8:15 a.m.
Dear all,
I am not sure this behavior is intended, but to me it looks some
unintended behavior.
In the following example one of the scalars in the vector of realized
values y is zero which drives the division "(e./y)=-inf" and which
affects the later sum calculation as well. I guess gretl should ignore
this "-inf"-value, or do I miss anything here?
<hansl>
matrix y = {0; 1}
matrix f = {0.5;2}
matrix e = y-f
eval fcstats(y,f)
# MPE by hand
eval e./y
mat = (0.5).*sumc(100*e./y )
mat
</hansl>
Best,
Artur
9:11 a.m.
Am 15.04.2016 um 10:15 schrieb Artur T.:
I am not sure this behavior is intended, but to me it looks some
unintended behavior.
In the following example one of the scalars in the vector of realized
values y is zero which drives the division "(e./y)=-inf" and which
affects the later sum calculation as well. I guess gretl should ignore
this "-inf"-value, or do I miss anything here?
I've had the same problem in the past, but my conclusion was that it's
basically unavoidable given the definitions of some of the FC eval measures.
I cooked up my own alternative forecast measure function back then,
which I could share (after I've found it).
cheers,
sven
9:28 a.m.
It would be fine if you would share it. Thanks in advance, Sven.
Artur
Am 15.04.2016 um 11:11 schrieb Sven Schreiber:
Am 15.04.2016 um 10:15 schrieb Artur T.:
> I am not sure this behavior is intended, but to me it looks some
> unintended behavior.
>
> In the following example one of the scalars in the vector of realized
> values y is zero which drives the division "(e./y)=-inf" and which
> affects the later sum calculation as well. I guess gretl should ignore
> this "-inf"-value, or do I miss anything here?
>
I've had the same problem in the past, but my conclusion was that it's
basically unavoidable given the definitions of some of the FC eval
measures.
I cooked up my own alternative forecast measure function back then,
which I could share (after I've found it).
cheers,
sven
9:44 a.m.
Am 15.04.2016 um 11:28 schrieb Artur T.:
It would be fine if you would share it. Thanks in advance, Sven.
I hope this is what we're talking about:
<hansl>
function matrix otherfcstats(series y, series fc)
list all = y fc
if sum(missing(all))
print "--otherfcstats(): found missings, aborting"
return {1}
endif
matrix mtemp = fcstats(y,fc)
# Root mean square error, mean abs. error
matrix out = sqrt(mtemp[2]) | mtemp[3]
# now the manual TheilU calculation
matrix my = {y}
matrix mfc = {fc}
numerator = sqrt(mean((my-mfc).^2))
denom = sqrt(mean(mfc.^2)) + sqrt(mean(my.^2))
matrix out |= numerator/denom
return out
end function
</hansl>
hth,
sven
9:47 a.m.
Thank you Sven!
I also wrote down the following little function as an attempt to
circumvent the zero-division problem.
<hansl>
function matrix FcStat(series y, series fc)
list all = y fc
if sum(missing(all))
print "found missings, aborting"
return {1}
endif
set warnings off
matrix out = zeros(6,1)
matrix mtemp = fcstats(y,fc)
out[1:3] = mtemp[1:3] # These criteria always work fine
series e = y-fc
# MPE
series ey = e/y
ey = misszero(ey)
matrix out[4] = 100*meanc({ey}) # contains only valid values
# MAPE
series ey = abs(e)/y
ey = misszero(ey)
matrix out[5] = 100*meanc({ey})
# Theil's U
series num = (e/y(-1))^2
num = misszero(num)
num[1] = NA # 1st obs. is NA
scalar mnum = mean({num})
series den = ( (y-y(-1)) / y )^2
den = misszero(den)
mden = mean({den})
matrix out[6] = mnum/mden
return out
end function
</hansl>s
Am 15.04.2016 um 11:44 schrieb Sven Schreiber:
Am 15.04.2016 um 11:28 schrieb Artur T.:
> It would be fine if you would share it. Thanks in advance, Sven.
