Forecasting Define yt+h|t as the forecast of yt+h based on It known

parameters. The forecast error is

Let {yt} be a covariance stationary are ergodic process,
εt+h|t = yt+h − yt+h|t
e.g. an ARMA(p, q) process with Wold representation
∞ and the mean squared error of the forecast is
X
yt = μ + ψ j εt−j , εt ~W N (0, σ 2)
j=0 M SE(εt+h|t) = E[ε2t+h|t]
= μ + εt + ψ 1εt−1 + ψ 2εt−2 + · · · = E[(yt+h − yt+h|t)2]
Let It = {yt, yt−1, . . .} denote the information set avail- Theorem: The minimum MSE forecast (best forecast)
able at time t. Recall, of yt+h based on It is
E[yt] = μ yt+h|t = E[yt+h|It]
∞
X
var(yt) = σ 2 ψ 2j Proof: See Hamilton pages 72-73.
j=0
Goal: Using It produce optimal forecasts of yt+h for
h = 1, 2, . . . , s

Note:

yt+h = μ + εt+h + ψ 1εt+h−1 + · · ·

+ψ h−1εt+1 + ψ hεt + ψ h+1εt−1 + · · ·
Remarks Linear Predictors

1. The computation of E[yt+h|It] depends on the distri- A linear predictor of yt+h|t is a linear function of the
bution of {εt} and may be a very complicated nonlinear variables in It.
function of the history of {εt}. Even if {εt} is an un-
Theorem: The minimum MSE linear forecast (best linear
correlated process (e.g. white noise) it may be the case
predictor) of yt+h based on It is
that
yt+h|t = μ + ψ hεt + ψ h+1εt−1 + · · ·
E[εt+1|It] 6= 0

Proof. See Hamilton page 74.

2. If {εt} is independent white noise, then E[εt+1|It] =
0 and E[yt+h|It] will be a simple linear function of {εt} The forecast error of the best linear predictor is

yt+h|t = μ + ψ hεt + ψ h+1εt−1 + · · · εt+h|t = yt+h − yt+h|t

= μ + εt+h + ψ 1εt+h−1 + · · ·
+ψ h−1εt+1 + ψ hεt + · · ·
−(μ + ψ hεt + ψ h+1εt−1 + · · · )
= εt+h + ψ 1εt+h−1 + · · · + ψ h−1εt+1
and the MSE of the forecast error is
MSE(εt+h|t) = σ 2(1 + ψ 21 + · · · + ψ 2h−1)
Remarks Example: BLP for MA(1) process

yt = μ + εt + θεt−1, εt ∼ WN(0, σ 2)
1. E[εt+h|t] = 0
Here
2. εt+h|t is uncorrelated with any element in It ψ 1 = θ, ψ h = 0 for h > 1
Therefore,
3. The form of yt+h|t is closely related to the IRF
yt+1|t = μ + θεt
4. M SE(εt+h|t) = var(εt+h|t) ≤ var(yt) yt+2|t = μ
yt+h|t = μ for h > 1
5. limh→∞ yt+h|t = μ The forecast errors and MSEs are

6. limh→∞ MSE(εt+h|t) = var(yt) εt+1|t = εt+1, MSE(εt+1|t) = σ 2

εt+2|t = εt+2 + θεt+1, MSE(εt+2|t) = σ 2(1 + θ2)
Prediction Confidence Intervals Predictions with Estimated Parameters

If {εt} is Gaussian then Let ŷt+h|t denote the BLP with estimated parameters:

yt+h|It ∼ N (yt+h|t, σ 2(1 + ψ 21 + · · · + ψ 2h−1)) ŷt+h|t = μ̂ + ψ̂ hε̂t + ψ̂ h+1ε̂t−1 + · · ·

