<snip> That is similar to Pardo's WFE (Walk forward efficiency), or a measure 
of how much curve fitting inflated test results.  Pardo suggests taking the 
concatenated out of sample returns and divide by the result treating the entire 
combined data set as in sample.  Anything below 0.65 will probably not trade 
well live.  The higher, the better.<snip>

There is where the subject becomes very interesting ... it is hard to find an 
absolute answer i.e. find a firm footing that doesn't slip out from under us 
e.g.

- is 0.65 an empirical number ... where did Pardo get it from ... experience or 
by calculation (based on observations)?
- this is one place where Howard and I disagree .. Howard is emphatic that the 
metrics from the IS are worthless whereas I think that if the OOS encourages us 
to continue then the combined IS and OOS provides us with useful stats ... I 
expect Howard reiterates that point, without compromise, as a teaching device.
- examining the principles ... using W/L as the example ... say our IS test 
returns W/L 65/35 and our PayOffRatio is 1/1 ... this translates into equity 
curves etc eventually i.e. once we introduce MM. We use the OOS metric to 
decide if the IS development process was successful but how exactly do we do 
that? Presumably if the IS metric was good and the OOS is the same +- X% we are 
happy i.e. if the OOS W/L == 65/35 and the PR == 1/1 i.e. matches the IS, then 
we can realistically expect to achieve the same eq curves, providing we trade 
over a stationary market (stationary == the same mean and dispersion as the 
market used for our IS dataset?).

How much X% should we tolerate and does it vary from metric to metric?

Comparing it to our benchmark, the null hypothesis (a fair coin toss) ... from 
my labtests, variance == sample error.
So the ratio of the IS metric/OOS metric depends on the size of the sample:

IS W/L +- sample error/OOS W/L +- sample error == ratio;

The concatenated IS and OOS creates a larger sample  and converges on the mean, 
as N approaches infinity (so this is a valid method afterall?).

The IS is always smaller than the combined IS and OOS and sometimes very small 
(according to user preference).

Since W/L always == 1, for a fair coin toss, the ratio of IS/(IS + OOS) metric 
will always equal the sample error difference.

+- sample error *1 == 1 StDev frequency
+- sample error *2 == 2 StDev frequency
etc

So, 100N IS/ infinite N(IS + OOS) ==
1 +- 10%/1 +- 0% == 0.9/1;// 66% of the time
0.8;// with 2 StDev significance
0.7;// for 3 StDev significance.
etc

For biased systems (65/35 W/L) it works in a similar way (give or take a 
bit)..... it looks like the 0.65 number is an approximation of significant 
variance. 

The really interesting part is that I can't (mathematically) tell the 
difference between market variance and non-stationarity.  





--- In [email protected], "dloyer123" <dloyer...@...> wrote:
>
> --- In [email protected], Rajiv Arya <rajivarya87@> wrote:
> >
> > 
> > I like to compute a ratio of the out-sample metric and divide it by the 
> > in-sample metric. 
> > 
> > And I like to look for multiple runs of out-sample/in-sample ratio to be 
> > above 0.5 and with little fluctuation.
> > 
> 
> That is similar to Pardo's WFE (Walk forward efficiency), or a measure of how 
> much curve fitting inflated test results.  Pardo suggests taking the 
> concatenated out of sample returns and divide by the result treating the 
> entire combined data set as in sample.  Anything below 0.65 will probably not 
> trade well live.  The higher, the better.
>


Reply via email to