The Functional Central Limit Theorem and Testing For Time Varying Parameters
The Functional Central Limit Theorem and Testing For Time Varying Parameters
Lecture 2:
1. FCLT
2. Overview of TVP topics (Models, Testing, Estimation)
3. Testing problems
4. Tests
T
function of (x1, x2, x3, … xT), say ∑x
t =1
2
t .
1. W(0) = 0
Three key theorems (Hall and Heyde (1980) and Davidson (1994, part
VI):
(i) Let 0 ≤ s1 < s2 … < sk ≤ 1, a set of k points. Suppose that (ξT(s1), ξT(s2),
d
… , ξT(sk)) → (ξ(s1), ξ(s2), … , ξ(sk)) for any set of k points, {si}.
(ii) The function ξT(s) is “tight” (or more generally satisfies “stochastic
equicontinuity” as discussed in Andrews (1994)), meaning
Suppose εt ~ mds with variance σ ε2 and bounded 2+δ moments for some δ
> 0.
(a) Let ξT ( s ) denote the function that linearly interpolates between the
t
1
points ξ(t/T) =
T
∑ ε . Then ξT ⇒ σεW, where W is a Wiener process
i =1
i
Then
1 T
1 T ⎡ 1 t ⎤
3/ 2 ∑ t
x = ∑ ⎢ 1/2 ∑ ε i ⎥ = σ ε ∫ ξT ( s )ds ⇒ σ ε ∫ W ( s )ds
1 1
T t =1 T t =1 ⎣ T i=1 ⎦ 0 0
Then
t t t
T−1/2
xt = T −1/2
∑ ai = T
i =1
−1/2
∑ (ε i − θε i−1 ) = (1 − θ )T
i =1
−1/2
∑i
ε +
i =1
θ T −1/2
(ε t − ε 0 )
Note: θ(1)σε is the “long-run” standard deviation of a.
Suppose I want to approximate the 95th quantile of the distribution of, say,
T
1
3/ 2 ∑ t
1
vT = x . Because vT v = σ ε ∫ W ( s )ds , I can use the 95th quantile of v
T t =1
0
This approximation works well when T is reasonably large, and does not
require knowledge of the distribution of x.
Models
yt = βt + εt
Discrete Breaks:
⎧ β for t ≤ τ
Single Break: βt = ⎨
⎩ β + δ for t > τ
⎧ β for t ≤ τ 1
⎪
Two Breaks: βt = ⎨ β + δ1 for τ 1 < t ≤ τ 2
⎪
⎩ β + δ1 + δ 2 for t > τ 2
Multiple Breaks: ….
2-Regime Model:
⎧ β when st = 0 ⎫
βt = ⎨ ⎬,
⎩ β + δ when st = 1⎭
Multiple Regimes …
(b) Elliott and Müller (2006). Optimal tests are very similar if number
of “breaks” is large (say, 3).
vs.
• VARs
o SW (1996): 5700 Bivariate VARs involving Post-war U.S. macro
series. Reject null of constant VAR coefficient (Φ) in over 50% of
cases using tests with size = 10%.
o Many others …
⎧ β for t ≤ τ
βt = ⎨
⎩ β + δ for t > τ
τ T
1 1
where Y1 = ∑ yt and Y2 = ∑τ y
τ t =1 T −τ t = +1
t
2
1
Wald statistic: ξW = 2 1 δ 1
σˆ ε ( τ + T −τ )
a σ e2 a σ ε2
Follows from Y1 ~ N ( β , ) and Y2 ~ N ( β + δ , ) and they are
τ T −τ
independent.
Quandt (1960) suggested computing the Chow statistic for a large number
of possible values of τ and using the largest of these as the test statistics.
The problem is then to find the distribution of ξQ under the null (it will not
be χ2), so that the critical value for the test can be determined.
