Entries statistically significant compared to random backtested entries?

[dom993]

Quote from abattia:

-1

Performance statistics like these are typical of a mean reversion strategy (as opposed to a trend following/momentum strategy). See http://arxiv.org/abs/physics/0508104

It is true that a string of consecutive losers creates a big drawdown that it takes some time to win back; so the trick is not to trade too leveraged, so that you are still around for when the good times come back.

Thank you abattia for the paper & your comments - much appreciated.

Here is the trade distribution for that prototype - I removed from it the largest win (+15235), which was way too far to the right. Not sure what it proves though


HurricaneUS, I would agree with you 100%, if it wasn't for the duration of the backtest (6 years, 2007-2012) and the number of trades (3500+).


Anyway, I am not saying this is a viable system as is, simply that there are other ways than always using stops & targets.

 

[HurricaneUS]

Quote from abattia:

-1

Performance statistics like these are typical of a mean reversion strategy (as opposed to a trend following/momentum strategy). See http://arxiv.org/abs/physics/0508104

It is true that a string of consecutive losers creates a big drawdown that it takes some time to win back; so the trick is not to trade too leveraged, so that you are still around for when the good times come back.

Have you ever traded with real money?

I'm assuming you haven't or else your experience would clearly tell you that marginally profitable strategies never hold up in the real world and the more trades you examine the more the performance of the system will degrade.

I looked at that paper and all I can say is that once academics start down the road of reducing trading down to a series of equations then I immediately know that the author hasn't likely traded profitably for any credible length of time.

Those types of short-sighted systems that seek to bargain quick and small profits in exchange for typically a high win rate (or in this case, a marginally profitable win rate) and look good in hindsight and on paper will work well until they don't.

 

[HurricaneUS]

Quote from dom993:

Thank you abattia for the paper & your comments - much appreciated.

Here is the trade distribution for that prototype - I removed from it the largest win (+15235), which was way too far to the right. Not sure what it proves though


HurricaneUS, I would agree with you 100%, if it wasn't for the duration of the backtest (6 years, 2007-2012) and the number of trades (3500+).


Anyway, I am not saying this is a viable system as is, simply that there are other ways than always using stops & targets.

survivorship bias?....ok now run your test for a different but equal time period.....oooh say....1999 to 2004 and tell me what you come up with. I'd be willing to bet the house that it blows up.

 

[abattia]

Quote from dom993:
I am currently toying with a system prototype which uses the same signal for entry/exits ... that signal tells me "Long / Short / Flat" ... consecutive same-side signals can be used to scale-in (or be ignored, as in the attached summary).

No stop whatsoever.

P/F isn't the greatest, but after 1-tick slippage per entry / same per exit / and commissions, the net avg/trade is still pretty decent.

You could try something similar, as an alternative to using stops.

Apologies to the OP, as this is marginally off topic (but is related to dom993's post above) ...

dom993,
Can you explain again the workings of your "BetterThanDefault" fill type. You have mentioned it in other threads, and my understanding is that it provides a case somehow in-between the "default" and "liberal" fill types ... you make some assumption about queue position that you have "calibrated" for the instruments you trade? Is that correct?

 

[dom993]

Quote from abattia:

Apologies to the OP, as this is marginally off topic (but is related to dom993's post above) ...

dom993,
Can you explain again the workings of your "BetterThanDefault" fill type. You have mentioned it in other threads, and my understanding is that it provides a case somehow in-between the "default" and "liberal" fill types ... you make some assumption about queue position that you have "calibrated" for the instruments you trade? Is that correct?

BetterThanDefault guarantees the requested slippage, even when that mean the fill given is beyond the H/L boundary of the bar it is executed on. The rest of it is pure Default.

I commented out the original lines (from DefaultFillType), you can easily see the difference.

 

[dom993]

Quote from HurricaneUS:

survivorship bias?....ok now run your test for a different but equal time period.....oooh say....1999 to 2004 and tell me what you come up with. I'd be willing to bet the house that it blows up.

This is CL ... it started electronic trading 24/7 in the fall of 2006, and completed that migration as of Jan-2, 2007. The model I use depends on volume, which wasn't tracked for intraday data for the pit.

I just ran the exact same algo on ES (of course, the volume-related params are different, but I am using the exact same values as for my reversal system on ES - this prototype is using its trend-identification mechanism). It would clearly not be tradable.

 

[dom993]

Interestingly, I just ran it on ES using CL parameter values (my mistake) ... got better results doing so, still far from tradable.

 

[HurricaneUS]

Quote from dom993:

I just ran the exact same algo on ES (of course, the volume-related params are different, but I am using the exact same values as for my reversal system on ES - this prototype is using its trend-identification mechanism). It would clearly not be tradable.

lack of robustness across markets is a dead giveway for overfit, nine times out of ten, particularly if the model is based on some form of price pattern recognition.

 

[abattia]

Quote from dom993:

BetterThanDefault guarantees the requested slippage, even when that mean the fill given is beyond the H/L boundary of the bar it is executed on. The rest of it is pure Default.

I commented out the original lines (from DefaultFillType), you can easily see the difference.

Many thanks.

 

[Dhalsim]

Okay here is one system that relies on some form of pattern recognition and mean reversion, this is similar strategy to one i traded as discretionary trader for years.

I am finding it far more difficult to create short term intraday strategies as obviously the shorter time in market and lack of overnight risk makes and obvious intraday noise by markets makes it very difficult to finding anything that works - sometimes i feel intraday (particularly futures which i trade) is simply random data.

Here are the results of my system - it is currently only a Short side only and Longs have yet to be implemented.

Pros of this system:
1. Intraday.
2. Stays in market for very limited time.
3. I already know what my max loss will be on a given day.
4. Tested with just system entry and random stops and targets and still performs 2 sigma greater than random entry.
5. Tested system stops with random stops and performs well, also system targets performs better than random targets.
6. System is simple - Occam's Razor.
7. Very few optimizing parameters.
8. Any system developer/discretionary trader who thinks volatility should not impact gains is crazy. Volatility is what trading is all about and where most of our profits are made - the system was initially intended to take advantage of vol when designing it.
9. Should allow a lot of leverage in future and large trading size without causing any distortion whatsoever.
10. No slippage as all tests carried out on worst case scenario limit orders where price has to trade through for entry and trade through for target. Therefore, i expect better performance than back-test as at-least 20% of either target or entry will fill with multiple touches and no trade through.
11. Resting limit orders placed very quickly after signal so gives me better chance of fill.
12. Apart from vol filter it has very few optimizing parameters which - it trades from 9:30 to 16:15 and other filters include:
i. basic entry time stop for signal
ii. Max loss in a day filter.
iii. Number of trades in a day filter.
13. ZERO compounding shown in results.
14. Strat uses no indicators at all - just pure price action (maybe a pro/con depending how you look at it).

Cons of this system:
1. Developed for one market -ES. Tested on YM/NQ and works well but not as great as ES. I know a lot of you think market structures are all the same but as a discretionary trader i feel every market has its own characteristics (anyway i am not debating this with anyone so please don't try guys).
2. Not massive gains - more of a grind out strategy.
3. Performs way better under volatile conditions.
4. Not that many trades when volatility filter is added.

Anyway guys here are the results - tell me if you think this is worth trading at all??

I have attached the optimized results (which is simply with volatility filter) and attached a non-optimized version with no volatility filter.