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

Intraday Realtime vs. Historical Results



PureBytes Links

Trading Reference Links

 
  I just put a mechanical system on a realtime feed (different computer/
different feed/ same data vendor----Why? is not important or relevant to the
issue). 
  Using this week as my example of realtime performance, the realtime
feed kicked in 6 trades(realtime) and the same 6 trades registered after
review on the data after trading hours on the same computer.
  The computer where the system was developed, kicked in 3 trades after
review on data after hours(offline). The 3-trade results were consistant with
past
system characteristics and performance during the system
developement(offline).
  Only 2 of the trades between machines was exactly the same or even similar.
  The historical testing is not bouncing-tick sensitive. In other words,
settings
for bouncing ticks covering the full range from 0-100 yielded no alteration
to any trade over 100,000 bars of 5min data.
  I don't expect any 2 sets of results to emulate eachother exactly but this
is 
rediculous and far from acceptable. The anticipated results yield +$3000
at 67% on one machine . The realtime results yield -$250 at 40% on the other.
  I can think of 3 ways to attempt to bring the realtime results closer to
anticipated...

1) Check the clock on the realtime computer and make certain it is alligned 
   with the correct time. Thus the bars are formed correctly.

  Only two TS built-in functions that create indicator variables are used in
the 
system code.(i.e. average). The average function is defined in TS under 2 
separate function files---both with the same name (one is simple, one is
series).
Which one is called by the system files? And is a different one called for
realtime as opposed to offline?

2) Re-write all built-in function files under different names and declare them
    as series, since I don't know which is being called when. Or will
hardcoding
    them into the system files eliminate any questionable differences more
    efficiently? I'm skeptical about this one because, so far, creating
indicators
    which, in part, use custom-named average functions as either simple or
series
    yield the same results on the screen indicators.

3) Refer to midpoint of bars for function calculations as opposed to close.
    This assumes the theory that if both sets of data have been recorded on
    computers with the exact same times, that both sets of data will have
    the similar highs and lows per bar irregardless of bouncing ticks.

  I'll be posting any results that narrow the gap between realtime and
historical
tests.
  Any other ideas are welcome.

dbs