Beruflich Dokumente
Kultur Dokumente
Appendix
Notation
Actual series: yt , forecast for period t: yt
Pf
Pf
Average of yt , y = f1 t=1 yt , average of forecast series yt , y = f1 t=1 yt ,
q P
f
Standard error of yt : sy = f1 t=1 (yt y)2 , standard error of yt : sy =
q P
f
1
yt y)2
t=1 (
f
Forecast error: F Et = yt yt
Pf
Bias: BIAS = f1 t=1 F Et =
1
f
q P
f
1
Pf
yt
t=1 (
t=1 (F Et
1
f
Pf
t=1
1
f
yt ) =
Pf
1
f
Pf
t=1
yt
1
f
BIAS)2
F Et2
t=1
M SE =
q P
f
1
t=1
F Et2
|F Et |
1
f
Pf
t=1
| FyEt t |
f
1X
(F Et BIAS)2
f t=1
f
1X
(F Et2 2BIAS F Et + BIAS 2 )
f t=1
f
f
f
1X
1X
1X
F Et2
2BIAS F Et +
BIAS 2
f t=1
f t=1
f t=1
Pf
t=1
yt
= M SE 2BIAS
f
1X
F Et + BIAS 2
f t=1
= M SE BIAS 2
=
=
Cov(y
yt )
t ,
sy sy
1
f
Pf
t=1
(yt
y )(
yt y
)
sy sy
=
=
f
f
f
1X
1X
1X
(yt y)2 +
(
yt y)2 2
(yt y)(
yt y) =
f t=1
f t=1
f t=1
f
o
1 Xn
(yt y)2 + (
yt y)2 2(yt y)(
yt y)
f t=1
f
o2
1 Xn
(yt y) (
yt y)
f t=1
f
o2
1 Xn
(
y y) (
yt yt )
f t=1
f
1X
(BIAS F Et )2 = SE 2
f t=1
Therefore bias proportion plus variance proportion plus covariance proportion sum uip to unity.
larger than for the second. In other words, if one makes his judgement solely
based on U1 statistic, she will incorrectly prefer the inferior forecast.
Suppose we know that the actual series yt are distributed with zero mean
and variance 2 . Consider the following forecasts for this test period:
1. Set yt = 0 for t = 1..f . Then
q P
f
1
U1
q P
f
1
yt )2
q P
t=1 (yt
t=1
t=1 t
t=1 yt
f
f
2 2 0 + 2
2
=
<1
2
2
2
+
Clearly, the second forecast is worse
the first. The root
squared error
qthan
q mean
Pf
Pf
1
1
2
of the first forecast is RM SE1 = f t=1 (yt yt ) = f t=1 yt 2 . For
q P
f
the second forecast RM SE2 = f1 t=1 (yt yt )2 2 > RM SE1 (see the
numerator of the second U statistic). At the same time the first forecast recieves
a worse ranking based on the U1 statistic.