Greetings --

h3po wrote:
"Actully this gives idea for another interesting question: If one is able to
model *stockprices* by means of a simple model, could this (monte carlo)
simulated dataseries be valuable as additional data and be used as more OOS
data or will it not give any new information in the testing process?
(Asuming one is doing a WF test. If Yes, then this would be valuable for
end-of-day trading system development as the amount of data here is usually
limited). Has this been  discussed on this forum before?"

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

In my opinion, synthetic or simulated prices have only one purpose.  That is
to investigate the superposition of signals -- such as combining a trend
with one or more cycles with noise -- at the most basic level in order to
understand superposition and decomposition.

Synthetic price series have no value in testing trading systems.  The
algorithm portion of a trading system -- the model -- is the logic that is
attempting to distinguish the signal from the noise.  Recognizing the signal
portion and issuing buy and sell orders is what makes the system
profitable.  Everything that the model does not specifically recognize and
use to issue buy and sell orders is noise -- even if it is clearly
identifiable as signal by some other method.  The ratio of signal to noise
in financial data is very low.  Identifying the signal component among all
the noise is very difficult.

If I can generate synthetic price series that are adequate for system
development, then I do not need to generate synthetic price series.  If I am
able to synthesize the signal well enough that I can develop a trading
system based on it and trade that system profitable when using the actual
data, then the logic I used to generate the synthetic price series describes
the signal.  I simply code that logic into a real trading system and use
it.

Synthetic price series based on random values will be random.  No trading
system can reliably and repeatedly extract profit from random series.  The
second law of thermodynamics applies here.

My recommendation to avoid synthetic price series extends to reordering
observed prices.  Reordering, for example reordering the daily prices of
IBM, in an attempt to give additional data for system testing destroys the
time related components of the original series.  If there was information in
trends, patterns, seasonality, autocorrelation, intermarket correlation,
etc, none of that exists in the newly created series.  What could be left?

In any event, these questions are easy to test.  Develop a trading system on
synthetic or reordered price series.  Then test on out-of-sample actual
data.  If it is profitable, then the value is in the logic used to generate
the synthetic data.

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

It is valuable to add random noise to actual price series in order to test
how well a trading system differentiates between noise and signal.  But that
is the answer to a different question.

-----------

Thanks,
Howard

Reply via email to