Professional Documents
Culture Documents
Univariate Time Series Model Notes
Univariate Time Series Model Notes
• The non-stationary decrease slowly
2. White noise process: a white noise process (for disturbance errors) is a
stationary process, with no discernible (perceptible) structure:
1 𝐸 𝑢K = 𝜇f ; 𝑡 = 1, 2, . . . , ∞
White noise process has constant mean
2 𝑣𝑎𝑟(𝑢K ) = 𝜎f O < ∞
White noise process has constant variance
𝜎f O 𝑖𝑓 𝑠 = 0
3 𝛾X =
0 𝑖𝑓 𝑠 ≠ 0
Each observation is uncorrelated with all other values in the sequence. Thus a
white noise process has zero auto-co-variances, except at lag zero. Another
way to state this last condition would be to say that each observation is
uncorrelated with all other values in the sequence.
Hence the autocorrelation function for a white noise process will be zero apart
from a single peak of 1 at 𝑠 = 0.
𝛾X 1 𝑖𝑓 𝑠 = 0
𝜏X = =
𝛾] 0 𝑖𝑓 𝑠 ≠ 0
If 𝜇f = 0, and the three conditions hold, the process is known as zero mean white
noise.
Furthermore, If it is assumed that 𝑢K is distributed normally, then the sample
autocorrelation coefficients are also approximately normally distributed
1
𝜏X ~𝑁 0,
𝑇
where T is the sample size, and 𝜏X denotes the autocorrelation coefficient at lag s
estimated from a sample
3. Hypothesis testing for 𝜏X :
3.1 Constructing a non-rejection region:
𝐻] : 𝜏X = 0
𝐻u : 𝜏X ≠ 0
V V
95% 𝑛𝑜𝑛 − 𝑟𝑒𝑗𝑒𝑐𝑡𝑖𝑜𝑛 𝑟𝑒𝑔𝑖𝑜𝑛: (−1.96×
•
, 1.96×
•
)
If the sample autocorrelation coefficient 𝜏X falls outside this region for a given
value of s, then the null hypothesis that the true value of the coefficient at that
lag s is zero is rejected.
Alternatively, we could calculate 𝑇𝑆 = 𝜏X × 𝑇
𝑇𝑆 > 𝑍… → 𝑅𝑒𝑗𝑒𝑐𝑡 𝐻]
R
𝑇𝑆 ≤ 𝑍… → 𝐹𝑎𝑖𝑙 𝑡𝑜 𝑟𝑒𝑗𝑒𝑐𝑡 𝐻]
R
It should be clear from the form of the statistic that asymptotically (that is, as
the sample size increases towards infinity), the (T + 2) and (T − k) terms in
the Ljung–Box formulation will cancel out, so that the statistic is equivalent to
the Box–Pierce test. This statistic is very useful as a portmanteau (general)
test of linear dependence in time series.
3.4 Example:
Suppose that a researcher had estimated the first five autocorrelation coefficients
using a series of length 100 observations, and found them to be
Lag 1 2 3 4 5
Autocorrelation 0.207 -0.013 0.086 0.005 -0.022
coefficient
Test each of the individual correlation coefficients for significance, and test all five
jointly using the Box–Pierce and Ljung–Box tests.
1) Constructing a non-rejection region:
1 1
95% 𝑛𝑜𝑛 − 𝑟𝑒𝑗𝑒𝑐𝑡𝑖𝑜𝑛 𝑟𝑒𝑔𝑖𝑜𝑛: (−1.96× , 1.96× )
𝑇 𝑇
where T = 100 in this case. The decision rule is thus to reject the null hypothesis that a
given coefficient is zero in the cases where the coefficient lies outside the range
(−0.196,0.196). For this example, it would be concluded that only the first
autocorrelation coefficient is significantly different from zero at the 5% level.
2) Box-Pierce and Ljung-Box test:
turning to the joint tests, the null hypothesis is that all of the first five autocorrelation
coefficients are jointly zero, i.e.
𝐻] : 𝜏XQ = 0, 𝜏XR = 0, 𝜏X = 0, 𝜏X¡ = 0, 𝜏X¢ = 0
or simply
𝐻] : 𝜏V = 0, 𝜏O = 0, 𝜏£ = 0, 𝜏¤ = 0, 𝜏¥ = 0
The test statistics for the Box–Pierce and Ljung–Box tests are given respectively, as
’
= 5.09
’
∗
𝜏X ‘ O
𝑄 = 𝑇× 𝑇 + 2 ×
𝑇−𝑘
“”V
= 𝜇 + 𝑢K × 𝜃 𝐿
4.1 Properties:
1 𝐸 𝑦K = 𝜇;
2 𝑣𝑎𝑟 𝑦K = 𝛾] = [1 + 𝜃V O + 𝜃O O + ⋯ + 𝜃´ O ]×𝜎 O < ∞
𝜃X + 𝜃´¸V 𝜃V + 𝜃´¸O 𝜃O + ⋯ 𝜃´ 𝜃´TX 𝑓𝑜𝑟 𝑠 = 1, 2, … , 𝑞
3 𝑐𝑜𝑣𝑎𝑟𝑖𝑎𝑛𝑐𝑒 𝛾X =
0 𝑓𝑜𝑟 𝑠 > 𝑞
So, a moving average process has constant mean, constant variance, and auto-co-
variances which may be non-zero to lag q and will always be zero thereafter.
