Linear Combination of Forecasts -- A Comment
-
Upload
simon-french -
Category
Documents
-
view
220 -
download
0
Transcript of Linear Combination of Forecasts -- A Comment
Linear Combination of Forecasts -- A CommentAuthor(s): Simon FrenchSource: The Journal of the Operational Research Society, Vol. 32, No. 10 (Oct., 1981), pp. 937-938Published by: Palgrave Macmillan Journals on behalf of the Operational Research SocietyStable URL: http://www.jstor.org/stable/2581241 .
Accessed: 25/06/2014 08:55
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .http://www.jstor.org/page/info/about/policies/terms.jsp
.JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range ofcontent in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new formsof scholarship. For more information about JSTOR, please contact [email protected].
.
Palgrave Macmillan Journals and Operational Research Society are collaborating with JSTOR to digitize,preserve and extend access to The Journal of the Operational Research Society.
http://www.jstor.org
This content downloaded from 62.122.78.91 on Wed, 25 Jun 2014 08:55:29 AMAll use subject to JSTOR Terms and Conditions
Letters and Viewpoints
In applying the simplified computational scheme, we first compute T^ = Tf2 =0.05
years, T^i = T\2 = 0.1 years, T\^ = 0.11547 years, Tf21 = 0.06124 years, Tfn =
0.1225 years, Tf21 = 0.0707 years and, Tmin = 0.05 years. Other results and the two iterations are given in Table 1.
Table 1. Results of the simplified solution
REFERENCE 1 M. G. Korgaonker (1979) Integrated production inventory policies for multistage multiproduct batch production systems. J. Opl Res Soc. 30, 355-362.
Diosd, Ifjusdg U.8/a.2 Hany A. Makroum
2049?Hungary
LINEAR COMBINATION OF FORECAST&-A COMMENT
Readers of JORS will know that Bunn and I disagree quite sharply about the method?
ology of combining distinct forecasts into a composite forecast.1,2 Even though I am still
unconvinced by Bunn's arguments, I had decided not to continue the discussion further; debates about methodology can decay quickly into entrenched restatements of the
opposing positions. However, in a recent paper3 Bunn again compares his outranking
approach to combination of forecasts with the Bayesian or, as he terms it, veridical
approach. I have no intention of commenting upon the methodological and philosophi? cal points in his discussion; but he does report the results of a simulation study in which
he shows that his method performs signifieantly better than the Bayesian. Although it is
quite possible to remain unconvinced by a battery of methodological arguments, it is an
act of extreme bigotry to ignore the results of a valid simulation study. Indeed, I assure
Bunn that I will adopt his forecasting method and stop championing the Bayesian?if he
can convince me that his simulation study is valid. However, from my reading of his
paper I believe that it contains a significant flaw.
It will be helpful, I think, to return to Bates and Granger's original suggestion for the
combination of forecasts.4 Suppose that we are concerned with a time series
Yu Y2,.., 7t,... For the present assume that we are considering the series from time
zero when none of the Ts have been observed. We have available two forecasting
methods/S1} and/<2); i.e.f\i){Y1, Y2,..,Yt) (i = 1,2) are forecasts of Yt+1. The forecast
errors at time (t + 1) are
e\%i = Yt+,-f?{YuY2,..,Yt). (i=l,2) (1)
Now for t ^ 1, but considering the knowledge available at time zero, there are at least
two unobserved random variables, viz. Yu Y2,.., Yt+ u involved in the definition of e{t+\
and e{2+\. Thus it is conceptually possible for the two forecast errors e{^i and e{2+\ to be
mutually independent. It is this assumption of mutual independence of forecast errors
that Bates and Granger make in deducing their formula for the weighted combination of
forecasts. (See Bunn3 equations (1) and (2).)
937
This content downloaded from 62.122.78.91 on Wed, 25 Jun 2014 08:55:29 AMAll use subject to JSTOR Terms and Conditions
Journal ofthe Operational Research Society Vol. 32, No. 10
However, now consider the forecasts of Yt+1 when we are at time t, i.e. when we have observed Y^ = yu Y2 = y2,..,Yt = yt. The one step ahead forecast errors are:
etlli = Yt+1-f^{yuy29..9yt)
e$2=Yt+1-f2\yuy2,..,yt). (2)
Note that there is only one unobserved random variable in their definition, namely Yt+1. Thus efti and e{2^ cannot be independent conditional on the information available at time t. In fact, ej!^ and e^ are totally dependent random variables. (To see this, suppose that e(ti\ is known, then e{2+\ = eH\ + /|1)(}>i,}>2,.. ,}>,) -f\2)(yi,y2,-.,yt) ? also known exactly.) In other words the random variables, which may be independent conditional on the information at time zero, must be fully dependent conditional on the information at time t.
This point is vital to Bunn's simulation study. He describes this in two sentences
(Bunn,3 p. 218).
The forecaster has two predictors available which need not be modelled explicitly since the combinations are only a function of their forecast errors. He assumes these one-step ahead forecast errors to be independently normally distributed with unknown mean and variances.
If by this Bunn means that he has generated efti and e^^ as independent conditional on the information at time t, then, as we have seen above, he has generated an impossible stream of data. He would have effectively told the methods that Yt+1 was realised both as the quantity (ej+i +/?1)()>i,J>2> ? ?>)>?)) and als? as the quantity (e{2+\ +f{t2)(yuy2,..,yn)\ quantities that with probability 1 cannot be equal. Such con?
tradictory information would naturally confuse the Bayesian forecasting system and lead to its poor performance. Indeed, I would be worried about any forecasting method that did not become confused by such impossible data.
On the other hand Bunn's assumption of independence may mean that sequential errors, i.e. e\ll2 and e\%u are generated independently. However, I cannot conceive of how this may be done without specifying the predictors explicitly; for then he would be
generating the Y's and calculating the forecast errors from (2).
Perhaps Bunn would explain his simulation study in a little more detail than given in his paper. I can assure him that if he can convince me that Bayesian forecasting performs very poorly on a sensible stream of data, then I shall cease to advocate it.
Department of Decision Theory Simon French
University of Manchester
REFERENCES
1S. French and D. W. Bunn (1980) Outranking probabilities and the synthesis of forecasts. J. Opl Res. Soc. 31, 545-551.
2R. I. Phelps, S. French and D. W. Bunn (1981) Forecast Combination. J. Opl Res. Soc. 32, 68-70. 3D. W. Bunn (1981) Two methodologies for the linear combination of forecasts. J. Opl Res. Soc. 32, 213-222. 4M. Bates and C. W. J. Granger (1969) The combination of forecasts. Opl Res. Q. 20, 451-468.
938
This content downloaded from 62.122.78.91 on Wed, 25 Jun 2014 08:55:29 AMAll use subject to JSTOR Terms and Conditions