Hi Brian, and all --

You wrote:
"Theoretically the OOS results should be better/worse than the IS tests on a
50/50 basis."

------------------------------------------

In my experience, the out-of-sample results are seldom better than the
in-sample results.

Here are a couple of reasons that might explain why.
1.  In-sample, the system learns both the signal and a portion of the
noise.  In the out-of-sample period, even if the signal remains the same,
the noise has changed, so the performance associated with the noise
disappears.
2.  The signal changes over time -- the model loses sync with the data.  The
signal in the out-of-sample period is different than the signal in the
in-sample period, so the performance due to the signal declines.

Thanks,
Howard



On Sun, May 10, 2009 at 7:37 PM, brian_z111 <[email protected]> wrote:

>
>
> > Near Random ?
> > You're kidding right ? ... I'd thought you had beeen around longer >than
> that ...
>
> All of your points aren't entirely clear to me but:
>
> I'm talking about daily bars (say the US S&P500) - my testing shows that
> non-randomness is not significant.
>
> Inefficiencies do exist but we have to look for them elsewhere.
> Opportunities do exist but we have to exploit them in ways other than using
> structural models.
>
>
> > On another level the number of trades has no relevance in calculating
> K-Ratio
>
> I take it you want me to think again about the K-ratio.
>
> However... K-ratio is a slope/stderror comparison ... so slope doesn't vary
> with N or the order of the trades? .... stderror is stdev inclusive ...
> stdev tails out in both directions with time?
>
> >and as calculated in AB is useless for anything but systems that >trade
> constant dollar amounts ...
>
> I haven't delved into the K-ratio in detail.
> I am not a fan of equity curve analysis so I haven't thought about the
> K-ratio in depth.
>
> I noticed that Kestner based his calcs on a constant contract basis and
> then used logN to equate compounded eqcurves to a constant contract basis
> ... are you saying that there is a mistake in his thinking or math at that
> point?
>
> I recall you mentioned this before but I didn't follow up because last time
> I looked the K-ration in AB was exactly as written in KL's last book?
>
> How do you calculate it and how do you use it?
>
> >By definition that would imply no degree of success over a long >period of
> time.
>
> With methods that are dependent on predicting the future? .... yes it does!
>
>
> --- In [email protected] <amibroker%40yahoogroups.com>, ftone...@...
> wrote:
> >
> > Near Random ?
> >
> > You're kidding right ? ... I'd thought you had beeen around longer then
> that ...
> >
> > On another level the number of trades has no relevance in calculating
> K-Ratio and as calculated in AB is useless for anything but systems that
> trade constant dollar amounts ... By definition that would imply no degree
> of success over a long period of time.
> >
> > ----- Original Message -----
> > From: brian_z11
> > Date: Sunday, May 10, 2009 8:58 pm
> > Subject: [amibroker] Re: Expectancy - and related--specifically K-rato
> > To: [email protected] <amibroker%40yahoogroups.com>
> >
> > > My understanding of Bob Pardo's book is that he feels
> > > that the length of the
> > > out-of-sample period can be determined by a calculation based on
> > > the length
> > > of the in-sample period. If that were true, then the ratio
> > > would be a
> > > little easier to compute. But, in my experience, there is no
> > > relationshipbetween the length of the out-of-sample period and
> > > the length of the
> > > in-sample period. And gathering the data and performing the
> > > calculation of
> > > the ratio for some objective functions would be difficult.
> > >
> > > I don't believe that it would be worth the effort but setting
> > > tbe BT the task of collecting the same number of trades in the
> > > OOS test as were collected in the IS test is the way to
> > > standardise and simplify calculating this ratio.
> > >
> > > N == the number of closed trades in a test run;
> > > Where IS(N) == OOS(N) the ratio IS(N bars)/OOS(N bars) will
> > > always be defined as a probability distribution.
> > >
> > > No constant relationsip, between IS and OOS length (in bars) can
> > > exist because of the (near) random nature of the markets.