A 95% confidence interval for the h−step prediction has where ε̂t is the estimated residual from the fitted model.
the form The forecast error with estimated parameters is
q
yt+h|t ± 1.96 · σ 2(1 + ψ 21 + · · · + ψ 2h−1) ε̂t+h|t = yt+h − ŷt+h|t
= (μ − μ̂) + εt+h + ψ 1εt+h−1 + · · · + ψ h−1εt+1
³ ´ ³ ´
+ ψ hεt − ψ̂ hε̂t + ψ h+1εt−1 − ψ̂ h+1ε̂t−1
+···
Obviously,

MSE(ε̂t+h|t) 6= MSE(εt+h|t) = σ 2(1+ψ 21+· · ·+ψ 2h−1)

Note: Most software computes
d 2 2 2
MSE(εt+h|t) = σ̂ (1 + ψ̂ 1 + · · · + ψ̂ h−1)
Computing the Best Linear Predictor The best linear forecasts of yt+1, yt+2, . . . , yt+h are com-
puted using the chain-rule of forecasting (law of iterated
The BLP yt+h|t may be computed in many different but projections)
equivalent ways. The algorithm for computing yt+h|t
yt+1|t = μ + φ(yt − μ)
from an AR(1) model is simple and the methodology al-
yt+2|t = μ + φ(yt+1|t − μ) = μ + φ(φ(yt − μ))
lows for the computation of forecasts for general ARMA
models as well as multivariate models. = μ + φ2(yt − μ)
..

Example: AR(1) Model yt+h|t = μ + φ(yt+h−1|t − μ) = μ + φh(yt − μ)

yt − μ = φ(yt−1 − μ) + εt The corresponding forecast errors are

εt ~W N (0, σ 2) εt+1|t = yt+1 − yt+1|t = εt+1
μ, φ, σ 2 are known εt+2|t = yt+2 − yt+2|t = εt+2 + φεt+1
In the Wold representation ψ j = φj . Starting at t and = εt+2 + ψ 1εt+1
iterating forward h periods gives ..
εt+h|t = yt+h − yt+h|t = εt+h + φεt+h−1 + · · ·
yt+h = μ + φh(yt − μ) + εt+h + φεt+h−1 + · · ·
+φh−1εt+1 +φh−1εt+1
= μ + φh(yt − μ) + εt+h + ψ 1εt+h−1 + · · · = εt+h + ψ 1εt+h−1 + · · · + ψ h−1εt+1
+ψ h−1εt+1
The forecast error variances are AR(p) Models

var(εt+1|t) = σ 2
Consider the AR(p) model
var(εt+2|t) = σ 2(1 + φ2) = σ 2(1 + ψ 21)
.. φ(L)(yt − μ) = εt, εt ∼ W N (0, σ 2)
2 2 2(h−1) 2 1 − φ2h φ(L) = 1 − φ1L − · · · φpLp
var(εt+h|t) = σ (1 + φ + · · · + φ )=σ 2 1−φ The forecasting algorithm for the AR(p) models is essen-
= σ 2(1 + ψ 21 + · · · + ψ 2h−1) tially the same as that for AR(1) models once we put the
Clearly, AR(p) model in state space form. Let Xt = yt − μ. The
AR(p) in state space form is
lim yt+h|t = μ = E[yt] ⎛ ⎞ ⎛ ⎞⎛ ⎞ ⎛ ⎞
h→∞
Xt φ1 φ2 · · · φp Xt−1 εt
σ2 ⎜ Xt−1 ⎟ ⎜ ⎟⎜ X ⎟ ⎜ 0 ⎟
lim var(εt+h|t) = ⎜ ⎟ ⎜ 1 0 · · · 0 ⎟ ⎜ t−2 ⎟ ⎜ ⎟
⎜ .. ⎟=⎜ .. ⎟ ⎜ .. ⎟+⎜ .. ⎟
h→∞ 1 − φ2 ⎝ ⎠ ⎝ ... ⎠⎝ ⎠ ⎝ ⎠
∞
X Xt−p+1 0 1 0 Xt−p 0
= σ2 ψ 2h = var(yt)
h=0 or

