Slides-Chapter-8 - TSA

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 57

Time Series Analysis

Chapter 8 - Nonlinearity

Annika Camehl
Econometric Institute
Erasmus University Rotterdam
Chapter 8 - Nonlinearity

Key concepts in this chapter:

• Regime-switching
• Threshold models
• Smooth transition models
• Markov-Switching models
• Structural change
• Testing for threshold nonlinearity
• Estimation of threshold models
• Forecasting with nonlinear models

Time Series Analysis: Chapter 8 1


Nonlinearity

(Autoregressive) Time series models take the general form


yt = g(yt−1, yt−2, . . . , yt−p; θ) + εt, (1)

where g(·) is some parametric function, θ denotes a vector of


unknown parameters, and εt is a white noise “shock” (“news”).

So far, we only considered linear AR models, where


g(yt−1, yt−2, . . . , yt−p; θ) = α + φ1yt−1 + φ2yt−2 + · · · + φpyt−p,
possibly augmented with additional components for trend,
seasonality and outliers.

⇒ If we allow nonlinear specifications for g(yt−1, yt−2, . . . , yt−p; θ),


the number of possibilities is infinite.
For economic and financial time series, most successful nonlinear
time series models incorporate the idea of regime-switching.
Time Series Analysis: Chapter 8 2
Regime-switching

• What do we mean by regime-switching, and why might this be


relevant for economic and financial variables?

Intuitively: The “world” (“the economy”/“financial markets”) can


be in different states or regimes, with switches between different
regimes occurring on a regular basis.

Examples of regimes:

• business cycle regimes: expansions vs. recessions;


• “bull” vs. “bear” markets;
• periods of low vs. high volatility.

⇒ Economic and financial variables (may) behave differently in


different regimes.
Time Series Analysis: Chapter 8 3
US Unemployment

11

10

3
1960 1970 1980 1990 2000 2010

Monthly US unemployment rate

Time Series Analysis: Chapter 8 4


Regime-switching models

Regime-switching models allow the dynamic behaviour of


economic and financial variables to depend on the regime which
occurs at any given point in time.

• Suppose we want to describe/forecast a financial or economic


variable yt with xt = (1, yt−1, . . . , yt−p, x1t, . . . , xkt)0 where xjt,
j = 1, . . . , k are exogenous variables.

• All regime-switching models assume that within each regime, a


linear model can describe the relationship between yt and xt.
But, across regimes, the parameters in the linear model can be
different.

Time Series Analysis: Chapter 8 5


Regime-switching models

In case of 2 regimes:

yt = α1 + φ1,1yt−1 + . . . + φ1,pyt−p
+ β1,1x1,t + . . . + β1,k xk,t + εt if regime 1 occurs,
and

yt = α2 + φ2,1yt−1 + . . . + φ2,pyt−p
+ β2,1x1,t + . . . + β2,k xk,t + εt if regime 2 occurs.

Written more compactly:



φ0 x + ε if regime 1 occurs,
yt = 1 t t
(2)
φ0 xt + εt if regime 2 occurs,
2

where εt ∼ (0, σ 2).

Time Series Analysis: Chapter 8 6


Regime-switching models

• Model needs to be completed by specification of the way the


regime is determined at each point in time, or of the
“regime-switching” process.

All regime-switching models assume that regime-switching is


stochastic, that is, regimes that occur in the future are not
known with certainty.

Two main classes of models:

• Regimes can be characterized by an observable variable:


Threshold models and smooth transition models;
• Regimes are latent and cannot be characterized by an
observable variable: Markov-Switching model.

Time Series Analysis: Chapter 8 7


Threshold models

• The regimes can be characterized by the value of a certain


(observable) variable qt relative to a threshold c:

φ0 x + ε if qt ≤ c,
yt = 1 t t
(3)
φ0 xt + εt if qt > c,
2

where εt ∼ (0, σ 2).

