Simulation question

[petrovich54]

Hi, I am trying to figure out profitable strategy by simulation trading.

For example, I found one strategy that returns on average 0.6% per trade without optimization. This is not enough to put in practice, considering commissions and slippage. So I am trying to vary strategy parameters, like trading hours, volume, volatility etc.

By optimizing parameters, I achieve 2% per trade, on 80 trades between 2007-06-01 to 2008-04-30. This already looks tradable.

Then I test these parameters on different time period, 2007-01-01 to 2007-05-30. Results are back to 0.6% /trade.

And so on, parameters optimized for 2006 do not work for 2007 and 2008 and vice versa.

The question is: is the strategy total s__t ? Or may be we should speak of market conditions change, like market went from bull to bear in the middle of 2007, and if I continuously adjust parameters, I can trade it?

 

[MrGecko]

I was going to suggest bootstrapping, but having spent 10 mins exlplaining how to do it, it became clear that it is probably quite useless in this case. Sorry mate, can't help you.

 

[petrovich54]

Does bootstrapping mean "go f__k yourself" ? Sorry I'm not too proficient in English.

I realize I asked a silly question. Just hoped to stimulate a discussion. Its lonely and boring out there.

 

[gooseman]

check the volatility across time period. i suspect this might be the problem.

liquidity issues have hurt me but not sure what size you are trading.

 

[intradaybill]

So I am trying to vary strategy parameters, like trading hours, volume, volatility etc.

These are not parameters that people normally optimize. Playing around with these parameters often results in "selection bias".

Your question is too general, like: "I broke up with my boyfriend, can you offer some help?", the type of question asked during morning radio shows featuring an astrologer

Bill

 

[MrGecko]

Bootstrapping is normally found in bond pricing and VaR calculations, where it is basically a Monte Carlo with returns drawn at random (normally distributed) from the history. It might be of some use, if return (N+1) is drawn at random from a normalised distribution around N, but you'd have to do it too many times to make it worthwhile.

It doesn't mean "go f*ck yourself".

lol.

 

[grantx]

Mr Gecko,

I think I've commented on your language skills previously.

Grant.

 

[MrGecko]

Bootstrap.

 

[grantx]

It isn't clever, you know.

 

[petrovich54]

So, shoud we adjust parameters ? Or the system should always work in all time periods with the same parameters ?

I recall Jim Simons mentioned in some interview that they readjust their programs every two weeks.

 

[BerkshireBull]

So, shoud we adjust parameters ? Or the system should always work in all time periods with the same parameters ?

I recall Jim Simons mentioned in some interview that they readjust their programs every two weeks.

It looks like you're curve-fitting your strategy to the historic data.

Try optimising is over a long time period, say 1/1/2002 to 1/1/2007, and then do an out of sample test over a short time 1/1/2007 to 1/12008.