Consider the following simple MA(2) model:
𝑦K = 𝑢K + 𝜃V 𝑢KTV + 𝜃O 𝑢KTO
where 𝑢K is a zero mean white noise process with variance σ2.
(1) Calculate the mean and variance of 𝑦K .
(2) Derive the autocorrelation function for this process (i.e., express the
autocorrelations 𝜏V , 𝜏O , 𝜏£ as functions of the parameters 𝜃V and 𝜃O
(3) If 𝜃V = −0.5 and 𝜃O = 0.25, sketch the acf of yt.
(1) 𝐸 𝑢K = 0 𝑡ℎ𝑒𝑛 𝐸 𝑢KTµ = 0, ∀𝑖
So the expected value of the error term is zero for all time periods.
Taking expectations of both sides of given equation gives
𝐸 𝑦K = 𝐸 𝑢K + 𝜃V 𝑢KTV + 𝜃O 𝑢KTO = 𝐸(𝑢K) + 𝐸 𝜃V 𝑢KTV + 𝐸 𝜃O 𝑢KTO = 0
𝑣𝑎𝑟 𝑦K = 𝐸(𝑦K − 𝐸 𝑦K ) (𝑦K − 𝐸 𝑦K ) = 𝐸(𝑦K )(𝑦K )
=E(𝑢K + 𝜃V 𝑢KTV + 𝜃O 𝑢KTO )(𝑢K + 𝜃V 𝑢KTV + 𝜃O 𝑢KTO )
=E(𝑢K O +𝜃V 𝑢KTV O + 𝜃O 𝑢KTO O + 𝑐𝑟𝑜𝑠𝑠 − 𝑝𝑟𝑜𝑑𝑢𝑐𝑡s)
But E(𝑐𝑟𝑜𝑠𝑠 − 𝑝𝑟𝑜𝑑𝑢𝑐𝑡s)=0 since 𝑐𝑜𝑣(𝑢K , 𝑢KTX ) = 0 for 𝑠 ≠ 0. ‘Cross-products’ is
thus a catchall expression for all of the terms in u which have different time
subscripts, such as 𝑢K −1 𝑢K −2 or 𝑢K −5 𝑢K −20, etc.
Again, one does not need to worry about these cross-product terms, since these
are effectively the auto-co-variances of 𝑢K , which will all be
zero by definition since 𝑢K is a random error process, which will have zero auto-
co-variances (except at lag zero).
𝛾] can also be interpreted as the auto-co-variance at lag zero.
So 𝑣𝑎𝑟 𝑦K = 𝛾]
= 𝐸 𝑢K O + 𝜃V 𝑢KTV O + 𝜃O 𝑢KTO O
= 𝜎 O + (𝜃V 𝜎)O + (𝜃O 𝜎)O )
= 𝜎 O (1 + 𝜃V O + 𝜃O O )
(2) Calculating now the acf of 𝑦K , first determine the auto-co-variances and then
the autocorrelations by dividing the auto-co-variances by the variance.
The auto-co-variance at lag 1 is given by
𝛾V = 𝐸(𝑦K − 𝐸 𝑦K ) (𝑦KTV − 𝐸 𝑦KTV ) = 𝐸(𝑦K )(𝑦KTV )
= 𝐸(𝑢K + 𝜃V 𝑢KTV + 𝜃O 𝑢KTO )(𝑢KTV + 𝜃V 𝑢KTO + 𝜃O 𝑢KT£ )
=E(𝜃V 𝑢KTV O + 𝜃V 𝜃O 𝑢KTO O + 𝑐𝑟𝑜𝑠𝑠 − 𝑝𝑟𝑜𝑑𝑢𝑐𝑡s)
= 𝐸(𝜃V 𝑢KTV O + 𝜃V 𝜃O 𝑢KTO O )
= 𝜃V 𝜎 O + 𝜃V 𝜃O 𝜎 O
The auto-co-variance at lag 1 is given by
𝛾O = 𝐸(𝑦K − 𝐸 𝑦K ) (𝑦KTO − 𝐸 𝑦KTO ) = 𝐸(𝑦K )(𝑦KTO )
= 𝐸(𝑢K + 𝜃V 𝑢KTV + 𝜃O 𝑢KTO )(𝑢KTO + 𝜃V 𝑢KT£ + 𝜃O 𝑢KT¤ )
=E(𝜃O 𝑢KTO O + 𝑐𝑟𝑜𝑠𝑠 − 𝑝𝑟𝑜𝑑𝑢𝑐𝑡s)
= 𝐸(𝜃O 𝑢KTO O )
= 𝜃O 𝜎 O
The auto-co-variance at lag 3 is given by
𝛾£ = 𝐸(𝑦K − 𝐸 𝑦K ) (𝑦KTO − 𝐸 𝑦KTO ) = 𝐸(𝑦K )(𝑦KT£ )
= 𝐸(𝑢K + 𝜃V 𝑢KTV + 𝜃O 𝑢KTO )(𝑢KT£ + 𝜃V 𝑢KT¤ + 𝜃O 𝑢KT¥ )
= 0
So 𝛾X = 0 for s > 2. All autocovariances for the MA(2) process will be
zero for any lag length, s, greater than 2.