• Note that the model can also be written as

yt = φ01xt(1 − I[qt > c]) + φ02xtI[qt > c] + εt, (4)


where I[A] = 1 if the event A occurs and I[A] = 0 otherwise.

Time Series Analysis: Chapter 8 8


Smooth transition models

• Crucial assumptions:
1. the regime that occurs at time t can be characterized by the
value of a certain (observable) variable qt, but
2. the “border” or threshold between the two regimes is not
“sharp” (as in the threshold model), but rather the transition
between two regimes is gradual or smooth:

The smooth transition model is given by

yt = φ01xt(1 − G(qt; γ, c)) + φ02xtG(qt; γ, c) + εt, (5)


where εt ∼ (0, σ 2) and G(qt; γ, c) is the logistic function
1
G(qt; γ, c) = , γ > 0. (6)
1 + exp(−γ[qt − c])

Time Series Analysis: Chapter 8 9


Logistic transition function

1.0

0.8

G(q(t);gamma,c)
0.6

0.4

gamma=1
0.2 gamma=2.5
gamma=5
gamma=25
0.0

-4 -3 -2 -1 0 1 2 3 4

q(t)

Examples of the logistic function G(qt; γ, c) as given in (6) for


various values of the smoothness parameter γ and threshold c = 0.

Time Series Analysis: Chapter 8 10


Smooth transition models

The smooth transition model

yt = φ01xt(1 − G(qt; γ, c)) + φ02xtG(qt; γ, c) + εt,


with the logistic function
1
G(qt; γ, c) = , γ > 0.
1 + exp(−γ[qt − c])
is such that:

• For very small values of qt, G(qt; γ, c) ≈ 0 and yt ≈ φ01xt + εt.


• For very large values of qt, G(qt; γ, c) ≈ 1 and yt ≈ φ02xt + εt.
• For intermediate values of qt, the parameters in the linear
model linking yt and xt are a linear combination of φ1 and φ2,
with the weights determined by the value of G(qt; γ, c).

Time Series Analysis: Chapter 8 11


Smooth transition models

The smooth transition model contains both threshold and linear


models as special cases:
1
G(qt; γ, c) = , γ > 0.
1 + exp(−γ[qt − c])

• As γ → ∞, G(qt; γ, c) → I[qt > c]: threshold model.


• As γ → 0, G(qt; γ, c) → 0.5 for all qt: linear model with
coefficients (φ1 + φ2)/2.

Time Series Analysis: Chapter 8 12


Threshold and smooth transition models

Threshold models and smooth transition models assume that


regime can be characterized by value of an observable variable qt.
This is convenient, because

• regimes which have occurred in past and present are known


with certainty,
• estimation of models and testing for regime-switching
behaviour is “not too difficult”.

Sometimes it might not be easy to obtain a “reliable”


regime-indicator qt.

Alternative: assume that qt does not exist, or, that regimes can
only be associated with an unobservable (stochastic) variable.

Time Series Analysis: Chapter 8 13


Markov-switching models

Alternative: assume that qt does not exist, or, that regimes can
not only be associated with an unobservable variable:

φ0 x + ε if St = 1,
yt = 1 t t
(7)
φ0 xt + εt if St = 2,
2
where St is unobserved.

• To complete the model, the properties of the process St need to


be specified (necessary for estimation, etc).

• Markov-Switching model: St is first-order Markov process


⇒ the current regime St only depends on the regime one period
ago, St−1.

Time Series Analysis: Chapter 8 14


Markov-switching models

Transition probabilities

Pr[St = 1|St−1 = 1] = p,
Pr[St = 2|St−1 = 1] = 1 − p,
Pr[St = 2|St−1 = 2] = q,
Pr[St = 1|St−1 = 2] = 1 − q.

⇒ 1 − p is equal to the probability that the Markov chain moves


from state 1 at time t − 1 to state 2 at time t or, put differently,
the probability that regime 1 at time t − 1 is followed by regime 2
at time t.

