|
PureBytes Links
Trading Reference Links
|
Come to think of it, anchored (expanding) window, exponentially damped
window and rolling window are just, er, three sides of the same coin. The
difference between them is a difference of degree, the "abruptness" if you
will, with which old data is discarded.
If a parameter stabilizes with increased volume of data in-sample, I'd
first be tempted to do some more backtesting using the value that it seems
to stabilize on. You may have found a characteristic of the market
concerned, and further optimization may not even be
necessary. Alternatively, if 32 months of data give you a stable, working
value, start using a 32-month sliding window from now on. If that leads to
slow changes in the parameter and the system stays profitable, fine. If it
leads to abrupt changes in the parameter, then you've probably overoptimized.
At 15:02 14/11/02 +0000, mark.keenan@xxxxxxxxxxxxxx wrote:
>Over the last few weeks I have spent a lot of time, & received a lot of
>advice form the list about walk forward optimization - After weeks of
>testing and reading I have arrived at the following conclusions:
>
>The 1st principal of optimization is that if you are going to do it - it
>should be done over as much historical data as possible.
>
>Rolling Walk forward optimization, (shifting the in-sample window, say two
>years, forward to encapsulate the most recent out of sample period, say six
>months, and re-opimazing the new shifted two year window) although sounds
>good - and mimics a lot of the currently popular adaptive techniques
>(adaptive moving averages etc) is unstable as far as i have investigated.
>Furthermore how does one know how much in-sample data to use versus how
>much out-of sample data to trade on etc etc (too many variables)
>
>Anchored Walk forward optimization seems very stable - it has been
>described in several books on optimization as is based on the following
>principal:
>
>A contract is selected - I am using the DAX future:
>
>The initial starting point is selected as a function of the contract
>characteristics - I have chosen 1/4/99 as this was the time the contract
>was denominated in euros and traded on the eurex platform. (if using the
>S&P for example a good place to start would be when the tick size changed
>from $500 to $250)
>
>The contract is optimized for the best parameter (my system has one , just
>length)
>
>The market is traded over the next month of unseen data
>
>At the end of the month the new optimization period is performed from the
>original point again 1/4/99 up to the end of the most recently traded
>month, the new parameter is recorded, the next month traded with the new
>parameter, before repeating the whole process at the end of the month.
>
>I have performed the this testing on two contracts over 32 unseen monthly
>periods and the results are good. (total 64 months each month having an
>average of 3 trades in it)
>
>I suspect the following problem will arise and am very interested in
>knowing any potential work arounds:
>
>The stability of the parameter increases dramatically as a function of how
>much data I use - at the beginning of the 32 month period the parameter
>changed quite frequently and towards the end it seems to have stabilised
>into the same value month after month.
>
>This is retrospectively predictable, and am now however expecting for the
>system to start breaking down in the near future as the sensitivity of the
>optimization process decreases as more data is used. I do not know how
>sensitive the optimization process will become in order to adjust for
>recent changes in the market
>
>What I figured I need is the following:
>
>A type of weighted anchored forward optimization process that pays more
>attention to the profitability of the more recent trades yet still takes
>all the data into account.
>
>Does anyone have any ideas on how to do this - I feel that the process has
>to involve anchored walk forward optimization and the more common rolling
>window methods are really not that reliable;
>
>The system is based changes in slope in a type of regression line - it has
>1 parameter (length).
>
>Thank you very much in advance
>
>Mark
>---------------------------------------------------------------------------
>This message (including any attachments) is confidential and may be
>privileged. If you have received it by mistake please notify the sender by
>return e-mail and delete this message from your system. Any unauthorised
>use or dissemination of this message in whole or in part is strictly
>prohibited. Please note that e-mails are susceptible to change.
>ABN AMRO Bank N.V. (including its group companies) shall not be liable for
>the improper or incomplete transmission of the information contained in
>this communication nor for any delay in its receipt or damage to your
>system. ABN AMRO Bank N.V. (or its group companies) does not guarantee that
>the integrity of this communication has been maintained nor that this
>communication is free of viruses, interceptions or interference.
>---------------------------------------------------------------------------
>
>
>
>
|