The autocorrelation at lag 0 is given by
ÀÁ
𝜏] =
ÀÁ
= 1
4.2 Process invertibility MA(1) ~ AR(∞):
MA(1) model is defined as:
𝑦K = 𝜇 + 𝑢K + 𝜃𝑢KTV = 𝜇 + 𝑢K + 𝜃𝐿×𝑢K
→ 𝑦K − 𝜇 = 𝑢K + 𝜃𝐿×𝑢K = 𝑢K (1 + 𝜃𝐿)
→ 𝑦K − 𝜇 = 𝑢K + 𝜃𝐿×𝑢K = 𝑢K (1 − −𝜃𝐿 )
® TÄ O
→ 𝑢K = VT ¯TÂÅ =(𝑦K − 𝜇)(1 + (−𝜃𝐿) + −𝜃𝐿 + (−𝜃𝐿)£ + ⋯ )
𝑦K = 𝜇 + 𝜙V 𝑦KTV + 𝑢K
(1) Calculate the (unconditional) mean of yt.
For the remainder of the question, set the constant to zero (𝜇 = 0) for
simplicity.
(2) Calculate the (unconditional) variance of 𝑦K .
(3) Derive the autocorrelation function for this process.
(1)
The unconditional mean will be given by the expected value of
Expression
𝐸 𝑦K = 𝐸 𝜇 + 𝜙V 𝑦KTV + 𝑢K
= 𝜇 + 𝜙V 𝐸 𝑦KTV
= 𝜇 + 𝜙V (𝜇 + 𝜙V 𝐸 𝑦KTO
= 𝜇 + 𝜙V 𝜇 + 𝜙V O 𝐸 𝑦KTO
= 𝜇 + 𝜙V 𝜇 + 𝜙V O (𝜇 + 𝜙V 𝐸 𝑦KT£
= ⋯
= 𝜇(1 + 𝜙V + 𝜙V O + ⋯ 𝜙V ÎTV )+ 𝜙V Î 𝐸 𝑦KTÎ
= 𝜇 1 + 𝜙V + 𝜙V O + ⋯ 𝜙V ÎTV
𝜇
=
1 − 𝜙V
Thus the expected or mean value of an autoregressive process of order one is
given by the intercept parameter divided by one minus the autoregressive
coefficient.
(2)
Calculating now the variance of yt, with µ set to zero
𝑦K = 𝜙V 𝑦KTV + 𝑢K
This can be written equivalently as
𝑦K (1 − 𝜙V 𝐿) = 𝑢K
𝑦K = 1 − 𝜙V 𝐿 TV
𝑢K = 1 + 𝜙V 𝐿 + 𝜙V O 𝐿O + ⋯ 𝑢K
= 𝑢K + 𝜙V 𝑢KTV + 𝜙V O 𝑢KTO +𝜙V £ 𝑢KT£ + ⋯
So long as |𝜙V | < 1, i.e., so long as the process for yt is stationary, this
sum will converge.
From the definition of the variance of any random variable y, it is
possible to write
𝑣𝑎𝑟(𝑦K ) = 𝐸(𝑦K − 𝐸 𝑦K ) (𝑦K − 𝐸 𝑦K )
= 𝐸 𝑦K 𝑦K
=E(𝑢K + 𝜙V 𝑢KTV + 𝜙V O 𝑢KTO +𝜙V £ 𝑢KT£ … )(𝑢K + 𝜙V 𝑢KTV 𝜙V O 𝑢KTO +𝜙V £ 𝑢KT£ … )
=E(𝑢K O + 𝜙V O 𝑢KTV O + 𝜙V ¤ 𝑢KTO O + ⋯ )
= 𝜎 O (1 + 𝜙V O + 𝜙V ¤ + ⋯ )
Provided that |𝜙V | < 1, the infinite sum in equation can be written as
𝜎O
𝑣𝑎𝑟(𝑦K ) =
1 − 𝜙V O
(3)
Turning now to the calculation of the autocorrelation function, the auto-
covariances must first be calculated. This is achieved by following similar
algebraic manipulations as for the variance above, starting with the definition
of the autocovariances for a random variable. The autocovariances for lags 1,
2, 3, ..., s, will be denoted by γ1, γ2, γ3, ..., γs, as previously.
