PureBytes Links
Trading Reference Links
|
Eric, another comment - for nets specifically. We noted a good
improvement when we also include fundamental data in the net. So
if we build a net for trading currencies say, then we include the
federal funds rate for one. Now we have to standardise this as well.
We found an article somewhere (they were building a net for trading
gold and got good results with it) and we used the following approach
as described there, suppose you want to include F, the federal funds
rate.
F / Mov(F,5,E) - 1
F / Mov(F,10,E) - 1
Again, it is standardised to a percentage, and we include a couple of
these in the net. We toy around with the lengths, but do not formally
optimise them. We also try things like
Mov(F,5,E) / Mov(F,50,E) - 1
to get a feel for where the rate, sans noise, would be relative to
it recent history (basically, is it currently high or low) and also
things like
Mov(F,10,E) / Mov(F,50,E) - 1
to measure if the rate is rising or falling. Note that we are trying
to measure and standardise observations a trader would make. A
trader would make a decision based on wheter the rate is high or low,
and on wheter it is rising or falling. So we try to formulate that
somehow, and also to standardise it so the net can digest it.
Although this is a net based answer, you can use a similar approach
to standardise other stuff.
Regards
MG Ferreira
TsaTsa EOD Programmer and trading model builder
http://tsatsaeod.ferra4models.com
http://www.ferra4models.com
--- In equismetastock@xxxxxxxxxxxxxxx, "MG Ferreira" <quant@xxxx> wrote:
>
> Hi Eric,
>
> We normalise based on the indicator. So, if it is an RSI, it is
> very easy, as this is always between 0 and 100, and we just map
> that range to -1 to +1. If it is some trendline that we fitted,
> we may use -1 or +1, based on wheter it is above or below the
> trendline. Also, for highest high/lowest low type stuff, we use
> a +1/-1 approach - +1 if say above and -1 if below some previous
> high. But for this last example, we typically will also add an
> indicator that is a percent, so that we have two inputs, the +1/-1
> as well as the percent above/below that value.
>
> We often use percentages to normalise. When we calculate the MACD
> for example, we then express it as a percent of the close, and so
> we have something that is between -1 and +1. We also use standard
> deviations, such when we divide the most recent percentage change
> by its standard deviation. But we don't use standard deviations
> that much.
>
> If all else fails, there is the tanh function, that will map
> any value to somewhere between -1 and +1!
>
> So, when we prepare inputs for the net, we use a variety of
> techniques, whichever makes most sense (to us!) for that indicator.
>
> When we do the one-size fits all optimisation, that does not end
> up in a net, we use a similar approach. Something like the RSI
> can easily be standardised, but something like the MACD or CCI
> we simply normalise to its own historical range. This sometimes
> leads to confusion, as you get a +1 or 100% reading a couple of
> days as the say CCI reach new highs and is normalised back to 100%
> every time. But it is good enough for our purposes.
>
> If you are going to feed the stuff into a net, you must pay some
> attention to this as it adds lots of value once you start training
> the net. We used to just throw values at the net, but got MUCH
> better results once we first standardised it. If it is just to
> compare with other standardised values, I think it may be a bit of
> an overkill and you can just use historical or standard deviation
> as a quick fix.
>
> Regards
> MG Ferreira
> TsaTsa EOD Programmer and trading model builder
> http://tsatsaeod.ferra4models.com
> http://www.ferra4models.com
>
>
> --- In equismetastock@xxxxxxxxxxxxxxx, chichungchoi <no_reply@xxxx>
wrote:
> >
> >
> > Hi MG Ferreira:
> >
> > > One thing we have also used, for a more thorough, automated
> > solution,
> > > is to take a number of 'good' systems, say what we feel are the best
> > > 30 of the 1 ... 100 systems. Then we do not optimise, but
> > > standardise all of these 30 inputs. Again, this is not trivial and
> > > there is no simple answer, but we try to generate a value between -1
> > > and +1 for each of the 30 systems. So in stead of taking the say
> > > RSI (between 0 and 100), we divide by 50 and subtract 1 to get it to
> > > between -1 and +1, and do something similar with all the rest. So
> > it
> > > is easy to compare and easy to optimise.
> >
> > This is the process of normalization as you mention, do you have any
> > suggestion on how to select the appropriate normalization please?
> > Referring to the article "Normalization" on Oct 2000, TASC, p.58-68.
> > There are three different types of normalization, such as
> >
> > 1) normalized to standard deviation
> > 2) normalized to average true range
> > 3) normalized to its own historical range
> >
> > Should I only apply one type of normalization for all indicators, or
> > should I apply different type of normalization for different
> > indicators based on their own best performance?
> >
> > On the other hands, if I don't normalize the indicators, and try to
> > define +1 for buy signal and -1 for sell signal for each indicator,
> > then do you know the drawback for this approach?
> >
> > Thank you
> > Eric
------------------------ Yahoo! Groups Sponsor --------------------~-->
In low income neighborhoods, 84% do not own computers.
At Network for Good, help bridge the Digital Divide!
http://us.click.yahoo.com/EpW3eD/3MnJAA/cosFAA/BefplB/TM
--------------------------------------------------------------------~->
Yahoo! Groups Links
<*> To visit your group on the web, go to:
http://groups.yahoo.com/group/equismetastock/
<*> To unsubscribe from this group, send an email to:
equismetastock-unsubscribe@xxxxxxxxxxxxxxx
<*> Your use of Yahoo! Groups is subject to:
http://docs.yahoo.com/info/terms/
|