Time Series Analysis: Chapter 8 15


Markov-switching models

⇒ Important difference between threshold and Markov-Switching


models:

• In threshold models, the probability of a regime-switch varies


over time (depending on the properties of qt);
• In Markov-Switching models this probability is constant (and
determined by p and q).

⇒ But, p and q can be made time-varying by making them


functions of “duration” or of observable variables x1t, . . . , xkt.

Time Series Analysis: Chapter 8 16


Specification procedure for regime-switching models

“Specific-to-general”:
1. Specify a linear model (ARX) to describe yt in terms of
xt = (1, yt−1, . . . , yt−p, x1t, . . . , xkt)0)
2. (As part of model evaluation) Test linearity against threshold,
smooth transition and/or Markov-Switching alternatives.
3. If linearity is rejected, estimate parameters in the
regime-switching model;
4. Evaluate the model by applying misspecification tests
(parameter constancy, no remaining nonlinearity, ...) and other
diagnostic measures;
5. Modify the model if necessary;
6. Use the model for description or forecasting.

Time Series Analysis: Chapter 8 17


Specification procedure for regime-switching models

• Why specific-to-general? ⇒ Nonlinear models are more difficult


to work with than linear models (e.g. parameter estimation,
forecasting), hence only use nonlinear models if they are really
necessary!

Time Series Analysis: Chapter 8 18


Structural change

What do we mean by structural change?

⇒ Permanent change in statistical properties of time series


variables, such as its mean, variance or autocorrelations.

Why should we care about structural change?

⇒ Because time series analysis is based on the assumption of


stationarity: the constancy of means, variances and
autocorrelations...

Time Series Analysis: Chapter 8 19


Structural change in mean?

0
Feb
Mar

Feb
Mar
Jan

Jun

Jan

Jun
Jul

Jul
May

May
Aug
Sep

Nov
Dec

Aug
Sep

Nov
Dec
Apr

Apr
Oct

Oct
1999 2000 2001

Daily closing prices of AAPL stock


(31/12/1999 - 28/12/2001)

Time Series Analysis: Chapter 8 20


Structural change in trend?

4.8

4.4

4.0

3.6

3.2

2.8

2.4
1960 1965 1970 1975 1980 1985 1990 1995 2000

Quarterly industrial production Japan

Time Series Analysis: Chapter 8 21


Structural change in seasonality?

.6

.4

.2

.0

-.2

-.4
Q1
Q2
-.6 Q3
Q4
-.8
1985 1990 1995 2000 2005 2010 2015

Quarterly foreign tourist expenditures in the UK (growth rates).

Time Series Analysis: Chapter 8 22


Structural change in volatility?

15

10

-5

-10
1960 1970 1980 1990 2000 2010

Quarterly growth US real GDP


σ̂(59-83)=4.3; σ̂(84-07)=2.0; [σ̂(84-18)=2.3]

Time Series Analysis: Chapter 8 23


Modelling structural change

Structural change can take different forms, e.g. it could be


immediate, or take some time to take effect.

⇒ Here we focus on instantaneous change.

We can then consider many different types of structural change as


breaks in parameters in a regression model

φ0 x + ε t = 1, . . . , τ,
yt = 1 t t
(8)
φ0 xt + εt t = τ + 1, . . . , T,
2

where εt ∼ (0, σ 2). Parameters change at the breakdate τ .

For example, to describe a change in trend growth, take xt = (1, t).

Time Series Analysis: Chapter 8 24


Modelling structural change

Consider the regression model with a single structural change



φ0 x + ε t = 1, . . . , τ,
yt = 1 t t
(9)
φ0 xt + εt t = τ + 1, . . . , T,
2

where εt ∼ (0, σ 2).

(At least) two relevant questions:

1. How do we test whether a structural change really occurred?


2. How do we estimate the timing of structural change, that is,
the breakdate τ ?

Time Series Analysis: Chapter 8 25


Testing for structural change

If we assume the breakdate τ is known, or if we fix it a priori to a


certain date, testing for structural change is straightforward.

