Just note! The stuff I gave is not the 'official' Kalman filter
and will probably be frowned upon by say a NASA engineer. But it
is something similar and we use such devices in our own work. We
sometimes also smooth the weights, so that one formula
w1 = w1 + ...
becomes
w1 = 0.9 * w1 + 0.1 * ( w1 + ... )
and we play around with the 0.9 amd 0.1 They should always sum to
1, and the higher the 0.9, the more smooth the weights will be. This
is just an exponential moving average....
For the rest, see below.
I am interested on Kalman filter too, in term of training periods
for the net, do you have any suggestion on how to determine the
length of periods? If the period is too long, then could it be
overtrained? If I can determine it, then the periods for rescaling
based on HHV(Abs(S),periods) could be found, will it be a good
approach?
OK, we are not really training a net here, and even if we were, it
is arbitrary. Also see my comment on the next one. I suggest you
plot your indicator and visually inspect it. If it makes cycles,
see how long it takes to make one full cycle. Hopefully it is not
several years, but say 1 or 0.5 years. Then pick a period that will
contain three of the previous cycles, say 3 or 1.5 years, and use
that. This is a long time, but we are only using it in the rescaling,
so it should not be too much of a problem.
I remember someone mentioned the ratio for period between training
and out of sample should be following
Training periods: 8
Out of Sample periods: 1
But he never mentioned about the rationale, and used it as a rule of
thumb, do you have any idea?
This is arbitrary. A real good, theoretical way to do this, is to
split the sample in ten periods. Then randomly choose 9, train it
on those 9, and test it on the rest. Now choose randomly a different
9 sections, train and test on the remaining 1 and so on. It takes
forever to do something like this and we've used it only rarely, but
it will make your Prof very happy! We more often use 20, 10 or even
the last 5%, depending on the length of data. The more data we have,
the more we can cut off to test on. But it is arbitrary. Again,
eyeball the data, and if it has a nice, complete cycle or two in the
last say 10%, then you can use that to test and you will get a good
feel for how it will do in real life.
According to your Message 16564, I have search for Woodes Rogers on
library and amazon, and found a book called "The speculative
strategist", which did mention about Woodes Rogers' approach, but it
is too brief, does it the one you read?
That is the book!
I think I need a few days to digest Kalman filter and to do some
coding on it, and will reply to this topic soon under the same
subject title
Again, those formulas may have bugs in them. So rather try to graph
the ideas they embody and then you can use them as guides to implement
your own. I look forward to hear from you!
Thank you :>
Eric
No problem!
Regards
MG Ferreira
TsaTsa EOD Programmer and trading model builder
http://www.ferra4models.com
http://fun.ferra4models.com
Yahoo! Groups Links