Another challenging subject (3 weekends in a row).

Initially I thought you were asking a question, or questions, about 
benchmarking, in its various forms, but I see now that you also seem to be 
asking about which evaluation metric is the most suitable to measure the 
'goodness' of your stops?

I think it will help if the discussion is not so hypothetical, so I will make 
up an example to save you revealing your trading secrets to the world ;-)

Will this do for an example filter and signal?

Signal = Ref(ROC(MA(C,5),1),-1) < 0 AND ROC(MA(C,5),1) > 0;//the fast MA turns 
up

Filter = ROC(MA(C,20),1) > 0;//the slow MA is up (an uptrend in this model)

disclaimer = string(I hope I got that right);

I think I backtested this once but I can't remember what the result was.

IMO you should think again about optimising your filter because the signal and 
the filter are both part of the system entry:

entry = signal == 1 AND filter == 1;

So 'entry' has to be true to action the stops, which are based on the price at 
entry) .... so you have two entry variables to optimise.


Noise is a concept that came from physics and possibly maths ... I know that 
optimalF was an extension of the maths component of a solution to interference 
in telephone lines (developed by Bell?).

Howard uses it with understanding and confidence (it means something to him) 
but I seldom use it because I don't understand its meaning, when using in a 
trading context etc .... all price action is real to me ... my broker seems to 
agree (Hey its Bob the Broker here with a margin call .... Don't worry, that 
was only noise!)

--- In [email protected], "Yofa" <jtoth...@...> wrote:
>
> Hi All,
> 
> 
> 
> Aron:
> you got better results by removing your original entries, because your 
> original entries were not better then random and you got more time in the 
> market by using simple random entries (my guess).
> 
> 
> 
> To All:
> Thanks for all the thoughts and consideration.
> 
> 
> 
> To give some more hints and encourage thoughts here is a bit more info. 
> My general idea is to divide a complete trading system into smaller 
> independently testable/optimizable pieces. I'm building a single equity, 
> intraday, automated trading system. To make it simple let's say it consists 
> of a filter (when not to trade) an entry & timing logic (generate buy and 
> short signals) and a trade management logic (initial stop, trailing logic, 
> profit taking exits, etc.)
> If we accept that the price movements consists of noise and real price 
> movements than the trade management logic's only job is to keep my stops 
> (initial and trailing) out of the noise level, while minimizing initial loss 
> and maximizing profit. It has to accomplish this REGARDLESS OF THE QUALITY OF 
> THE ENTRIES AND FILTERS. If all my entries are bad it has to produce the 
> least amount of loss. If all my entries are excellent it has to collect the 
> most profit.
> If I run a number of backtest runs with random entries while keeping the 
> settings of trade management logic constant I get a "sample" of what might 
> happen using the settings if my entries are not better then chance. This 
> sample has a distribution of profits, CARs, system drawdowns, etc. All the 
> attributes of a backtest runs or a series of real life trades!
> If I run similar test with each possible setting (optimization) and compare 
> the samples of each settings, I'm able to select settings that produce the 
> best performance distribution (defined by my objective).
> So if my trade management logic is up to its job (using the best settings) it 
> has to produce the best distribution of drawdowns and profits of backtest 
> runs with random entries.
>  
> 
> Similarly, the filter's job is to keep me out of market when trading is not 
> profitable. It's not profitable because there are more noise than real price 
> movement (so initial stop is going to be hit sooner or later) OR because of 
> entering the market in the wrong direction. If using random entries (in 
> random directions) and the filter is bad, the initial stop is hit because of 
> either cause. If the filter is good, the numbers of initial losses are 
> minimized because initial stop is hit if I try to ride the market in wrong 
> direction but noise is appropriately addressed.
> So if I use random entries and use the same initial loss with no trailing and 
> add the "perfect" filter to it,the filtered system has to provide the 
> smallest loss and the smallest drawdown. By running a number of random 
> backtests for each possible filter settings, I produce a sample of that 
> filter settings. These samples can be somehow compared and the best selected.
> 
> 
> 
> Any opinion, thoughts or experience is appreciated.
> I don't really know what the best way of comparing "samples" is. Any idea?
> 
> 
> 
> Regards,
> 
> 
> 
> Y
>


Reply via email to