⇒ The null hypothesis of no structural change can then be tested


by means of the Chow test (Wald test) F (τ ),
σ̃ 2 − σ̂ 2(τ )
!
F (τ ) = T , (10)
σ̂ 2(τ )
where
1
PT
• σ̃ 2 is an estimate of the residual variance under H0 , that is σ̃ 2 = T
2
t=1 ε̃t
0
with ε̃t = yt − φ̃ xt .
• σ̂ 2 (τ ) an estimate of the residual variance under the alternative of a
structural break occurring at t = τ .

F (τ ) has asymptotic χ2-distribution with k degrees of freedom,


where k is the dimension of xt.
Time Series Analysis: Chapter 8 26
Testing for structural change

Possible drawbacks

• Result may be sensitive to the specified breakdate τ ;


• We may not have a clear idea of when a structural change has
occurred.

⇒ In general, we may want to treat the breakdate as unknown


when testing for the possible presence of a structural change.

Quandt (1960): take the maximum of the Chow statistics over all
possible breakdates, that is, the test statistic is

SupF = sup F (τ ).
τ
⇒ Problem: distribution of SupF under H0 is non-standard, and
remained unknown until the early 1990’s...
Time Series Analysis: Chapter 8 27
US Labor productivity

Empirical example (Hansen, 2001): Monthly observations on US


labor productivity in the manufacturing /durables sector, for the
period 1947.1-2001.4

Conventional wisdom: “US labor productivity experienced a


slowdown around 1973 and a speedup in the second half of the
1990s.”

Is this reflected in time series models for labor productivity growth?


Can structural change be detected by means of formal tests?

Consider tests for structural change in an AR(1) model for


monthly growth rate

yt = α + ρyt−1 + εt.

Time Series Analysis: Chapter 8 28


US Labor productivity

-7.6

-8.0

-8.4

-8.8

-9.2

-9.6

-10.0
50 55 60 65 70 75 80 85 90 95 00

Monthly (log) US labor productivity, 1947.1-2001.4

Time Series Analysis: Chapter 8 29


US Labor productivity

16

12

-4

-8
50 55 60 65 70 75 80 85 90 95 00

Annual growth US labor productivity, 1947.1-2001.4

Time Series Analysis: Chapter 8 30


US Labor productivity

24

20

16

12

0
50 55 60 65 70 75 80 85 90 95 00

Chow test sequence as a function of breakdate

Time Series Analysis: Chapter 8 31


Testing for structural change

The Quandt statistic is given by

SupF = sup F (τ ),
τ
where F (τ ) is a Wald (or LR or LM) statistic testing for a
structural break occurring at t = τ .

Each individual F (τ ) statistic has asymptotic χ2(k)-distribution.

Statistics F (τ1) and F (τ2) will be correlated, but exact form of


dependence is unclear.

⇒ This makes deriving the asymptotic distribution of SupF


complicated.

Time Series Analysis: Chapter 8 32


Testing for structural change

Alternative interpretation: Consider the problem of testing for a


single structural change, that is, the model under the alternative is

φ0 x + ε t = 1, . . . , τ,
yt = 1 t t
(11)
φ0 xt + εt t = τ + 1, . . . , T,
2

where εt ∼ (0, σ 2).

• The null hypothesis can be expressed as equality of the


parameters before and after the break, that is, H0 : φ1 = φ2.

Time Series Analysis: Chapter 8 33


Testing for structural change

Problem: “unidentified nuisance parameters under H0”

• Intuitively: the model under the alternative contains certain


parameters which are not restricted by the null hypothesis and are
not present in the model under the null.

Here, the unidentified parameter is the breakdate τ .

⇒ Consequence: (asymptotic) distribution of (LR, LM, Wald) test


statistics is non-standard.

Time Series Analysis: Chapter 8 34


Testing for structural change

Andrews (1993) derived the asymptotic distribution of

SupF = sup F (τ ),
λT ≤τ ≤(1−λ)T
where 0 < λ < 0.5 is called the trimming fraction.

