57
8 -5.795 -3.43 -2.86 -2.57
9 -4.92 -3.43 -2.86 -2.57
10 -4.249 -3.43 -2.86 -2.57

Toivanen ECON-C4210 Lecture 9A 57 / 59

Some issues in testing for a unit root

• Consider a break in the series: A one-time change in the mean of a series (e.g. collapse
of Lehman Brothers in 2008). Such a shock would bias conclusions towards a unit root.
• We will consider how to test for a break in the next lecture.

• Consider a large (few large) outlier. The series may then look as if it mean-reverting
although it is not. Test results may be biased towards stationarity.

Toivanen ECON-C4210 Lecture 9A 58 / 59

Summary on trends in time series

• The random walk model is the workhorse model for trends in economic time series.

• Always first plot the series.

• Then compute the Dickey-Fuller test (either with or without a trend).

• If Yt has a unit root, transform data by differencing, i.e., use ∆Yt .

• If Yt does not have a unit root, move ahead with your analysis.

Toivanen ECON-C4210 Lecture 9A 59 / 59