|
PureBytes Links
Trading Reference Links
|
On Wed, 10 Jan 2001 20:42:46 -0600, you wrote:
>I am curious as to what the common wisdom is on the minimum number of trades
>a system needs in order to consider the resulting data valid.
Because under certain conditions a valid system should not show any
trade (also for a long time period), not the _number of trades_, but
rather the _back testing period_ can be seen as a measure here.
The back testing period should be large enough to "see" the typical
behavior of a market, but small enough to avoid too much
"contradictions" to be picked up in non-stationary changing markets.
To give an example:
For an option market, I found a five years time period too large and a
one year time period too small. A compromise led to a _weighted_ five
years period with an adaptively updated trading model.
>Second item for discussion. Once I have some backtested data together, is
>it appropriate to separate above/below average winners/losers. To take this
>one step further, if I were to group the trades as indicated (above/below
>average) and average key measures for each group, how much leeway should I
>give the averages if I want to filter out distinguishing characteristics of
>the above average winners?
>
>For example, if I find on above average winners that volume for the entry
>day is on average +85% of the previous day, would using a volume filter of
>+85% be appropriate for real-time trades taken going forward? Is there a
>more representative calculation than average to tell me how a specific group
>performs for a given measure (volume, etc.)?
Imo, we should not restrict ourselves to the _average_. A more
powerful approach is to look at the _distribution_ of wins & losses
and to understand this as a _probability distribution_ for future
trades. This way, we can measure (and avoid ?) things like "big
drawbacks" etc, by including important features of this distribution
into our system optimization's goal function. - Of coarse, _powerful_
distribution functions should be used here.
But, imo, there are more topics to discuss, if it comes to "how robust
is a trading system", where all of them should be modeled into the
"goal function" of the trading system optimization. - What I use here
(beside of profit) is: number of winning / number of loosing trades,
volume of wins / volume of losses, IN-rate, tail-volumes of win & loss
distributions, all together "mixed up" in a goal function which fits
_my_ style of trading, not optimized for maximum profit only.
The main problem left for me is, " how to improve generalization"
(extrapolation capability), i.e. what can I do to separate more
accurately hidden but significant market effects from (e.g. external)
market disturbances.
Any hints?
mfg rudolf stricker
| Disclaimer: The views of this user are strictly his own.
|