⇒ In principle the supremum could be taken over all values of τ


such that k + 1 ≤ τ ≤ T − k. The idea of ‘trimming’ is that you
would like to get reliable/accurate parameter estimates both
before and after the break. Therefore, we impose a minimum
fraction λ of observations in both subsamples.

⇒ Distribution of SupF depends on the number of restrictions k


that are tested, as well as on the trimming fraction λ.

Time Series Analysis: Chapter 8 35


Estimating the timing of structural change

How do we estimate the parameters φ1, φ2 and τ in the structural


change model

φ0 x + ε t = 1, . . . , τ,
yt = 1 t t
(12)
φ0 xt + εt t = τ + 1, . . . , T,
2
and how do we obtain confidence intervals for the breakdate τ ?

⇒ We can apply least squares to minimize the residual variance

(τ̂ , φ̂1, φ̂2) = argmin σ 2(τ, φ1, φ2).

In practice, we do a ‘grid search’ over τ : we estimate φ1 and φ2


for all possible break-dates, compute σ 2(τ, φ1, φ2) and examine for
which breakdate this is minimized.

Time Series Analysis: Chapter 8 36


US Labor productivity

124.5

124.0

123.5

123.0

122.5

122.0

121.5

121.0

120.5

120.0
50 55 60 65 70 75 80 85 90 95 00

Residual variance as a function of breakdate

Time Series Analysis: Chapter 8 37


US Labor productivity

125

124

123

122

25 Residual variance 121


Chow test
20 120

15

10

0
50 55 60 65 70 75 80 85 90 95 00

Residual variance and Chow test as a function of breakdate

Time Series Analysis: Chapter 8 38


Estimating the timing of structural change

The structural change model



φ0 x + ε t = 1, . . . , τ,
yt = 1 t t
(13)
φ0 xt + εt t = τ + 1, . . . , T,
2
can also be written as

yt = φ01xtI[t ≤ τ ] + φ02xtI[t > τ ] + εt, (14)


where I[A] = 1 if the event A occurs and I[A] = 0 otherwise.

⇒ For fixed or known breakdate τ , the model is linear in the


remaining parameters φ1 and φ2.

Time Series Analysis: Chapter 8 39


Estimating the timing of structural change

Estimates φ̂(τ ) of φ = (φ01, φ02)0 are obtained by least squares


regression of yt on xt(τ ) = (x0tI[t ≤ τ ], x0tI[t > τ ])0, that is
 −1  
T T
0
X X
φ̂(τ ) =  xt(τ )xt(τ )   xt(τ )yt . (15)
t=1 t=1

Define the corresponding residuals ε̂t(τ ) = yt − φ̂(τ )0xt(τ ), which


2 1 PT
have variance σ̂ (τ ) = T t=1 ε̂t(τ )2.

Time Series Analysis: Chapter 8 40


Estimating the timing of structural change

An estimate of τ can be obtained by minimizing this residual


variance, that is
τ̂ = argmin σ̂ 2(τ ), (16)
λT ≤τ ≤(1−λ)T
where again we assume a ‘trimming fraction’ λ.

Time Series Analysis: Chapter 8 41


Threshold Models

• Consider the threshold model with two regimes:



φ0 x + ε if qt ≤ c,
yt = 1 t t
(17)
φ0 xt + εt if qt > c,
2

where εt ∼ (0, σ 2).

⇒ If qt = t, this is in fact a structural change model!!

Hence, for testing and estimation we can adopt the same methods
as used in the context of structural change.

Time Series Analysis: Chapter 8 42


Testing for nonlinearity

Consider the problem of testing the null hypothesis of linearity


against the alternative of a 2-regime threshold model,

φ0 x + ε if qt ≤ c,
yt = 1 t t
(18)
φ0 xt + εt if qt > c,
2

where εt ∼ (0, σ 2).

• The null hypothesis can be expressed as equality of the


conditional mean parameters in the two regimes, that is,

H0 : φ1 = φ2,
while the alternative hypothesis is

H1 : φ1 6= φ2.