𝛾V = 𝑐𝑜𝑣 𝑦K , 𝑦KTV = 𝐸(𝑦K − 𝐸 𝑦K )(𝑦KTV − 𝐸 𝑦KTV )=E(𝑦K )(𝑦KTV )
=E(𝑢K + 𝜙V 𝑢KTV + 𝜙V O 𝑢KTO +𝜙V £ 𝑢KT£ + ⋯ )(𝑢KTV + 𝜙V 𝑢KTO + 𝜙V O 𝑢KT£ +𝜙V £ 𝑢KT¤ +. . . )
=E(𝜙V 𝑢KTV O +𝜙V £ 𝑢KTO O +𝜙V ¥ 𝑢KT£ O +…)
=𝜙V 𝜎 O 1 + 𝜙V O + 𝜙V ¤ + ⋯
𝜙V 𝜎 O
𝑠𝑜 𝛾V =
1 − 𝜙V O
𝛾O = 𝑐𝑜𝑣 𝑦K , 𝑦KTO = 𝐸(𝑦K − 𝐸 𝑦K )(𝑦KTO − 𝐸 𝑦KTO )=E(𝑦K )(𝑦KTO )
=E(𝑢K + 𝜙V 𝑢KTV + 𝜙V O 𝑢KTO +𝜙V £ 𝑢KT£ + ⋯ )(𝑢KTO + 𝜙V 𝑢KT£ + 𝜙V O 𝑢KT¤ +𝜙V £ 𝑢KT¥ +. . . )
=E(𝜙V O 𝑢KTO O +𝜙V ¤ 𝑢KT£ O +𝜙V Ò 𝑢KT¤ O +…)
=𝜙V O 𝜎 O 1 + 𝜙V O + 𝜙V ¤ + ⋯
𝜙V O 𝜎 O
𝑠𝑜 𝛾O =
1 − 𝜙V O
…
𝜙V £ 𝜎 O
𝛾£ =
1 − 𝜙V O
𝜙V X 𝜎 O
𝛾X =
1 − 𝜙V O
The acf can now be obtained by dividing the covariances by the variance, so
that
𝛾]
𝜏] = = 1
𝛾]
𝜙V 𝜎 O
𝛾V 1 − 𝜙V O
𝜏V = =
𝛾] 𝜎O
1 − 𝜙V O
= 𝜙V
𝜙V O 𝜎 O
𝛾O 1 − 𝜙V O
𝜏O = =
𝛾] 𝜎O
1 − 𝜙V O
= 𝜙V O
…
𝜏X = 𝜙X O
which means that 𝑐𝑜𝑣 𝑦K , 𝑦KTO = 𝜙X O . Note that use of the Yule–Walker
equations would have given the same answer.
6. The Partial Autocorrelation Function
The PACF is useful for telling the difference between an AR process and
an ARMA process.
In the case of an AR(p), there are direct connections between 𝑦K and 𝑦KTX only for
𝑠 ≤ 𝑝. So for an AR(p), the theoretical PACF will be zero after lag p.
In the case of an MA(q), if it is invertible (roots of characteristic equations
𝜃(𝑧)=0 lie outside the unit circle), it can be written as an AR(∞), so there are
direct connections between 𝑦K and all its previous values. For an MA(q), the
theoretical PACF will be geometrically declining.
At lag 1: 𝜏VV = 𝜏V
•R T•Q R
At lag 2: 𝜏OO = VT•Q R
7. ARMA(p,q):
By combining the AR(p) and MA(q) models, an ARMA(p, q) model is obtained.
Such a model states that the current value of some series y depends linearly on
its own previous values plus a combination of current and previous values of a
white noise error term. The model could be written
For 𝜙 𝐿 = 1 − 𝜙V 𝐿 − 𝜙O 𝐿O − ⋯ − 𝜙È 𝐿È
𝑎𝑛𝑑 𝜃 𝐿 = 1 + 𝜃V 𝐿 + 𝜃O 𝐿O + ⋯ + 𝜃´ 𝐿´ , we have
𝜙 𝐿 𝑦K = 𝜇 + 𝜃(𝐿)𝑢K , or
𝑦K = 𝜇 + 𝜙V 𝑦KTV + 𝜙O 𝑦KTO + ⋯ + 𝜙È 𝑦KTÈ +𝜃V 𝑢KTV + 𝜃O 𝑢KTO + ⋯ + 𝜃´ 𝑢KT´ +𝑢K