|
PureBytes Links
Trading Reference Links
|
Comments below...
"Bengtsson, Mats" wrote:
> All sound good, and similar to advice I got from others as well, so this is
> the route I will take.
>
> Any thought on what is "significantly better", 2 percent profitability? .05
> better Kelly value? What measurement is a good one?
>
If you try this approach, you'll wind up with a distribution of returns for the
"naive" forecast and a distribution of the "signal event" returns (ie returns
subsequent to your signals). You could calculate the averages and variances for
both sets of returns and calculate a test statistic (a "z" stat) to give you an
idea of whether there's a statistical difference in the averages at some
significance level. There are some fairly restrictive assumptions that are
inherent to this method (normal populations, finite and known variances) that
are most probably violated, but it's a start. It's certainly better than making
comparisons for "significance" based on a fixed number. Maybe some of the
rocket scientists on the list know of a non parametric test for comparing means
of financial returns.
In the description given in Colby and Meyers' book, the author used a technique
he called "variance reduction" to determine significance. I could never
decipher what the hell he was talking about though.
>
> This is something that should be interesting for everyone to do. Is there a
> package that has this kind of feature in it anywhere? I mean, you can do it
> in TS, but obviously that will take a lot of hard manual labor. Anything
> that automates it?
>
This technique can be programmed in EL. You simply program it as a "system"
which allows multiple passes thru the data during an optimization.
>
> > -----Original Message-----
> > From: Bill Vedder [mailto:bved01@xxxxxxx]
> > Sent: den 7 september 2001 03:55
> > To: Bengtsson, Mats
> > Cc: omega-list@xxxxxxxxxx
> > Subject: Re: A question on how to compare buying signals
> >
> >
> > You might try this. I use EOD data only so all references are daily.
> >
> > For each day, calculate future 1, 3, 5, 10, x day returns.
> > Average these. These are your "naive" returns; ie over a
> > given time period, you expect these returns w/o any signal.
> > Note that you'll have to calc log returns so the averages
> > come out properly.
> >
> > For each signal day, calculate the future 1, 3, 5, 10, x day
> > returns. Average these. Compare these returns to the "naive"
> > returns. If significantly different (better), you can say
> > with some degree of certainty that your signals have some
> > degree of predictive ability. You might find that your signal
> > returns are significantly better over a particular time span
> > (eg 1-5 days). You can then tailor an exit signal accordingly.
> >
> > This has been described in Colby and Meyers' book "The
> > Encyclopedia of Technical Market Indicators", 1988
> >
> > Best Regards,
> > Bill
> >
> > "Bengtsson, Mats" wrote:
> >
> > > I wondered if there was someone willing to share a number of good
> > > ideas on this topic. Testing a strategy is pretty straight forward,
> > > but comparing buy signals is a little tougher, since it is
> > likely the
> > > best sell signal to use is depending on buy signal used.
> > >
> > > Is there anyone who has a good way of looking at buy signals more
> > > isolated, to see which of them that deilvers good results?
> > Any tips on
> > > how to do this and what to look for when comparing (like for
> > > strategies, I compare kelly value and Sharpe ratio).
> > >
> > > --- Mats ---
> > >
> > > This message contains information that may be privileged or
> > > confidential and is the property of the Cap Gemini Ernst & Young
> > > Group. It is intended only for the person to whom it is
> > addressed. If
> > > you are not the intended recipient, you are not authorized to read,
> > > print, retain, copy, disseminate, distribute, or use this
> > message or
> > > any part thereof. If you receive this message in error,
> > please notify
> > > the sender immediately and delete all copies of this message.
> >
|