Time Series Analysis: Chapter 8 43


Testing for nonlinearity

• This testing problem of course resembles the problem of testing


for structural change. The same complication of “unidentified
nuisance parameters under the null hypothesis” arises here if we
want to leave the threshold c unspecified.

⇒ Linearity can then be tested against a threshold model with


fixed c with a standard the “pointwise” Wald or F statistic F (c)

• If we want to leave c unspecified, we can use the SupF statistic.

Time Series Analysis: Chapter 8 44


Estimation of threshold models

Consider the 2-regime threshold model

yt = φ01xt(1 − I[qt > c]) + φ02xtI[qt > c] + εt.

⇒ For fixed or known threshold c, the model is linear in the


remaining parameters φ1 and φ2.

Estimates φ̂(c) of φ = (φ01, φ02)0 are obtained by least squares


regression of yt on xt(c) = (x0t(1 − I[qt > c]), x0tI[qt > c])0, that is
 −1  
T T
xt(c)xt(c)0
X X
φ̂(c) =   xt(c)yt . (19)
t=1 t=1

Define the corresponding residuals ε̂t(c) = yt − φ̂(c)0xt(c), which


2 1 PT
have variance σ̂ (c) = T t=1 ε̂t(c)2.
Time Series Analysis: Chapter 8 45
Estimation of threshold models

An estimate of c can be obtained by minimizing this residual


variance, that is
ĉ = argmin σ̂ 2(c), (20)
c∈C
where C denotes the set of all allowable threshold values.

For example, C may be defined such that each regime contains at


least a fraction λ of the observations, that is,

C = {c | q([λT ]) ≤ c ≤ q([(1−λ)T ])}. (21)

Time Series Analysis: Chapter 8 46


Estimation of threshold models

• Is this a feasible procedure, as the set C may contain a very large


number of possible threshold values?

Note: for given c, the estimates of φ1 (φ2 ) are based only on


those observations in the lower (upper) regime, i.e. those
observations for which qt ≤ c (qt > c).

As c is varied between two consecutive order statistics of qt, no


observations move from one regime to the other.
⇒ φ̂(c) does not change
⇒ the value of σ̂ 2(c) does not change

Hence, σ̂ 2(c) need only be computed for values of c equal to the


order statistics of qt.

Time Series Analysis: Chapter 8 47


Estimation of threshold models

-2

-4

-6

-8

-10
-6 -4 -2 0 2 4 6

Example of estimation of threshold model: varying the


threshold c between the two vertical lines does not change σ̂ 2(c).

Time Series Analysis: Chapter 8 48


Forecasting with regime-switching models

• Out-of-sample forecasting with regime-switching models is more


complicated than forecasting with linear models.

• Denote the point forecast of yT +h made at time T as ŷT +h|T


and the corresponding forecast error as eT +h|T , that is
eT +h|T = yT +h − ŷT +h|T . (22)

• If we want to minimize the mean squared prediction error


[MSPE]
MSPE(h) ≡ E[e2
T +h|T ] = E [(yT +h − ŷT +h|T ) 2 ]. (23)
the optimal forecast ŷT +h|T is the conditional expectation of
yT +h at time T ,
ŷT +h|T = E[yT +h|YT ], (24)
where YT denotes the information set at time T .
Time Series Analysis: Chapter 8 49
Forecasting with regime-switching models

Example: AR(1) model


yt = φ1yt−1 + εt, (25)
with E[εt|Yt−1] = 0 and E[ε2 2
t |Yt−1 ] = σ .

• The 1-step ahead forecast at time T is given by


ŷT +1|T = E[yT +1|YT ]
= E[φ1yT + εT +1|YT ]
= φ1yT .

• In general, the h-step ahead forecast can be computed directly as


ŷT +h|T = φh
1 yT ,
or recursively as
ŷT +h|T = φ1ŷT +h−1|T .