> > >
> > > Believe that such a relationship does exist, and hence can be
> > > calculated/exploited, is an artefact of the belief that the
> > > markets are non random i.e. have a structure that we can
> > > identify, model and therefore use to make predictions about the
> > > future of the markets.
> > >
> > > Any attempt to improve OOS performance by 'tuning' the length of
> > > the sample periods is a futile excercise.
> > >
> > > The exception to this, as you point out in your book, is that
> > > the markets do change (they are not quite random .... human
> > > behaviour leaves a faint trail of non randomness) and so current
> > > data may be more relevant than non- current data e.g. increased
> > > use of computers, and access to information, by traders may have
> > > changed the markets in the last decade or two.
> > >
> > > Whether these changes are strong enough, and occur over short
> > > enough time periods or identifiable time periods, to justify the
> > > view that we can/must account for this, in our design processes,
> > > is arguable i.e. it is questionable that we can "synchronise the
> > > market and our trading system by shortening (varying) the time
> > > periods used in our walk forward testing", as you suggested in
> > > your QTS book (P261).
> > >
> > > Granted that we are attempting to synchronise our systems to
> > > market behaviour but this is done within our system rules, not
> > > by altering the N bars tested.
> > >
> > > The fundamental premise of trading is that the system will
> > > perform, in the future, irrespective of market conditions or the
> > > time period involved.
> > >
> > > The OOS walk through is a test to find out if the patterns we
> > > identified in the IS data still exist in the future and that the
> > > code we used to synchronise to that pattern(s) is effective.
> > >
> > > The only time that will be time dependent is when the patterns
> > > are time based e.g. the number of times the price of oil goes up
> > > on the first day of the month is statistically significant.
> > >
> > > Very few exploitable inefficiencies in the markets are time based.
> > >
> > >
> > > On another point:
> > >
> > > Theoretically the OOS results should be better/worse than the IS
> > > tests on a 50/50 basis.
> > >
> > > Continual skewing of that ratio to the downside is an indicator
> > > of curve fitting at the IS stage?
> > >
> > >
> > >
> > >
> > >
> > > --- In [email protected] <amibroker%40yahoogroups.com>, Howard
> B wrote:
> > > >
> > > > Greetings all --
> > > >
> > > > In-sample results and out-of-sample results can be, and
> > > usually are, very
> > > > different in their characteristics.
> > > >
> > > > My experience is that the ratio (OOS/(IS+OOS)), where these
> > > are the
> > > > in-sample and out-of-sample results, is difficult to compute
> > > and often even
> > > > difficult to define. I have not found it to be of value in
> > > estimating the
> > > > future performance of the system.
> > > >
> > > > My understanding of Bob Pardo's book is that he feels that the
> > > length of the
> > > > out-of-sample period can be determined by a calculation based
> > > on the length
> > > > of the in-sample period. If that were true, then the ratio
> > > would be a
> > > > little easier to compute. But, in my experience, there is no
> > > relationship> between the length of the out-of-sample period and
> > > the length of the
> > > > in-sample period. And gathering the data and performing the
> > > calculation of
> > > > the ratio for some objective functions would be difficult.
> > > >
> > > > Isn't it the out-of-sample results we are trying to estimate?
> > > Do we care
> > > > what the in-sample results look like? If the out-of-sample
> > > results are
> > > > terrible, why bother computing the ratio. If the out-of-
> > > sample results are
> > > > good, why bother computing the ratio -- how will that
> > > information be used to
> > > > improve the system? And if it is used to modify the system,
> > > then the
> > > > previously out-of-sample data has become in-sample.
> > > >
> > > > Do any of the forum member have examples they can contribute
> > > where computing
> > > > the ratio is helpful?
> > > >
> > > > Thanks,
> > > > Howard
> > > >
> > >
> > >
> > >
> >
>
>  
>

Reply via email to