|
PureBytes Links
Trading Reference Links
|
>4. If you remove the labels from the axes of a chart, it is not possible to
>tell the timescale of the chart. Therefore, claims by systems vendors that
>their system will work on all time scales are probably correct, with the
>possible exception of very short timescales where transaction costs are too
>high a percentage of the gross return.
This is very interesting and everybody believes it to be true (especially
the Chaos people). But is it? Consider a system that is shown to be
profitable in out-of-sample data. Furthermore, lets assume that
profitability and robustness (stability of profits as a function of system
parameters) are demonstrated in the hourly, daily and weekly timeframes. So
the system is profitable and robust in these 3 time frames. Does this imply
that it is equally profitable and robust in all other time frames? What do
we do if it isn't? Should such system be discarded if it is found breakeven
at 45-minutes and a loser in the 2-day time frame? Of course it would be
great to design a system equally profitable and robust in all time frames
(demonstrated out-of-sample) but even when one is found is it over-optimized
to the data? Suppose that the rules are changed slightly so that the 45-min
and 2-day frames are also profitable but by doing so we find these mods
considerably reduce profitability in the previous time frames
(hour-day-week) which we want to trade. What then?
These are some of the questions I had to face (and still facing today) and
would appreciate comments from anyone working down this road.
|