Time Series Analysis: Chapter 8 50


Forecasting with regime-switching models

Example: Threshold model



−2 + ε if yt−1 ≤ 0,
t
yt = (26)
 2 + εt if yt−1 > 0,

with E[εt|Yt−1] = 0 and E[ε2


t |Y t−1 ] = σ 2 , or

yt = −2I(yt−1 ≤ 0) + 2I(yt−1 > 0) + εt.

Suppose yT = −1, then the 1-step ahead forecast is equal to

ŷT +1|T = E[yT +1|YT ] = −2.

Time Series Analysis: Chapter 8 51


Forecasting with regime-switching models

• For the 2-step ahead forecast, it follows that:


ŷT +2|T = E[yT +2|YT ]
= E[−2I(yT +1 ≤ 0) + 2I(yT +1 > 0)|YT ]
= −2P (yT +1 ≤ 0|YT ) + 2P (yT +1 > 0|YT )
= −2P (−2 + εT +1 ≤ 0) + 2P (−2 + εT +1 > 0)
= −2P (εT +1 ≤ 2) + 2P (εT +1 > 2),
because yT < 0 (such that yT +1 = −2 + εT +1), and
E[I(A)] = P (A).

• Computing the 3-step ahead forecast is even more complicated:


ŷT +3|T = E[yT +3|YT ]
= E[−2I(yT +2 ≤ 0) + 2I(yT +2 > 0)|YT ]
= −2P (yT +2 ≤ 0|YT ) + 2P (yT +2 > 0|YT ),
because P (yT +2 > 0|YT ) is difficult to compute.
Time Series Analysis: Chapter 8 52
Forecasting with regime-switching models

Notice that:

ŷT +2|T = E[yT +2|YT ]


= E[−2I(yT +1 ≤ 0) + 2I(yT +1 > 0)|YT ]
6= −2I(E[yT +1|YT ] ≤ 0) + 2I(E[yT +1|YT ] > 0)
= −2I(ŷT +1|T ≤ 0) + 2I(ŷT +1|T > 0)
= −2
Hence, it appears that in this case, multiple step ahead forecasts
cannot be computed recursively...

Time Series Analysis: Chapter 8 53


Forecasting with regime-switching models

• The reason the above computation fails is that the expected


value of a nonlinear function is not equal to the function evaluated
at the expected value of its argument, that is

E[g(x)] 6= g(E[x]).

⇒ Example: g(x) = x2 and x uniformly distributed on the interval


[0, 1]. Then E[x] = 0.5, so g(E[x]) = 0.25, but E[g(x)] = 1/3.

⇒ As a consequence, multiple step ahead forecasts can not be


computed recursively (as in the linear AR model).

In addition, analytic expressions for multiple step ahead forecasts


are not available...

Time Series Analysis: Chapter 8 54


Forecasting with regime-switching models

• For the general nonlinear autoregressive model of order 1,

yt = g(yt−1; θ) + εt, (27)


the 2-step ahead forecast cannot be computed as

ŷT +2|T = E[yT +2|YT ] = E[g(yT +1; θ) + εT +2|YT ]


= E[g(yT +1; θ)|YT ]
6= g(E[yT +1|YT ]; θ) = g(ŷT +1|T ; θ).

• The relation between 1- and 2-step ahead forecasts is given by

ŷT +2|T = E[yT +2|YT ] = E[g(yT +1; θ) + εT +2|YT ]


= E[g(g(yT ; θ) + εT +1; θ)|YT ]
= E[g(ŷT +1|T + εT +1; θ)|YT ].

Time Series Analysis: Chapter 8 55


Forecasting with regime-switching models

• There are several different ways to compute multiple step ahead


forecasts from nonlinear models - the easiest of which is using
simulation:
k
(mc) 1 X
ŷT +2|T = g(ŷT +1|T + εi; θ). (28)
k i=1

⇒ εi can be drawn either from an assumed distribution, e.g.


normal (“Monte Carlo”), or from the residuals of the estimated
model (“bootstrap”).

⇒ easy to compute!

Time Series Analysis: Chapter 8 56

You might also like