∑ ∑ t =[ sT ]+1 t
[ sT ] T
Ho
1 [
1
[ sT ] t =1
εt − 1
[(1− s )T ] ε ]2
ξW ,T ( s ) = 2
σˆ e 1
[ sT ] + [(1−1s )T ]
∑ ∑
[ sT ] T
1 s [ 1 1
T t =1
ε t − (1−1s ) 1
T t =[ sT ]+1
ε t ]2
= 2
σˆ e 1
s + (1−1s )
[ 1s WTa ( s ) − (1−1s ) (WTa (1) − WTa ( s ))]2 [WTa ( s ) − sWTa (1)]2
= =
1
s + (1−1s ) s (1 − s )
1 1
∑
[ sT ]
where WTa ( s ) = ε t , and the last equality follows from
σˆ ε T t =1
[W ( s ) − sW (1)]2
Thus, using FCLT, ξW,T ( ) ⇒ ξ( ), where ξ(s) = .
s (1 − s )
The results have been derived here for the case of a single constant
regressor. Exensions to the case of multiple (non-constant) regressors can
be found in Andrews (1993) (Critical values for the test statistic are also
given in Andrews (1993) with corrections in Andrews (2003), reprinted in
Stock and Watson (2006).)
The QLR test seems very sensible, but is there a more powerful
procedure? Andrews and Ploberger (1993) develop optimal tests (most
powerful) for this (and related) problems.
then the most powerful test rejects Ho for large values of the Likelihood
Ratio, LR = fa(Y)/fo(Y), the ratio of the densities evaluated at the realized
value of the random variable.
LR = Eτ[ fa(y|τ)]/fo(Y).
When δ is “small,” the test statistic turns out to be simple average of ξW(τ)
over all possible break dates. Sometimes this test statistic is called the
“Mean Wald” statistic.
Let Y = (y1, … yT)′, so that Y ~ N(0,σ ε2 Ω(γ)), where Ω(γ) = I + γ2A, where
A = [aij] with aij = min(i,j).
Y ′Ω(γ a ) −1Y
so that the test rejects the null for large values of ′
.
YY
Optimal tests require a choice of γa, which measures the amount of time
variation under the alternative. A common way to choose γa is to use a
value so that, whenγ = γa, the test has a pre-specified power, often 50%.
Ω(γ a ) −1 = [ I + γ a2 A]−1 ≈ 1 − γ a2 A .
Y ′ AY
ψ= ′
YY
YY
T
∑ t
p 2
∑ t =1 ∑ i=t i
T
( y ) 2T
∑ t =1 ( ∑
T T
1 1
yi ) 2 Ho 1
T ψ= ⇒ ∫ (W (1) − W ( s )) 2 ds .
=1 T T i =t
∑ t =1 y
T 2 0
1
T t
∑ t =1 ( ∑ i =t i
T T
1 1
ε
ˆ ) 2
T =1ψ =
T T
, where εˆt = yt − βˆ is the OLS residual. In this
∑ t =1 t
T
1
T εˆ 2
Ho 1
case, one can show that T ψ ⇒ ∫ (W ( s ) − sW (1)) 2 ds
=1
0
Regressors: yt = xt′βt + εt
= Σxxβt + et + mt β t
t t t t
Test statistic depends on T −1/2
∑yx
i =1
t t = Σ xxT −1/2
∑β
i =1
t +T −1/2
∑e +T ∑m β
i =1
t
−1/2
i =1
t t
Tests for martingale variation (qLL, Ny) have power that is similar to tests
for discrete break (SupF=QLR, AP).
Lecture 2 - 36, July 21, 2008
2. Stock-Watson (1998): Martingale TVP DGP: yt = βt + εt
Tests for a discrete break (QLR, EW, MW) have power that is similar to
tests for martingale variation (L, POI)
1.FCLT (tool)
2.Overview of TVP topics
a. Persistent vs. mean reverting TVP
3.Testing problems
a. Discrete break and AP approach
4.Tests
a. Little difference between tests for discrete break and persistent
“continuous” variation. What do you conclude when the null of
stability is rejected?