ξt = Fξt−1+wt
var(wt) = Σw
Starting at t and iterating forward h periods gives The forecast errors are given by

ξt+h = Fhξt + wt+h + Fwt+h−1 + · · · + Fh−1wt+1 wt+1|t = ξt+1 − ξt+1|t = wt+1

Then the best linear forecasts of yt+1, yt+2, . . . , yt+h wt+2|t = ξt+2 − ξt+2|t = wt+2 + Fwt+1
..
are computed using the chain-rule of forecasting are
wt+h|t = ξt+h − ξt+h|t = wt+h + Fwt+h−1 + · · ·
ξt+1|t = Fξt
+Fh−1wt+1
ξt+2|t = Fξt+1|t= F2ξt
.. and the corresponding forecast MSE matrices are
ξt+h|t = Fξt+h−1|t= Fhξt var(wt+1|t) = var(wt) = Σw
The forecast for yt+h is given by μ plus the first row of var(wt+2|t) = var(wt+2) + Fvar(wt+1)F0
ξt+h|t = Fhξt : = Σw + FΣw F0
..
⎛ ⎞h ⎛ ⎞
φ1 φ2 · · · φp yt − μ h−1
X 0
⎜ ⎟ ⎜ y ⎟ var(wt+h|t) = Fj Σw Fj
1 0 · · · 0 ⎟ ⎜ t−1 − μ
ξt+h|t = ⎜
⎜ ... .. ⎟ ⎜ ..
⎟
⎟ j=0
⎝ ⎠ ⎝ ⎠
0 1 0 yt−p+1 − μ Notice that

var(wt+h|t) = Σw + Fvar(wt+h−1|t)F0
Forecast Evaluation Diebold-Mariano Test for Equal Predictive Accuracy

1
Let {yt} denote the series to be forecast and let yt+h|t
2
and yt+h|t denote two competing forecasts of yt+h based
1
on It. For example, yt+h|t could be computed from an
2
AR(p) model and yt+h|t could be computed from an
ARMA(p, q) model. The forecast errors from the two
models are

ε1t+h|t = yt+h − yt+h|t

1

ε2t+h|t = yt+h − yt+h|t

2

The h−step forecasts are assumed to be computed for

t = t0, . . . , T for a total of T0 forecasts giving

{ε1t+h|t}T 2 T
t0 , {εt+h|t}t0
Because the h-step forecasts use overlapping data the
forecast errors in {ε1t+h|t}T 2 T
t0 and {εt+h|t}t0 will be seri-
ally correlated.
The accuracy of each forecast is measured by a particular The Diebold-Mariano test is based on the loss differential
loss function
dt = L(ε1t+h|t) − L(ε2t+h|t)
i
L(yt+h, yt+h|t ) = L(εit+h|t), i = 1, 2 The null of equal predictive accuracy is then
Some popular loss functions are:
H0 : E[dt] = 0
³ ´2
L(εit+h|t) = εit+h|t : squared error loss The Diebold-Mariano test statistic is
¯ ¯
¯ ¯
L(εit+h|t) = ¯εit+h|t¯ : absolute value loss d¯ d¯
S=³ ´1/2 = ³ ´
To determine if one model predicts better than another d d)
avar( ¯ d /T 1/2
LRV d¯
we may test null hypotheses
where
H0 : E[L(ε1t+h|t)] = E[L(ε2t+h|t)] 1 X T
d¯ = dt
against the alternative T0 t=t
0
∞
X
H1 : E[L(ε1t+h|t)] 6= E[L(ε2t+h|t)] LRVd¯ = γ 0 + 2 γ j , γ j = cov(dt, dt−j )
j=1
Note: The long-run variance is used in the statistic be-
cause the sample of loss differentials {dt}T
t0 are serially
correlated for h > 1.
Diebold and Mariano (1995) show that under the null of
equal predictive accuracy
A
S ~ N (0, 1)
So we reject the null of equal predictive accuracy at the
5% level if
|S| > 1.96
One sided tests may also be computed.