>
I hope this is what we're talking about:
<hansl>
function matrix otherfcstats(series y, series fc)
list all = y fc
if sum(missing(all))
print "--otherfcstats(): found missings, aborting"
return {1}
endif
matrix mtemp = fcstats(y,fc)
# Root mean square error, mean abs. error
matrix out = sqrt(mtemp[2]) | mtemp[3]
# now the manual TheilU calculation
matrix my = {y}
matrix mfc = {fc}
numerator = sqrt(mean((my-mfc).^2))
denom = sqrt(mean(mfc.^2)) + sqrt(mean(my.^2))
matrix out |= numerator/denom
return out
end function
</hansl>
hth,
sven
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
10:06 a.m.
Actually, I just see that their is some error in the function.
Artur
Am 16.04.2016 um 11:47 schrieb Artur T.:
Thank you Sven!
I also wrote down the following little function as an attempt to
circumvent the zero-division problem.
<hansl>
function matrix FcStat(series y, series fc)
list all = y fc
if sum(missing(all))
print "found missings, aborting"
return {1}
endif
set warnings off
matrix out = zeros(6,1)
matrix mtemp = fcstats(y,fc)
out[1:3] = mtemp[1:3] # These criteria always work fine
series e = y-fc
# MPE
series ey = e/y
ey = misszero(ey)
matrix out[4] = 100*meanc({ey}) # contains only valid values
# MAPE
series ey = abs(e)/y
ey = misszero(ey)
matrix out[5] = 100*meanc({ey})
# Theil's U
series num = (e/y(-1))^2
num = misszero(num)
num[1] = NA # 1st obs. is NA
scalar mnum = mean({num})
series den = ( (y-y(-1)) / y )^2
den = misszero(den)
mden = mean({den})
matrix out[6] = mnum/mden
return out
end function
</hansl>s
Am 15.04.2016 um 11:44 schrieb Sven Schreiber:
> Am 15.04.2016 um 11:28 schrieb Artur T.:
>> It would be fine if you would share it. Thanks in advance, Sven.
>>
>
> I hope this is what we're talking about:
>
> <hansl>
> function matrix otherfcstats(series y, series fc)
> list all = y fc
> if sum(missing(all))
> print "--otherfcstats(): found missings, aborting"
> return {1}
> endif
>
> matrix mtemp = fcstats(y,fc)
> # Root mean square error, mean abs. error
> matrix out = sqrt(mtemp[2]) | mtemp[3]
>
> # now the manual TheilU calculation
> matrix my = {y}
> matrix mfc = {fc}
> numerator = sqrt(mean((my-mfc).^2))
> denom = sqrt(mean(mfc.^2)) + sqrt(mean(my.^2))
> matrix out |= numerator/denom
>
> return out
> end function
> </hansl>
>
> hth,
> sven
> _______________________________________________
> Gretl-users mailing list
> Gretl-users(a)lists.wfu.edu
> http://lists.wfu.edu/mailman/listinfo/gretl-users
10:30 a.m.
For those who are interested in:
<hansl>
function matrix FcStat(series y, series fc)
list all = y fc
if sum(missing(all))
print "--FcStat(): found missings, aborting"
return {1}
endif
set warnings off
matrix out = zeros(6,1)
matrix mtemp = fcstats(y,fc)
out[1:3] = mtemp[1:3] # These criteria always work fine
series e = y-fc
# MPE
series ey = e/y
ey = misszero(ey)
out[4] = meanc(100*{ey}) # contains only valid values
# MAPE
series ey = abs(e/y)
ey = misszero(ey)
out[5] = meanc(100*{ey})
# Theil's U
scalar T = nobs(fc)
series num = ( (fc-y)/y(-1) )^2
num = misszero(num)
num[1] = NA
scalar mnum = sqrt( (1/T)*sumc({num}) )
series den = ( (y-y(-1)) / y(-1) )^2
den = misszero(den)
mden = sqrt( (1/T)*sumc({den}) )
out[6] = mnum/mden
return out
end function
</hansl>
Artur
Am 16.04.2016 um 12:06 schrieb Artur T.:
Actually, I just see that their is some error in the function.
3236
days inactive
3237
days old
6 comments
2 participants
participants (2)
-
Artur T. -
Sven Schreiber