[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Another way to answer to Mr Gaius



PureBytes Links

Trading Reference Links

Peter2150@xxxxxxx writes:
>In a message dated 98-10-04 11:59:48 EDT, trade_jack@xxxxxxxxx writes:
>
>> i plot close to 13,000 tick bars per chart, so i really don't worry
>>  about that problem since it's 13,000 bars in the past. now i'm
>>  wondering should i worry?
>
>Practically speaking probably not, but theoritically yes.    The
>initialization may be 6 days ago, which is removed from today, but tomorrow,
>the initialization starts with a new day.  This could make minor changes in
>something computed with exponential average.  True might be in the 5 or 6th
>decimal place, but if you are testing numbers it might change something.

I haven't followed the rest of this thread, but I'll just interject my
uninformed observation from reading Pete's notes:

Exponential moving averages are a class of filter known as "Infinite
Impulse Response" which, as the name implies, has infinite memory.
In a simple XMA, the effect of an "impulse" dies gracefully away,
more or less slowly depending on the XMA length, and assuming there
are no roundoff errors within the filter.

The effects of roundoff error become very much more pronounced the
more complex the filter.  If, as Pete implied, the EL code involves
cascaded XMAs, roundoff error could cause the initialization effects
to persist for a very long time indeed.  Only a close numerical analysis
of the code itself would tell you how many significant digits you'd
need for a given timeframe.

The subject thread could indeed be another example of a situation 
where single-precision arithmetic is insufficient.

Jim